Add metric

cleaner error propagation in thread creation
optimize extension download:
2026-02-12 15:10:38 +00:00 · 2023-07-05 13:50:40 -04:00 · 2023-07-05 09:56:03 -04:00 · 2023-07-05 15:04:16 +03:00 · 2023-07-04 21:15:54 +03:00 · 2023-07-04 18:16:34 +03:00
165 changed files with 7913 additions and 3676 deletions
--- a/.github/workflows/benchmarking.yml
+++ b/.github/workflows/benchmarking.yml
@@ -180,7 +180,8 @@ jobs:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
      options: --init

-    timeout-minutes: 360 # 6h
+    # Increase timeout to 8h, default timeout is 6h
+    timeout-minutes: 480

    steps:
    - uses: actions/checkout@v3
@@ -321,8 +322,6 @@ jobs:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
      options: --init

-    timeout-minutes: 360 # 6h
-
    steps:
    - uses: actions/checkout@v3

@@ -414,8 +413,6 @@ jobs:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
      options: --init

-    timeout-minutes: 360 # 6h
-
    steps:
    - uses: actions/checkout@v3

@@ -501,8 +498,6 @@ jobs:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
      options: --init

-    timeout-minutes: 360 # 6h
-
    steps:
    - uses: actions/checkout@v3

--- a/.github/workflows/build_and_test.yml
+++ b/.github/workflows/build_and_test.yml
@@ -264,7 +264,7 @@ jobs:
          export REMOTE_STORAGE_S3_BUCKET=neon-github-public-dev
          export REMOTE_STORAGE_S3_REGION=eu-central-1
          # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
-          ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test pagination_tests -- s3_pagination_should_work --exact
+          ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_s3

      - name: Install rust binaries
        run: |
@@ -623,51 +623,6 @@ jobs:
      - name: Cleanup ECR folder
        run: rm -rf ~/.ecr

-
-  neon-image-depot:
-    # For testing this will run side-by-side for a few merges.
-    # This action is not really optimized yet, but gets the job done
-    runs-on: [ self-hosted, gen3, large ]
-    needs: [ tag ]
-    container: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/base:pinned
-    permissions:
-      contents: read
-      id-token: write
-
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v3
-        with:
-          submodules: true
-          fetch-depth: 0
-
-      - name: Setup go
-        uses: actions/setup-go@v3
-        with:
-          go-version: '1.19'
-
-      - name: Set up Depot CLI
-        uses: depot/setup-action@v1
-
-      - name: Install Crane & ECR helper
-        run: go install github.com/awslabs/amazon-ecr-credential-helper/ecr-login/cli/docker-credential-ecr-login@69c85dc22db6511932bbf119e1a0cc5c90c69a7f # v0.6.0
-
-      - name: Configure ECR login
-        run: |
-          mkdir /github/home/.docker/
-          echo "{\"credsStore\":\"ecr-login\"}" > /github/home/.docker/config.json
-
-      - name: Build and push
-        uses: depot/build-push-action@v1
-        with:
-          # if no depot.json file is at the root of your repo, you must specify the project id
-          project: nrdv0s4kcs
-          push: true
-          tags: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/neon:depot-${{needs.tag.outputs.build-tag}}
-          build-args: |
-            GIT_VERSION=${{ github.sha }}
-            REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
-
  compute-tools-image:
    runs-on: [ self-hosted, gen3, large ]
    needs: [ tag ]
@@ -704,6 +659,7 @@ jobs:
                           --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache
                           --context .
                           --build-arg GIT_VERSION=${{ github.sha }}
+                           --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}}
                           --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
                           --dockerfile Dockerfile.compute-tools
                           --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{needs.tag.outputs.build-tag}}
@@ -761,10 +717,40 @@ jobs:
                           --context .
                           --build-arg GIT_VERSION=${{ 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
                           --dockerfile Dockerfile.compute-node
                           --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
                           --destination neondatabase/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
+                           --cleanup
+
+      # Due to a kaniko bug, we can't use cache for extensions image, thus it takes about the same amount of time as compute-node image to build (~10 min)
+      # During the transition period we need to have extensions in both places (in S3 and in compute-node image),
+      # so we won't build extension twice, but extract them from compute-node.
+      #
+      # For now we use extensions image only for new custom extensitons
+      - name: Kaniko build extensions only
+        run: |
+          # Kaniko is suposed to clean up after itself if --cleanup flag is set, but it doesn't.
+          # Despite some fixes were made in https://github.com/GoogleContainerTools/kaniko/pull/2504 (in kaniko v1.11.0),
+          # it still fails with error:
+          #   error building image: could not save file: copying file: symlink postgres /kaniko/1/usr/local/pgsql/bin/postmaster: file exists
+          #
+          # Ref https://github.com/GoogleContainerTools/kaniko/issues/1406
+          find /kaniko -maxdepth 1 -mindepth 1 -type d -regex "/kaniko/[0-9]*" -exec rm -rv {} \;
+
+          /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=${{ github.sha }} \
+                           --build-arg PG_VERSION=${{ matrix.version }} \
+                           --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}} \
+                           --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com \
+                           --dockerfile Dockerfile.compute-node \
+                           --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-${{ matrix.version }}:${{needs.tag.outputs.build-tag}} \
+                           --destination neondatabase/extensions-${{ matrix.version }}:${{needs.tag.outputs.build-tag}} \
+                           --cleanup \
+                           --target postgres-extensions

      # Cleanup script fails otherwise - rm: cannot remove '/nvme/actions-runner/_work/_temp/_github_home/.ecr': Permission denied
      - name: Cleanup ECR folder
@@ -781,7 +767,7 @@ jobs:
      run:
        shell: sh -eu {0}
    env:
-      VM_BUILDER_VERSION: v0.8.0
+      VM_BUILDER_VERSION: v0.11.1

    steps:
      - name: Checkout
@@ -883,8 +869,10 @@ jobs:
          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{needs.tag.outputs.build-tag}} latest
          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v14:${{needs.tag.outputs.build-tag}} latest
          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} latest
+          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-v14:${{needs.tag.outputs.build-tag}} latest
          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v15:${{needs.tag.outputs.build-tag}} latest
          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} latest
+          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-v15:${{needs.tag.outputs.build-tag}} latest

      - name: Push images to production ECR
        if: |
@@ -895,8 +883,10 @@ jobs:
          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:latest
          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v14:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v14:latest
          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:latest
+          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-v14:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/extensions-v14:latest
          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v15:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v15:latest
          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:latest
+          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-v15:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/extensions-v15:latest

      - name: Configure Docker Hub login
        run: |
@@ -918,16 +908,93 @@ jobs:
          crane tag neondatabase/compute-tools:${{needs.tag.outputs.build-tag}} latest
          crane tag neondatabase/compute-node-v14:${{needs.tag.outputs.build-tag}} latest
          crane tag neondatabase/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} latest
+          crane tag neondatabase/extensions-v14:${{needs.tag.outputs.build-tag}} latest
          crane tag neondatabase/compute-node-v15:${{needs.tag.outputs.build-tag}} latest
          crane tag neondatabase/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} latest
+          crane tag neondatabase/extensions-v15:${{needs.tag.outputs.build-tag}} latest

      - name: Cleanup ECR folder
        run: rm -rf ~/.ecr

+  upload-postgres-extensions-to-s3:
+    if: |
+      (github.ref_name == 'main' || github.ref_name == 'release') &&
+       github.event_name != 'workflow_dispatch'
+    runs-on: ${{ github.ref_name == 'release' && fromJSON('["self-hosted", "prod", "x64"]') || fromJSON('["self-hosted", "gen3", "small"]') }}
+    needs: [ tag, promote-images ]
+    strategy:
+      fail-fast: false
+      matrix:
+        version: [ v14, v15 ]
+
+    env:
+      # While on transition period we extract public extensions from compute-node image and custom extensions from extensions image.
+      # Later all the extensions will be moved to extensions image.
+      EXTENSIONS_IMAGE: ${{ github.ref_name == 'release' && '093970136003' || '369495373322'}}.dkr.ecr.eu-central-1.amazonaws.com/extensions-${{ matrix.version }}:latest
+      COMPUTE_NODE_IMAGE: ${{ github.ref_name == 'release' && '093970136003' || '369495373322'}}.dkr.ecr.eu-central-1.amazonaws.com/compute-node-${{ matrix.version }}:latest
+      AWS_ACCESS_KEY_ID: ${{ github.ref_name == 'release' && secrets.AWS_ACCESS_KEY_PROD || secrets.AWS_ACCESS_KEY_DEV }}
+      AWS_SECRET_ACCESS_KEY: ${{ github.ref_name == 'release' && secrets.AWS_SECRET_KEY_PROD || secrets.AWS_SECRET_KEY_DEV }}
+      S3_BUCKETS: |
+        ${{ github.ref_name == 'release' &&
+          'neon-prod-extensions-ap-southeast-1 neon-prod-extensions-eu-central-1 neon-prod-extensions-us-east-1 neon-prod-extensions-us-east-2 neon-prod-extensions-us-west-2' ||
+          'neon-dev-extensions-eu-central-1 neon-dev-extensions-eu-west-1 neon-dev-extensions-us-east-2' }}
+
+    steps:
+      - name: Pull postgres-extensions image
+        run: |
+          docker pull ${EXTENSIONS_IMAGE}
+          docker pull ${COMPUTE_NODE_IMAGE}
+
+      - name: Create postgres-extensions container
+        id: create-container
+        run: |
+          EID=$(docker create ${EXTENSIONS_IMAGE} true)
+          echo "EID=${EID}" >> $GITHUB_OUTPUT
+
+          CID=$(docker create ${COMPUTE_NODE_IMAGE} true)
+          echo "CID=${CID}" >> $GITHUB_OUTPUT
+
+      - name: Extract postgres-extensions from container
+        run: |
+          rm -rf ./extensions-to-upload ./custom-extensions # Just in case
+
+          # In compute image we have a bit different directory layout
+          mkdir -p extensions-to-upload/share
+          docker cp ${{ steps.create-container.outputs.CID }}:/usr/local/share/extension ./extensions-to-upload/share/extension
+          docker cp ${{ steps.create-container.outputs.CID }}:/usr/local/lib             ./extensions-to-upload/lib
+
+          # Delete Neon extensitons (they always present on compute-node image)
+          rm -rf ./extensions-to-upload/share/extension/neon*
+          rm -rf ./extensions-to-upload/lib/neon*
+
+          # Delete leftovers from the extension build step
+          rm -rf ./extensions-to-upload/lib/pgxs
+          rm -rf ./extensions-to-upload/lib/pkgconfig
+
+          docker cp ${{ steps.create-container.outputs.EID }}:/extensions ./custom-extensions
+          for EXT_NAME in $(ls ./custom-extensions); do
+            mkdir -p ./extensions-to-upload/${EXT_NAME}/share
+
+            mv ./custom-extensions/${EXT_NAME}/share/extension ./extensions-to-upload/${EXT_NAME}/share/extension
+            mv ./custom-extensions/${EXT_NAME}/lib             ./extensions-to-upload/${EXT_NAME}/lib
+          done
+
+      - name: Upload postgres-extensions to S3
+        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 }}
+          done
+
+      - name: Cleanup
+        if: ${{ always() && (steps.create-container.outputs.CID || steps.create-container.outputs.EID) }}
+        run: |
+          docker rm ${{ steps.create-container.outputs.CID }} || true
+          docker rm ${{ steps.create-container.outputs.EID }} || true
+
  deploy:
    runs-on: [ self-hosted, gen3, small ]
    container: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/ansible:latest
-    needs: [ promote-images, tag, regress-tests ]
+    needs: [ upload-postgres-extensions-to-s3, promote-images, tag, regress-tests ]
    if: ( github.ref_name == 'main' || github.ref_name == 'release' ) && github.event_name != 'workflow_dispatch'
    steps:
      - name: Fix git ownership
@@ -959,6 +1026,20 @@ jobs:
            exit 1
          fi

+      - name: Create git tag
+        if: github.ref_name == 'release'
+        uses: actions/github-script@v6
+        with:
+          # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
+          retries: 5
+          script: |
+            github.rest.git.createRef({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              ref: "refs/tags/${{ needs.tag.outputs.build-tag }}",
+              sha: context.sha,
+            })
+
  promote-compatibility-data:
    runs-on: [ self-hosted, gen3, small ]
    container:
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -3,6 +3,7 @@ name: Create Release Branch
 on:
  schedule:
    - cron: '0 10 * * 2'
+  workflow_dispatch:

 jobs:
  create_release_branch:
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -200,17 +200,6 @@ dependencies = [
 "critical-section",
 ]

-[[package]]
-name = "atty"
-version = "0.2.14"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d9b39be18770d11421cdb1b9947a45dd3f37e93092cbf377614828a319d5fee8"
-dependencies = [
- "hermit-abi 0.1.19",
- "libc",
- "winapi",
-]
-
 [[package]]
 name = "autocfg"
 version = "1.1.0"
@@ -805,18 +794,6 @@ dependencies = [
 "libloading",
 ]

-[[package]]
-name = "clap"
-version = "3.2.25"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4ea181bf566f71cb9a5d17a59e1871af638180a18fb0035c92ae62b705207123"
-dependencies = [
- "bitflags",
- "clap_lex 0.2.4",
- "indexmap",
- "textwrap",
-]
-
 [[package]]
 name = "clap"
 version = "4.3.0"
@@ -837,7 +814,7 @@ dependencies = [
 "anstream",
 "anstyle",
 "bitflags",
- "clap_lex 0.5.0",
+ "clap_lex",
 "strsim",
 ]

@@ -853,15 +830,6 @@ dependencies = [
 "syn 2.0.16",
 ]

-[[package]]
-name = "clap_lex"
-version = "0.2.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2850f2f5a82cbf437dd5af4d49848fbdfc27c157c3d010345776f952765261c5"
-dependencies = [
- "os_str_bytes",
-]
-
 [[package]]
 name = "clap_lex"
 version = "0.5.0"
@@ -915,7 +883,7 @@ version = "0.1.0"
 dependencies = [
 "anyhow",
 "chrono",
- "clap 4.3.0",
+ "clap",
 "compute_api",
 "futures",
 "hyper",
@@ -924,12 +892,14 @@ dependencies = [
 "opentelemetry",
 "postgres",
 "regex",
+ "remote_storage",
 "reqwest",
 "serde",
 "serde_json",
 "tar",
 "tokio",
 "tokio-postgres",
+ "toml_edit",
 "tracing",
 "tracing-opentelemetry",
 "tracing-subscriber",
@@ -977,7 +947,7 @@ name = "control_plane"
 version = "0.1.0"
 dependencies = [
 "anyhow",
- "clap 4.3.0",
+ "clap",
 "comfy-table",
 "compute_api",
 "git-version",
@@ -997,6 +967,7 @@ dependencies = [
 "tar",
 "thiserror",
 "toml",
+ "tracing",
 "url",
 "utils",
 "workspace_hack",
@@ -1047,19 +1018,19 @@ dependencies = [

 [[package]]
 name = "criterion"
-version = "0.4.0"
+version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e7c76e09c1aae2bc52b3d2f29e13c6572553b30c4aa1b8a49fd70de6412654cb"
+checksum = "f2b12d017a929603d80db1831cd3a24082f8137ce19c69e6447f54f5fc8d692f"
 dependencies = [
 "anes",
- "atty",
 "cast",
 "ciborium",
- "clap 3.2.25",
+ "clap",
 "criterion-plot",
+ "is-terminal",
 "itertools",
- "lazy_static",
 "num-traits",
+ "once_cell",
 "oorandom",
 "plotters",
 "rayon",
@@ -1140,7 +1111,7 @@ dependencies = [
 "crossterm_winapi",
 "libc",
 "mio",
- "parking_lot",
+ "parking_lot 0.12.1",
 "signal-hook",
 "signal-hook-mio",
 "winapi",
@@ -1210,7 +1181,7 @@ dependencies = [
 "hashbrown 0.12.3",
 "lock_api",
 "once_cell",
- "parking_lot_core",
+ "parking_lot_core 0.9.7",
 ]

 [[package]]
@@ -1676,15 +1647,6 @@ version = "0.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "95505c38b4572b2d910cecb0281560f54b440a19336cbbcb27bf6ce6adc6f5a8"

-[[package]]
-name = "hermit-abi"
-version = "0.1.19"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "62b467343b94ba476dcb2500d242dadbb39557df889310ac77c5d99100aaac33"
-dependencies = [
- "libc",
-]
-
 [[package]]
 name = "hermit-abi"
 version = "0.2.6"
@@ -1939,6 +1901,9 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7a5bbe824c507c5da5956355e86a746d82e0e1464f65d862cc5e71da70e94b2c"
 dependencies = [
 "cfg-if",
+ "js-sys",
+ "wasm-bindgen",
+ "web-sys",
 ]

 [[package]]
@@ -2267,16 +2232,6 @@ dependencies = [
 "windows-sys 0.45.0",
 ]

-[[package]]
-name = "nu-ansi-term"
-version = "0.46.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "77a8165726e8236064dbb45459242600304b42a5ea24ee2948e18e023bf7ba84"
-dependencies = [
- "overload",
- "winapi",
-]
-
 [[package]]
 name = "num-bigint"
 version = "0.4.3"
@@ -2349,9 +2304,9 @@ checksum = "0ab1bc2a289d34bd04a330323ac98a1b4bc82c9d9fcb1e66b63caa84da26b575"

 [[package]]
 name = "openssl"
-version = "0.10.52"
+version = "0.10.55"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "01b8574602df80f7b85fdfc5392fa884a4e3b3f4f35402c070ab34c3d3f78d56"
+checksum = "345df152bc43501c5eb9e4654ff05f794effb78d4efe3d53abc158baddc0703d"
 dependencies = [
 "bitflags",
 "cfg-if",
@@ -2381,9 +2336,9 @@ checksum = "ff011a302c396a5197692431fc1948019154afc178baf7d8e37367442a4601cf"

 [[package]]
 name = "openssl-sys"
-version = "0.9.87"
+version = "0.9.90"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8e17f59264b2809d77ae94f0e1ebabc434773f370d6ca667bd223ea10e06cc7e"
+checksum = "374533b0e45f3a7ced10fcaeccca020e66656bc03dac384f852e4e5a7a8104a6"
 dependencies = [
 "cc",
 "libc",
@@ -2504,31 +2459,19 @@ dependencies = [
 "winapi",
 ]

-[[package]]
-name = "os_str_bytes"
-version = "6.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ceedf44fb00f2d1984b0bc98102627ce622e083e49a5bacdb3e514fa4238e267"
-
 [[package]]
 name = "outref"
 version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4030760ffd992bef45b0ae3f10ce1aba99e33464c90d14dd7c039884963ddc7a"

-[[package]]
-name = "overload"
-version = "0.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b15813163c1d831bf4a13c3610c05c0d03b39feb07f7e09fa234dac9b15aaf39"
-
 [[package]]
 name = "pagectl"
 version = "0.1.0"
 dependencies = [
 "anyhow",
 "bytes",
- "clap 4.3.0",
+ "clap",
 "git-version",
 "pageserver",
 "postgres_ffi",
@@ -2547,7 +2490,7 @@ dependencies = [
 "byteorder",
 "bytes",
 "chrono",
- "clap 4.3.0",
+ "clap",
 "close_fds",
 "const_format",
 "consumption_metrics",
@@ -2629,6 +2572,17 @@ dependencies = [
 "workspace_hack",
 ]

+[[package]]
+name = "parking_lot"
+version = "0.11.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7d17b78036a60663b797adeaee46f5c9dfebb86948d1255007a1d6be0271ff99"
+dependencies = [
+ "instant",
+ "lock_api",
+ "parking_lot_core 0.8.6",
+]
+
 [[package]]
 name = "parking_lot"
 version = "0.12.1"
@@ -2636,7 +2590,21 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3742b2c103b9f06bc9fff0a37ff4912935851bee6d36f3c02bcc755bcfec228f"
 dependencies = [
 "lock_api",
- "parking_lot_core",
+ "parking_lot_core 0.9.7",
+]
+
+[[package]]
+name = "parking_lot_core"
+version = "0.8.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "60a2cfe6f0ad2bfc16aefa463b497d5c7a5ecd44a23efa72aa342d90177356dc"
+dependencies = [
+ "cfg-if",
+ "instant",
+ "libc",
+ "redox_syscall 0.2.16",
+ "smallvec",
+ "winapi",
 ]

 [[package]]
@@ -2770,7 +2738,7 @@ dependencies = [
 [[package]]
 name = "postgres"
 version = "0.19.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9#2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9"
+source = "git+https://github.com/neondatabase/rust-postgres.git?rev=1aaedab101b23f7612042850d8f2036810fa7c7f#1aaedab101b23f7612042850d8f2036810fa7c7f"
 dependencies = [
 "bytes",
 "fallible-iterator",
@@ -2783,7 +2751,7 @@ dependencies = [
 [[package]]
 name = "postgres-native-tls"
 version = "0.5.0"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9#2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9"
+source = "git+https://github.com/neondatabase/rust-postgres.git?rev=1aaedab101b23f7612042850d8f2036810fa7c7f#1aaedab101b23f7612042850d8f2036810fa7c7f"
 dependencies = [
 "native-tls",
 "tokio",
@@ -2794,7 +2762,7 @@ dependencies = [
 [[package]]
 name = "postgres-protocol"
 version = "0.6.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9#2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9"
+source = "git+https://github.com/neondatabase/rust-postgres.git?rev=1aaedab101b23f7612042850d8f2036810fa7c7f#1aaedab101b23f7612042850d8f2036810fa7c7f"
 dependencies = [
 "base64 0.20.0",
 "byteorder",
@@ -2812,7 +2780,7 @@ dependencies = [
 [[package]]
 name = "postgres-types"
 version = "0.2.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9#2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9"
+source = "git+https://github.com/neondatabase/rust-postgres.git?rev=1aaedab101b23f7612042850d8f2036810fa7c7f#1aaedab101b23f7612042850d8f2036810fa7c7f"
 dependencies = [
 "bytes",
 "fallible-iterator",
@@ -2957,7 +2925,7 @@ dependencies = [
 "lazy_static",
 "libc",
 "memchr",
- "parking_lot",
+ "parking_lot 0.12.1",
 "procfs",
 "thiserror",
 ]
@@ -3022,12 +2990,11 @@ version = "0.1.0"
 dependencies = [
 "anyhow",
 "async-trait",
- "atty",
 "base64 0.13.1",
 "bstr",
 "bytes",
 "chrono",
- "clap 4.3.0",
+ "clap",
 "consumption_metrics",
 "futures",
 "git-version",
@@ -3045,7 +3012,7 @@ dependencies = [
 "native-tls",
 "once_cell",
 "opentelemetry",
- "parking_lot",
+ "parking_lot 0.12.1",
 "pin-project-lite",
 "postgres-native-tls",
 "postgres_backend",
@@ -3056,6 +3023,7 @@ dependencies = [
 "regex",
 "reqwest",
 "reqwest-middleware",
+ "reqwest-retry",
 "reqwest-tracing",
 "routerify",
 "rstest",
@@ -3291,6 +3259,29 @@ dependencies = [
 "thiserror",
 ]

+[[package]]
+name = "reqwest-retry"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "48d0fd6ef4c6d23790399fe15efc8d12cd9f3d4133958f9bd7801ee5cbaec6c4"
+dependencies = [
+ "anyhow",
+ "async-trait",
+ "chrono",
+ "futures",
+ "getrandom",
+ "http",
+ "hyper",
+ "parking_lot 0.11.2",
+ "reqwest",
+ "reqwest-middleware",
+ "retry-policies",
+ "task-local-extensions",
+ "tokio",
+ "tracing",
+ "wasm-timer",
+]
+
 [[package]]
 name = "reqwest-tracing"
 version = "0.4.4"
@@ -3309,6 +3300,17 @@ dependencies = [
 "tracing-opentelemetry",
 ]

+[[package]]
+name = "retry-policies"
+version = "0.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e09bbcb5003282bcb688f0bae741b278e9c7e8f378f561522c9806c58e075d9b"
+dependencies = [
+ "anyhow",
+ "chrono",
+ "rand",
+]
+
 [[package]]
 name = "ring"
 version = "0.16.20"
@@ -3507,7 +3509,7 @@ dependencies = [
 "byteorder",
 "bytes",
 "chrono",
- "clap 4.3.0",
+ "clap",
 "const_format",
 "crc32c",
 "fs2",
@@ -3518,7 +3520,7 @@ dependencies = [
 "hyper",
 "metrics",
 "once_cell",
- "parking_lot",
+ "parking_lot 0.12.1",
 "postgres",
 "postgres-protocol",
 "postgres_backend",
@@ -3937,7 +3939,7 @@ dependencies = [
 "anyhow",
 "async-stream",
 "bytes",
- "clap 4.3.0",
+ "clap",
 "const_format",
 "futures",
 "futures-core",
@@ -3947,7 +3949,7 @@ dependencies = [
 "hyper",
 "metrics",
 "once_cell",
- "parking_lot",
+ "parking_lot 0.12.1",
 "prost",
 "tokio",
 "tokio-stream",
@@ -4118,12 +4120,6 @@ dependencies = [
 "syn 1.0.109",
 ]

-[[package]]
-name = "textwrap"
-version = "0.16.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "222a222a5bfe1bba4a77b45ec488a741b3cb8872e5e499451fd7d0129c9c7c3d"
-
 [[package]]
 name = "thiserror"
 version = "1.0.40"
@@ -4272,7 +4268,7 @@ dependencies = [
 [[package]]
 name = "tokio-postgres"
 version = "0.7.7"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9#2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9"
+source = "git+https://github.com/neondatabase/rust-postgres.git?rev=1aaedab101b23f7612042850d8f2036810fa7c7f#1aaedab101b23f7612042850d8f2036810fa7c7f"
 dependencies = [
 "async-trait",
 "byteorder",
@@ -4281,7 +4277,7 @@ dependencies = [
 "futures-channel",
 "futures-util",
 "log",
- "parking_lot",
+ "parking_lot 0.12.1",
 "percent-encoding",
 "phf",
 "pin-project-lite",
@@ -4539,7 +4535,7 @@ name = "trace"
 version = "0.1.0"
 dependencies = [
 "anyhow",
- "clap 4.3.0",
+ "clap",
 "pageserver_api",
 "utils",
 "workspace_hack",
@@ -4641,7 +4637,6 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "30a651bc37f915e81f087d86e62a18eec5f79550c7faff886f7090b4ea757c77"
 dependencies = [
 "matchers",
- "nu-ansi-term",
 "once_cell",
 "regex",
 "serde",
@@ -4810,7 +4805,6 @@ version = "0.1.0"
 dependencies = [
 "anyhow",
 "async-trait",
- "atty",
 "bincode",
 "byteorder",
 "bytes",
@@ -4887,7 +4881,7 @@ name = "wal_craft"
 version = "0.1.0"
 dependencies = [
 "anyhow",
- "clap 4.3.0",
+ "clap",
 "env_logger",
 "log",
 "once_cell",
@@ -4991,6 +4985,21 @@ version = "0.2.86"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ed9d5b4305409d1fc9482fee2d7f9bcbf24b3972bf59817ef757e23982242a93"

+[[package]]
+name = "wasm-timer"
+version = "0.2.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "be0ecb0db480561e9a7642b5d3e4187c128914e58aa84330b9493e3eb68c5e7f"
+dependencies = [
+ "futures",
+ "js-sys",
+ "parking_lot 0.11.2",
+ "pin-utils",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "web-sys",
+]
+
 [[package]]
 name = "web-sys"
 version = "0.3.63"
@@ -5252,7 +5261,7 @@ dependencies = [
 "anyhow",
 "bytes",
 "chrono",
- "clap 4.3.0",
+ "clap",
 "clap_builder",
 "crossbeam-utils",
 "either",
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -34,7 +34,6 @@ license = "Apache-2.0"
 anyhow = { version = "1.0", features = ["backtrace"] }
 async-stream = "0.3"
 async-trait = "0.1"
-atty = "0.2.14"
 aws-config = { version = "0.55", default-features = false, features=["rustls"] }
 aws-sdk-s3 = "0.27"
 aws-smithy-http = "0.55"
@@ -95,6 +94,7 @@ regex = "1.4"
 reqwest = { version = "0.11", default-features = false, features = ["rustls-tls"] }
 reqwest-tracing = { version = "0.4.0", features = ["opentelemetry_0_18"] }
 reqwest-middleware = "0.2.0"
+reqwest-retry = "0.2.2"
 routerify = "3"
 rpds = "0.13"
 rustls = "0.20"
@@ -128,7 +128,7 @@ tonic = {version = "0.9", features = ["tls", "tls-roots"]}
 tracing = "0.1"
 tracing-error = "0.2.0"
 tracing-opentelemetry = "0.18.0"
-tracing-subscriber = { version = "0.3", features = ["env-filter"] }
+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"] }
 walkdir = "2.3.2"
@@ -140,11 +140,11 @@ env_logger = "0.10"
 log = "0.4"

 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9" }
-postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9" }
-postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9" }
-postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9" }
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9" }
+postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
+postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
+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
@@ -170,7 +170,7 @@ utils = { version = "0.1", path = "./libs/utils/" }
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }

 ## Build dependencies
-criterion = "0.4"
+criterion = "0.5.1"
 rcgen = "0.10"
 rstest = "0.17"
 tempfile = "3.4"
@@ -180,7 +180,7 @@ tonic-build = "0.9"

 # This is only needed for proxy's tests.
 # TODO: we should probably fork `tokio-postgres-rustls` instead.
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="2e9b5f1ddc481d1a98fa79f6b9378ac4f170b7c9" }
+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,
--- a/Dockerfile.compute-node
+++ b/Dockerfile.compute-node
@@ -2,6 +2,7 @@ ARG PG_VERSION
 ARG REPOSITORY=neondatabase
 ARG IMAGE=rust
 ARG TAG=pinned
+ARG BUILD_TAG

 #########################################################################################
 #
@@ -67,7 +68,7 @@ RUN apt update && \
 RUN wget https://gitlab.com/Oslandia/SFCGAL/-/archive/v1.3.10/SFCGAL-v1.3.10.tar.gz -O SFCGAL.tar.gz && \
    echo "4e39b3b2adada6254a7bdba6d297bb28e1a9835a9f879b74f37e2dab70203232 SFCGAL.tar.gz" | sha256sum --check && \
    mkdir sfcgal-src && cd sfcgal-src && tar xvzf ../SFCGAL.tar.gz --strip-components=1 -C . && \
-    cmake . && make -j $(getconf _NPROCESSORS_ONLN) && \
+    cmake -DCMAKE_BUILD_TYPE=Release . && make -j $(getconf _NPROCESSORS_ONLN) && \
    DESTDIR=/sfcgal make install -j $(getconf _NPROCESSORS_ONLN) && \
    make clean && cp -R /sfcgal/* /

@@ -95,7 +96,7 @@ RUN wget https://github.com/pgRouting/pgrouting/archive/v3.4.2.tar.gz -O pgrouti
    mkdir pgrouting-src && cd pgrouting-src && tar xvzf ../pgrouting.tar.gz --strip-components=1 -C . && \
    mkdir build && \
    cd build && \
-    cmake .. && \
+    cmake -DCMAKE_BUILD_TYPE=Release .. && \
    make -j $(getconf _NPROCESSORS_ONLN) && \
    make -j $(getconf _NPROCESSORS_ONLN) install && \
    echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgrouting.control
@@ -355,7 +356,7 @@ RUN apt-get update && \
    wget https://github.com/timescale/timescaledb/archive/refs/tags/2.10.1.tar.gz -O timescaledb.tar.gz && \
    echo "6fca72a6ed0f6d32d2b3523951ede73dc5f9b0077b38450a029a5f411fdb8c73 timescaledb.tar.gz" | sha256sum --check && \
    mkdir timescaledb-src && cd timescaledb-src && tar xvzf ../timescaledb.tar.gz --strip-components=1 -C . && \
-    ./bootstrap -DSEND_TELEMETRY_DEFAULT:BOOL=OFF -DUSE_TELEMETRY:BOOL=OFF -DAPACHE_ONLY:BOOL=ON && \
+    ./bootstrap -DSEND_TELEMETRY_DEFAULT:BOOL=OFF -DUSE_TELEMETRY:BOOL=OFF -DAPACHE_ONLY:BOOL=ON -DCMAKE_BUILD_TYPE=Release && \
    cd build && \
    make -j $(getconf _NPROCESSORS_ONLN) && \
    make install -j $(getconf _NPROCESSORS_ONLN) && \
@@ -410,7 +411,7 @@ RUN apt-get update && \
    mkdir kq_imcx-src && cd kq_imcx-src && tar xvzf ../kq_imcx.tar.gz --strip-components=1 -C . && \
    mkdir build && \
    cd build && \
-    cmake .. && \
+    cmake -DCMAKE_BUILD_TYPE=Release .. && \
    make -j $(getconf _NPROCESSORS_ONLN) && \
    make -j $(getconf _NPROCESSORS_ONLN) install && \
    echo 'trusted = true' >> /usr/local/pgsql/share/extension/kq_imcx.control
@@ -432,6 +433,108 @@ RUN wget https://github.com/citusdata/pg_cron/archive/refs/tags/v1.5.2.tar.gz -O
    make -j $(getconf _NPROCESSORS_ONLN) install && \
    echo 'trusted = true' >> /usr/local/pgsql/share/extension/pg_cron.control

+#########################################################################################
+#
+# Layer "rdkit-pg-build"
+# compile rdkit extension
+#
+#########################################################################################
+FROM build-deps AS rdkit-pg-build
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+RUN apt-get update && \
+    apt-get install -y \
+        cmake \
+        libboost-iostreams1.74-dev \
+        libboost-regex1.74-dev \
+        libboost-serialization1.74-dev \
+        libboost-system1.74-dev \
+        libeigen3-dev \
+        libfreetype6-dev
+
+ENV PATH "/usr/local/pgsql/bin/:/usr/local/pgsql/:$PATH"
+RUN wget https://github.com/rdkit/rdkit/archive/refs/tags/Release_2023_03_1.tar.gz -O rdkit.tar.gz && \
+    echo "db346afbd0ba52c843926a2a62f8a38c7b774ffab37eaf382d789a824f21996c rdkit.tar.gz" | sha256sum --check && \
+    mkdir rdkit-src && cd rdkit-src && tar xvzf ../rdkit.tar.gz --strip-components=1 -C . && \
+    cmake \
+        -D RDK_BUILD_CAIRO_SUPPORT=OFF \
+        -D RDK_BUILD_INCHI_SUPPORT=ON \
+        -D RDK_BUILD_AVALON_SUPPORT=ON \
+        -D RDK_BUILD_PYTHON_WRAPPERS=OFF \
+        -D RDK_BUILD_DESCRIPTORS3D=OFF \
+        -D RDK_BUILD_FREESASA_SUPPORT=OFF \
+        -D RDK_BUILD_COORDGEN_SUPPORT=ON \
+        -D RDK_BUILD_MOLINTERCHANGE_SUPPORT=OFF \
+        -D RDK_BUILD_YAEHMOP_SUPPORT=OFF \
+        -D RDK_BUILD_STRUCTCHECKER_SUPPORT=OFF \
+        -D RDK_USE_URF=OFF \
+        -D RDK_BUILD_PGSQL=ON \
+        -D RDK_PGSQL_STATIC=ON \
+        -D PostgreSQL_CONFIG=pg_config \
+        -D PostgreSQL_INCLUDE_DIR=`pg_config --includedir` \
+        -D PostgreSQL_TYPE_INCLUDE_DIR=`pg_config --includedir-server` \
+        -D PostgreSQL_LIBRARY_DIR=`pg_config --libdir` \
+        -D RDK_INSTALL_INTREE=OFF \
+        -D CMAKE_BUILD_TYPE=Release \
+        . && \
+    make -j $(getconf _NPROCESSORS_ONLN) && \
+    make -j $(getconf _NPROCESSORS_ONLN) install && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/rdkit.control
+
+#########################################################################################
+#
+# Layer "pg-uuidv7-pg-build"
+# compile pg_uuidv7 extension
+#
+#########################################################################################
+FROM build-deps AS pg-uuidv7-pg-build
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+ENV PATH "/usr/local/pgsql/bin/:$PATH"
+RUN wget https://github.com/fboulnois/pg_uuidv7/archive/refs/tags/v1.0.1.tar.gz -O pg_uuidv7.tar.gz && \
+    echo "0d0759ab01b7fb23851ecffb0bce27822e1868a4a5819bfd276101c716637a7a pg_uuidv7.tar.gz" | sha256sum --check && \
+    mkdir pg_uuidv7-src && cd pg_uuidv7-src && tar xvzf ../pg_uuidv7.tar.gz --strip-components=1 -C . && \
+    make -j $(getconf _NPROCESSORS_ONLN) && \
+    make -j $(getconf _NPROCESSORS_ONLN) install && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/pg_uuidv7.control
+
+#########################################################################################
+#
+# Layer "pg-roaringbitmap-pg-build"
+# compile pg_roaringbitmap extension
+#
+#########################################################################################
+FROM build-deps AS pg-roaringbitmap-pg-build
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+ENV PATH "/usr/local/pgsql/bin/:$PATH"
+RUN wget https://github.com/ChenHuajun/pg_roaringbitmap/archive/refs/tags/v0.5.4.tar.gz -O pg_roaringbitmap.tar.gz && \
+    echo "b75201efcb1c2d1b014ec4ae6a22769cc7a224e6e406a587f5784a37b6b5a2aa pg_roaringbitmap.tar.gz" | sha256sum --check && \
+    mkdir pg_roaringbitmap-src && cd pg_roaringbitmap-src && tar xvzf ../pg_roaringbitmap.tar.gz --strip-components=1 -C . && \
+    make -j $(getconf _NPROCESSORS_ONLN) && \
+    make -j $(getconf _NPROCESSORS_ONLN) install && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/roaringbitmap.control
+
+#########################################################################################
+#
+# Layer "pg-anon-pg-build"
+# compile anon extension
+#
+#########################################################################################
+FROM build-deps AS pg-anon-pg-build
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+# Kaniko doesn't allow to do `${from#/usr/local/pgsql/}`, so we use `${from:17}` instead
+ENV PATH "/usr/local/pgsql/bin/:$PATH"
+RUN wget https://gitlab.com/dalibo/postgresql_anonymizer/-/archive/1.1.0/postgresql_anonymizer-1.1.0.tar.gz -O pg_anon.tar.gz && \
+    echo "08b09d2ff9b962f96c60db7e6f8e79cf7253eb8772516998fc35ece08633d3ad pg_anon.tar.gz" | sha256sum --check && \
+    mkdir pg_anon-src && cd pg_anon-src && tar xvzf ../pg_anon.tar.gz --strip-components=1 -C . && \
+    find /usr/local/pgsql -type f | sort  > /before.txt && \
+    make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/anon.control && \
+    find /usr/local/pgsql -type f | sort  > /after.txt && \
+    /bin/bash -c 'for from in $(comm -13 /before.txt /after.txt); do to=/extensions/anon/${from:17} && mkdir -p $(dirname ${to}) && cp -a ${from} ${to}; done'
+
 #########################################################################################
 #
 # Layer "rust extensions"
@@ -540,6 +643,7 @@ RUN wget https://github.com/pksunkara/pgx_ulid/archive/refs/tags/v0.1.0.tar.gz -
 #
 #########################################################################################
 FROM build-deps AS neon-pg-ext-build
+# Public extensions
 COPY --from=postgis-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=postgis-build /sfcgal/* /
 COPY --from=plv8-build /usr/local/pgsql/ /usr/local/pgsql/
@@ -564,6 +668,9 @@ COPY --from=pg-hint-plan-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=kq-imcx-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-cron-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-pgx-ulid-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=rdkit-pg-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=pg-uuidv7-pg-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=pg-roaringbitmap-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY pgxn/ pgxn/

 RUN make -j $(getconf _NPROCESSORS_ONLN) \
@@ -585,6 +692,9 @@ RUN make -j $(getconf _NPROCESSORS_ONLN) \
 #
 #########################################################################################
 FROM $REPOSITORY/$IMAGE:$TAG AS compute-tools
+ARG BUILD_TAG
+ENV BUILD_TAG=$BUILD_TAG
+
 USER nonroot
 # Copy entire project to get Cargo.* files with proper dependencies for the whole project
 COPY --chown=nonroot . .
@@ -609,6 +719,22 @@ RUN rm -r /usr/local/pgsql/include
 # if they were to be used by other libraries.
 RUN rm /usr/local/pgsql/lib/lib*.a

+#########################################################################################
+#
+# Extenstion only
+#
+#########################################################################################
+FROM scratch AS postgres-extensions
+# After the transition this layer will include all extensitons.
+# As for now, it's only for new custom ones
+#
+# # Default extensions
+# COPY --from=postgres-cleanup-layer /usr/local/pgsql/share/extension /usr/local/pgsql/share/extension
+# COPY --from=postgres-cleanup-layer /usr/local/pgsql/lib             /usr/local/pgsql/lib
+# Custom extensions
+COPY --from=pg-anon-pg-build /extensions/anon/lib/ /extensions/anon/lib
+COPY --from=pg-anon-pg-build /extensions/anon/share/extension /extensions/anon/share/extension
+
 #########################################################################################
 #
 # Final layer
@@ -637,14 +763,19 @@ COPY --from=compute-tools --chown=postgres /home/nonroot/target/release-line-deb
 # libgeos, libgdal, libsfcgal1, libproj and libprotobuf-c1 for PostGIS
 # libxml2, libxslt1.1 for xml2
 # libzstd1 for zstd
+# libboost*, libfreetype6, and zlib1g for rdkit
 RUN apt update &&  \
    apt install --no-install-recommends -y \
        gdb \
-        locales \
        libicu67 \
        liblz4-1 \
        libreadline8 \
+        libboost-iostreams1.74.0 \
+        libboost-regex1.74.0 \
+        libboost-serialization1.74.0 \
+        libboost-system1.74.0 \
        libossp-uuid16 \
+        libfreetype6 \
        libgeos-c1v5 \
        libgdal28 \
        libproj19 \
@@ -654,7 +785,9 @@ RUN apt update &&  \
        libxslt1.1 \
        libzstd1 \
        libcurl4-openssl-dev \
-        procps && \
+        locales \
+        procps \
+        zlib1g && \
    rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/* && \
    localedef -i en_US -c -f UTF-8 -A /usr/share/locale/locale.alias en_US.UTF-8

--- a/Dockerfile.compute-tools
+++ b/Dockerfile.compute-tools
@@ -3,6 +3,7 @@
 ARG REPOSITORY=neondatabase
 ARG IMAGE=rust
 ARG TAG=pinned
+ARG BUILD_TAG

 FROM $REPOSITORY/$IMAGE:$TAG AS rust-build
 WORKDIR /home/nonroot
@@ -16,6 +17,8 @@ ENV CACHEPOT_S3_KEY_PREFIX=cachepot
 ARG CACHEPOT_BUCKET=neon-github-dev
 #ARG AWS_ACCESS_KEY_ID
 #ARG AWS_SECRET_ACCESS_KEY
+ARG BUILD_TAG
+ENV BUILD_TAG=$BUILD_TAG

 COPY . .

--- a/README.md
+++ b/README.md
@@ -132,13 +132,13 @@ Python (3.9 or higher), and install python3 packages using `./scripts/pysync` (r
 # Create repository in .neon with proper paths to binaries and data
 # Later that would be responsibility of a package install script
 > cargo neon init
-Starting pageserver at '127.0.0.1:64000' in '.neon'.
+Initializing pageserver node 1 at '127.0.0.1:64000' in ".neon"

 # start pageserver, safekeeper, and broker for their intercommunication
 > cargo neon start
-Starting neon broker at 127.0.0.1:50051
+Starting neon broker at 127.0.0.1:50051.
 storage_broker started, pid: 2918372
-Starting pageserver at '127.0.0.1:64000' in '.neon'.
+Starting pageserver node 1 at '127.0.0.1:64000' in ".neon".
 pageserver started, pid: 2918386
 Starting safekeeper at '127.0.0.1:5454' in '.neon/safekeepers/sk1'.
 safekeeper 1 started, pid: 2918437
@@ -152,8 +152,7 @@ Setting tenant 9ef87a5bf0d92544f6fafeeb3239695c as a default one
 # start postgres compute node
 > cargo neon endpoint start main
 Starting new endpoint main (PostgreSQL v14) on timeline de200bd42b49cc1814412c7e592dd6e9 ...
-Extracting base backup to create postgres instance: path=.neon/pgdatadirs/tenants/9ef87a5bf0d92544f6fafeeb3239695c/main port=55432
-Starting postgres at 'host=127.0.0.1 port=55432 user=cloud_admin dbname=postgres'
+Starting postgres at 'postgresql://cloud_admin@127.0.0.1:55432/postgres'

 # check list of running postgres instances
 > cargo neon endpoint list
@@ -189,18 +188,17 @@ Created timeline 'b3b863fa45fa9e57e615f9f2d944e601' at Lsn 0/16F9A00 for tenant:
 # start postgres on that branch
 > cargo neon endpoint start migration_check --branch-name migration_check
 Starting new endpoint migration_check (PostgreSQL v14) on timeline b3b863fa45fa9e57e615f9f2d944e601 ...
-Extracting base backup to create postgres instance: path=.neon/pgdatadirs/tenants/9ef87a5bf0d92544f6fafeeb3239695c/migration_check port=55433
-Starting postgres at 'host=127.0.0.1 port=55433 user=cloud_admin dbname=postgres'
+Starting postgres at 'postgresql://cloud_admin@127.0.0.1:55434/postgres'

 # check the new list of running postgres instances
 > cargo neon endpoint list
 ENDPOINT         ADDRESS          TIMELINE                          BRANCH NAME      LSN        STATUS
 main             127.0.0.1:55432  de200bd42b49cc1814412c7e592dd6e9  main             0/16F9A38  running
- migration_check  127.0.0.1:55433  b3b863fa45fa9e57e615f9f2d944e601  migration_check  0/16F9A70  running
+ migration_check  127.0.0.1:55434  b3b863fa45fa9e57e615f9f2d944e601  migration_check  0/16F9A70  running

 # this new postgres instance will have all the data from 'main' postgres,
 # but all modifications would not affect data in original postgres
-> psql -p55433 -h 127.0.0.1 -U cloud_admin postgres
+> psql -p55434 -h 127.0.0.1 -U cloud_admin postgres
 postgres=# select * from t;
 key | value
 -----+-------
--- a/compute_tools/Cargo.toml
+++ b/compute_tools/Cargo.toml
@@ -30,3 +30,5 @@ url.workspace = true
 compute_api.workspace = true
 utils.workspace = true
 workspace_hack.workspace = true
+toml_edit.workspace = true
+remote_storage = { version = "0.1", path = "../libs/remote_storage/" }
--- a/compute_tools/src/bin/compute_ctl.rs
+++ b/compute_tools/src/bin/compute_ctl.rs
@@ -5,6 +5,8 @@
 //! - `compute_ctl` accepts cluster (compute node) specification as a JSON file.
 //! - Every start is a fresh start, so the data directory is removed and
 //!   initialized again on each run.
+//! - If remote_extension_config is provided, it will be used to fetch extensions list
+//!  and download `shared_preload_libraries` from the remote storage.
 //! - Next it will put configuration files into the `PGDATA` directory.
 //! - Sync safekeepers and get commit LSN.
 //! - Get `basebackup` from pageserver using the returned on the previous step LSN.
@@ -27,7 +29,8 @@
 //! compute_ctl -D /var/db/postgres/compute \
 //!             -C 'postgresql://cloud_admin@localhost/postgres' \
 //!             -S /var/db/postgres/specs/current.json \
-//!             -b /usr/local/bin/postgres
+//!             -b /usr/local/bin/postgres \
+//!             -r {"bucket": "my-bucket", "region": "eu-central-1", "endpoint": "http:://localhost:9000"} \
 //! ```
 //!
 use std::collections::HashMap;
@@ -35,7 +38,7 @@ use std::fs::File;
 use std::panic;
 use std::path::Path;
 use std::process::exit;
-use std::sync::{mpsc, Arc, Condvar, Mutex};
+use std::sync::{mpsc, Arc, Condvar, Mutex, OnceLock};
 use std::{thread, time::Duration};

 use anyhow::{Context, Result};
@@ -48,16 +51,32 @@ use compute_api::responses::ComputeStatus;

 use compute_tools::compute::{ComputeNode, ComputeState, ParsedSpec};
 use compute_tools::configurator::launch_configurator;
+use compute_tools::extension_server::launch_download_extensions;
+use compute_tools::extension_server::{get_pg_version, init_remote_storage};
 use compute_tools::http::api::launch_http_server;
 use compute_tools::logger::*;
 use compute_tools::monitor::launch_monitor;
 use compute_tools::params::*;
 use compute_tools::spec::*;

+const BUILD_TAG_DEFAULT: &str = "local";
+
 fn main() -> Result<()> {
    init_tracing_and_logging(DEFAULT_LOG_LEVEL)?;

+    let build_tag = option_env!("BUILD_TAG").unwrap_or(BUILD_TAG_DEFAULT);
+
+    info!("build_tag: {build_tag}");
+
    let matches = cli().get_matches();
+    let pgbin_default = String::from("postgres");
+    let pgbin = matches.get_one::<String>("pgbin").unwrap_or(&pgbin_default);
+
+    let remote_ext_config = matches.get_one::<String>("remote-ext-config");
+    let ext_remote_storage = remote_ext_config.map(|x| {
+        init_remote_storage(x, build_tag)
+            .expect("cannot initialize remote extension storage from config")
+    });

    let http_port = *matches
        .get_one::<u16>("http-port")
@@ -122,9 +141,6 @@ fn main() -> Result<()> {
    let compute_id = matches.get_one::<String>("compute-id");
    let control_plane_uri = matches.get_one::<String>("control-plane-uri");

-    // Try to use just 'postgres' if no path is provided
-    let pgbin = matches.get_one::<String>("pgbin").unwrap();
-
    let spec;
    let mut live_config_allowed = false;
    match spec_json {
@@ -162,6 +178,7 @@ fn main() -> Result<()> {

    let mut new_state = ComputeState::new();
    let spec_set;
+
    if let Some(spec) = spec {
        let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
        new_state.pspec = Some(pspec);
@@ -173,9 +190,13 @@ fn main() -> Result<()> {
        connstr: Url::parse(connstr).context("cannot parse connstr as a URL")?,
        pgdata: pgdata.to_string(),
        pgbin: pgbin.to_string(),
+        pgversion: get_pg_version(pgbin),
        live_config_allowed,
        state: Mutex::new(new_state),
        state_changed: Condvar::new(),
+        ext_remote_storage,
+        available_libraries: OnceLock::new(),
+        available_extensions: OnceLock::new(),
    };
    let compute = Arc::new(compute_node);

@@ -184,6 +205,8 @@ fn main() -> Result<()> {
    let _http_handle =
        launch_http_server(http_port, &compute).expect("cannot launch http endpoint thread");

+    let extension_server_port: u16 = http_port;
+
    if !spec_set {
        // No spec provided, hang waiting for it.
        info!("no compute spec provided, waiting");
@@ -221,10 +244,13 @@ fn main() -> Result<()> {
    let _configurator_handle =
        launch_configurator(&compute).expect("cannot launch configurator thread");

+    let _download_extensions_handle =
+        launch_download_extensions(&compute).expect("cannot launch download extensions thread");
+
    // Start Postgres
    let mut delay_exit = false;
    let mut exit_code = None;
-    let pg = match compute.start_compute() {
+    let pg = match compute.start_compute(extension_server_port) {
        Ok(pg) => Some(pg),
        Err(err) => {
            error!("could not start the compute node: {:?}", err);
@@ -250,6 +276,16 @@ fn main() -> Result<()> {
        exit_code = ecode.code()
    }

+    // Maybe sync safekeepers again, to speed up next startup
+    let compute_state = compute.state.lock().unwrap().clone();
+    let pspec = compute_state.pspec.as_ref().expect("spec must be set");
+    if matches!(pspec.spec.mode, compute_api::spec::ComputeMode::Primary) {
+        info!("syncing safekeepers on shutdown");
+        let storage_auth_token = pspec.storage_auth_token.clone();
+        let lsn = compute.sync_safekeepers(storage_auth_token)?;
+        info!("synced safekeepers at lsn {lsn}");
+    }
+
    if let Err(err) = compute.check_for_core_dumps() {
        error!("error while checking for core dumps: {err:?}");
    }
@@ -343,6 +379,12 @@ fn cli() -> clap::Command {
                .long("control-plane-uri")
                .value_name("CONTROL_PLANE_API_BASE_URI"),
        )
+        .arg(
+            Arg::new("remote-ext-config")
+                .short('r')
+                .long("remote-ext-config")
+                .value_name("REMOTE_EXT_CONFIG"),
+        )
 }

 #[test]
--- a/compute_tools/src/compute.rs
+++ b/compute_tools/src/compute.rs
@@ -1,13 +1,15 @@
+use std::collections::HashMap;
 use std::fs;
 use std::os::unix::fs::PermissionsExt;
 use std::path::Path;
 use std::process::{Command, Stdio};
 use std::str::FromStr;
-use std::sync::{Condvar, Mutex};
+use std::sync::{Condvar, Mutex, OnceLock};

 use anyhow::{Context, Result};
 use chrono::{DateTime, Utc};
 use postgres::{Client, NoTls};
+use tokio;
 use tokio_postgres;
 use tracing::{info, instrument, warn};
 use utils::id::{TenantId, TimelineId};
@@ -16,9 +18,12 @@ use utils::lsn::Lsn;
 use compute_api::responses::{ComputeMetrics, ComputeStatus};
 use compute_api::spec::{ComputeMode, ComputeSpec};

-use crate::config;
+use remote_storage::{GenericRemoteStorage, RemotePath};
+
+use crate::extension_server::PathAndFlag;
 use crate::pg_helpers::*;
 use crate::spec::*;
+use crate::{config, extension_server};

 /// Compute node info shared across several `compute_ctl` threads.
 pub struct ComputeNode {
@@ -26,6 +31,7 @@ pub struct ComputeNode {
    pub connstr: url::Url,
    pub pgdata: String,
    pub pgbin: String,
+    pub pgversion: String,
    /// We should only allow live re- / configuration of the compute node if
    /// it uses 'pull model', i.e. it can go to control-plane and fetch
    /// the latest configuration. Otherwise, there could be a case:
@@ -45,6 +51,11 @@ pub struct ComputeNode {
    pub state: Mutex<ComputeState>,
    /// `Condvar` to allow notifying waiters about state changes.
    pub state_changed: Condvar,
+    ///  the S3 bucket that we search for extensions in
+    pub ext_remote_storage: Option<GenericRemoteStorage>,
+    // cached lists of available extensions and libraries
+    pub available_libraries: OnceLock<HashMap<String, Vec<RemotePath>>>,
+    pub available_extensions: OnceLock<HashMap<String, Vec<PathAndFlag>>>,
 }

 #[derive(Clone, Debug)]
@@ -133,6 +144,84 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
    }
 }

+/// Create special neon_superuser role, that's a slightly nerfed version of a real superuser
+/// that we give to customers
+fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
+    let roles = spec
+        .cluster
+        .roles
+        .iter()
+        .map(|r| format!("'{}'", escape_literal(&r.name)))
+        .collect::<Vec<_>>();
+
+    let dbs = spec
+        .cluster
+        .databases
+        .iter()
+        .map(|db| format!("'{}'", escape_literal(&db.name)))
+        .collect::<Vec<_>>();
+
+    let roles_decl = if roles.is_empty() {
+        String::from("roles text[] := NULL;")
+    } else {
+        format!(
+            r#"
+               roles text[] := ARRAY(SELECT rolname
+                                     FROM pg_catalog.pg_roles
+                                     WHERE rolname IN ({}));"#,
+            roles.join(", ")
+        )
+    };
+
+    let database_decl = if dbs.is_empty() {
+        String::from("dbs text[] := NULL;")
+    } else {
+        format!(
+            r#"
+               dbs text[] := ARRAY(SELECT datname
+                                   FROM pg_catalog.pg_database
+                                   WHERE datname IN ({}));"#,
+            dbs.join(", ")
+        )
+    };
+
+    // ALL PRIVILEGES grants CREATE, CONNECT, and TEMPORARY on all databases
+    // (see https://www.postgresql.org/docs/current/ddl-priv.html)
+    let query = format!(
+        r#"
+            DO $$
+                DECLARE
+                    r text;
+                    {}
+                    {}
+                BEGIN
+                    IF NOT EXISTS (
+                        SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
+                    THEN
+                        CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN IN ROLE pg_read_all_data, pg_write_all_data;
+                        IF array_length(roles, 1) IS NOT NULL THEN
+                            EXECUTE format('GRANT neon_superuser TO %s',
+                                           array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(roles) as x), ', '));
+                            FOREACH r IN ARRAY roles LOOP
+                                EXECUTE format('ALTER ROLE %s CREATEROLE CREATEDB', quote_ident(r));
+                            END LOOP;
+                        END IF;
+                        IF array_length(dbs, 1) IS NOT NULL THEN
+                            EXECUTE format('GRANT ALL PRIVILEGES ON DATABASE %s TO neon_superuser',
+                                           array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(dbs) as x), ', '));
+                        END IF;
+                    END IF;
+                END
+            $$;"#,
+        roles_decl, database_decl,
+    );
+    info!("Neon superuser created:\n{}", &query);
+    client
+        .simple_query(&query)
+        .map_err(|e| anyhow::anyhow!(e).context(query))?;
+    Ok(())
+}
+
 impl ComputeNode {
    pub fn set_status(&self, status: ComputeStatus) {
        let mut state = self.state.lock().unwrap();
@@ -157,7 +246,7 @@ impl ComputeNode {

    // Get basebackup from the libpq connection to pageserver using `connstr` and
    // unarchive it to `pgdata` directory overriding all its previous content.
-    #[instrument(skip(self, compute_state))]
+    #[instrument(skip_all, fields(%lsn))]
    fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
        let spec = compute_state.pspec.as_ref().expect("spec must be set");
        let start_time = Utc::now();
@@ -199,8 +288,8 @@ impl ComputeNode {

    // Run `postgres` in a special mode with `--sync-safekeepers` argument
    // and return the reported LSN back to the caller.
-    #[instrument(skip(self, storage_auth_token))]
-    fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
+    #[instrument(skip_all)]
+    pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
        let start_time = Utc::now();

        let sync_handle = Command::new(&self.pgbin)
@@ -244,15 +333,23 @@ impl ComputeNode {

    /// Do all the preparations like PGDATA directory creation, configuration,
    /// safekeepers sync, basebackup, etc.
-    #[instrument(skip(self, compute_state))]
-    pub fn prepare_pgdata(&self, compute_state: &ComputeState) -> Result<()> {
+    #[instrument(skip_all)]
+    pub fn prepare_pgdata(
+        &self,
+        compute_state: &ComputeState,
+        extension_server_port: u16,
+    ) -> Result<()> {
        let pspec = compute_state.pspec.as_ref().expect("spec must be set");
        let spec = &pspec.spec;
        let pgdata_path = Path::new(&self.pgdata);

        // Remove/create an empty pgdata directory and put configuration there.
        self.create_pgdata()?;
-        config::write_postgres_conf(&pgdata_path.join("postgresql.conf"), &pspec.spec)?;
+        config::write_postgres_conf(
+            &pgdata_path.join("postgresql.conf"),
+            &pspec.spec,
+            Some(extension_server_port),
+        )?;

        // Syncing safekeepers is only safe with primary nodes: if a primary
        // is already connected it will be kicked out, so a secondary (standby)
@@ -302,7 +399,7 @@ impl ComputeNode {

    /// Start Postgres as a child process and manage DBs/roles.
    /// After that this will hang waiting on the postmaster process to exit.
-    #[instrument(skip(self))]
+    #[instrument(skip_all)]
    pub fn start_postgres(
        &self,
        storage_auth_token: Option<String>,
@@ -326,7 +423,7 @@ impl ComputeNode {
    }

    /// Do initial configuration of the already started Postgres.
-    #[instrument(skip(self, compute_state))]
+    #[instrument(skip_all)]
    pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
        // If connection fails,
        // it may be the old node with `zenith_admin` superuser.
@@ -347,6 +444,8 @@ impl ComputeNode {
                    .map_err(|_| anyhow::anyhow!("invalid connstr"))?;

                let mut client = Client::connect(zenith_admin_connstr.as_str(), NoTls)?;
+                // Disable forwarding so that users don't get a cloud_admin role
+                client.simple_query("SET neon.forward_ddl = false")?;
                client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
                client.simple_query("GRANT zenith_admin TO cloud_admin")?;
                drop(client);
@@ -357,31 +456,28 @@ impl ComputeNode {
            Ok(client) => client,
        };

-        // Proceed with post-startup configuration. Note, that order of operations is important.
        // Disable DDL forwarding because control plane already knows about these roles/databases.
        client.simple_query("SET neon.forward_ddl = false")?;
+
+        // Proceed with post-startup configuration. Note, that order of operations is important.
        let spec = &compute_state.pspec.as_ref().expect("spec must be set").spec;
+        create_neon_superuser(spec, &mut client)?;
        handle_roles(spec, &mut client)?;
        handle_databases(spec, &mut client)?;
        handle_role_deletions(spec, self.connstr.as_str(), &mut client)?;
-        handle_grants(spec, self.connstr.as_str(), &mut client)?;
+        handle_grants(spec, self.connstr.as_str())?;
        handle_extensions(spec, &mut client)?;

        // 'Close' connection
        drop(client);

-        info!(
-            "finished configuration of compute for project {}",
-            spec.cluster.cluster_id.as_deref().unwrap_or("None")
-        );
-
        Ok(())
    }

    // We could've wrapped this around `pg_ctl reload`, but right now we don't use
    // `pg_ctl` for start / stop, so this just seems much easier to do as we already
    // have opened connection to Postgres and superuser access.
-    #[instrument(skip(self, client))]
+    #[instrument(skip_all)]
    fn pg_reload_conf(&self, client: &mut Client) -> Result<()> {
        client.simple_query("SELECT pg_reload_conf()")?;
        Ok(())
@@ -389,13 +485,13 @@ impl ComputeNode {

    /// Similar to `apply_config()`, but does a bit different sequence of operations,
    /// as it's used to reconfigure a previously started and configured Postgres node.
-    #[instrument(skip(self))]
+    #[instrument(skip_all)]
    pub fn reconfigure(&self) -> Result<()> {
        let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;

        // Write new config
        let pgdata_path = Path::new(&self.pgdata);
-        config::write_postgres_conf(&pgdata_path.join("postgresql.conf"), &spec)?;
+        config::write_postgres_conf(&pgdata_path.join("postgresql.conf"), &spec, None)?;

        let mut client = Client::connect(self.connstr.as_str(), NoTls)?;
        self.pg_reload_conf(&mut client)?;
@@ -407,7 +503,7 @@ impl ComputeNode {
            handle_roles(&spec, &mut client)?;
            handle_databases(&spec, &mut client)?;
            handle_role_deletions(&spec, self.connstr.as_str(), &mut client)?;
-            handle_grants(&spec, self.connstr.as_str(), &mut client)?;
+            handle_grants(&spec, self.connstr.as_str())?;
            handle_extensions(&spec, &mut client)?;
        }

@@ -424,25 +520,45 @@ impl ComputeNode {
        Ok(())
    }

-    #[instrument(skip(self))]
-    pub fn start_compute(&self) -> Result<std::process::Child> {
+    #[instrument(skip_all)]
+    pub fn start_compute(&self, extension_server_port: u16) -> Result<std::process::Child> {
        let compute_state = self.state.lock().unwrap().clone();
-        let spec = compute_state.pspec.as_ref().expect("spec must be set");
+        let pspec = compute_state.pspec.as_ref().expect("spec must be set");
        info!(
            "starting compute for project {}, operation {}, tenant {}, timeline {}",
-            spec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
-            spec.spec.operation_uuid.as_deref().unwrap_or("None"),
-            spec.tenant_id,
-            spec.timeline_id,
+            pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
+            pspec.spec.operation_uuid.as_deref().unwrap_or("None"),
+            pspec.tenant_id,
+            pspec.timeline_id,
        );

-        self.prepare_pgdata(&compute_state)?;
+        // This part is sync, because we need to download
+        // remote shared_preload_libraries before postgres start (if any)
+        let library_load_start_time = Utc::now();
+        {
+            self.prepare_extenal_libraries(&compute_state)?;
+
+            let library_load_time = Utc::now()
+                .signed_duration_since(library_load_start_time)
+                .to_std()
+                .unwrap()
+                .as_millis() as u64;
+
+            let mut state = self.state.lock().unwrap();
+            state.metrics.load_libraries_ms = library_load_time;
+            info!(
+                "Loading shared_preload_libraries took {:?}ms",
+                library_load_time
+            );
+        }
+
+        self.prepare_pgdata(&compute_state, extension_server_port)?;

        let start_time = Utc::now();

-        let pg = self.start_postgres(spec.storage_auth_token.clone())?;
+        let pg = self.start_postgres(pspec.storage_auth_token.clone())?;

-        if spec.spec.mode == ComputeMode::Primary {
+        if pspec.spec.mode == ComputeMode::Primary && !pspec.spec.skip_pg_catalog_updates {
            self.apply_config(&compute_state)?;
        }

@@ -462,6 +578,11 @@ impl ComputeNode {
        }
        self.set_status(ComputeStatus::Running);

+        info!(
+            "finished configuration of compute for project {}",
+            pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None")
+        );
+
        Ok(pg)
    }

@@ -567,4 +688,150 @@ LIMIT 100",
            "{{\"pg_stat_statements\": []}}".to_string()
        }
    }
+
+    // If remote extension storage is configured,
+    // download shared preload libraries.
+    #[tokio::main]
+    pub async fn prepare_extenal_libraries(&self, compute_state: &ComputeState) -> Result<()> {
+        if let Some(ref ext_remote_storage) = self.ext_remote_storage {
+            let pspec = compute_state.pspec.as_ref().expect("spec must be set");
+            // download preload shared libraries before postgres start (if any)
+            let spec = &pspec.spec;
+
+            // 1. parse custom extension paths from spec
+            let custom_ext_prefixes = match &spec.custom_extensions {
+                Some(custom_extensions) => custom_extensions.clone(),
+                None => Vec::new(),
+            };
+
+            info!("custom_ext_prefixes: {:?}", &custom_ext_prefixes);
+
+            // parse shared_preload_libraries from spec
+            let mut libs_vec = Vec::new();
+
+            if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
+                libs_vec = libs
+                    .split(&[',', '\'', ' '])
+                    .filter(|s| *s != "neon" && !s.is_empty())
+                    .map(str::to_string)
+                    .collect();
+            }
+
+            info!(
+                "shared_preload_libraries parsed from spec.cluster.settings: {:?}",
+                libs_vec
+            );
+
+            // also parse shared_preload_libraries from provided postgresql.conf
+            // that is used in neon_local and python tests
+            if let Some(conf) = &spec.cluster.postgresql_conf {
+                let conf_lines = conf.split('\n').collect::<Vec<&str>>();
+
+                let mut shared_preload_libraries_line = "";
+                for line in conf_lines {
+                    if line.starts_with("shared_preload_libraries") {
+                        shared_preload_libraries_line = line;
+                    }
+                }
+
+                let mut preload_libs_vec = Vec::new();
+                if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
+                    preload_libs_vec = libs
+                        .split(&[',', '\'', ' '])
+                        .filter(|s| *s != "neon" && !s.is_empty())
+                        .map(str::to_string)
+                        .collect();
+                }
+
+                info!(
+                    "shared_preload_libraries parsed from spec.cluster.postgresql_conf: {:?}",
+                    preload_libs_vec
+                );
+
+                libs_vec.extend(preload_libs_vec);
+            }
+
+            info!("Libraries to download: {:?}", &libs_vec);
+            // download shared_preload_libraries
+            let available_libraries = extension_server::get_available_libraries(
+                ext_remote_storage,
+                &self.pgbin,
+                &self.pgversion,
+                &custom_ext_prefixes,
+                &libs_vec,
+            )
+            .await?;
+
+            self.available_libraries
+                .set(available_libraries)
+                .expect("available_libraries.set error");
+        }
+        Ok(())
+    }
+
+    // If remote extension storage is configured,
+    // download extension control files
+    #[tokio::main]
+    pub async fn prepare_external_extensions(&self, compute_state: &ComputeState) -> Result<()> {
+        if let Some(ref ext_remote_storage) = self.ext_remote_storage {
+            let pspec = compute_state.pspec.as_ref().expect("spec must be set");
+            let spec = &pspec.spec;
+
+            // 1. parse custom extension paths from spec
+            let custom_ext_prefixes = match &spec.custom_extensions {
+                Some(custom_extensions) => custom_extensions.clone(),
+                None => Vec::new(),
+            };
+
+            info!("custom_ext_prefixes: {:?}", &custom_ext_prefixes);
+
+            // download extension control files
+            let available_extensions = extension_server::get_available_extensions(
+                ext_remote_storage,
+                &self.pgbin,
+                &self.pgversion,
+                &custom_ext_prefixes,
+            )
+            .await?;
+
+            self.available_extensions
+                .set(available_extensions)
+                .expect("available_extensions.set error");
+        }
+        Ok(())
+    }
+
+    pub async fn download_extension_files(&self, filename: String) -> Result<()> {
+        match &self.ext_remote_storage {
+            None => anyhow::bail!("No remote extension storage"),
+            Some(remote_storage) => {
+                extension_server::download_extension_files(
+                    &filename,
+                    remote_storage,
+                    &self.pgbin,
+                    self.available_extensions
+                        .get()
+                        .context("available_extensions broke")?,
+                )
+                .await
+            }
+        }
+    }
+
+    pub async fn download_library_file(&self, filename: String) -> Result<()> {
+        match &self.ext_remote_storage {
+            None => anyhow::bail!("No remote extension storage"),
+            Some(remote_storage) => {
+                extension_server::download_library_file(
+                    &filename,
+                    remote_storage,
+                    &self.pgbin,
+                    self.available_libraries
+                        .get()
+                        .context("available_libraries broke")?,
+                )
+                .await
+            }
+        }
+    }
 }
--- a/compute_tools/src/config.rs
+++ b/compute_tools/src/config.rs
@@ -33,7 +33,11 @@ pub fn line_in_file(path: &Path, line: &str) -> Result<bool> {
 }

 /// Create or completely rewrite configuration file specified by `path`
-pub fn write_postgres_conf(path: &Path, spec: &ComputeSpec) -> Result<()> {
+pub fn write_postgres_conf(
+    path: &Path,
+    spec: &ComputeSpec,
+    extension_server_port: Option<u16>,
+) -> Result<()> {
    // File::create() destroys the file content if it exists.
    let mut file = File::create(path)?;

@@ -95,5 +99,9 @@ pub fn write_postgres_conf(path: &Path, spec: &ComputeSpec) -> Result<()> {
        writeln!(file, "# Managed by compute_ctl: end")?;
    }

+    if let Some(port) = extension_server_port {
+        writeln!(file, "neon.extension_server_port={}", port)?;
+    }
+
    Ok(())
 }
--- a/compute_tools/src/configurator.rs
+++ b/compute_tools/src/configurator.rs
@@ -8,7 +8,7 @@ use compute_api::responses::ComputeStatus;

 use crate::compute::ComputeNode;

-#[instrument(skip(compute))]
+#[instrument(skip_all)]
 fn configurator_main_loop(compute: &Arc<ComputeNode>) {
    info!("waiting for reconfiguration requests");
    loop {
@@ -42,13 +42,15 @@ fn configurator_main_loop(compute: &Arc<ComputeNode>) {
    }
 }

-pub fn launch_configurator(compute: &Arc<ComputeNode>) -> Result<thread::JoinHandle<()>> {
+pub fn launch_configurator(
+    compute: &Arc<ComputeNode>,
+) -> Result<thread::JoinHandle<()>, std::io::Error> {
    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");
-        })?)
+        })
 }
--- a/compute_tools/src/extension_server.rs
+++ b/compute_tools/src/extension_server.rs
@@ -0,0 +1,447 @@
+// Download extension files from the extension store
+// and put them in the right place in the postgres directory
+use crate::compute::ComputeNode;
+use anyhow::{self, bail, Context, Result};
+use futures::future::join_all;
+use remote_storage::*;
+use serde_json::{self, Value};
+use std::collections::HashMap;
+use std::fs::File;
+use std::io::{BufWriter, Write};
+use std::num::{NonZeroU32, NonZeroUsize};
+use std::path::{Path, PathBuf};
+use std::str;
+use std::sync::Arc;
+use std::thread;
+use tokio::io::AsyncReadExt;
+use tracing::info;
+
+// remote!
+const SHARE_EXT_PATH: &str = "share/extension";
+
+fn pass_any_error(results: Vec<Result<()>>) -> Result<()> {
+    for result in results {
+        result?;
+    }
+    Ok(())
+}
+
+fn get_pg_config(argument: &str, pgbin: &str) -> String {
+    // gives the result of `pg_config [argument]`
+    // where argument is a flag like `--version` or `--sharedir`
+    let pgconfig = pgbin.replace("postgres", "pg_config");
+    let config_output = std::process::Command::new(pgconfig)
+        .arg(argument)
+        .output()
+        .expect("pg_config error");
+    std::str::from_utf8(&config_output.stdout)
+        .expect("pg_config error")
+        .trim()
+        .to_string()
+}
+
+pub fn get_pg_version(pgbin: &str) -> String {
+    // pg_config --version returns a (platform specific) human readable string
+    // such as "PostgreSQL 15.4". We parse this to v14/v15
+    let human_version = get_pg_config("--version", pgbin);
+    if human_version.contains("15") {
+        return "v15".to_string();
+    } else if human_version.contains("14") {
+        return "v14".to_string();
+    }
+    panic!("Unsuported postgres version {human_version}");
+}
+
+async fn download_helper(
+    remote_storage: &GenericRemoteStorage,
+    remote_from_path: RemotePath,
+    sub_directory: Option<&str>,
+    download_location: &Path,
+) -> anyhow::Result<()> {
+    // downloads file at remote_from_path to
+    // `download_location/[optional: subdirectory]/[remote_storage.object_name()]`
+    // Note: the subdirectory commmand is needed when there is an extension that
+    // depends on files in a subdirectory.
+    // For example, v14/share/extension/some_ext.control
+    // might depend on v14/share/extension/some_ext/some_ext--1.1.0.sql
+    // and v14/share/extension/some_ext/xxx.csv
+    // Note: it is the caller's responsibility to create the appropriate subdirectory
+
+    let local_path = match sub_directory {
+        Some(subdir) => download_location
+            .join(subdir)
+            .join(remote_from_path.object_name().expect("bad object")),
+        None => download_location.join(remote_from_path.object_name().expect("bad object")),
+    };
+    if local_path.exists() {
+        info!("File {:?} already exists. Skipping download", &local_path);
+        return Ok(());
+    }
+    info!(
+        "Downloading {:?} to location {:?}",
+        &remote_from_path, &local_path
+    );
+    let mut download = remote_storage.download(&remote_from_path).await?;
+    let mut write_data_buffer = Vec::new();
+    download
+        .download_stream
+        .read_to_end(&mut write_data_buffer)
+        .await?;
+    let mut output_file = BufWriter::new(File::create(local_path)?);
+    output_file.write_all(&write_data_buffer)?;
+    info!("Download {:?} completed successfully", &remote_from_path);
+    Ok(())
+}
+
+// download extension control files
+//
+// if custom_ext_prefixes is provided - search also in custom extension paths
+//
+pub async fn get_available_extensions(
+    remote_storage: &GenericRemoteStorage,
+    pgbin: &str,
+    pg_version: &str,
+    custom_ext_prefixes: &Vec<String>,
+) -> anyhow::Result<HashMap<String, Vec<PathAndFlag>>> {
+    let local_sharedir = Path::new(&get_pg_config("--sharedir", pgbin)).join("extension");
+
+    // public path, plus any private paths to download extensions from
+    let mut paths: Vec<RemotePath> = Vec::new();
+    paths.push(RemotePath::new(
+        &Path::new(pg_version).join(SHARE_EXT_PATH),
+    )?);
+    for custom_prefix in custom_ext_prefixes {
+        paths.push(RemotePath::new(
+            &Path::new(pg_version)
+                .join(custom_prefix)
+                .join(SHARE_EXT_PATH),
+        )?);
+    }
+
+    let (extension_files, control_files) =
+        organized_extension_files(remote_storage, &paths).await?;
+
+    let mut control_file_download_tasks = Vec::new();
+    // download all control files
+    for control_file in control_files {
+        control_file_download_tasks.push(download_helper(
+            remote_storage,
+            control_file.clone(),
+            None,
+            &local_sharedir,
+        ));
+    }
+    pass_any_error(join_all(control_file_download_tasks).await)?;
+    Ok(extension_files)
+}
+
+// Download requested shared_preload_libraries
+//
+// Note that tenant_id is not optional here, because we only download libraries
+// after we know the tenant spec and the tenant_id.
+//
+// return list of all library files to use it in the future searches
+pub async fn get_available_libraries(
+    remote_storage: &GenericRemoteStorage,
+    pgbin: &str,
+    pg_version: &str,
+    custom_ext_prefixes: &Vec<String>,
+    preload_libraries: &Vec<String>,
+) -> anyhow::Result<HashMap<String, Vec<RemotePath>>> {
+    // Construct a hashmap of all available libraries
+    // example (key, value) pair: test_lib0: [RemotePath(v14/lib/test_lib0.so), RemotePath(v14/lib/test_lib0.so.3)]
+    let mut paths: Vec<RemotePath> = Vec::new();
+    // public libraries
+    paths.push(
+        RemotePath::new(&Path::new(&pg_version).join("lib/"))
+            .expect("The hard coded path here is valid"),
+    );
+    // custom libraries
+    for custom_prefix in custom_ext_prefixes {
+        paths.push(
+            RemotePath::new(&Path::new(&pg_version).join(custom_prefix).join("lib"))
+                .expect("The hard coded path here is valid"),
+        );
+    }
+    let all_available_libraries = organized_library_files(remote_storage, &paths).await?;
+
+    info!("list of library files {:?}", &all_available_libraries);
+    // download all requested libraries
+    let mut download_tasks = Vec::new();
+    for lib_name in preload_libraries {
+        download_tasks.push(download_library_file(
+            lib_name,
+            remote_storage,
+            pgbin,
+            &all_available_libraries,
+        ));
+    }
+    pass_any_error(join_all(download_tasks).await)?;
+    Ok(all_available_libraries)
+}
+
+// download all sqlfiles (and possibly data files) for a given extension name
+//
+pub async fn download_extension_files(
+    ext_name: &str,
+    remote_storage: &GenericRemoteStorage,
+    pgbin: &str,
+    all_available_files: &HashMap<String, Vec<PathAndFlag>>,
+) -> Result<()> {
+    let local_sharedir = Path::new(&get_pg_config("--sharedir", pgbin)).join("extension");
+    let mut downloaded_something = false;
+    let mut made_subdir = false;
+
+    info!("EXTENSION {:?}", ext_name);
+    info!("{:?}", all_available_files.get(ext_name));
+
+    info!("start download");
+    let mut download_tasks = Vec::new();
+    if let Some(files) = all_available_files.get(ext_name) {
+        info!("Downloading files for extension {:?}", &ext_name);
+        for path_and_flag in files {
+            let file = &path_and_flag.path;
+            let subdir_flag = path_and_flag.subdir_flag;
+            info!(
+                "--- Downloading {:?} (for {:?} as subdir? = {:?})",
+                &file, &ext_name, subdir_flag
+            );
+            let mut subdir = None;
+            if subdir_flag {
+                subdir = Some(ext_name);
+                if !made_subdir {
+                    made_subdir = true;
+                    std::fs::create_dir_all(local_sharedir.join(ext_name))?;
+                }
+            }
+            download_tasks.push(download_helper(
+                remote_storage,
+                file.clone(),
+                subdir,
+                &local_sharedir,
+            ));
+            downloaded_something = true;
+        }
+    }
+    if !downloaded_something {
+        bail!("Files for extension {ext_name} are not found in the extension store");
+    }
+    pass_any_error(join_all(download_tasks).await)?;
+    info!("finish download");
+    Ok(())
+}
+
+// appends an .so suffix to libname if it does not already have one
+fn enforce_so_end(libname: &str) -> String {
+    if !libname.contains(".so") {
+        format!("{}.so", libname)
+    } else {
+        libname.to_string()
+    }
+}
+
+// download shared library file
+pub async fn download_library_file(
+    lib_name: &str,
+    remote_storage: &GenericRemoteStorage,
+    pgbin: &str,
+    all_available_libraries: &HashMap<String, Vec<RemotePath>>,
+) -> Result<()> {
+    let lib_name = get_library_name(lib_name);
+    let local_libdir: PathBuf = Path::new(&get_pg_config("--pkglibdir", pgbin)).into();
+    info!("looking for library {:?}", &lib_name);
+    match all_available_libraries.get(&*lib_name) {
+        Some(remote_paths) => {
+            let mut library_download_tasks = Vec::new();
+            for remote_path in remote_paths {
+                let file_path = local_libdir.join(remote_path.object_name().expect("bad object"));
+                if file_path.exists() {
+                    info!("File {:?} already exists. Skipping download", &file_path);
+                } else {
+                    library_download_tasks.push(download_helper(
+                        remote_storage,
+                        remote_path.clone(),
+                        None,
+                        &local_libdir,
+                    ));
+                }
+            }
+            pass_any_error(join_all(library_download_tasks).await)?;
+        }
+        None => {
+            // minor TODO: this logic seems to be somewhat faulty for .so.3 type files?
+            let lib_name_with_ext = enforce_so_end(&lib_name);
+            let file_path = local_libdir.join(lib_name_with_ext);
+            if file_path.exists() {
+                info!("File {:?} already exists. Skipping download", &file_path);
+            } else {
+                bail!("Library file {lib_name} not found")
+            }
+        }
+    }
+    Ok(())
+}
+
+// This function initializes the necessary structs to use remmote storage (should be fairly cheap)
+pub fn init_remote_storage(
+    remote_ext_config: &str,
+    default_prefix: &str,
+) -> anyhow::Result<GenericRemoteStorage> {
+    let remote_ext_config: serde_json::Value = serde_json::from_str(remote_ext_config)?;
+
+    let remote_ext_bucket = match &remote_ext_config["bucket"] {
+        Value::String(x) => x,
+        _ => bail!("remote_ext_config missing bucket"),
+    };
+    let remote_ext_region = match &remote_ext_config["region"] {
+        Value::String(x) => x,
+        _ => bail!("remote_ext_config missing region"),
+    };
+    let remote_ext_endpoint = match &remote_ext_config["endpoint"] {
+        Value::String(x) => Some(x.clone()),
+        _ => None,
+    };
+    let remote_ext_prefix = match &remote_ext_config["prefix"] {
+        Value::String(x) => Some(x.clone()),
+        // if prefix is not provided, use default, which is the build_tag
+        _ => Some(default_prefix.to_string()),
+    };
+
+    // load will not be large, so default parameters are fine
+    let config = S3Config {
+        bucket_name: remote_ext_bucket.to_string(),
+        bucket_region: remote_ext_region.to_string(),
+        prefix_in_bucket: remote_ext_prefix,
+        endpoint: remote_ext_endpoint,
+        concurrency_limit: NonZeroUsize::new(100).expect("100 != 0"),
+        max_keys_per_list_response: None,
+    };
+    let config = RemoteStorageConfig {
+        max_concurrent_syncs: NonZeroUsize::new(100).expect("100 != 0"),
+        max_sync_errors: NonZeroU32::new(100).expect("100 != 0"),
+        storage: RemoteStorageKind::AwsS3(config),
+    };
+    GenericRemoteStorage::from_config(&config)
+}
+
+fn get_library_name(path: &str) -> String {
+    let path_suffix: Vec<&str> = path.split('/').collect();
+    let path_suffix = path_suffix.last().expect("bad ext name").to_string();
+    if let Some(index) = path_suffix.find(".so") {
+        return path_suffix[..index].to_string();
+    }
+    path_suffix
+}
+
+// asyncrounously lists files in all necessary directories
+// TODO: potential optimization: do a single list files on the entire bucket
+// and then filter out the files we don't need
+async fn list_all_files(
+    remote_storage: &GenericRemoteStorage,
+    paths: &Vec<RemotePath>,
+) -> Result<Vec<RemotePath>> {
+    let mut list_tasks = Vec::new();
+    let mut all_files = Vec::new();
+    for path in paths {
+        list_tasks.push(remote_storage.list_files(Some(path)));
+    }
+    for list_result in join_all(list_tasks).await {
+        all_files.extend(list_result?);
+    }
+    Ok(all_files)
+}
+
+// helper to collect all libraries, grouped by library name
+// Returns a hashmap of (library name: [paths]})
+// example entry: {libpgtypes: [libpgtypes.so.3, libpgtypes.so]}
+async fn organized_library_files(
+    remote_storage: &GenericRemoteStorage,
+    paths: &Vec<RemotePath>,
+) -> Result<HashMap<String, Vec<RemotePath>>> {
+    let mut library_groups = HashMap::new();
+    for file in list_all_files(remote_storage, paths).await? {
+        let lib_name = get_library_name(file.get_path().to_str().context("invalid path")?);
+        let lib_list = library_groups.entry(lib_name).or_insert(Vec::new());
+        lib_list.push(file.to_owned());
+    }
+    Ok(library_groups)
+}
+
+// store a path, paired with a flag indicating whether the path is to a file in
+// the root or subdirectory
+#[derive(Debug)]
+pub struct PathAndFlag {
+    path: RemotePath,
+    subdir_flag: bool,
+}
+
+// get_ext_name extracts the extension name, and returns a flag indicating
+// whether this file is in a subdirectory or not.
+//
+// extension files can be in subdirectories of the extension store.
+// examples of layout:
+// v14//share//extension/extension_name--1.0.sql,
+// v14//share//extension/extension_name/extension_name--1.0.sql,
+// v14//share//extension/extension_name/extra_data.csv
+// Note: we *assume* that the  extension files is in one of these formats.
+// If it is not, this code's behavior is *undefined*.
+fn get_ext_name(path: &str) -> Result<(&str, bool)> {
+    let path_suffix: Vec<&str> = path.split(&format!("{SHARE_EXT_PATH}/")).collect();
+    let ext_name = path_suffix.last().expect("bad ext name");
+
+    if let Some(index) = ext_name.find('/') {
+        return Ok((&ext_name[..index], true));
+    } else if let Some(index) = ext_name.find("--") {
+        return Ok((&ext_name[..index], false));
+    }
+    Ok((ext_name, false))
+}
+
+// helper to collect files of given prefixes for extensions and group them by extension
+// returns a hashmap of (extension_name, Vector of remote paths for all files needed for this extension)
+// and a list of control files
+// For example, an entry in the hashmap could be
+// {"anon": [RemotePath("v14/anon/share/extension/anon/address.csv"),
+// RemotePath("v14/anon/share/extension/anon/anon--1.1.0.sql")]},
+// with corresponding list of control files entry being
+// {"anon.control": RemotePath("v14/anon/share/extension/anon.control")}
+async fn organized_extension_files(
+    remote_storage: &GenericRemoteStorage,
+    paths: &Vec<RemotePath>,
+) -> Result<(HashMap<String, Vec<PathAndFlag>>, Vec<RemotePath>)> {
+    let mut grouped_dependencies = HashMap::new();
+    let mut control_files = Vec::new();
+
+    for file in list_all_files(remote_storage, paths).await? {
+        if file.extension().context("bad file name")? == "control" {
+            control_files.push(file.to_owned());
+        } else {
+            let (file_ext_name, subdir_flag) =
+                get_ext_name(file.get_path().to_str().context("invalid path")?)?;
+            let ext_file_list = grouped_dependencies
+                .entry(file_ext_name.to_string())
+                .or_insert(Vec::new());
+            ext_file_list.push(PathAndFlag {
+                path: file.to_owned(),
+                subdir_flag,
+            });
+        }
+    }
+    Ok((grouped_dependencies, control_files))
+}
+
+pub fn launch_download_extensions(
+    compute: &Arc<ComputeNode>,
+) -> Result<thread::JoinHandle<()>, std::io::Error> {
+    let compute = Arc::clone(compute);
+    thread::Builder::new()
+        .name("download-extensions".into())
+        .spawn(move || {
+            info!("start download_extension_files");
+            let compute_state = compute.state.lock().expect("error unlocking compute.state");
+            compute
+                .prepare_external_extensions(&compute_state)
+                .expect("error preparing extensions");
+            info!("download_extension_files done, exiting thread");
+        })
+}
--- a/compute_tools/src/http/api.rs
+++ b/compute_tools/src/http/api.rs
@@ -121,6 +121,55 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
            }
        }

+        // download extension files from S3 on demand
+        (&Method::POST, route) if route.starts_with("/extension_server/") => {
+            info!("serving {:?} POST request", route);
+            info!("req.uri {:?}", req.uri());
+
+            let mut is_library = false;
+
+            if let Some(params) = req.uri().query() {
+                info!("serving {:?} POST request with params: {}", route, params);
+
+                if params == "is_library=true" {
+                    is_library = true;
+                } else {
+                    let mut resp = Response::new(Body::from("Wrong request parameters"));
+                    *resp.status_mut() = StatusCode::BAD_REQUEST;
+                    return resp;
+                }
+            }
+
+            let filename = route.split('/').last().unwrap().to_string();
+
+            info!(
+                "serving /extension_server POST request, filename: {:?} is_library: {}",
+                filename, is_library
+            );
+
+            if is_library {
+                match compute.download_library_file(filename.to_string()).await {
+                    Ok(_) => Response::new(Body::from("OK")),
+                    Err(e) => {
+                        error!("library download failed: {}", e);
+                        let mut resp = Response::new(Body::from(e.to_string()));
+                        *resp.status_mut() = StatusCode::INTERNAL_SERVER_ERROR;
+                        resp
+                    }
+                }
+            } else {
+                match compute.download_extension_files(filename.to_string()).await {
+                    Ok(_) => Response::new(Body::from("OK")),
+                    Err(e) => {
+                        error!("extension download failed: {}", e);
+                        let mut resp = Response::new(Body::from(e.to_string()));
+                        *resp.status_mut() = StatusCode::INTERNAL_SERVER_ERROR;
+                        resp
+                    }
+                }
+            }
+        }
+
        // Return the `404 Not Found` for any other routes.
        _ => {
            let mut not_found = Response::new(Body::from("404 Not Found"));
--- a/compute_tools/src/http/openapi_spec.yaml
+++ b/compute_tools/src/http/openapi_spec.yaml
@@ -139,6 +139,34 @@ paths:
            application/json:
              schema:
                $ref: "#/components/schemas/GenericError"
+  /extension_server:
+    post:
+      tags:
+      - Extension
+      summary: Download extension from S3 to local folder.
+      description: ""
+      operationId: downloadExtension
+      responses:
+        200:
+          description: Extension downloaded
+          content:
+            text/plain:
+              schema:
+                type: string
+                description: Error text or 'OK' if download succeeded.
+                example: "OK"
+        400:
+        description: Request is invalid.
+        content:
+          application/json:
+            schema:
+              $ref: "#/components/schemas/GenericError"
+        500:
+        description: Extension download request failed.
+        content:
+          application/json:
+            schema:
+              $ref: "#/components/schemas/GenericError"

 components:
  securitySchemes:
--- a/compute_tools/src/lib.rs
+++ b/compute_tools/src/lib.rs
@@ -9,6 +9,7 @@ pub mod http;
 #[macro_use]
 pub mod logger;
 pub mod compute;
+pub mod extension_server;
 pub mod monitor;
 pub mod params;
 pub mod pg_helpers;
--- a/compute_tools/src/logger.rs
+++ b/compute_tools/src/logger.rs
@@ -18,6 +18,7 @@ pub fn init_tracing_and_logging(default_log_level: &str) -> anyhow::Result<()> {
        .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new(default_log_level));

    let fmt_layer = tracing_subscriber::fmt::layer()
+        .with_ansi(false)
        .with_target(false)
        .with_writer(std::io::stderr);

--- a/compute_tools/src/monitor.rs
+++ b/compute_tools/src/monitor.rs
@@ -105,10 +105,10 @@ 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>) -> Result<thread::JoinHandle<()>, std::io::Error> {
    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))
 }
--- a/compute_tools/src/pg_helpers.rs
+++ b/compute_tools/src/pg_helpers.rs
@@ -17,7 +17,7 @@ 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
-fn escape_literal(s: &str) -> String {
+pub fn escape_literal(s: &str) -> String {
    s.replace('\'', "''").replace('\\', "\\\\")
 }

@@ -215,7 +215,7 @@ pub fn get_existing_dbs(client: &mut Client) -> Result<Vec<Database>> {
 /// Wait for Postgres to become ready to accept connections. It's ready to
 /// accept connections when the state-field in `pgdata/postmaster.pid` says
 /// 'ready'.
-#[instrument(skip(pg))]
+#[instrument(skip_all, fields(pgdata = %pgdata.display()))]
 pub fn wait_for_postgres(pg: &mut Child, pgdata: &Path) -> Result<()> {
    let pid_path = pgdata.join("postmaster.pid");

--- a/compute_tools/src/spec.rs
+++ b/compute_tools/src/spec.rs
@@ -124,7 +124,7 @@ pub fn get_spec_from_control_plane(
 pub fn handle_configuration(spec: &ComputeSpec, pgdata_path: &Path) -> Result<()> {
    // File `postgresql.conf` is no longer included into `basebackup`, so just
    // always write all config into it creating new file.
-    config::write_postgres_conf(&pgdata_path.join("postgresql.conf"), spec)?;
+    config::write_postgres_conf(&pgdata_path.join("postgresql.conf"), spec, None)?;

    update_pg_hba(pgdata_path)?;

@@ -269,17 +269,13 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                xact.execute(query.as_str(), &[])?;
            }
            RoleAction::Create => {
-                let mut query: String = format!("CREATE ROLE {} ", name.pg_quote());
+                let mut query: String = format!(
+                    "CREATE ROLE {} CREATEROLE CREATEDB IN ROLE neon_superuser",
+                    name.pg_quote()
+                );
                info!("role create query: '{}'", &query);
                query.push_str(&role.to_pg_options());
                xact.execute(query.as_str(), &[])?;
-
-                let grant_query = format!(
-                    "GRANT pg_read_all_data, pg_write_all_data TO {}",
-                    name.pg_quote()
-                );
-                xact.execute(grant_query.as_str(), &[])?;
-                info!("role grant query: '{}'", &grant_query);
            }
        }

@@ -476,6 +472,11 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                query.push_str(&db.to_pg_options());
                let _guard = info_span!("executing", query).entered();
                client.execute(query.as_str(), &[])?;
+                let grant_query: String = format!(
+                    "GRANT ALL PRIVILEGES ON DATABASE {} TO neon_superuser",
+                    name.pg_quote()
+                );
+                client.execute(grant_query.as_str(), &[])?;
            }
        };

@@ -495,35 +496,9 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
 /// Grant CREATE ON DATABASE to the database owner and do some other alters and grants
 /// to allow users creating trusted extensions and re-creating `public` schema, for example.
 #[instrument(skip_all)]
-pub fn handle_grants(spec: &ComputeSpec, connstr: &str, client: &mut Client) -> Result<()> {
+pub fn handle_grants(spec: &ComputeSpec, connstr: &str) -> Result<()> {
    info!("cluster spec grants:");

-    // We now have a separate `web_access` role to connect to the database
-    // via the web interface and proxy link auth. And also we grant a
-    // read / write all data privilege to every role. So also grant
-    // create to everyone.
-    // XXX: later we should stop messing with Postgres ACL in such horrible
-    // ways.
-    let roles = spec
-        .cluster
-        .roles
-        .iter()
-        .map(|r| r.name.pg_quote())
-        .collect::<Vec<_>>();
-
-    for db in &spec.cluster.databases {
-        let dbname = &db.name;
-
-        let query: String = format!(
-            "GRANT CREATE ON DATABASE {} TO {}",
-            dbname.pg_quote(),
-            roles.join(", ")
-        );
-        info!("grant query {}", &query);
-
-        client.execute(query.as_str(), &[])?;
-    }
-
    // Do some per-database access adjustments. We'd better do this at db creation time,
    // but CREATE DATABASE isn't transactional. So we cannot create db + do some grants
    // atomically.
--- a/control_plane/Cargo.toml
+++ b/control_plane/Cargo.toml
@@ -32,3 +32,4 @@ utils.workspace = true

 compute_api.workspace = true
 workspace_hack.workspace = true
+tracing.workspace = true
--- a/control_plane/src/background_process.rs
+++ b/control_plane/src/background_process.rs
@@ -180,6 +180,11 @@ pub fn stop_process(immediate: bool, process_name: &str, pid_file: &Path) -> any
    }

    // Wait until process is gone
+    wait_until_stopped(process_name, pid)?;
+    Ok(())
+}
+
+pub fn wait_until_stopped(process_name: &str, pid: Pid) -> anyhow::Result<()> {
    for retries in 0..RETRIES {
        match process_has_stopped(pid) {
            Ok(true) => {
--- a/control_plane/src/bin/neon_local.rs
+++ b/control_plane/src/bin/neon_local.rs
@@ -308,7 +308,8 @@ fn handle_init(init_match: &ArgMatches) -> anyhow::Result<LocalEnv> {

    let mut env =
        LocalEnv::parse_config(&toml_file).context("Failed to create neon configuration")?;
-    env.init(pg_version)
+    let force = init_match.get_flag("force");
+    env.init(pg_version, force)
        .context("Failed to initialize neon repository")?;

    // Initialize pageserver, create initial tenant and timeline.
@@ -657,6 +658,8 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                .get_one::<String>("endpoint_id")
                .ok_or_else(|| anyhow!("No endpoint ID was provided to start"))?;

+            let remote_ext_config = sub_args.get_one::<String>("remote-ext-config");
+
            // If --safekeepers argument is given, use only the listed safekeeper nodes.
            let safekeepers =
                if let Some(safekeepers_str) = sub_args.get_one::<String>("safekeepers") {
@@ -698,7 +701,7 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                    _ => {}
                }
                println!("Starting existing endpoint {endpoint_id}...");
-                endpoint.start(&auth_token, safekeepers)?;
+                endpoint.start(&auth_token, safekeepers, remote_ext_config)?;
            } else {
                let branch_name = sub_args
                    .get_one::<String>("branch-name")
@@ -742,7 +745,7 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                    pg_version,
                    mode,
                )?;
-                ep.start(&auth_token, safekeepers)?;
+                ep.start(&auth_token, safekeepers, remote_ext_config)?;
            }
        }
        "stop" => {
@@ -1002,6 +1005,12 @@ fn cli() -> Command {
        .help("Additional pageserver's configuration options or overrides, refer to pageserver's 'config-override' CLI parameter docs for more")
        .required(false);

+    let remote_ext_config_args = Arg::new("remote-ext-config")
+        .long("remote-ext-config")
+        .num_args(1)
+        .help("Configure the S3 bucket that we search for extensions in.")
+        .required(false);
+
    let lsn_arg = Arg::new("lsn")
        .long("lsn")
        .help("Specify Lsn on the timeline to start from. By default, end of the timeline would be used.")
@@ -1013,6 +1022,13 @@ fn cli() -> Command {
        .help("If set, the node will be a hot replica on the specified timeline")
        .required(false);

+    let force_arg = Arg::new("force")
+        .value_parser(value_parser!(bool))
+        .long("force")
+        .action(ArgAction::SetTrue)
+        .help("Force initialization even if the repository is not empty")
+        .required(false);
+
    Command::new("Neon CLI")
        .arg_required_else_help(true)
        .version(GIT_VERSION)
@@ -1028,6 +1044,7 @@ fn cli() -> Command {
                        .value_name("config"),
                )
                .arg(pg_version_arg.clone())
+                .arg(force_arg)
        )
        .subcommand(
            Command::new("timeline")
@@ -1152,6 +1169,7 @@ fn cli() -> Command {
                    .arg(pg_version_arg)
                    .arg(hot_standby_arg)
                    .arg(safekeepers_arg)
+                    .arg(remote_ext_config_args)
                )
                .subcommand(
                    Command::new("stop")
--- a/control_plane/src/endpoint.rs
+++ b/control_plane/src/endpoint.rs
@@ -67,6 +67,7 @@ pub struct EndpointConf {
    pg_port: u16,
    http_port: u16,
    pg_version: u32,
+    skip_pg_catalog_updates: bool,
 }

 //
@@ -135,6 +136,7 @@ impl ComputeControlPlane {
            mode,
            tenant_id,
            pg_version,
+            skip_pg_catalog_updates: false,
        });

        ep.create_endpoint_dir()?;
@@ -148,6 +150,7 @@ impl ComputeControlPlane {
                http_port,
                pg_port,
                pg_version,
+                skip_pg_catalog_updates: false,
            })?,
        )?;
        std::fs::write(
@@ -183,6 +186,9 @@ pub struct Endpoint {
    // the endpoint runs in.
    pub env: LocalEnv,
    pageserver: Arc<PageServerNode>,
+
+    // Optimizations
+    skip_pg_catalog_updates: bool,
 }

 impl Endpoint {
@@ -216,6 +222,7 @@ impl Endpoint {
            mode: conf.mode,
            tenant_id: conf.tenant_id,
            pg_version: conf.pg_version,
+            skip_pg_catalog_updates: conf.skip_pg_catalog_updates,
        })
    }

@@ -304,7 +311,7 @@ impl Endpoint {

                // TODO: use future host field from safekeeper spec
                // Pass the list of safekeepers to the replica so that it can connect to any of them,
-                // whichever is availiable.
+                // whichever is available.
                let sk_ports = self
                    .env
                    .safekeepers
@@ -398,10 +405,25 @@ impl Endpoint {
                String::from_utf8_lossy(&pg_ctl.stderr),
            );
        }
+
+        // Also wait for the compute_ctl process to die. It might have some cleanup
+        // work to do after postgres stops, like syncing safekeepers, etc.
+        //
+        // TODO use background_process::stop_process instead
+        let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
+        let pid: u32 = std::fs::read_to_string(pidfile_path)?.parse()?;
+        let pid = nix::unistd::Pid::from_raw(pid as i32);
+        crate::background_process::wait_until_stopped("compute_ctl", pid)?;
+
        Ok(())
    }

-    pub fn start(&self, auth_token: &Option<String>, safekeepers: Vec<NodeId>) -> Result<()> {
+    pub fn start(
+        &self,
+        auth_token: &Option<String>,
+        safekeepers: Vec<NodeId>,
+        remote_ext_config: Option<&String>,
+    ) -> Result<()> {
        if self.status() == "running" {
            anyhow::bail!("The endpoint is already running");
        }
@@ -450,6 +472,7 @@ impl Endpoint {

        // Create spec file
        let spec = ComputeSpec {
+            skip_pg_catalog_updates: self.skip_pg_catalog_updates,
            format_version: 1.0,
            operation_uuid: None,
            cluster: Cluster {
@@ -468,6 +491,13 @@ impl Endpoint {
            pageserver_connstring: Some(pageserver_connstring),
            safekeeper_connstrings,
            storage_auth_token: auth_token.clone(),
+            // TODO FIXME: This is a hack to test custom extensions locally.
+            // In test_download_extensions, we assume that the custom extension
+            // prefix is the tenant ID. So we set it here.
+            //
+            // The proper way to implement this is to pass the custom extension
+            // in spec, but we don't have a way to do that yet in the python tests.
+            custom_extensions: Some(vec![self.tenant_id.to_string()]),
        };
        let spec_path = self.endpoint_path().join("spec.json");
        std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;
@@ -499,7 +529,18 @@ impl Endpoint {
            .stdin(std::process::Stdio::null())
            .stderr(logfile.try_clone()?)
            .stdout(logfile);
-        let _child = cmd.spawn()?;
+
+        if let Some(remote_ext_config) = remote_ext_config {
+            cmd.args(["--remote-ext-config", remote_ext_config]);
+        }
+
+        let child = cmd.spawn()?;
+
+        // Write down the pid so we can wait for it when we want to stop
+        // TODO use background_process::start_process instead
+        let pid = child.id();
+        let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
+        std::fs::write(pidfile_path, pid.to_string())?;

        // Wait for it to start
        let mut attempt = 0;
--- a/control_plane/src/local_env.rs
+++ b/control_plane/src/local_env.rs
@@ -364,7 +364,7 @@ impl LocalEnv {
    //
    // Initialize a new Neon repository
    //
-    pub fn init(&mut self, pg_version: u32) -> anyhow::Result<()> {
+    pub fn init(&mut self, pg_version: u32, force: bool) -> anyhow::Result<()> {
        // check if config already exists
        let base_path = &self.base_data_dir;
        ensure!(
@@ -372,11 +372,29 @@ impl LocalEnv {
            "repository base path is missing"
        );

-        ensure!(
-            !base_path.exists(),
-            "directory '{}' already exists. Perhaps already initialized?",
-            base_path.display()
-        );
+        if base_path.exists() {
+            if force {
+                println!("removing all contents of '{}'", base_path.display());
+                // instead of directly calling `remove_dir_all`, we keep the original dir but removing
+                // all contents inside. This helps if the developer symbol links another directory (i.e.,
+                // S3 local SSD) to the `.neon` base directory.
+                for entry in std::fs::read_dir(base_path)? {
+                    let entry = entry?;
+                    let path = entry.path();
+                    if path.is_dir() {
+                        fs::remove_dir_all(&path)?;
+                    } else {
+                        fs::remove_file(&path)?;
+                    }
+                }
+            } else {
+                bail!(
+                    "directory '{}' already exists. Perhaps already initialized? (Hint: use --force to remove all contents)",
+                    base_path.display()
+                );
+            }
+        }
+
        if !self.pg_bin_dir(pg_version)?.join("postgres").exists() {
            bail!(
                "Can't find postgres binary at {}",
@@ -392,7 +410,9 @@ impl LocalEnv {
            }
        }

-        fs::create_dir(base_path)?;
+        if !base_path.exists() {
+            fs::create_dir(base_path)?;
+        }

        // Generate keypair for JWT.
        //
--- a/docs/rfcs/024-extension-loading.md
+++ b/docs/rfcs/024-extension-loading.md
@@ -0,0 +1,183 @@
+# Supporting custom user Extensions (Dynamic Extension Loading)
+Created 2023-05-03
+
+## Motivation
+
+There are many extensions in the PostgreSQL ecosystem, and not all extensions
+are of a quality that we can confidently support them. Additionally, our
+current extension inclusion mechanism has several problems because we build all
+extensions into the primary Compute image: We build the extensions every time
+we build the compute image regardless of whether we actually need to rebuild
+the image, and the inclusion of these extensions in the image adds a hard
+dependency on all supported extensions - thus increasing the image size, and
+with it the time it takes to download that image - increasing first start
+latency.
+
+This RFC proposes a dynamic loading mechanism that solves most of these
+problems.
+
+## Summary
+
+`compute_ctl` is made responsible for loading extensions on-demand into
+the container's file system for dynamically loaded extensions, and will also
+make sure that the extensions in `shared_preload_libraries` are downloaded
+before the compute node starts.
+
+## Components
+
+compute_ctl, PostgreSQL, neon (extension), Compute Host Node, Extension Store
+
+## Requirements
+
+Compute nodes with no extra extensions should not be negatively impacted by
+the existence of support for many extensions.
+
+Installing an extension into PostgreSQL should be easy.
+
+Non-preloaded extensions shouldn't impact startup latency.
+
+Uninstalled extensions shouldn't impact query latency.
+
+A small latency penalty for dynamically loaded extensions is acceptable in
+the first seconds of compute startup, but not in steady-state operations.
+
+## Proposed implementation
+
+### On-demand, JIT-loading of extensions
+
+Before postgres starts we download 
+- control files for all extensions available to that compute node;
+- all `shared_preload_libraries`;
+
+After postgres is running, `compute_ctl` listens for requests to load files.
+When PostgreSQL requests a file, `compute_ctl` downloads it.
+
+PostgreSQL requests files in the following cases:
+- When loading a preload library set in `local_preload_libraries`
+- When explicitly loading a library with `LOAD`
+- Wnen creating extension with `CREATE EXTENSION` (download sql scripts, (optional) extension data files and (optional) library files)))
+
+
+#### Summary
+
+Pros:
+ - Startup is only as slow as it takes to load all (shared_)preload_libraries
+ - Supports BYO Extension
+
+Cons:
+ - O(sizeof(extensions)) IO requirement for loading all extensions.
+
+### Alternative solutions
+
+1. Allow users to add their extensions to the base image
+   
+   Pros:
+    - Easy to deploy
+
+   Cons:
+    - Doesn't scale - first start size is dependent on image size;
+    - All extensions are shared across all users: It doesn't allow users to
+      bring their own restrictive-licensed extensions
+
+2. Bring Your Own compute image
+   
+   Pros:
+    - Still easy to deploy
+    - User can bring own patched version of PostgreSQL
+
+   Cons:
+    - First start latency is O(sizeof(extensions image))
+    - Warm instance pool for skipping pod schedule latency is not feasible with
+      O(n) custom images
+    - Support channels are difficult to manage
+
+3. Download all user extensions in bulk on compute start
+   
+   Pros:
+    - Easy to deploy
+    - No startup latency issues for "clean" users.
+    - Warm instance pool for skipping pod schedule latency is possible
+
+   Cons:
+    - Downloading all extensions in advance takes a lot of time, thus startup
+      latency issues
+
+4. Store user's extensions in persistent storage
+   
+   Pros:
+    - Easy to deploy
+    - No startup latency issues
+    - Warm instance pool for skipping pod schedule latency is possible
+
+   Cons:
+    - EC2 instances have only limited number of attachments shared between EBS
+      volumes, direct-attached NVMe drives, and ENIs.
+    - Compute instance migration isn't trivially solved for EBS mounts (e.g.
+      the device is unavailable whilst moving the mount between instances).
+    - EBS can only mount on one instance at a time (except the expensive IO2
+      device type).
+
+5. Store user's extensions in network drive
+   
+   Pros:
+    - Easy to deploy
+    - Few startup latency issues
+    - Warm instance pool for skipping pod schedule latency is possible
+
+   Cons:
+    - We'd need networked drives, and a lot of them, which would store many
+      duplicate extensions.
+    - **UNCHECKED:** Compute instance migration may not work nicely with
+      networked IOs
+
+
+### Idea extensions
+
+The extension store does not have to be S3 directly, but could be a Node-local
+caching service on top of S3. This would reduce the load on the network for
+popular extensions.
+
+## Extension Storage implementation
+
+Extension Storage in our case is an S3 bucket with a "directory" per build and postgres version,
+where extension files are stored as plain files in the bucket following the same directory structure as in the postgres.
+
+i.e.
+
+`s3://<the-bucket>/<build-version>/<postgres-version>/lib/postgis-3.1.so`
+`s3://<the-bucket>/<build-version>/<postgres-version>/share/extension/postgis.control`
+`s3://<the-bucket>/<build-version>/<postgres-version>/share/extension/postgis--3.1.sql`
+
+To handle custom extensions, that available only to specific users, we use per-extension subdirectories:
+
+i.e.
+`s3://<the-bucket>/<build-version>/<postgres-version>/<custom-ext-prefix>/lib/ext-name.so`, etc.
+`s3://<the-bucket>/<build-version>/<postgres-version>/<custom-ext-prefix>/share/extension/ext-name.control`, etc.
+
+On compute start, `compute_ctl` accepts a list of custom_ext_prefixes.
+
+To get the list of available extensions,`compute_ctl` downloads control files from all prefixes:
+
+`s3://<the-bucket>/<build-version>/<postgres-version>/share/extension/`
+`s3://<the-bucket>/<build-version>/<postgres-version>/<custom-ext-prefix1>/share/extension/`
+`s3://<the-bucket>/<build-version>/<postgres-version>/<custom-ext-prefix2>/share/extension/`
+
+
+
+### How to add new extension to the Extension Storage?
+
+Simply upload build artifacts to the S3 bucket.
+Implement a CI step for that. Splitting it from ompute-node-image build.
+
+### How do we deal with extension versions and updates?
+
+Currently, we rebuild extensions on every compute-node-image build and store them in the <build-version> prefix.
+This is needed to ensure that `/share` and `/lib` files are in sync.
+
+For extension updates, we rely on the PostgreSQL extension versioning mechanism (sql update scripts) and extension authors to not break backwards compatibility within one major version of PostgreSQL.
+
+### Alternatives
+
+For extensions written on trusted languages we can also adopt
+`dbdev` PostgreSQL Package Manager based on `pg_tle` by Supabase.
+This will increase the amount supported extensions and decrease the amount of work required to support them.
--- a/libs/compute_api/src/responses.rs
+++ b/libs/compute_api/src/responses.rs
@@ -73,6 +73,7 @@ pub struct ComputeMetrics {
    pub basebackup_ms: u64,
    pub config_ms: u64,
    pub total_startup_ms: u64,
+    pub load_libraries_ms: u64,
 }

 /// Response of the `/computes/{compute_id}/spec` control-plane API.
--- a/libs/compute_api/src/spec.rs
+++ b/libs/compute_api/src/spec.rs
@@ -27,6 +27,12 @@ pub struct ComputeSpec {
    pub cluster: Cluster,
    pub delta_operations: Option<Vec<DeltaOp>>,

+    /// An optinal hint that can be passed to speed up startup time if we know
+    /// that no pg catalog mutations (like role creation, database creation,
+    /// extension creation) need to be done on the actual database to start.
+    #[serde(default)] // Default false
+    pub skip_pg_catalog_updates: bool,
+
    // Information needed to connect to the storage layer.
    //
    // `tenant_id`, `timeline_id` and `pageserver_connstring` are always needed.
@@ -54,6 +60,9 @@ pub struct ComputeSpec {
    /// If set, 'storage_auth_token' is used as the password to authenticate to
    /// the pageserver and safekeepers.
    pub storage_auth_token: Option<String>,
+
+    // list of prefixes to search for custom extensions in remote extension storage
+    pub custom_extensions: Option<Vec<String>>,
 }

 #[serde_as]
@@ -142,4 +151,14 @@ mod tests {
        let file = File::open("tests/cluster_spec.json").unwrap();
        let _spec: ComputeSpec = serde_json::from_reader(file).unwrap();
    }
+
+    #[test]
+    fn parse_unknown_fields() {
+        // Forward compatibility test
+        let file = File::open("tests/cluster_spec.json").unwrap();
+        let mut json: serde_json::Value = serde_json::from_reader(file).unwrap();
+        let ob = json.as_object_mut().unwrap();
+        ob.insert("unknown_field_123123123".into(), "hello".into());
+        let _spec: ComputeSpec = serde_json::from_value(json).unwrap();
+    }
 }
--- a/libs/metrics/src/lib.rs
+++ b/libs/metrics/src/lib.rs
@@ -23,6 +23,7 @@ use prometheus::{Registry, Result};
 pub mod launch_timestamp;
 mod wrappers;
 pub use wrappers::{CountedReader, CountedWriter};
+pub mod metric_vec_duration;

 pub type UIntGauge = GenericGauge<AtomicU64>;
 pub type UIntGaugeVec = GenericGaugeVec<AtomicU64>;
--- a/libs/metrics/src/metric_vec_duration.rs
+++ b/libs/metrics/src/metric_vec_duration.rs
@@ -0,0 +1,23 @@
+//! Helpers for observing duration on HistogramVec / CounterVec / GaugeVec / MetricVec<T>.
+
+use std::{future::Future, time::Instant};
+
+pub trait DurationResultObserver {
+    fn observe_result<T, E>(&self, res: &Result<T, E>, duration: std::time::Duration);
+}
+
+pub async fn observe_async_block_duration_by_result<
+    T,
+    E,
+    F: Future<Output = Result<T, E>>,
+    O: DurationResultObserver,
+>(
+    observer: &O,
+    block: F,
+) -> Result<T, E> {
+    let start = Instant::now();
+    let result = block.await;
+    let duration = start.elapsed();
+    observer.observe_result(&result, duration);
+    result
+}
--- a/libs/pageserver_api/src/models.rs
+++ b/libs/pageserver_api/src/models.rs
@@ -152,7 +152,7 @@ pub enum ActivatingFrom {
 }

 /// A state of a timeline in pageserver's memory.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 pub enum TimelineState {
    /// The timeline is recognized by the pageserver but is not yet operational.
    /// In particular, the walreceiver connection loop is not running for this timeline.
@@ -165,7 +165,7 @@ pub enum TimelineState {
    /// It cannot transition back into any other state.
    Stopping,
    /// The timeline is broken and not operational (previous states: Loading or Active).
-    Broken,
+    Broken { reason: String, backtrace: String },
 }

 #[serde_as]
--- a/libs/remote_storage/src/lib.rs
+++ b/libs/remote_storage/src/lib.rs
@@ -70,6 +70,14 @@ impl RemotePath {
    pub fn join(&self, segment: &Path) -> Self {
        Self(self.0.join(segment))
    }
+
+    pub fn get_path(&self) -> &PathBuf {
+        &self.0
+    }
+
+    pub fn extension(&self) -> Option<&str> {
+        self.0.extension()?.to_str()
+    }
 }

 /// Storage (potentially remote) API to manage its state.
@@ -86,6 +94,19 @@ pub trait RemoteStorage: Send + Sync + 'static {
        prefix: Option<&RemotePath>,
    ) -> Result<Vec<RemotePath>, DownloadError>;

+    /// Lists all files in directory "recursively"
+    /// (not really recursively, because AWS has a flat namespace)
+    /// Note: This is subtely different than list_prefixes,
+    /// because it is for listing files instead of listing
+    /// names sharing common prefixes.
+    /// For example,
+    /// list_files("foo/bar") = ["foo/bar/cat123.txt",
+    /// "foo/bar/cat567.txt", "foo/bar/dog123.txt", "foo/bar/dog456.txt"]
+    /// whereas,
+    /// list_prefixes("foo/bar/") = ["cat", "dog"]
+    /// See `test_real_s3.rs` for more details.
+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>>;
+
    /// Streams the local file contents into remote into the remote storage entry.
    async fn upload(
        &self,
@@ -111,6 +132,8 @@ pub trait RemoteStorage: Send + Sync + 'static {
    ) -> Result<Download, DownloadError>;

    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()>;
+
+    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()>;
 }

 pub struct Download {
@@ -161,6 +184,20 @@ pub enum GenericRemoteStorage {
 }

 impl GenericRemoteStorage {
+    // A function for listing all the files in a "directory"
+    // Example:
+    // list_files("foo/bar") = ["foo/bar/a.txt", "foo/bar/b.txt"]
+    pub async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+        match self {
+            Self::LocalFs(s) => s.list_files(folder).await,
+            Self::AwsS3(s) => s.list_files(folder).await,
+            Self::Unreliable(s) => s.list_files(folder).await,
+        }
+    }
+
+    // lists common *prefixes*, if any of files
+    // Example:
+    // list_prefixes("foo123","foo567","bar123","bar432") = ["foo", "bar"]
    pub async fn list_prefixes(
        &self,
        prefix: Option<&RemotePath>,
@@ -223,6 +260,14 @@ impl GenericRemoteStorage {
            Self::Unreliable(s) => s.delete(path).await,
        }
    }
+
+    pub async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+        match self {
+            Self::LocalFs(s) => s.delete_objects(paths).await,
+            Self::AwsS3(s) => s.delete_objects(paths).await,
+            Self::Unreliable(s) => s.delete_objects(paths).await,
+        }
+    }
 }

 impl GenericRemoteStorage {
--- a/libs/remote_storage/src/local_fs.rs
+++ b/libs/remote_storage/src/local_fs.rs
@@ -17,7 +17,7 @@ use tokio::{
    io::{self, AsyncReadExt, AsyncSeekExt, AsyncWriteExt},
 };
 use tracing::*;
-use utils::crashsafe::path_with_suffix_extension;
+use utils::{crashsafe::path_with_suffix_extension, fs_ext::is_directory_empty};

 use crate::{Download, DownloadError, RemotePath};

@@ -48,6 +48,14 @@ impl LocalFs {
        Ok(Self { storage_root })
    }

+    // mirrors S3Bucket::s3_object_to_relative_path
+    fn local_file_to_relative_path(&self, key: PathBuf) -> RemotePath {
+        let relative_path = key
+            .strip_prefix(&self.storage_root)
+            .expect("relative path must contain storage_root as prefix");
+        RemotePath(relative_path.into())
+    }
+
    async fn read_storage_metadata(
        &self,
        file_path: &Path,
@@ -101,19 +109,63 @@ impl RemoteStorage for LocalFs {
            Some(prefix) => Cow::Owned(prefix.with_base(&self.storage_root)),
            None => Cow::Borrowed(&self.storage_root),
        };
-        Ok(get_all_files(path.as_ref(), false)
+
+        let prefixes_to_filter = get_all_files(path.as_ref(), false)
            .await
-            .map_err(DownloadError::Other)?
-            .into_iter()
-            .map(|path| {
-                path.strip_prefix(&self.storage_root)
-                    .context("Failed to strip preifix")
+            .map_err(DownloadError::Other)?;
+
+        let mut prefixes = Vec::with_capacity(prefixes_to_filter.len());
+
+        // filter out empty directories to mirror s3 behavior.
+        for prefix in prefixes_to_filter {
+            if prefix.is_dir()
+                && is_directory_empty(&prefix)
+                    .await
+                    .map_err(DownloadError::Other)?
+            {
+                continue;
+            }
+
+            prefixes.push(
+                prefix
+                    .strip_prefix(&self.storage_root)
+                    .context("Failed to strip prefix")
                    .and_then(RemotePath::new)
                    .expect(
                        "We list files for storage root, hence should be able to remote the prefix",
-                    )
-            })
-            .collect())
+                    ),
+            )
+        }
+
+        Ok(prefixes)
+    }
+
+    // recursively lists all files in a directory,
+    // mirroring the `list_files` for `s3_bucket`
+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+        let full_path = match folder {
+            Some(folder) => folder.with_base(&self.storage_root),
+            None => self.storage_root.clone(),
+        };
+        let mut files = vec![];
+        let mut directory_queue = vec![full_path.clone()];
+
+        while !directory_queue.is_empty() {
+            let cur_folder = directory_queue
+                .pop()
+                .expect("queue cannot be empty: we just checked");
+            let mut entries = fs::read_dir(cur_folder.clone()).await?;
+            while let Some(entry) = entries.next_entry().await? {
+                let file_name: PathBuf = entry.file_name().into();
+                let full_file_name = cur_folder.clone().join(&file_name);
+                let file_remote_path = self.local_file_to_relative_path(full_file_name.clone());
+                files.push(file_remote_path.clone());
+                if full_file_name.is_dir() {
+                    directory_queue.push(full_file_name);
+                }
+            }
+        }
+        Ok(files)
    }

    async fn upload(
@@ -291,11 +343,25 @@ 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() && file_path.is_file() {
-            Ok(fs::remove_file(file_path).await?)
-        } else {
-            bail!("File {file_path:?} either does not exist or is not a file")
+        if !file_path.exists() {
+            // 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(());
        }
+
+        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<()> {
+        for path in paths {
+            self.delete(path).await?
+        }
+        Ok(())
    }
 }

@@ -320,7 +386,7 @@ where
                    let file_type = dir_entry.file_type().await?;
                    let entry_path = dir_entry.path();
                    if file_type.is_symlink() {
-                        debug!("{entry_path:?} us a symlink, skipping")
+                        debug!("{entry_path:?} is a symlink, skipping")
                    } else if file_type.is_dir() {
                        if recursive {
                            paths.extend(get_all_files(&entry_path, true).await?.into_iter())
@@ -595,15 +661,11 @@ mod fs_tests {
        storage.delete(&upload_target).await?;
        assert!(storage.list().await?.is_empty());

-        match storage.delete(&upload_target).await {
-            Ok(()) => panic!("Should not allow deleting non-existing storage files"),
-            Err(e) => {
-                let error_string = e.to_string();
-                assert!(error_string.contains("does not exist"));
-                let expected_path = upload_target.with_base(&storage.storage_root);
-                assert!(error_string.contains(expected_path.to_str().unwrap()));
-            }
-        }
+        storage
+            .delete(&upload_target)
+            .await
+            .expect("Should allow deleting non-existing storage files");
+
        Ok(())
    }

--- a/libs/remote_storage/src/s3_bucket.rs
+++ b/libs/remote_storage/src/s3_bucket.rs
@@ -17,6 +17,7 @@ use aws_sdk_s3::{
    error::SdkError,
    operation::get_object::GetObjectError,
    primitives::ByteStream,
+    types::{Delete, ObjectIdentifier},
    Client,
 };
 use aws_smithy_http::body::SdkBody;
@@ -33,6 +34,8 @@ use crate::{
    Download, DownloadError, RemotePath, RemoteStorage, S3Config, REMOTE_STORAGE_PREFIX_SEPARATOR,
 };

+const MAX_DELETE_OBJECTS_REQUEST_SIZE: usize = 1000;
+
 pub(super) mod metrics {
    use metrics::{register_int_counter_vec, IntCounterVec};
    use once_cell::sync::Lazy;
@@ -81,12 +84,24 @@ pub(super) mod metrics {
            .inc();
    }

+    pub fn inc_delete_objects(count: u64) {
+        S3_REQUESTS_COUNT
+            .with_label_values(&["delete_object"])
+            .inc_by(count);
+    }
+
    pub fn inc_delete_object_fail() {
        S3_REQUESTS_FAIL_COUNT
            .with_label_values(&["delete_object"])
            .inc();
    }

+    pub fn inc_delete_objects_fail(count: u64) {
+        S3_REQUESTS_FAIL_COUNT
+            .with_label_values(&["delete_object"])
+            .inc_by(count);
+    }
+
    pub fn inc_list_objects() {
        S3_REQUESTS_COUNT.with_label_values(&["list_objects"]).inc();
    }
@@ -332,6 +347,58 @@ impl RemoteStorage for S3Bucket {
        Ok(document_keys)
    }

+    /// See the doc for `RemoteStorage::list_files`
+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+        let mut folder_name = folder
+            .map(|p| self.relative_path_to_s3_object(p))
+            .or_else(|| self.prefix_in_bucket.clone());
+
+        // remove leading "/" if one exists
+        if let Some(folder_name_slash) = folder_name.clone() {
+            if folder_name_slash.starts_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                folder_name = Some(folder_name_slash[1..].to_string());
+            }
+        }
+
+        // AWS may need to break the response into several parts
+        let mut continuation_token = None;
+        let mut all_files = vec![];
+        loop {
+            let _guard = self
+                .concurrency_limiter
+                .acquire()
+                .await
+                .context("Concurrency limiter semaphore got closed during S3 list_files")?;
+            metrics::inc_list_objects();
+
+            let response = self
+                .client
+                .list_objects_v2()
+                .bucket(self.bucket_name.clone())
+                .set_prefix(folder_name.clone())
+                .set_continuation_token(continuation_token)
+                .set_max_keys(self.max_keys_per_list_response)
+                .send()
+                .await
+                .map_err(|e| {
+                    metrics::inc_list_objects_fail();
+                    e
+                })
+                .context("Failed to list files in S3 bucket")?;
+
+            for object in response.contents().unwrap_or_default() {
+                let object_path = object.key().expect("response does not contain a key");
+                let remote_path = self.s3_object_to_relative_path(object_path);
+                all_files.push(remote_path);
+            }
+            match response.next_continuation_token {
+                Some(new_token) => continuation_token = Some(new_token),
+                None => break,
+            }
+        }
+        Ok(all_files)
+    }
+
    async fn upload(
        &self,
        from: impl io::AsyncRead + Unpin + Send + Sync + 'static,
@@ -396,6 +463,50 @@ impl RemoteStorage for S3Bucket {
        })
        .await
    }
+    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+        let _guard = self
+            .concurrency_limiter
+            .acquire()
+            .await
+            .context("Concurrency limiter semaphore got closed during S3 delete")?;
+
+        let mut delete_objects = Vec::with_capacity(paths.len());
+        for path in paths {
+            let obj_id = ObjectIdentifier::builder()
+                .set_key(Some(self.relative_path_to_s3_object(path)))
+                .build();
+            delete_objects.push(obj_id);
+        }
+
+        for chunk in delete_objects.chunks(MAX_DELETE_OBJECTS_REQUEST_SIZE) {
+            metrics::inc_delete_objects(chunk.len() as u64);
+
+            let resp = self
+                .client
+                .delete_objects()
+                .bucket(self.bucket_name.clone())
+                .delete(Delete::builder().set_objects(Some(chunk.to_vec())).build())
+                .send()
+                .await;
+
+            match resp {
+                Ok(resp) => {
+                    if let Some(errors) = resp.errors {
+                        metrics::inc_delete_objects_fail(errors.len() as u64);
+                        return Err(anyhow::format_err!(
+                            "Failed to delete {} objects",
+                            errors.len()
+                        ));
+                    }
+                }
+                Err(e) => {
+                    metrics::inc_delete_objects_fail(chunk.len() as u64);
+                    return Err(e.into());
+                }
+            }
+        }
+        Ok(())
+    }

    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
        let _guard = self
--- a/libs/remote_storage/src/simulate_failures.rs
+++ b/libs/remote_storage/src/simulate_failures.rs
@@ -24,6 +24,7 @@ enum RemoteOp {
    Upload(RemotePath),
    Download(RemotePath),
    Delete(RemotePath),
+    DeleteObjects(Vec<RemotePath>),
 }

 impl UnreliableWrapper {
@@ -82,6 +83,11 @@ impl RemoteStorage for UnreliableWrapper {
        self.inner.list_prefixes(prefix).await
    }

+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+        self.attempt(RemoteOp::ListPrefixes(folder.cloned()))?;
+        self.inner.list_files(folder).await
+    }
+
    async fn upload(
        &self,
        data: impl tokio::io::AsyncRead + Unpin + Send + Sync + 'static,
@@ -119,4 +125,21 @@ impl RemoteStorage for UnreliableWrapper {
        self.attempt(RemoteOp::Delete(path.clone()))?;
        self.inner.delete(path).await
    }
+
+    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+        self.attempt(RemoteOp::DeleteObjects(paths.to_vec()))?;
+        let mut error_counter = 0;
+        for path in paths {
+            if (self.delete(path).await).is_err() {
+                error_counter += 1;
+            }
+        }
+        if error_counter > 0 {
+            return Err(anyhow::anyhow!(
+                "failed to delete {} objects",
+                error_counter
+            ));
+        }
+        Ok(())
+    }
 }
--- a/libs/remote_storage/tests/pagination_tests.rs
+++ b/libs/remote_storage/tests/pagination_tests.rs
@@ -1,274 +0,0 @@
-use std::collections::HashSet;
-use std::env;
-use std::num::{NonZeroU32, NonZeroUsize};
-use std::ops::ControlFlow;
-use std::path::{Path, PathBuf};
-use std::sync::Arc;
-use std::time::UNIX_EPOCH;
-
-use anyhow::Context;
-use remote_storage::{
-    GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind, S3Config,
-};
-use test_context::{test_context, AsyncTestContext};
-use tokio::task::JoinSet;
-use tracing::{debug, error, info};
-
-const ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_S3_REMOTE_STORAGE";
-
-/// Tests that S3 client can list all prefixes, even if the response come paginated and requires multiple S3 queries.
-/// Uses real S3 and requires [`ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME`] and related S3 cred env vars specified.
-/// See the client creation in [`create_s3_client`] for details on the required env vars.
-/// If real S3 tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
-/// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
-///
-/// First, the test creates a set of S3 objects with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_s3_data`]
-/// where
-/// * `random_prefix_part` is set for the entire S3 client during the S3 client creation in [`create_s3_client`], to avoid multiple test runs interference
-/// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
-///
-/// Then, verifies that the client does return correct prefixes when queried:
-/// * with no prefix, it lists everything after its `${random_prefix_part}/` — that should be `${base_prefix_str}` value only
-/// * with `${base_prefix_str}/` prefix, it lists every `sub_prefix_${i}`
-///
-/// With the real S3 enabled and `#[cfg(test)]` Rust configuration used, the S3 client test adds a `max-keys` param to limit the response keys.
-/// This way, we are able to test the pagination implicitly, by ensuring all results are returned from the remote storage and avoid uploading too many blobs to S3,
-/// since current default AWS S3 pagination limit is 1000.
-/// (see https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax)
-///
-/// Lastly, the test attempts to clean up and remove all uploaded S3 files.
-/// If any errors appear during the clean up, they get logged, but the test is not failed or stopped until clean up is finished.
-#[test_context(MaybeEnabledS3)]
-#[tokio::test]
-async fn s3_pagination_should_work(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()> {
-    let ctx = match ctx {
-        MaybeEnabledS3::Enabled(ctx) => ctx,
-        MaybeEnabledS3::Disabled => return Ok(()),
-        MaybeEnabledS3::UploadsFailed(e, _) => anyhow::bail!("S3 init failed: {e:?}"),
-    };
-
-    let test_client = Arc::clone(&ctx.client_with_excessive_pagination);
-    let expected_remote_prefixes = ctx.remote_prefixes.clone();
-
-    let base_prefix =
-        RemotePath::new(Path::new(ctx.base_prefix_str)).context("common_prefix construction")?;
-    let root_remote_prefixes = test_client
-        .list_prefixes(None)
-        .await
-        .context("client list root prefixes failure")?
-        .into_iter()
-        .collect::<HashSet<_>>();
-    assert_eq!(
-        root_remote_prefixes, HashSet::from([base_prefix.clone()]),
-        "remote storage root prefixes list mismatches with the uploads. Returned prefixes: {root_remote_prefixes:?}"
-    );
-
-    let nested_remote_prefixes = test_client
-        .list_prefixes(Some(&base_prefix))
-        .await
-        .context("client list nested prefixes failure")?
-        .into_iter()
-        .collect::<HashSet<_>>();
-    let remote_only_prefixes = nested_remote_prefixes
-        .difference(&expected_remote_prefixes)
-        .collect::<HashSet<_>>();
-    let missing_uploaded_prefixes = expected_remote_prefixes
-        .difference(&nested_remote_prefixes)
-        .collect::<HashSet<_>>();
-    assert_eq!(
-        remote_only_prefixes.len() + missing_uploaded_prefixes.len(), 0,
-        "remote storage nested prefixes list mismatches with the uploads. Remote only prefixes: {remote_only_prefixes:?}, missing uploaded prefixes: {missing_uploaded_prefixes:?}",
-    );
-
-    Ok(())
-}
-
-enum MaybeEnabledS3 {
-    Enabled(S3WithTestBlobs),
-    Disabled,
-    UploadsFailed(anyhow::Error, S3WithTestBlobs),
-}
-
-struct S3WithTestBlobs {
-    client_with_excessive_pagination: Arc<GenericRemoteStorage>,
-    base_prefix_str: &'static str,
-    remote_prefixes: HashSet<RemotePath>,
-    remote_blobs: HashSet<RemotePath>,
-}
-
-#[async_trait::async_trait]
-impl AsyncTestContext for MaybeEnabledS3 {
-    async fn setup() -> Self {
-        utils::logging::init(
-            utils::logging::LogFormat::Test,
-            utils::logging::TracingErrorLayerEnablement::Disabled,
-        )
-        .expect("logging init failed");
-        if env::var(ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
-            info!(
-                "`{}` env variable is not set, skipping the test",
-                ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME
-            );
-            return Self::Disabled;
-        }
-
-        let max_keys_in_list_response = 10;
-        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
-
-        let client_with_excessive_pagination = create_s3_client(max_keys_in_list_response)
-            .context("S3 client creation")
-            .expect("S3 client creation failed");
-
-        let base_prefix_str = "test/";
-        match upload_s3_data(
-            &client_with_excessive_pagination,
-            base_prefix_str,
-            upload_tasks_count,
-        )
-        .await
-        {
-            ControlFlow::Continue(uploads) => {
-                info!("Remote objects created successfully");
-                Self::Enabled(S3WithTestBlobs {
-                    client_with_excessive_pagination,
-                    base_prefix_str,
-                    remote_prefixes: uploads.prefixes,
-                    remote_blobs: uploads.blobs,
-                })
-            }
-            ControlFlow::Break(uploads) => Self::UploadsFailed(
-                anyhow::anyhow!("One or multiple blobs failed to upload to S3"),
-                S3WithTestBlobs {
-                    client_with_excessive_pagination,
-                    base_prefix_str,
-                    remote_prefixes: uploads.prefixes,
-                    remote_blobs: uploads.blobs,
-                },
-            ),
-        }
-    }
-
-    async fn teardown(self) {
-        match self {
-            Self::Disabled => {}
-            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
-                cleanup(&ctx.client_with_excessive_pagination, ctx.remote_blobs).await;
-            }
-        }
-    }
-}
-
-fn create_s3_client(max_keys_per_list_response: i32) -> anyhow::Result<Arc<GenericRemoteStorage>> {
-    let remote_storage_s3_bucket = env::var("REMOTE_STORAGE_S3_BUCKET")
-        .context("`REMOTE_STORAGE_S3_BUCKET` env var is not set, but real S3 tests are enabled")?;
-    let remote_storage_s3_region = env::var("REMOTE_STORAGE_S3_REGION")
-        .context("`REMOTE_STORAGE_S3_REGION` env var is not set, but real S3 tests are enabled")?;
-    let random_prefix_part = std::time::SystemTime::now()
-        .duration_since(UNIX_EPOCH)
-        .context("random s3 test prefix part calculation")?
-        .as_millis();
-    let remote_storage_config = RemoteStorageConfig {
-        max_concurrent_syncs: NonZeroUsize::new(100).unwrap(),
-        max_sync_errors: NonZeroU32::new(5).unwrap(),
-        storage: RemoteStorageKind::AwsS3(S3Config {
-            bucket_name: remote_storage_s3_bucket,
-            bucket_region: remote_storage_s3_region,
-            prefix_in_bucket: Some(format!("pagination_should_work_test_{random_prefix_part}/")),
-            endpoint: None,
-            concurrency_limit: NonZeroUsize::new(100).unwrap(),
-            max_keys_per_list_response: Some(max_keys_per_list_response),
-        }),
-    };
-    Ok(Arc::new(
-        GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
-    ))
-}
-
-struct Uploads {
-    prefixes: HashSet<RemotePath>,
-    blobs: HashSet<RemotePath>,
-}
-
-async fn upload_s3_data(
-    client: &Arc<GenericRemoteStorage>,
-    base_prefix_str: &'static str,
-    upload_tasks_count: usize,
-) -> ControlFlow<Uploads, Uploads> {
-    info!("Creating {upload_tasks_count} S3 files");
-    let mut upload_tasks = JoinSet::new();
-    for i in 1..upload_tasks_count + 1 {
-        let task_client = Arc::clone(client);
-        upload_tasks.spawn(async move {
-            let prefix = PathBuf::from(format!("{base_prefix_str}/sub_prefix_{i}/"));
-            let blob_prefix = RemotePath::new(&prefix)
-                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
-            let blob_path = blob_prefix.join(Path::new(&format!("blob_{i}")));
-            debug!("Creating remote item {i} at path {blob_path:?}");
-
-            let data = format!("remote blob data {i}").into_bytes();
-            let data_len = data.len();
-            task_client
-                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
-                .await?;
-
-            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
-        });
-    }
-
-    let mut upload_tasks_failed = false;
-    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
-    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
-    while let Some(task_run_result) = upload_tasks.join_next().await {
-        match task_run_result
-            .context("task join failed")
-            .and_then(|task_result| task_result.context("upload task failed"))
-        {
-            Ok((upload_prefix, upload_path)) => {
-                uploaded_prefixes.insert(upload_prefix);
-                uploaded_blobs.insert(upload_path);
-            }
-            Err(e) => {
-                error!("Upload task failed: {e:?}");
-                upload_tasks_failed = true;
-            }
-        }
-    }
-
-    let uploads = Uploads {
-        prefixes: uploaded_prefixes,
-        blobs: uploaded_blobs,
-    };
-    if upload_tasks_failed {
-        ControlFlow::Break(uploads)
-    } else {
-        ControlFlow::Continue(uploads)
-    }
-}
-
-async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
-    info!(
-        "Removing {} objects from the remote storage during cleanup",
-        objects_to_delete.len()
-    );
-    let mut delete_tasks = JoinSet::new();
-    for object_to_delete in objects_to_delete {
-        let task_client = Arc::clone(client);
-        delete_tasks.spawn(async move {
-            debug!("Deleting remote item at path {object_to_delete:?}");
-            task_client
-                .delete(&object_to_delete)
-                .await
-                .with_context(|| format!("{object_to_delete:?} removal"))
-        });
-    }
-
-    while let Some(task_run_result) = delete_tasks.join_next().await {
-        match task_run_result {
-            Ok(task_result) => match task_result {
-                Ok(()) => {}
-                Err(e) => error!("Delete task failed: {e:?}"),
-            },
-            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
-        }
-    }
-}
--- a/libs/remote_storage/tests/test_real_s3.rs
+++ b/libs/remote_storage/tests/test_real_s3.rs
@@ -0,0 +1,542 @@
+use std::collections::HashSet;
+use std::env;
+use std::num::{NonZeroU32, NonZeroUsize};
+use std::ops::ControlFlow;
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+use std::time::UNIX_EPOCH;
+
+use anyhow::Context;
+use once_cell::sync::OnceCell;
+use remote_storage::{
+    GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind, S3Config,
+};
+use test_context::{test_context, AsyncTestContext};
+use tokio::task::JoinSet;
+use tracing::{debug, error, info};
+
+static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+
+const ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_S3_REMOTE_STORAGE";
+
+const BASE_PREFIX: &str = "test/";
+
+/// Tests that S3 client can list all prefixes, even if the response come paginated and requires multiple S3 queries.
+/// Uses real S3 and requires [`ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME`] and related S3 cred env vars specified.
+/// See the client creation in [`create_s3_client`] for details on the required env vars.
+/// If real S3 tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
+/// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
+///
+/// First, the test creates a set of S3 objects with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_s3_data`]
+/// where
+/// * `random_prefix_part` is set for the entire S3 client during the S3 client creation in [`create_s3_client`], to avoid multiple test runs interference
+/// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
+///
+/// Then, verifies that the client does return correct prefixes when queried:
+/// * with no prefix, it lists everything after its `${random_prefix_part}/` — that should be `${base_prefix_str}` value only
+/// * with `${base_prefix_str}/` prefix, it lists every `sub_prefix_${i}`
+///
+/// With the real S3 enabled and `#[cfg(test)]` Rust configuration used, the S3 client test adds a `max-keys` param to limit the response keys.
+/// This way, we are able to test the pagination implicitly, by ensuring all results are returned from the remote storage and avoid uploading too many blobs to S3,
+/// since current default AWS S3 pagination limit is 1000.
+/// (see https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax)
+///
+/// Lastly, the test attempts to clean up and remove all uploaded S3 files.
+/// If any errors appear during the clean up, they get logged, but the test is not failed or stopped until clean up is finished.
+#[test_context(MaybeEnabledS3WithTestBlobs)]
+#[tokio::test]
+async fn s3_pagination_should_work(ctx: &mut MaybeEnabledS3WithTestBlobs) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledS3WithTestBlobs::Enabled(ctx) => ctx,
+        MaybeEnabledS3WithTestBlobs::Disabled => return Ok(()),
+        MaybeEnabledS3WithTestBlobs::UploadsFailed(e, _) => anyhow::bail!("S3 init failed: {e:?}"),
+    };
+
+    let test_client = Arc::clone(&ctx.enabled.client);
+    let expected_remote_prefixes = ctx.remote_prefixes.clone();
+
+    let base_prefix = RemotePath::new(Path::new(ctx.enabled.base_prefix))
+        .context("common_prefix construction")?;
+    let root_remote_prefixes = test_client
+        .list_prefixes(None)
+        .await
+        .context("client list root prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_remote_prefixes, HashSet::from([base_prefix.clone()]),
+        "remote storage root prefixes list mismatches with the uploads. Returned prefixes: {root_remote_prefixes:?}"
+    );
+
+    let nested_remote_prefixes = test_client
+        .list_prefixes(Some(&base_prefix))
+        .await
+        .context("client list nested prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let remote_only_prefixes = nested_remote_prefixes
+        .difference(&expected_remote_prefixes)
+        .collect::<HashSet<_>>();
+    let missing_uploaded_prefixes = expected_remote_prefixes
+        .difference(&nested_remote_prefixes)
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        remote_only_prefixes.len() + missing_uploaded_prefixes.len(), 0,
+        "remote storage nested prefixes list mismatches with the uploads. Remote only prefixes: {remote_only_prefixes:?}, missing uploaded prefixes: {missing_uploaded_prefixes:?}",
+    );
+
+    Ok(())
+}
+
+/// Tests that S3 client can list all files in a folder, even if the response comes paginated and requirees multiple S3 queries.
+/// Uses real S3 and requires [`ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME`] and related S3 cred env vars specified. Test will skip real code and pass if env vars not set.
+/// See `s3_pagination_should_work` for more information.
+///
+/// First, create a set of S3 objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_s3_data`]
+/// Then performs the following queries:
+///    1. `list_files(None)`. This should return all files `random_prefix/folder{j}/blob_{i}.txt`
+///    2. `list_files("folder1")`.  This  should return all files `random_prefix/folder1/blob_{i}.txt`
+#[test_context(MaybeEnabledS3WithSimpleTestBlobs)]
+#[tokio::test]
+async fn s3_list_files_works(ctx: &mut MaybeEnabledS3WithSimpleTestBlobs) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledS3WithSimpleTestBlobs::Enabled(ctx) => ctx,
+        MaybeEnabledS3WithSimpleTestBlobs::Disabled => return Ok(()),
+        MaybeEnabledS3WithSimpleTestBlobs::UploadsFailed(e, _) => {
+            anyhow::bail!("S3 init failed: {e:?}")
+        }
+    };
+    let test_client = Arc::clone(&ctx.enabled.client);
+    let base_prefix =
+        RemotePath::new(Path::new("folder1")).context("common_prefix construction")?;
+    let root_files = test_client
+        .list_files(None)
+        .await
+        .context("client list root files failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_files,
+        ctx.remote_blobs.clone(),
+        "remote storage list_files on root mismatches with the uploads."
+    );
+    let nested_remote_files = test_client
+        .list_files(Some(&base_prefix))
+        .await
+        .context("client list nested files failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let trim_remote_blobs: HashSet<_> = ctx
+        .remote_blobs
+        .iter()
+        .map(|x| x.get_path().to_str().expect("must be valid name"))
+        .filter(|x| x.starts_with("folder1"))
+        .map(|x| RemotePath::new(Path::new(x)).expect("must be valid name"))
+        .collect();
+    assert_eq!(
+        nested_remote_files, trim_remote_blobs,
+        "remote storage list_files on subdirrectory mismatches with the uploads."
+    );
+    Ok(())
+}
+
+#[test_context(MaybeEnabledS3)]
+#[tokio::test]
+async fn s3_delete_non_exising_works(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledS3::Enabled(ctx) => ctx,
+        MaybeEnabledS3::Disabled => return Ok(()),
+    };
+
+    let path = RemotePath::new(&PathBuf::from(format!(
+        "{}/for_sure_there_is_nothing_there_really",
+        ctx.base_prefix,
+    )))
+    .with_context(|| "RemotePath conversion")?;
+
+    ctx.client.delete(&path).await.expect("should succeed");
+
+    Ok(())
+}
+
+#[test_context(MaybeEnabledS3)]
+#[tokio::test]
+async fn s3_delete_objects_works(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledS3::Enabled(ctx) => ctx,
+        MaybeEnabledS3::Disabled => return Ok(()),
+    };
+
+    let path1 = RemotePath::new(&PathBuf::from(format!("{}/path1", ctx.base_prefix,)))
+        .with_context(|| "RemotePath conversion")?;
+
+    let path2 = RemotePath::new(&PathBuf::from(format!("{}/path2", ctx.base_prefix,)))
+        .with_context(|| "RemotePath conversion")?;
+
+    let path3 = RemotePath::new(&PathBuf::from(format!("{}/path3", ctx.base_prefix,)))
+        .with_context(|| "RemotePath conversion")?;
+
+    let data1 = "remote blob data1".as_bytes();
+    let data1_len = data1.len();
+    let data2 = "remote blob data2".as_bytes();
+    let data2_len = data2.len();
+    let data3 = "remote blob data3".as_bytes();
+    let data3_len = data3.len();
+    ctx.client
+        .upload(std::io::Cursor::new(data1), data1_len, &path1, None)
+        .await?;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data2), data2_len, &path2, None)
+        .await?;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data3), data3_len, &path3, None)
+        .await?;
+
+    ctx.client.delete_objects(&[path1, path2]).await?;
+
+    let prefixes = ctx.client.list_prefixes(None).await?;
+
+    assert_eq!(prefixes.len(), 1);
+
+    ctx.client.delete_objects(&[path3]).await?;
+
+    Ok(())
+}
+
+fn ensure_logging_ready() {
+    LOGGING_DONE.get_or_init(|| {
+        utils::logging::init(
+            utils::logging::LogFormat::Test,
+            utils::logging::TracingErrorLayerEnablement::Disabled,
+        )
+        .expect("logging init failed");
+    });
+}
+
+struct EnabledS3 {
+    client: Arc<GenericRemoteStorage>,
+    base_prefix: &'static str,
+}
+
+impl EnabledS3 {
+    async fn setup(max_keys_in_list_response: Option<i32>) -> Self {
+        let client = create_s3_client(max_keys_in_list_response)
+            .context("S3 client creation")
+            .expect("S3 client creation failed");
+
+        EnabledS3 {
+            client,
+            base_prefix: BASE_PREFIX,
+        }
+    }
+}
+
+enum MaybeEnabledS3 {
+    Enabled(EnabledS3),
+    Disabled,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledS3 {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+
+        if env::var(ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        Self::Enabled(EnabledS3::setup(None).await)
+    }
+}
+
+enum MaybeEnabledS3WithTestBlobs {
+    Enabled(S3WithTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, S3WithTestBlobs),
+}
+
+struct S3WithTestBlobs {
+    enabled: EnabledS3,
+    remote_prefixes: HashSet<RemotePath>,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledS3WithTestBlobs {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+        if env::var(ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let enabled = EnabledS3::setup(Some(max_keys_in_list_response)).await;
+
+        match upload_s3_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+
+                Self::Enabled(S3WithTestBlobs {
+                    enabled,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to S3"),
+                S3WithTestBlobs {
+                    enabled,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.enabled.client, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+// NOTE: the setups for the list_prefixes test and the list_files test are very similar
+// However, they are not idential. The list_prefixes function is concerned with listing prefixes,
+// whereas the list_files function is concerned with listing files.
+// See `RemoteStorage::list_files` documentation for more details
+enum MaybeEnabledS3WithSimpleTestBlobs {
+    Enabled(S3WithSimpleTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, S3WithSimpleTestBlobs),
+}
+struct S3WithSimpleTestBlobs {
+    enabled: EnabledS3,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledS3WithSimpleTestBlobs {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+        if env::var(ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let enabled = EnabledS3::setup(Some(max_keys_in_list_response)).await;
+
+        match upload_simple_s3_data(&enabled.client, upload_tasks_count).await {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+
+                Self::Enabled(S3WithSimpleTestBlobs {
+                    enabled,
+                    remote_blobs: uploads,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to S3"),
+                S3WithSimpleTestBlobs {
+                    enabled,
+                    remote_blobs: uploads,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.enabled.client, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+fn create_s3_client(
+    max_keys_per_list_response: Option<i32>,
+) -> anyhow::Result<Arc<GenericRemoteStorage>> {
+    let remote_storage_s3_bucket = env::var("REMOTE_STORAGE_S3_BUCKET")
+        .context("`REMOTE_STORAGE_S3_BUCKET` env var is not set, but real S3 tests are enabled")?;
+    let remote_storage_s3_region = env::var("REMOTE_STORAGE_S3_REGION")
+        .context("`REMOTE_STORAGE_S3_REGION` env var is not set, but real S3 tests are enabled")?;
+    let random_prefix_part = std::time::SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .context("random s3 test prefix part calculation")?
+        .as_nanos();
+    let remote_storage_config = RemoteStorageConfig {
+        max_concurrent_syncs: NonZeroUsize::new(100).unwrap(),
+        max_sync_errors: NonZeroU32::new(5).unwrap(),
+        storage: RemoteStorageKind::AwsS3(S3Config {
+            bucket_name: remote_storage_s3_bucket,
+            bucket_region: remote_storage_s3_region,
+            prefix_in_bucket: Some(format!("pagination_should_work_test_{random_prefix_part}/")),
+            endpoint: None,
+            concurrency_limit: NonZeroUsize::new(100).unwrap(),
+            max_keys_per_list_response,
+        }),
+    };
+    Ok(Arc::new(
+        GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
+    ))
+}
+
+struct Uploads {
+    prefixes: HashSet<RemotePath>,
+    blobs: HashSet<RemotePath>,
+}
+
+async fn upload_s3_data(
+    client: &Arc<GenericRemoteStorage>,
+    base_prefix_str: &'static str,
+    upload_tasks_count: usize,
+) -> ControlFlow<Uploads, Uploads> {
+    info!("Creating {upload_tasks_count} S3 files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let prefix = PathBuf::from(format!("{base_prefix_str}/sub_prefix_{i}/"));
+            let blob_prefix = RemotePath::new(&prefix)
+                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
+            let blob_path = blob_prefix.join(Path::new(&format!("blob_{i}")));
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let data = format!("remote blob data {i}").into_bytes();
+            let data_len = data.len();
+            task_client
+                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
+                .await?;
+
+            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok((upload_prefix, upload_path)) => {
+                uploaded_prefixes.insert(upload_prefix);
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    let uploads = Uploads {
+        prefixes: uploaded_prefixes,
+        blobs: uploaded_blobs,
+    };
+    if upload_tasks_failed {
+        ControlFlow::Break(uploads)
+    } else {
+        ControlFlow::Continue(uploads)
+    }
+}
+
+async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
+    info!(
+        "Removing {} objects from the remote storage during cleanup",
+        objects_to_delete.len()
+    );
+    let mut delete_tasks = JoinSet::new();
+    for object_to_delete in objects_to_delete {
+        let task_client = Arc::clone(client);
+        delete_tasks.spawn(async move {
+            debug!("Deleting remote item at path {object_to_delete:?}");
+            task_client
+                .delete(&object_to_delete)
+                .await
+                .with_context(|| format!("{object_to_delete:?} removal"))
+        });
+    }
+
+    while let Some(task_run_result) = delete_tasks.join_next().await {
+        match task_run_result {
+            Ok(task_result) => match task_result {
+                Ok(()) => {}
+                Err(e) => error!("Delete task failed: {e:?}"),
+            },
+            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
+        }
+    }
+}
+
+// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
+async fn upload_simple_s3_data(
+    client: &Arc<GenericRemoteStorage>,
+    upload_tasks_count: usize,
+) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
+    info!("Creating {upload_tasks_count} S3 files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
+            let blob_path = RemotePath::new(&blob_path)
+                .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let data = format!("remote blob data {i}").into_bytes();
+            let data_len = data.len();
+            task_client
+                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
+                .await?;
+
+            Ok::<_, anyhow::Error>(blob_path)
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok(upload_path) => {
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    if upload_tasks_failed {
+        ControlFlow::Break(uploaded_blobs)
+    } else {
+        ControlFlow::Continue(uploaded_blobs)
+    }
+}
--- a/libs/utils/Cargo.toml
+++ b/libs/utils/Cargo.toml
@@ -5,7 +5,6 @@ edition.workspace = true
 license.workspace = true

 [dependencies]
-atty.workspace = true
 sentry.workspace = true
 async-trait.workspace = true
 anyhow.workspace = true
--- a/libs/utils/src/fs_ext.rs
+++ b/libs/utils/src/fs_ext.rs
@@ -1,6 +1,8 @@
 /// Extensions to `std::fs` types.
 use std::{fs, io, path::Path};

+use anyhow::Context;
+
 pub trait PathExt {
    /// Returns an error if `self` is not a directory.
    fn is_empty_dir(&self) -> io::Result<bool>;
@@ -15,10 +17,19 @@ where
    }
 }

+pub async fn is_directory_empty(path: impl AsRef<Path>) -> anyhow::Result<bool> {
+    let mut dir = tokio::fs::read_dir(&path)
+        .await
+        .context(format!("read_dir({})", path.as_ref().display()))?;
+    Ok(dir.next_entry().await?.is_none())
+}
+
 #[cfg(test)]
 mod test {
    use std::path::PathBuf;

+    use crate::fs_ext::is_directory_empty;
+
    #[test]
    fn is_empty_dir() {
        use super::PathExt;
@@ -42,4 +53,26 @@ mod test {
        std::fs::remove_file(&file_path).unwrap();
        assert!(file_path.is_empty_dir().is_err());
    }
+
+    #[tokio::test]
+    async fn is_empty_dir_async() {
+        let dir = tempfile::tempdir().unwrap();
+        let dir_path = dir.path();
+
+        // test positive case
+        assert!(
+            is_directory_empty(dir_path).await.expect("test failure"),
+            "new tempdir should be empty"
+        );
+
+        // invoke on a file to ensure it returns an error
+        let file_path: PathBuf = dir_path.join("testfile");
+        let f = std::fs::File::create(&file_path).unwrap();
+        drop(f);
+        assert!(is_directory_empty(&file_path).await.is_err());
+
+        // do it again on a path, we know to be nonexistent
+        std::fs::remove_file(&file_path).unwrap();
+        assert!(is_directory_empty(file_path).await.is_err());
+    }
 }
--- a/libs/utils/src/http/endpoint.rs
+++ b/libs/utils/src/http/endpoint.rs
@@ -1,19 +1,18 @@
 use crate::auth::{Claims, JwtAuth};
 use crate::http::error::{api_error_handler, route_error_handler, ApiError};
-use anyhow::{anyhow, Context};
+use anyhow::Context;
 use hyper::header::{HeaderName, AUTHORIZATION};
 use hyper::http::HeaderValue;
 use hyper::Method;
-use hyper::{header::CONTENT_TYPE, Body, Request, Response, Server};
+use hyper::{header::CONTENT_TYPE, Body, Request, Response};
 use metrics::{register_int_counter, Encoder, IntCounter, TextEncoder};
 use once_cell::sync::Lazy;
 use routerify::ext::RequestExt;
-use routerify::{Middleware, RequestInfo, Router, RouterBuilder, RouterService};
+use routerify::{Middleware, RequestInfo, Router, RouterBuilder};
 use tokio::task::JoinError;
 use tracing::{self, debug, info, info_span, warn, Instrument};

 use std::future::Future;
-use std::net::TcpListener;
 use std::str::FromStr;

 static SERVE_METRICS_COUNT: Lazy<IntCounter> = Lazy::new(|| {
@@ -348,40 +347,6 @@ pub fn check_permission_with(
    }
 }

-///
-/// Start listening for HTTP requests on given socket.
-///
-/// 'shutdown_future' can be used to stop. If the Future becomes
-/// ready, we stop listening for new requests, and the function returns.
-///
-pub fn serve_thread_main<S>(
-    router_builder: RouterBuilder<hyper::Body, ApiError>,
-    listener: TcpListener,
-    shutdown_future: S,
-) -> anyhow::Result<()>
-where
-    S: Future<Output = ()> + Send + Sync,
-{
-    info!("Starting an HTTP endpoint at {}", listener.local_addr()?);
-
-    // Create a Service from the router above to handle incoming requests.
-    let service = RouterService::new(router_builder.build().map_err(|err| anyhow!(err))?).unwrap();
-
-    // Enter a single-threaded tokio runtime bound to the current thread
-    let runtime = tokio::runtime::Builder::new_current_thread()
-        .enable_all()
-        .build()?;
-
-    let _guard = runtime.enter();
-
-    let server = Server::from_tcp(listener)?
-        .serve(service)
-        .with_graceful_shutdown(shutdown_future);
-
-    runtime.block_on(server)?;
-
-    Ok(())
-}
 #[cfg(test)]
 mod tests {
    use super::*;
--- a/libs/utils/src/http/error.rs
+++ b/libs/utils/src/http/error.rs
@@ -1,5 +1,6 @@
 use hyper::{header, Body, Response, StatusCode};
 use serde::{Deserialize, Serialize};
+use std::error::Error as StdError;
 use thiserror::Error;
 use tracing::error;

@@ -15,13 +16,13 @@ pub enum ApiError {
    Unauthorized(String),

    #[error("NotFound: {0}")]
-    NotFound(anyhow::Error),
+    NotFound(Box<dyn StdError + Send + Sync + 'static>),

    #[error("Conflict: {0}")]
    Conflict(String),

    #[error("Precondition failed: {0}")]
-    PreconditionFailed(&'static str),
+    PreconditionFailed(Box<str>),

    #[error(transparent)]
    InternalServerError(anyhow::Error),
--- a/libs/utils/src/logging.rs
+++ b/libs/utils/src/logging.rs
@@ -84,7 +84,7 @@ pub fn init(
    let r = r.with({
        let log_layer = tracing_subscriber::fmt::layer()
            .with_target(false)
-            .with_ansi(atty::is(atty::Stream::Stdout))
+            .with_ansi(false)
            .with_writer(std::io::stdout);
        let log_layer = match log_format {
            LogFormat::Json => log_layer.json().boxed(),
--- a/pageserver/benches/bench_layer_map.rs
+++ b/pageserver/benches/bench_layer_map.rs
@@ -1,22 +1,23 @@
 use pageserver::keyspace::{KeyPartitioning, KeySpace};
 use pageserver::repository::Key;
 use pageserver::tenant::layer_map::LayerMap;
-use pageserver::tenant::storage_layer::{Layer, LayerDescriptor, LayerFileName};
+use pageserver::tenant::storage_layer::{tests::LayerDescriptor, Layer, LayerFileName};
+use pageserver::tenant::storage_layer::{PersistentLayer, PersistentLayerDesc};
 use rand::prelude::{SeedableRng, SliceRandom, StdRng};
 use std::cmp::{max, min};
 use std::fs::File;
 use std::io::{BufRead, BufReader};
 use std::path::PathBuf;
 use std::str::FromStr;
-use std::sync::Arc;
 use std::time::Instant;
+use utils::id::{TenantId, TimelineId};

 use utils::lsn::Lsn;

 use criterion::{black_box, criterion_group, criterion_main, Criterion};

-fn build_layer_map(filename_dump: PathBuf) -> LayerMap<LayerDescriptor> {
-    let mut layer_map = LayerMap::<LayerDescriptor>::default();
+fn build_layer_map(filename_dump: PathBuf) -> LayerMap {
+    let mut layer_map = LayerMap::default();

    let mut min_lsn = Lsn(u64::MAX);
    let mut max_lsn = Lsn(0);
@@ -33,7 +34,7 @@ fn build_layer_map(filename_dump: PathBuf) -> LayerMap<LayerDescriptor> {
        min_lsn = min(min_lsn, lsn_range.start);
        max_lsn = max(max_lsn, Lsn(lsn_range.end.0 - 1));

-        updates.insert_historic(layer.get_persistent_layer_desc(), Arc::new(layer));
+        updates.insert_historic(layer.layer_desc().clone());
    }

    println!("min: {min_lsn}, max: {max_lsn}");
@@ -43,7 +44,7 @@ fn build_layer_map(filename_dump: PathBuf) -> LayerMap<LayerDescriptor> {
 }

 /// Construct a layer map query pattern for benchmarks
-fn uniform_query_pattern(layer_map: &LayerMap<LayerDescriptor>) -> Vec<(Key, Lsn)> {
+fn uniform_query_pattern(layer_map: &LayerMap) -> Vec<(Key, Lsn)> {
    // For each image layer we query one of the pages contained, at LSN right
    // before the image layer was created. This gives us a somewhat uniform
    // coverage of both the lsn and key space because image layers have
@@ -69,7 +70,7 @@ fn uniform_query_pattern(layer_map: &LayerMap<LayerDescriptor>) -> Vec<(Key, Lsn

 // Construct a partitioning for testing get_difficulty map when we
 // don't have an exact result of `collect_keyspace` to work with.
-fn uniform_key_partitioning(layer_map: &LayerMap<LayerDescriptor>, _lsn: Lsn) -> KeyPartitioning {
+fn uniform_key_partitioning(layer_map: &LayerMap, _lsn: Lsn) -> KeyPartitioning {
    let mut parts = Vec::new();

    // We add a partition boundary at the start of each image layer,
@@ -209,13 +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 {
-            key: zero.add(10 * i32)..zero.add(10 * i32 + 1),
-            lsn: Lsn(i)..Lsn(i + 1),
-            is_incremental: false,
-            short_id: format!("Layer {}", i),
-        };
-        updates.insert_historic(layer.get_persistent_layer_desc(), Arc::new(layer));
+        let layer = LayerDescriptor::from(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.flush();
    println!("Finished layer map init in {:?}", now.elapsed());
--- a/pageserver/src/bin/pageserver.rs
+++ b/pageserver/src/bin/pageserver.rs
@@ -495,50 +495,50 @@ fn start_pageserver(
                Ok(())
            },
        );
+    }

-        if let Some(metric_collection_endpoint) = &conf.metric_collection_endpoint {
-            let background_jobs_barrier = background_jobs_barrier;
-            let metrics_ctx = RequestContext::todo_child(
-                TaskKind::MetricsCollection,
-                // This task itself shouldn't download anything.
-                // The actual size calculation does need downloads, and
-                // creates a child context with the right DownloadBehavior.
-                DownloadBehavior::Error,
-            );
-            task_mgr::spawn(
-                MGMT_REQUEST_RUNTIME.handle(),
-                TaskKind::MetricsCollection,
-                None,
-                None,
-                "consumption metrics collection",
-                true,
-                async move {
-                    // first wait until background jobs are cleared to launch.
-                    //
-                    // this is because we only process active tenants and timelines, and the
-                    // Timeline::get_current_logical_size will spawn the logical size calculation,
-                    // which will not be rate-limited.
-                    let cancel = task_mgr::shutdown_token();
+    if let Some(metric_collection_endpoint) = &conf.metric_collection_endpoint {
+        let background_jobs_barrier = background_jobs_barrier;
+        let metrics_ctx = RequestContext::todo_child(
+            TaskKind::MetricsCollection,
+            // This task itself shouldn't download anything.
+            // The actual size calculation does need downloads, and
+            // creates a child context with the right DownloadBehavior.
+            DownloadBehavior::Error,
+        );
+        task_mgr::spawn(
+            crate::BACKGROUND_RUNTIME.handle(),
+            TaskKind::MetricsCollection,
+            None,
+            None,
+            "consumption metrics collection",
+            true,
+            async move {
+                // first wait until background jobs are cleared to launch.
+                //
+                // this is because we only process active tenants and timelines, and the
+                // Timeline::get_current_logical_size will spawn the logical size calculation,
+                // which will not be rate-limited.
+                let cancel = task_mgr::shutdown_token();

-                    tokio::select! {
-                        _ = cancel.cancelled() => { return Ok(()); },
-                        _ = background_jobs_barrier.wait() => {}
-                    };
+                tokio::select! {
+                    _ = cancel.cancelled() => { return Ok(()); },
+                    _ = background_jobs_barrier.wait() => {}
+                };

-                    pageserver::consumption_metrics::collect_metrics(
-                        metric_collection_endpoint,
-                        conf.metric_collection_interval,
-                        conf.cached_metric_collection_interval,
-                        conf.synthetic_size_calculation_interval,
-                        conf.id,
-                        metrics_ctx,
-                    )
-                    .instrument(info_span!("metrics_collection"))
-                    .await?;
-                    Ok(())
-                },
-            );
-        }
+                pageserver::consumption_metrics::collect_metrics(
+                    metric_collection_endpoint,
+                    conf.metric_collection_interval,
+                    conf.cached_metric_collection_interval,
+                    conf.synthetic_size_calculation_interval,
+                    conf.id,
+                    metrics_ctx,
+                )
+                .instrument(info_span!("metrics_collection"))
+                .await?;
+                Ok(())
+            },
+        );
    }

    // Spawn a task to listen for libpq connections. It will spawn further tasks
--- a/pageserver/src/config.rs
+++ b/pageserver/src/config.rs
@@ -96,12 +96,12 @@ pub mod defaults {

 #background_task_maximum_delay = '{DEFAULT_BACKGROUND_TASK_MAXIMUM_DELAY}'

-# [tenant_config]
+[tenant_config]
 #checkpoint_distance = {DEFAULT_CHECKPOINT_DISTANCE} # in bytes
 #checkpoint_timeout = {DEFAULT_CHECKPOINT_TIMEOUT}
 #compaction_target_size = {DEFAULT_COMPACTION_TARGET_SIZE} # in bytes
 #compaction_period = '{DEFAULT_COMPACTION_PERIOD}'
-#compaction_threshold = '{DEFAULT_COMPACTION_THRESHOLD}'
+#compaction_threshold = {DEFAULT_COMPACTION_THRESHOLD}

 #gc_period = '{DEFAULT_GC_PERIOD}'
 #gc_horizon = {DEFAULT_GC_HORIZON}
@@ -111,7 +111,8 @@ pub mod defaults {
 #min_resident_size_override = .. # in bytes
 #evictions_low_residence_duration_metric_threshold = '{DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD}'
 #gc_feedback = false
-# [remote_storage]
+
+[remote_storage]

 "###
    );
--- a/pageserver/src/consumption_metrics.rs
+++ b/pageserver/src/consumption_metrics.rs
@@ -24,6 +24,8 @@ const RESIDENT_SIZE: &str = "resident_size";
 const REMOTE_STORAGE_SIZE: &str = "remote_storage_size";
 const TIMELINE_LOGICAL_SIZE: &str = "timeline_logical_size";

+const DEFAULT_HTTP_REPORTING_TIMEOUT: Duration = Duration::from_secs(60);
+
 #[serde_as]
 #[derive(Serialize, Debug)]
 struct Ids {
@@ -73,7 +75,10 @@ pub async fn collect_metrics(
    );

    // define client here to reuse it for all requests
-    let client = reqwest::Client::new();
+    let client = reqwest::ClientBuilder::new()
+        .timeout(DEFAULT_HTTP_REPORTING_TIMEOUT)
+        .build()
+        .expect("Failed to create http client with timeout");
    let mut cached_metrics: HashMap<PageserverConsumptionMetricsKey, u64> = HashMap::new();
    let mut prev_iteration_time: std::time::Instant = std::time::Instant::now();

@@ -83,7 +88,7 @@ pub async fn collect_metrics(
                info!("collect_metrics received cancellation request");
                return Ok(());
            },
-            _ = ticker.tick() => {
+            tick_at = ticker.tick() => {

                // send cached metrics every cached_metric_collection_interval
                let send_cached = prev_iteration_time.elapsed() >= cached_metric_collection_interval;
@@ -93,6 +98,12 @@ pub async fn collect_metrics(
                }

                collect_metrics_iteration(&client, &mut cached_metrics, metric_collection_endpoint, node_id, &ctx, send_cached).await;
+
+                crate::tenant::tasks::warn_when_period_overrun(
+                    tick_at.elapsed(),
+                    metric_collection_interval,
+                    "consumption_metrics_collect_metrics",
+                );
            }
        }
    }
@@ -273,31 +284,42 @@ pub async fn collect_metrics_iteration(
        })
        .expect("PageserverConsumptionMetric should not fail serialization");

-        let res = client
-            .post(metric_collection_endpoint.clone())
-            .json(&chunk_json)
-            .send()
-            .await;
+        const MAX_RETRIES: u32 = 3;

-        match res {
-            Ok(res) => {
-                if res.status().is_success() {
-                    // update cached metrics after they were sent successfully
-                    for (curr_key, curr_val) in chunk.iter() {
-                        cached_metrics.insert(curr_key.clone(), *curr_val);
-                    }
-                } else {
-                    error!("metrics endpoint refused the sent metrics: {:?}", res);
-                    for metric in chunk_to_send.iter() {
-                        // Report if the metric value is suspiciously large
-                        if metric.value > (1u64 << 40) {
+        for attempt in 0..MAX_RETRIES {
+            let res = client
+                .post(metric_collection_endpoint.clone())
+                .json(&chunk_json)
+                .send()
+                .await;
+
+            match res {
+                Ok(res) => {
+                    if res.status().is_success() {
+                        // update cached metrics after they were sent successfully
+                        for (curr_key, curr_val) in chunk.iter() {
+                            cached_metrics.insert(curr_key.clone(), *curr_val);
+                        }
+                    } else {
+                        error!("metrics endpoint refused the sent metrics: {:?}", res);
+                        for metric in chunk_to_send
+                            .iter()
+                            .filter(|metric| metric.value > (1u64 << 40))
+                        {
+                            // Report if the metric value is suspiciously large
                            error!("potentially abnormal metric value: {:?}", metric);
                        }
                    }
+                    break;
+                }
+                Err(err) if err.is_timeout() => {
+                    error!(attempt, "timeout sending metrics, retrying immediately");
+                    continue;
+                }
+                Err(err) => {
+                    error!(attempt, ?err, "failed to send metrics");
+                    break;
                }
-            }
-            Err(err) => {
-                error!("failed to send metrics: {:?}", err);
            }
        }
    }
@@ -317,7 +339,7 @@ pub async fn calculate_synthetic_size_worker(
            _ = task_mgr::shutdown_watcher() => {
                return Ok(());
            },
-        _ = ticker.tick() => {
+        tick_at = ticker.tick() => {

                let tenants = match mgr::list_tenants().await {
                    Ok(tenants) => tenants,
@@ -343,6 +365,12 @@ pub async fn calculate_synthetic_size_worker(
                    }

                }
+
+                crate::tenant::tasks::warn_when_period_overrun(
+                    tick_at.elapsed(),
+                    synthetic_size_calculation_interval,
+                    "consumption_metrics_synthetic_size_worker",
+                );
            }
        }
    }
--- a/pageserver/src/disk_usage_eviction_task.rs
+++ b/pageserver/src/disk_usage_eviction_task.rs
@@ -516,7 +516,7 @@ async fn collect_eviction_candidates(
            if !tl.is_active() {
                continue;
            }
-            let info = tl.get_local_layers_for_disk_usage_eviction();
+            let info = tl.get_local_layers_for_disk_usage_eviction().await;
            debug!(tenant_id=%tl.tenant_id, timeline_id=%tl.timeline_id, "timeline resident layers count: {}", info.resident_layers.len());
            tenant_candidates.extend(
                info.resident_layers
--- a/pageserver/src/http/openapi_spec.yml
+++ b/pageserver/src/http/openapi_spec.yml
@@ -186,10 +186,8 @@ paths:
              schema:
                $ref: "#/components/schemas/Error"
    delete:
-      description: "Attempts to delete specified timeline. On 500 errors should be retried"
+      description: "Attempts to delete specified timeline. 500 and 409 errors should be retried"
      responses:
-        "200":
-          description: Ok
        "400":
          description: Error when no tenant id found in path or no timeline id
          content:
@@ -214,8 +212,14 @@ paths:
            application/json:
              schema:
                $ref: "#/components/schemas/NotFoundError"
+        "409":
+          description: Deletion is already in progress, continue polling
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ConflictError"
        "412":
-          description: Tenant is missing
+          description: Tenant is missing, or timeline has children
          content:
            application/json:
              schema:
@@ -386,6 +390,7 @@ paths:
        "202":
          description: Tenant attaching scheduled
        "400":
+          description: Bad Request
          content:
            application/json:
              schema:
@@ -717,6 +722,12 @@ paths:
            application/json:
              schema:
                $ref: "#/components/schemas/ForbiddenError"
+        "406":
+          description: Permanently unsatisfiable request, don't retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
        "409":
          description: Timeline already exists, creation skipped
          content:
@@ -945,7 +956,7 @@ components:
              type: string
              enum: [ "maybe", "attached", "failed" ]
            data:
-            - type: object
+              type: object
              properties:
                reason:
                  type: string
--- a/pageserver/src/http/routes.rs
+++ b/pageserver/src/http/routes.rs
@@ -23,7 +23,6 @@ use super::models::{
    TimelineCreateRequest, TimelineGcRequest, TimelineInfo,
 };
 use crate::context::{DownloadBehavior, RequestContext};
-use crate::disk_usage_eviction_task;
 use crate::metrics::{StorageTimeOperation, STORAGE_TIME_GLOBAL};
 use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::task_mgr::TaskKind;
@@ -35,6 +34,7 @@ use crate::tenant::size::ModelInputs;
 use crate::tenant::storage_layer::LayerAccessStatsReset;
 use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
 use crate::{config::PageServerConf, tenant::mgr};
+use crate::{disk_usage_eviction_task, tenant};
 use utils::{
    auth::JwtAuth,
    http::{
@@ -142,7 +142,7 @@ impl From<TenantMapInsertError> for ApiError {
 impl From<TenantStateError> for ApiError {
    fn from(tse: TenantStateError) -> ApiError {
        match tse {
-            TenantStateError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid)),
+            TenantStateError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid).into()),
            _ => ApiError::InternalServerError(anyhow::Error::new(tse)),
        }
    }
@@ -151,7 +151,7 @@ impl From<TenantStateError> for ApiError {
 impl From<GetTenantError> for ApiError {
    fn from(tse: GetTenantError) -> ApiError {
        match tse {
-            GetTenantError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid)),
+            GetTenantError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid).into()),
            e @ GetTenantError::NotActive(_) => {
                // Why is this not `ApiError::NotFound`?
                // Because we must be careful to never return 404 for a tenant if it does
@@ -169,7 +169,7 @@ impl From<SetNewTenantConfigError> for ApiError {
    fn from(e: SetNewTenantConfigError) -> ApiError {
        match e {
            SetNewTenantConfigError::GetTenant(tid) => {
-                ApiError::NotFound(anyhow!("tenant {}", tid))
+                ApiError::NotFound(anyhow!("tenant {}", tid).into())
            }
            e @ SetNewTenantConfigError::Persist(_) => {
                ApiError::InternalServerError(anyhow::Error::new(e))
@@ -182,10 +182,12 @@ impl From<crate::tenant::DeleteTimelineError> for ApiError {
    fn from(value: crate::tenant::DeleteTimelineError) -> Self {
        use crate::tenant::DeleteTimelineError::*;
        match value {
-            NotFound => ApiError::NotFound(anyhow::anyhow!("timeline not found")),
-            HasChildren => ApiError::BadRequest(anyhow::anyhow!(
-                "Cannot delete timeline which has child timelines"
-            )),
+            NotFound => ApiError::NotFound(anyhow::anyhow!("timeline not found").into()),
+            HasChildren(children) => ApiError::PreconditionFailed(
+                format!("Cannot delete timeline which has child timelines: {children:?}")
+                    .into_boxed_str(),
+            ),
+            a @ AlreadyInProgress => ApiError::Conflict(a.to_string()),
            Other(e) => ApiError::InternalServerError(e),
        }
    }
@@ -197,9 +199,9 @@ impl From<crate::tenant::mgr::DeleteTimelineError> for ApiError {
        match value {
            // Report Precondition failed so client can distinguish between
            // "tenant is missing" case from "timeline is missing"
-            Tenant(GetTenantError::NotFound(..)) => {
-                ApiError::PreconditionFailed("Requested tenant is missing")
-            }
+            Tenant(GetTenantError::NotFound(..)) => ApiError::PreconditionFailed(
+                "Requested tenant is missing".to_owned().into_boxed_str(),
+            ),
            Tenant(t) => ApiError::from(t),
            Timeline(t) => ApiError::from(t),
        }
@@ -214,7 +216,7 @@ async fn build_timeline_info(
 ) -> anyhow::Result<TimelineInfo> {
    crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id();

-    let mut info = build_timeline_info_common(timeline, ctx)?;
+    let mut info = build_timeline_info_common(timeline, ctx).await?;
    if include_non_incremental_logical_size {
        // XXX we should be using spawn_ondemand_logical_size_calculation here.
        // Otherwise, if someone deletes the timeline / detaches the tenant while
@@ -232,7 +234,7 @@ async fn build_timeline_info(
    Ok(info)
 }

-fn build_timeline_info_common(
+async fn build_timeline_info_common(
    timeline: &Arc<Timeline>,
    ctx: &RequestContext,
 ) -> anyhow::Result<TimelineInfo> {
@@ -263,7 +265,7 @@ fn build_timeline_info_common(
            None
        }
    };
-    let current_physical_size = Some(timeline.layer_size_sum());
+    let current_physical_size = Some(timeline.layer_size_sum().await);
    let state = timeline.current_state();
    let remote_consistent_lsn = timeline.get_remote_consistent_lsn().unwrap_or(Lsn(0));

@@ -326,14 +328,22 @@ async fn timeline_create_handler(
            &ctx,
        )
        .await {
-            Ok(Some(new_timeline)) => {
+            Ok(new_timeline) => {
                // Created. Construct a TimelineInfo for it.
                let timeline_info = build_timeline_info_common(&new_timeline, &ctx)
+                    .await
                    .map_err(ApiError::InternalServerError)?;
                json_response(StatusCode::CREATED, timeline_info)
            }
-            Ok(None) => json_response(StatusCode::CONFLICT, ()), // timeline already exists
-            Err(err) => Err(ApiError::InternalServerError(err)),
+            Err(tenant::CreateTimelineError::AlreadyExists) => {
+                json_response(StatusCode::CONFLICT, ())
+            }
+            Err(tenant::CreateTimelineError::AncestorLsn(err)) => {
+                json_response(StatusCode::NOT_ACCEPTABLE, HttpErrorBody::from_msg(
+                    format!("{err:#}")
+                ))
+            }
+            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))
@@ -395,7 +405,7 @@ async fn timeline_detail_handler(

        let timeline = tenant
            .get_timeline(timeline_id, false)
-            .map_err(ApiError::NotFound)?;
+            .map_err(|e| ApiError::NotFound(e.into()))?;

        let timeline_info = build_timeline_info(
            &timeline,
@@ -494,7 +504,8 @@ async fn timeline_delete_handler(
        .instrument(info_span!("timeline_delete", tenant = %tenant_id, timeline = %timeline_id))
        .await?;

-    json_response(StatusCode::OK, ())
+    // FIXME: needs to be an error for console to retry it. Ideally Accepted should be used and retried until 404.
+    json_response(StatusCode::ACCEPTED, ())
 }

 async fn tenant_detach_handler(
@@ -589,7 +600,7 @@ async fn tenant_status(
        // Calculate total physical size of all timelines
        let mut current_physical_size = 0;
        for timeline in tenant.list_timelines().iter() {
-            current_physical_size += timeline.layer_size_sum();
+            current_physical_size += timeline.layer_size_sum().await;
        }

        let state = tenant.current_state();
@@ -699,7 +710,7 @@ async fn layer_map_info_handler(
    check_permission(&request, Some(tenant_id))?;

    let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
-    let layer_map_info = timeline.layer_map_info(reset);
+    let layer_map_info = timeline.layer_map_info(reset).await;

    json_response(StatusCode::OK, layer_map_info)
 }
@@ -1058,7 +1069,7 @@ async fn timeline_download_remote_layers_handler_get(
    let info = timeline
        .get_download_all_remote_layers_task_info()
        .context("task never started since last pageserver process start")
-        .map_err(ApiError::NotFound)?;
+        .map_err(|e| ApiError::NotFound(e.into()))?;
    json_response(StatusCode::OK, info)
 }

@@ -1069,7 +1080,7 @@ async fn active_timeline_of_active_tenant(
    let tenant = mgr::get_tenant(tenant_id, true).await?;
    tenant
        .get_timeline(timeline_id, true)
-        .map_err(ApiError::NotFound)
+        .map_err(|e| ApiError::NotFound(e.into()))
 }

 async fn always_panic_handler(
@@ -1125,8 +1136,6 @@ async fn disk_usage_eviction_run(
        freed_bytes: 0,
    };

-    use crate::task_mgr::MGMT_REQUEST_RUNTIME;
-
    let (tx, rx) = tokio::sync::oneshot::channel();

    let state = get_state(&r);
@@ -1144,7 +1153,7 @@ async fn disk_usage_eviction_run(
    let _g = cancel.drop_guard();

    crate::task_mgr::spawn(
-        MGMT_REQUEST_RUNTIME.handle(),
+        crate::task_mgr::BACKGROUND_RUNTIME.handle(),
        TaskKind::DiskUsageEviction,
        None,
        None,
--- a/pageserver/src/import_datadir.rs
+++ b/pageserver/src/import_datadir.rs
@@ -75,12 +75,12 @@ pub async fn import_timeline_from_postgres_datadir(
            {
                pg_control = Some(control_file);
            }
-            modification.flush()?;
+            modification.flush().await?;
        }
    }

    // We're done importing all the data files.
-    modification.commit()?;
+    modification.commit().await?;

    // We expect the Postgres server to be shut down cleanly.
    let pg_control = pg_control.context("pg_control file not found")?;
@@ -148,17 +148,17 @@ async fn import_rel(
    // because there is no guarantee about the order in which we are processing segments.
    // ignore "relation already exists" error
    //
-    // FIXME: use proper error type for this, instead of parsing the error message.
-    // Or better yet, keep track of which relations we've already created
+    // FIXME: Keep track of which relations we've already created?
    // https://github.com/neondatabase/neon/issues/3309
    if let Err(e) = modification
        .put_rel_creation(rel, nblocks as u32, ctx)
        .await
    {
-        if e.to_string().contains("already exists") {
-            debug!("relation {} already exists. we must be extending it", rel);
-        } else {
-            return Err(e);
+        match e {
+            RelationError::AlreadyExists => {
+                debug!("Relation {} already exist. We must be extending it.", rel)
+            }
+            _ => return Err(e.into()),
        }
    }

@@ -359,7 +359,7 @@ pub async fn import_basebackup_from_tar(
                    // We found the pg_control file.
                    pg_control = Some(res);
                }
-                modification.flush()?;
+                modification.flush().await?;
            }
            tokio_tar::EntryType::Directory => {
                debug!("directory {:?}", file_path);
@@ -377,7 +377,7 @@ pub async fn import_basebackup_from_tar(
    // sanity check: ensure that pg_control is loaded
    let _pg_control = pg_control.context("pg_control file not found")?;

-    modification.commit()?;
+    modification.commit().await?;
    Ok(())
 }

@@ -594,7 +594,7 @@ async fn import_file(
        // zenith.signal is not necessarily the last file, that we handle
        // but it is ok to call `finish_write()`, because final `modification.commit()`
        // will update lsn once more to the final one.
-        let writer = modification.tline.writer();
+        let writer = modification.tline.writer().await;
        writer.finish_write(prev_lsn);

        debug!("imported zenith signal {}", prev_lsn);
--- a/pageserver/src/metrics.rs
+++ b/pageserver/src/metrics.rs
@@ -1,4 +1,4 @@
-use metrics::core::{AtomicU64, GenericCounter};
+use metrics::metric_vec_duration::DurationResultObserver;
 use metrics::{
    register_counter_vec, register_histogram, register_histogram_vec, register_int_counter,
    register_int_counter_vec, register_int_gauge, register_int_gauge_vec, register_uint_gauge_vec,
@@ -95,21 +95,19 @@ static READ_NUM_FS_LAYERS: Lazy<HistogramVec> = Lazy::new(|| {
 });

 // Metrics collected on operations on the storage repository.
-static RECONSTRUCT_TIME: Lazy<HistogramVec> = Lazy::new(|| {
-    register_histogram_vec!(
+pub static RECONSTRUCT_TIME: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
        "pageserver_getpage_reconstruct_seconds",
-        "Time spent in reconstruct_value",
-        &["tenant_id", "timeline_id"],
+        "Time spent in reconstruct_value (reconstruct a page from deltas)",
        CRITICAL_OP_BUCKETS.into(),
    )
    .expect("failed to define a metric")
 });

-static MATERIALIZED_PAGE_CACHE_HIT_DIRECT: Lazy<IntCounterVec> = Lazy::new(|| {
-    register_int_counter_vec!(
+pub static MATERIALIZED_PAGE_CACHE_HIT_DIRECT: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
        "pageserver_materialized_cache_hits_direct_total",
        "Number of cache hits from materialized page cache without redo",
-        &["tenant_id", "timeline_id"]
    )
    .expect("failed to define a metric")
 });
@@ -124,11 +122,10 @@ static GET_RECONSTRUCT_DATA_TIME: Lazy<HistogramVec> = Lazy::new(|| {
    .expect("failed to define a metric")
 });

-static MATERIALIZED_PAGE_CACHE_HIT: Lazy<IntCounterVec> = Lazy::new(|| {
-    register_int_counter_vec!(
+pub static MATERIALIZED_PAGE_CACHE_HIT: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
        "pageserver_materialized_cache_hits_total",
        "Number of cache hits from materialized page cache",
-        &["tenant_id", "timeline_id"]
    )
    .expect("failed to define a metric")
 });
@@ -207,11 +204,11 @@ pub static TENANT_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {

 pub static TENANT_SYNTHETIC_SIZE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
    register_uint_gauge_vec!(
-        "pageserver_tenant_synthetic_size",
-        "Synthetic size of each tenant",
+        "pageserver_tenant_synthetic_cached_size_bytes",
+        "Synthetic size of each tenant in bytes",
        &["tenant_id"]
    )
-    .expect("Failed to register pageserver_tenant_synthetic_size metric")
+    .expect("Failed to register pageserver_tenant_synthetic_cached_size_bytes metric")
 });

 // Metrics for cloud upload. These metrics reflect data uploaded to cloud storage,
@@ -428,6 +425,27 @@ pub static SMGR_QUERY_TIME: Lazy<HistogramVec> = Lazy::new(|| {
    .expect("failed to define a metric")
 });

+pub struct BasebackupQueryTime(HistogramVec);
+pub static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|| {
+    BasebackupQueryTime({
+        register_histogram_vec!(
+            "pageserver_basebackup_query_seconds",
+            "Histogram of basebackup queries durations, by result type",
+            &["result"],
+            CRITICAL_OP_BUCKETS.into(),
+        )
+        .expect("failed to define a metric")
+    })
+});
+
+impl DurationResultObserver for BasebackupQueryTime {
+    fn observe_result<T, E>(&self, res: &Result<T, E>, duration: std::time::Duration) {
+        let label_value = if res.is_ok() { "ok" } else { "error" };
+        let metric = self.0.get_metric_with_label_values(&[label_value]).unwrap();
+        metric.observe(duration.as_secs_f64());
+    }
+}
+
 pub static LIVE_CONNECTIONS_COUNT: Lazy<IntGaugeVec> = Lazy::new(|| {
    register_int_gauge_vec!(
        "pageserver_live_connections",
@@ -752,10 +770,7 @@ impl StorageTimeMetrics {
 pub struct TimelineMetrics {
    tenant_id: String,
    timeline_id: String,
-    pub reconstruct_time_histo: Histogram,
    pub get_reconstruct_data_time_histo: Histogram,
-    pub materialized_page_cache_hit_counter: GenericCounter<AtomicU64>,
-    pub materialized_page_cache_hit_upon_request_counter: GenericCounter<AtomicU64>,
    pub flush_time_histo: StorageTimeMetrics,
    pub compact_time_histo: StorageTimeMetrics,
    pub create_images_time_histo: StorageTimeMetrics,
@@ -783,15 +798,9 @@ impl TimelineMetrics {
    ) -> Self {
        let tenant_id = tenant_id.to_string();
        let timeline_id = timeline_id.to_string();
-        let reconstruct_time_histo = RECONSTRUCT_TIME
-            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
-            .unwrap();
        let get_reconstruct_data_time_histo = GET_RECONSTRUCT_DATA_TIME
            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
            .unwrap();
-        let materialized_page_cache_hit_counter = MATERIALIZED_PAGE_CACHE_HIT
-            .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 =
@@ -833,19 +842,13 @@ impl TimelineMetrics {
        let read_num_fs_layers = READ_NUM_FS_LAYERS
            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
            .unwrap();
-        let materialized_page_cache_hit_upon_request_counter = MATERIALIZED_PAGE_CACHE_HIT_DIRECT
-            .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,
-            reconstruct_time_histo,
            get_reconstruct_data_time_histo,
-            materialized_page_cache_hit_counter,
-            materialized_page_cache_hit_upon_request_counter,
            flush_time_histo,
            compact_time_histo,
            create_images_time_histo,
@@ -872,10 +875,7 @@ impl Drop for TimelineMetrics {
    fn drop(&mut self) {
        let tenant_id = &self.tenant_id;
        let timeline_id = &self.timeline_id;
-        let _ = RECONSTRUCT_TIME.remove_label_values(&[tenant_id, timeline_id]);
        let _ = GET_RECONSTRUCT_DATA_TIME.remove_label_values(&[tenant_id, timeline_id]);
-        let _ = MATERIALIZED_PAGE_CACHE_HIT.remove_label_values(&[tenant_id, timeline_id]);
-        let _ = MATERIALIZED_PAGE_CACHE_HIT_DIRECT.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]);
@@ -1319,4 +1319,8 @@ pub fn preinitialize_metrics() {

    // Same as above for this metric, but, it's a Vec-type metric for which we don't know all the labels.
    BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT.reset();
+
+    // Python tests need these.
+    MATERIALIZED_PAGE_CACHE_HIT_DIRECT.get();
+    MATERIALIZED_PAGE_CACHE_HIT.get();
 }
--- a/pageserver/src/page_service.rs
+++ b/pageserver/src/page_service.rs
@@ -390,7 +390,9 @@ impl PageServerHandler {
        };

        // Check that the timeline exists
-        let timeline = tenant.get_timeline(timeline_id, true)?;
+        let timeline = tenant
+            .get_timeline(timeline_id, true)
+            .map_err(|e| anyhow::anyhow!(e))?;

        // switch client to COPYBOTH
        pgb.write_message_noflush(&BeMessage::CopyBothResponse)?;
@@ -902,7 +904,7 @@ where

            self.check_permission(Some(tenant_id))?;

-            let lsn = if params.len() == 3 {
+            let lsn = if params.len() >= 3 {
                Some(
                    Lsn::from_str(params[2])
                        .with_context(|| format!("Failed to parse Lsn from {}", params[2]))?,
@@ -911,10 +913,24 @@ where
                None
            };

-            // Check that the timeline exists
-            self.handle_basebackup_request(pgb, tenant_id, timeline_id, lsn, None, false, ctx)
-                .await?;
-            pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
+            metrics::metric_vec_duration::observe_async_block_duration_by_result(
+                &*crate::metrics::BASEBACKUP_QUERY_TIME,
+                async move {
+                    self.handle_basebackup_request(
+                        pgb,
+                        tenant_id,
+                        timeline_id,
+                        lsn,
+                        None,
+                        false,
+                        ctx,
+                    )
+                    .await?;
+                    pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
+                    anyhow::Ok(())
+                },
+            )
+            .await?;
        }
        // return pair of prev_lsn and last_lsn
        else if query_string.starts_with("get_last_record_rlsn ") {
@@ -1230,6 +1246,6 @@ async fn get_active_tenant_timeline(
        .map_err(GetActiveTimelineError::Tenant)?;
    let timeline = tenant
        .get_timeline(timeline_id, true)
-        .map_err(GetActiveTimelineError::Timeline)?;
+        .map_err(|e| GetActiveTimelineError::Timeline(anyhow::anyhow!(e)))?;
    Ok(timeline)
 }
--- a/pageserver/src/pgdatadir_mapping.rs
+++ b/pageserver/src/pgdatadir_mapping.rs
@@ -43,6 +43,16 @@ pub enum CalculateLogicalSizeError {
    Other(#[from] anyhow::Error),
 }

+#[derive(Debug, thiserror::Error)]
+pub enum RelationError {
+    #[error("Relation Already Exists")]
+    AlreadyExists,
+    #[error("invalid relnode")]
+    InvalidRelnode,
+    #[error(transparent)]
+    Other(#[from] anyhow::Error),
+}
+
 ///
 /// This impl provides all the functionality to store PostgreSQL relations, SLRUs,
 /// and other special kinds of files, in a versioned key-value store. The
@@ -101,9 +111,9 @@ impl Timeline {
        ctx: &RequestContext,
    ) -> Result<Bytes, PageReconstructError> {
        if tag.relnode == 0 {
-            return Err(PageReconstructError::Other(anyhow::anyhow!(
-                "invalid relnode"
-            )));
+            return Err(PageReconstructError::Other(
+                RelationError::InvalidRelnode.into(),
+            ));
        }

        let nblocks = self.get_rel_size(tag, lsn, latest, ctx).await?;
@@ -148,9 +158,9 @@ impl Timeline {
        ctx: &RequestContext,
    ) -> Result<BlockNumber, PageReconstructError> {
        if tag.relnode == 0 {
-            return Err(PageReconstructError::Other(anyhow::anyhow!(
-                "invalid relnode"
-            )));
+            return Err(PageReconstructError::Other(
+                RelationError::InvalidRelnode.into(),
+            ));
        }

        if let Some(nblocks) = self.get_cached_rel_size(&tag, lsn) {
@@ -193,9 +203,9 @@ impl Timeline {
        ctx: &RequestContext,
    ) -> Result<bool, PageReconstructError> {
        if tag.relnode == 0 {
-            return Err(PageReconstructError::Other(anyhow::anyhow!(
-                "invalid relnode"
-            )));
+            return Err(PageReconstructError::Other(
+                RelationError::InvalidRelnode.into(),
+            ));
        }

        // first try to lookup relation in cache
@@ -699,6 +709,20 @@ impl<'a> DatadirModification<'a> {
        Ok(())
    }

+    #[cfg(test)]
+    pub fn init_empty_test_timeline(&mut self) -> anyhow::Result<()> {
+        self.init_empty()?;
+        self.put_control_file(bytes::Bytes::from_static(
+            b"control_file contents do not matter",
+        ))
+        .context("put_control_file")?;
+        self.put_checkpoint(bytes::Bytes::from_static(
+            b"checkpoint_file contents do not matter",
+        ))
+        .context("put_checkpoint_file")?;
+        Ok(())
+    }
+
    /// Put a new page version that can be constructed from a WAL record
    ///
    /// NOTE: this will *not* implicitly extend the relation, if the page is beyond the
@@ -710,7 +734,7 @@ impl<'a> DatadirModification<'a> {
        blknum: BlockNumber,
        rec: NeonWalRecord,
    ) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, "invalid relnode");
+        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
        self.put(rel_block_to_key(rel, blknum), Value::WalRecord(rec));
        Ok(())
    }
@@ -737,7 +761,7 @@ impl<'a> DatadirModification<'a> {
        blknum: BlockNumber,
        img: Bytes,
    ) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, "invalid relnode");
+        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
        self.put(rel_block_to_key(rel, blknum), Value::Image(img));
        Ok(())
    }
@@ -861,32 +885,38 @@ impl<'a> DatadirModification<'a> {
        rel: RelTag,
        nblocks: BlockNumber,
        ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, "invalid relnode");
+    ) -> Result<(), RelationError> {
+        if rel.relnode == 0 {
+            return Err(RelationError::InvalidRelnode);
+        }
        // It's possible that this is the first rel for this db in this
        // tablespace.  Create the reldir entry for it if so.
-        let mut dbdir = DbDirectory::des(&self.get(DBDIR_KEY, ctx).await?)?;
+        let mut dbdir = DbDirectory::des(&self.get(DBDIR_KEY, ctx).await.context("read db")?)
+            .context("deserialize db")?;
        let rel_dir_key = rel_dir_to_key(rel.spcnode, rel.dbnode);
        let mut rel_dir = if dbdir.dbdirs.get(&(rel.spcnode, rel.dbnode)).is_none() {
            // Didn't exist. Update dbdir
            dbdir.dbdirs.insert((rel.spcnode, rel.dbnode), false);
-            let buf = DbDirectory::ser(&dbdir)?;
+            let buf = DbDirectory::ser(&dbdir).context("serialize db")?;
            self.put(DBDIR_KEY, Value::Image(buf.into()));

            // and create the RelDirectory
            RelDirectory::default()
        } else {
            // reldir already exists, fetch it
-            RelDirectory::des(&self.get(rel_dir_key, ctx).await?)?
+            RelDirectory::des(&self.get(rel_dir_key, ctx).await.context("read db")?)
+                .context("deserialize db")?
        };

        // Add the new relation to the rel directory entry, and write it back
        if !rel_dir.rels.insert((rel.relnode, rel.forknum)) {
-            anyhow::bail!("rel {rel} already exists");
+            return Err(RelationError::AlreadyExists);
        }
        self.put(
            rel_dir_key,
-            Value::Image(Bytes::from(RelDirectory::ser(&rel_dir)?)),
+            Value::Image(Bytes::from(
+                RelDirectory::ser(&rel_dir).context("serialize")?,
+            )),
        );

        // Put size
@@ -911,7 +941,7 @@ impl<'a> DatadirModification<'a> {
        nblocks: BlockNumber,
        ctx: &RequestContext,
    ) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, "invalid relnode");
+        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
        let last_lsn = self.tline.get_last_record_lsn();
        if self.tline.get_rel_exists(rel, last_lsn, true, ctx).await? {
            let size_key = rel_size_to_key(rel);
@@ -942,7 +972,7 @@ impl<'a> DatadirModification<'a> {
        nblocks: BlockNumber,
        ctx: &RequestContext,
    ) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, "invalid relnode");
+        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);

        // Put size
        let size_key = rel_size_to_key(rel);
@@ -963,7 +993,7 @@ impl<'a> DatadirModification<'a> {

    /// Drop a relation.
    pub async fn put_rel_drop(&mut self, rel: RelTag, ctx: &RequestContext) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, "invalid relnode");
+        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);

        // Remove it from the directory entry
        let dir_key = rel_dir_to_key(rel.spcnode, rel.dbnode);
@@ -1108,7 +1138,7 @@ impl<'a> DatadirModification<'a> {
    /// retains all the metadata, but data pages are flushed. That's again OK
    /// for bulk import, where you are just loading data pages and won't try to
    /// modify the same pages twice.
-    pub fn flush(&mut self) -> anyhow::Result<()> {
+    pub async fn flush(&mut self) -> anyhow::Result<()> {
        // Unless we have accumulated a decent amount of changes, it's not worth it
        // to scan through the pending_updates list.
        let pending_nblocks = self.pending_nblocks;
@@ -1116,19 +1146,20 @@ impl<'a> DatadirModification<'a> {
            return Ok(());
        }

-        let writer = self.tline.writer();
+        let writer = self.tline.writer().await;

        // Flush relation and  SLRU data blocks, keep metadata.
-        let mut result: anyhow::Result<()> = Ok(());
-        self.pending_updates.retain(|&key, value| {
-            if result.is_ok() && (is_rel_block_key(key) || is_slru_block_key(key)) {
-                result = writer.put(key, self.lsn, value);
-                false
+        let mut retained_pending_updates = HashMap::new();
+        for (key, value) in self.pending_updates.drain() {
+            if is_rel_block_key(key) || is_slru_block_key(key) {
+                // This bails out on first error without modifying pending_updates.
+                // That's Ok, cf this function's doc comment.
+                writer.put(key, self.lsn, &value).await?;
            } else {
-                true
+                retained_pending_updates.insert(key, value);
            }
-        });
-        result?;
+        }
+        self.pending_updates.extend(retained_pending_updates);

        if pending_nblocks != 0 {
            writer.update_current_logical_size(pending_nblocks * i64::from(BLCKSZ));
@@ -1143,17 +1174,17 @@ impl<'a> DatadirModification<'a> {
    /// underlying timeline.
    /// All the modifications in this atomic update are stamped by the specified LSN.
    ///
-    pub fn commit(&mut self) -> anyhow::Result<()> {
-        let writer = self.tline.writer();
+    pub async fn commit(&mut self) -> anyhow::Result<()> {
+        let writer = self.tline.writer().await;
        let lsn = self.lsn;
        let pending_nblocks = self.pending_nblocks;
        self.pending_nblocks = 0;

        for (key, value) in self.pending_updates.drain() {
-            writer.put(key, lsn, &value)?;
+            writer.put(key, lsn, &value).await?;
        }
        for key_range in self.pending_deletions.drain(..) {
-            writer.delete(key_range, lsn)?;
+            writer.delete(key_range, lsn).await?;
        }

        writer.finish_write(lsn);
@@ -1593,20 +1624,6 @@ fn is_slru_block_key(key: Key) -> bool {
        && key.field6 != 0xffffffff // and not SlruSegSize
 }

-#[cfg(test)]
-pub fn create_test_timeline(
-    tenant: &crate::tenant::Tenant,
-    timeline_id: utils::id::TimelineId,
-    pg_version: u32,
-    ctx: &RequestContext,
-) -> anyhow::Result<std::sync::Arc<Timeline>> {
-    let tline = tenant.create_test_timeline(timeline_id, Lsn(8), pg_version, ctx)?;
-    let mut m = tline.begin_modification(Lsn(8));
-    m.init_empty()?;
-    m.commit()?;
-    Ok(tline)
-}
-
 #[allow(clippy::bool_assert_comparison)]
 #[cfg(test)]
 mod tests {
--- a/pageserver/src/task_mgr.rs
+++ b/pageserver/src/task_mgr.rs
@@ -257,6 +257,9 @@ pub enum TaskKind {
    // task that handles attaching a tenant
    Attach,

+    // Used mostly for background deletion from s3
+    TimelineDeletionWorker,
+
    // task that handhes metrics collection
    MetricsCollection,

@@ -503,17 +506,17 @@ pub async fn shutdown_tasks(
                    warn!(name = task.name, tenant_id = ?tenant_id, timeline_id = ?timeline_id, kind = ?task_kind, "stopping left-over");
                }
            }
-            let completed = tokio::select! {
+            let join_handle = tokio::select! {
                biased;
-                _ = &mut join_handle => { true },
+                _ = &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);
-                    false
+                    Some(join_handle)
                }
            };
-            if !completed {
+            if let Some(join_handle) = join_handle {
                // we never handled this return value, but:
                // - we don't deschedule which would lead to is_cancelled
                // - panics are already logged (is_panicked)
--- a/pageserver/src/tenant.rs
+++ b/pageserver/src/tenant.rs
--- a/pageserver/src/tenant/config.rs
+++ b/pageserver/src/tenant/config.rs
@@ -38,8 +38,8 @@ pub mod defaults {
    pub const DEFAULT_GC_PERIOD: &str = "1 hr";
    pub const DEFAULT_IMAGE_CREATION_THRESHOLD: usize = 3;
    pub const DEFAULT_PITR_INTERVAL: &str = "7 days";
-    pub const DEFAULT_WALRECEIVER_CONNECT_TIMEOUT: &str = "2 seconds";
-    pub const DEFAULT_WALRECEIVER_LAGGING_WAL_TIMEOUT: &str = "3 seconds";
+    pub const DEFAULT_WALRECEIVER_CONNECT_TIMEOUT: &str = "10 seconds";
+    pub const DEFAULT_WALRECEIVER_LAGGING_WAL_TIMEOUT: &str = "10 seconds";
    pub const DEFAULT_MAX_WALRECEIVER_LSN_WAL_LAG: u64 = 10 * 1024 * 1024;
    pub const DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD: &str = "24 hour";
 }
--- a/pageserver/src/tenant/layer_map.rs
+++ b/pageserver/src/tenant/layer_map.rs
@@ -51,25 +51,23 @@ use crate::keyspace::KeyPartitioning;
 use crate::repository::Key;
 use crate::tenant::storage_layer::InMemoryLayer;
 use crate::tenant::storage_layer::Layer;
-use anyhow::Context;
 use anyhow::Result;
-use std::collections::HashMap;
 use std::collections::VecDeque;
 use std::ops::Range;
 use std::sync::Arc;
 use utils::lsn::Lsn;

 use historic_layer_coverage::BufferedHistoricLayerCoverage;
-pub use historic_layer_coverage::Replacement;
+pub use historic_layer_coverage::LayerKey;

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

 ///
 /// LayerMap tracks what layers exist on a timeline.
 ///
-pub struct LayerMap<L: ?Sized> {
+#[derive(Default)]
+pub struct LayerMap {
    //
    // 'open_layer' holds the current InMemoryLayer that is accepting new
    // records. If it is None, 'next_open_layer_at' will be set instead, indicating
@@ -95,24 +93,6 @@ pub struct LayerMap<L: ?Sized> {
    /// L0 layers have key range Key::MIN..Key::MAX, and locating them using R-Tree search is very inefficient.
    /// So L0 layers are held in l0_delta_layers vector, in addition to the R-tree.
    l0_delta_layers: Vec<Arc<PersistentLayerDesc>>,
-
-    /// Mapping from persistent layer key to the actual layer object. Currently, it stores delta, image, and
-    /// remote layers. In future refactors, this will be eventually moved out of LayerMap into Timeline, and
-    /// RemoteLayer will be removed.
-    mapping: HashMap<PersistentLayerKey, Arc<L>>,
-}
-
-impl<L: ?Sized> Default for LayerMap<L> {
-    fn default() -> Self {
-        Self {
-            open_layer: None,
-            next_open_layer_at: None,
-            frozen_layers: VecDeque::default(),
-            l0_delta_layers: Vec::default(),
-            historic: BufferedHistoricLayerCoverage::default(),
-            mapping: HashMap::default(),
-        }
-    }
 }

 /// The primary update API for the layer map.
@@ -120,24 +100,21 @@ impl<L: ?Sized> Default for LayerMap<L> {
 /// Batching historic layer insertions and removals is good for
 /// performance and this struct helps us do that correctly.
 #[must_use]
-pub struct BatchedUpdates<'a, L: ?Sized + Layer> {
+pub struct BatchedUpdates<'a> {
    // While we hold this exclusive reference to the layer map the type checker
    // will prevent us from accidentally reading any unflushed updates.
-    layer_map: &'a mut LayerMap<L>,
+    layer_map: &'a mut LayerMap,
 }

 /// Provide ability to batch more updates while hiding the read
 /// API so we don't accidentally read without flushing.
-impl<L> BatchedUpdates<'_, L>
-where
-    L: ?Sized + Layer,
-{
+impl BatchedUpdates<'_> {
    ///
    /// Insert an on-disk layer.
    ///
    // TODO remove the `layer` argument when `mapping` is refactored out of `LayerMap`
-    pub fn insert_historic(&mut self, layer_desc: PersistentLayerDesc, layer: Arc<L>) {
-        self.layer_map.insert_historic_noflush(layer_desc, layer)
+    pub fn insert_historic(&mut self, layer_desc: PersistentLayerDesc) {
+        self.layer_map.insert_historic_noflush(layer_desc)
    }

    ///
@@ -145,31 +122,8 @@ where
    ///
    /// This should be called when the corresponding file on disk has been deleted.
    ///
-    pub fn remove_historic(&mut self, layer_desc: PersistentLayerDesc, layer: Arc<L>) {
-        self.layer_map.remove_historic_noflush(layer_desc, layer)
-    }
-
-    /// Replaces existing layer iff it is the `expected`.
-    ///
-    /// If the expected layer has been removed it will not be inserted by this function.
-    ///
-    /// Returned `Replacement` describes succeeding in replacement or the reason why it could not
-    /// be done.
-    ///
-    /// TODO replacement can be done without buffering and rebuilding layer map updates.
-    ///      One way to do that is to add a layer of indirection for returned values, so
-    ///      that we can replace values only by updating a hashmap.
-    pub fn replace_historic(
-        &mut self,
-        expected_desc: PersistentLayerDesc,
-        expected: &Arc<L>,
-        new_desc: PersistentLayerDesc,
-        new: Arc<L>,
-    ) -> anyhow::Result<Replacement<Arc<L>>> {
-        fail::fail_point!("layermap-replace-notfound", |_| Ok(Replacement::NotFound));
-
-        self.layer_map
-            .replace_historic_noflush(expected_desc, expected, new_desc, new)
+    pub fn remove_historic(&mut self, layer_desc: PersistentLayerDesc) {
+        self.layer_map.remove_historic_noflush(layer_desc)
    }

    // We will flush on drop anyway, but this method makes it
@@ -185,25 +139,19 @@ where
 // than panic later or read without flushing.
 //
 // TODO maybe warn if flush hasn't explicitly been called
-impl<L> Drop for BatchedUpdates<'_, L>
-where
-    L: ?Sized + Layer,
-{
+impl Drop for BatchedUpdates<'_> {
    fn drop(&mut self) {
        self.layer_map.flush_updates();
    }
 }

 /// Return value of LayerMap::search
-pub struct SearchResult<L: ?Sized> {
-    pub layer: Arc<L>,
+pub struct SearchResult {
+    pub layer: Arc<PersistentLayerDesc>,
    pub lsn_floor: Lsn,
 }

-impl<L> LayerMap<L>
-where
-    L: ?Sized + Layer,
-{
+impl LayerMap {
    ///
    /// Find the latest layer (by lsn.end) that covers the given
    /// 'key', with lsn.start < 'end_lsn'.
@@ -235,7 +183,7 @@ where
    /// NOTE: This only searches the 'historic' layers, *not* the
    /// 'open' and 'frozen' layers!
    ///
-    pub fn search(&self, key: Key, end_lsn: Lsn) -> Option<SearchResult<L>> {
+    pub fn search(&self, key: Key, end_lsn: Lsn) -> Option<SearchResult> {
        let version = self.historic.get().unwrap().get_version(end_lsn.0 - 1)?;
        let latest_delta = version.delta_coverage.query(key.to_i128());
        let latest_image = version.image_coverage.query(key.to_i128());
@@ -244,7 +192,6 @@ where
            (None, None) => None,
            (None, Some(image)) => {
                let lsn_floor = image.get_lsn_range().start;
-                let image = self.get_layer_from_mapping(&image.key()).clone();
                Some(SearchResult {
                    layer: image,
                    lsn_floor,
@@ -252,7 +199,6 @@ where
            }
            (Some(delta), None) => {
                let lsn_floor = delta.get_lsn_range().start;
-                let delta = self.get_layer_from_mapping(&delta.key()).clone();
                Some(SearchResult {
                    layer: delta,
                    lsn_floor,
@@ -263,7 +209,6 @@ where
                let image_is_newer = image.get_lsn_range().end >= delta.get_lsn_range().end;
                let image_exact_match = img_lsn + 1 == end_lsn;
                if image_is_newer || image_exact_match {
-                    let image = self.get_layer_from_mapping(&image.key()).clone();
                    Some(SearchResult {
                        layer: image,
                        lsn_floor: img_lsn,
@@ -271,7 +216,6 @@ where
                } else {
                    let lsn_floor =
                        std::cmp::max(delta.get_lsn_range().start, image.get_lsn_range().start + 1);
-                    let delta = self.get_layer_from_mapping(&delta.key()).clone();
                    Some(SearchResult {
                        layer: delta,
                        lsn_floor,
@@ -282,7 +226,7 @@ where
    }

    /// Start a batch of updates, applied on drop
-    pub fn batch_update(&mut self) -> BatchedUpdates<'_, L> {
+    pub fn batch_update(&mut self) -> BatchedUpdates<'_> {
        BatchedUpdates { layer_map: self }
    }

@@ -292,48 +236,32 @@ where
    /// Helper function for BatchedUpdates::insert_historic
    ///
    /// TODO(chi): remove L generic so that we do not need to pass layer object.
-    pub(self) fn insert_historic_noflush(
-        &mut self,
-        layer_desc: PersistentLayerDesc,
-        layer: Arc<L>,
-    ) {
-        self.mapping.insert(layer_desc.key(), layer.clone());
-
+    pub(self) fn insert_historic_noflush(&mut self, layer_desc: PersistentLayerDesc) {
        // TODO: See #3869, resulting #4088, attempted fix and repro #4094

-        if Self::is_l0(&layer) {
+        if Self::is_l0(&layer_desc) {
            self.l0_delta_layers.push(layer_desc.clone().into());
        }

        self.historic.insert(
-            historic_layer_coverage::LayerKey::from(&*layer),
+            historic_layer_coverage::LayerKey::from(&layer_desc),
            layer_desc.into(),
        );
    }

-    fn get_layer_from_mapping(&self, key: &PersistentLayerKey) -> &Arc<L> {
-        let layer = self
-            .mapping
-            .get(key)
-            .with_context(|| format!("{key:?}"))
-            .expect("inconsistent layer mapping");
-        layer
-    }
-
    ///
    /// Remove an on-disk layer from the map.
    ///
    /// Helper function for BatchedUpdates::remove_historic
    ///
-    pub fn remove_historic_noflush(&mut self, layer_desc: PersistentLayerDesc, layer: Arc<L>) {
+    pub fn remove_historic_noflush(&mut self, layer_desc: PersistentLayerDesc) {
        self.historic
-            .remove(historic_layer_coverage::LayerKey::from(&*layer));
-        if Self::is_l0(&layer) {
+            .remove(historic_layer_coverage::LayerKey::from(&layer_desc));
+        let layer_key = layer_desc.key();
+        if Self::is_l0(&layer_desc) {
            let len_before = self.l0_delta_layers.len();
            let mut l0_delta_layers = std::mem::take(&mut self.l0_delta_layers);
-            l0_delta_layers.retain(|other| {
-                !Self::compare_arced_layers(self.get_layer_from_mapping(&other.key()), &layer)
-            });
+            l0_delta_layers.retain(|other| other.key() != layer_key);
            self.l0_delta_layers = l0_delta_layers;
            // this assertion is related to use of Arc::ptr_eq in Self::compare_arced_layers,
            // there's a chance that the comparison fails at runtime due to it comparing (pointer,
@@ -344,69 +272,6 @@ where
                "failed to locate removed historic layer from l0_delta_layers"
            );
        }
-        self.mapping.remove(&layer_desc.key());
-    }
-
-    pub(self) fn replace_historic_noflush(
-        &mut self,
-        expected_desc: PersistentLayerDesc,
-        expected: &Arc<L>,
-        new_desc: PersistentLayerDesc,
-        new: Arc<L>,
-    ) -> anyhow::Result<Replacement<Arc<L>>> {
-        let key = historic_layer_coverage::LayerKey::from(&**expected);
-        let other = historic_layer_coverage::LayerKey::from(&*new);
-
-        let expected_l0 = Self::is_l0(expected);
-        let new_l0 = Self::is_l0(&new);
-
-        anyhow::ensure!(
-            key == other,
-            "expected and new must have equal LayerKeys: {key:?} != {other:?}"
-        );
-
-        anyhow::ensure!(
-            expected_l0 == new_l0,
-            "expected and new must both be l0 deltas or neither should be: {expected_l0} != {new_l0}"
-        );
-
-        let l0_index = if expected_l0 {
-            // find the index in case replace worked, we need to replace that as well
-            let pos = self.l0_delta_layers.iter().position(|slot| {
-                Self::compare_arced_layers(self.get_layer_from_mapping(&slot.key()), expected)
-            });
-
-            if pos.is_none() {
-                return Ok(Replacement::NotFound);
-            }
-            pos
-        } else {
-            None
-        };
-
-        let new_desc = Arc::new(new_desc);
-        let replaced = self.historic.replace(&key, new_desc.clone(), |existing| {
-            **existing == expected_desc
-        });
-
-        if let Replacement::Replaced { .. } = &replaced {
-            self.mapping.remove(&expected_desc.key());
-            self.mapping.insert(new_desc.key(), new);
-            if let Some(index) = l0_index {
-                self.l0_delta_layers[index] = new_desc;
-            }
-        }
-
-        let replaced = match replaced {
-            Replacement::Replaced { in_buffered } => Replacement::Replaced { in_buffered },
-            Replacement::NotFound => Replacement::NotFound,
-            Replacement::RemovalBuffered => Replacement::RemovalBuffered,
-            Replacement::Unexpected(x) => {
-                Replacement::Unexpected(self.get_layer_from_mapping(&x.key()).clone())
-            }
-        };
-
-        Ok(replaced)
    }

    /// Helper function for BatchedUpdates::drop.
@@ -454,10 +319,8 @@ where
        Ok(true)
    }

-    pub fn iter_historic_layers(&self) -> impl '_ + Iterator<Item = Arc<L>> {
-        self.historic
-            .iter()
-            .map(|x| self.get_layer_from_mapping(&x.key()).clone())
+    pub fn iter_historic_layers(&self) -> impl '_ + Iterator<Item = Arc<PersistentLayerDesc>> {
+        self.historic.iter()
    }

    ///
@@ -472,7 +335,7 @@ where
        &self,
        key_range: &Range<Key>,
        lsn: Lsn,
-    ) -> Result<Vec<(Range<Key>, Option<Arc<L>>)>> {
+    ) -> Result<Vec<(Range<Key>, Option<Arc<PersistentLayerDesc>>)>> {
        let version = match self.historic.get().unwrap().get_version(lsn.0) {
            Some(v) => v,
            None => return Ok(vec![]),
@@ -482,36 +345,26 @@ where
        let end = key_range.end.to_i128();

        // Initialize loop variables
-        let mut coverage: Vec<(Range<Key>, Option<Arc<L>>)> = vec![];
+        let mut coverage: Vec<(Range<Key>, Option<Arc<PersistentLayerDesc>>)> = vec![];
        let mut current_key = start;
        let mut current_val = version.image_coverage.query(start);

        // Loop through the change events and push intervals
        for (change_key, change_val) in version.image_coverage.range(start..end) {
            let kr = Key::from_i128(current_key)..Key::from_i128(change_key);
-            coverage.push((
-                kr,
-                current_val
-                    .take()
-                    .map(|l| self.get_layer_from_mapping(&l.key()).clone()),
-            ));
+            coverage.push((kr, current_val.take()));
            current_key = change_key;
            current_val = change_val.clone();
        }

        // Add the final interval
        let kr = Key::from_i128(current_key)..Key::from_i128(end);
-        coverage.push((
-            kr,
-            current_val
-                .take()
-                .map(|l| self.get_layer_from_mapping(&l.key()).clone()),
-        ));
+        coverage.push((kr, current_val.take()));

        Ok(coverage)
    }

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

@@ -537,7 +390,7 @@ where
    /// TODO The optimal number should probably be slightly higher than 1, but to
    ///      implement that we need to plumb a lot more context into this function
    ///      than just the current partition_range.
-    pub fn is_reimage_worthy(layer: &L, partition_range: &Range<Key>) -> bool {
+    pub fn is_reimage_worthy(layer: &PersistentLayerDesc, partition_range: &Range<Key>) -> bool {
        // Case 1
        if !Self::is_l0(layer) {
            return true;
@@ -595,9 +448,7 @@ where
                    let kr = Key::from_i128(current_key)..Key::from_i128(change_key);
                    let lr = lsn.start..val.get_lsn_range().start;
                    if !kr.is_empty() {
-                        let base_count =
-                            Self::is_reimage_worthy(self.get_layer_from_mapping(&val.key()), key)
-                                as usize;
+                        let base_count = Self::is_reimage_worthy(&val, key) as usize;
                        let new_limit = limit.map(|l| l - base_count);
                        let max_stacked_deltas_underneath =
                            self.count_deltas(&kr, &lr, new_limit)?;
@@ -620,9 +471,7 @@ where
                let lr = lsn.start..val.get_lsn_range().start;

                if !kr.is_empty() {
-                    let base_count =
-                        Self::is_reimage_worthy(self.get_layer_from_mapping(&val.key()), key)
-                            as usize;
+                    let base_count = Self::is_reimage_worthy(&val, key) as usize;
                    let new_limit = limit.map(|l| l - base_count);
                    let max_stacked_deltas_underneath = self.count_deltas(&kr, &lr, new_limit)?;
                    max_stacked_deltas = std::cmp::max(
@@ -772,12 +621,8 @@ where
    }

    /// Return all L0 delta layers
-    pub fn get_level0_deltas(&self) -> Result<Vec<Arc<L>>> {
-        Ok(self
-            .l0_delta_layers
-            .iter()
-            .map(|x| self.get_layer_from_mapping(&x.key()).clone())
-            .collect())
+    pub fn get_level0_deltas(&self) -> Result<Vec<Arc<PersistentLayerDesc>>> {
+        Ok(self.l0_delta_layers.to_vec())
    }

    /// debugging function to print out the contents of the layer map
@@ -802,72 +647,51 @@ where
        println!("End dump LayerMap");
        Ok(())
    }
-
-    /// Similar to `Arc::ptr_eq`, but only compares the object pointers, not vtables.
-    ///
-    /// Returns `true` if the two `Arc` point to the same layer, false otherwise.
-    #[inline(always)]
-    pub fn compare_arced_layers(left: &Arc<L>, right: &Arc<L>) -> bool {
-        // "dyn Trait" objects are "fat pointers" in that they have two components:
-        // - pointer to the object
-        // - pointer to the vtable
-        //
-        // rust does not provide a guarantee that these vtables are unique, but however
-        // `Arc::ptr_eq` as of writing (at least up to 1.67) uses a comparison where both the
-        // pointer and the vtable need to be equal.
-        //
-        // See: https://github.com/rust-lang/rust/issues/103763
-        //
-        // A future version of rust will most likely use this form below, where we cast each
-        // pointer into a pointer to unit, which drops the inaccessible vtable pointer, making it
-        // not affect the comparison.
-        //
-        // See: https://github.com/rust-lang/rust/pull/106450
-        let left = Arc::as_ptr(left) as *const ();
-        let right = Arc::as_ptr(right) as *const ();
-
-        left == right
-    }
 }

 #[cfg(test)]
 mod tests {
-    use super::{LayerMap, Replacement};
-    use crate::tenant::storage_layer::{Layer, LayerDescriptor, LayerFileName};
+    use super::LayerMap;
+    use crate::tenant::storage_layer::{tests::LayerDescriptor, LayerFileName};
    use std::str::FromStr;
    use std::sync::Arc;

    mod l0_delta_layers_updated {

+        use crate::tenant::{
+            storage_layer::{PersistentLayer, PersistentLayerDesc},
+            timeline::LayerFileManager,
+        };
+
        use super::*;

        #[test]
        fn for_full_range_delta() {
            // l0_delta_layers are used by compaction, and should observe all buffered updates
            l0_delta_layers_updated_scenario(
-                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000053423C21-0000000053424D69",
-                true
-            )
+                 "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000053423C21-0000000053424D69",
+                 true
+             )
        }

        #[test]
        fn for_non_full_range_delta() {
            // has minimal uncovered areas compared to l0_delta_layers_updated_on_insert_replace_remove_for_full_range_delta
            l0_delta_layers_updated_scenario(
-                "000000000000000000000000000000000001-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFE__0000000053423C21-0000000053424D69",
-                // because not full range
-                false
-            )
+                 "000000000000000000000000000000000001-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFE__0000000053423C21-0000000053424D69",
+                 // because not full range
+                 false
+             )
        }

        #[test]
        fn for_image() {
            l0_delta_layers_updated_scenario(
-                "000000000000000000000000000000000000-000000000000000000000000000000010000__0000000053424D69",
-                // code only checks if it is a full range layer, doesn't care about images, which must
-                // mean we should in practice never have full range images
-                false
-            )
+                 "000000000000000000000000000000000000-000000000000000000000000000000010000__0000000053424D69",
+                 // code only checks if it is a full range layer, doesn't care about images, which must
+                 // mean we should in practice never have full range images
+                 false
+             )
        }

        #[test]
@@ -883,16 +707,16 @@ mod tests {
            let not_found = Arc::new(layer.clone());
            let new_version = Arc::new(layer);

-            let mut map = LayerMap::default();
+            // 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 res = map.batch_update().replace_historic(
-                not_found.get_persistent_layer_desc(),
-                &not_found,
-                new_version.get_persistent_layer_desc(),
-                new_version,
-            );
+            let mut mapping = LayerFileManager::new();

-            assert!(matches!(res, Ok(Replacement::NotFound)), "{res:?}");
+            mapping
+                .replace_and_verify(not_found, new_version)
+                .unwrap_err();
        }

        fn l0_delta_layers_updated_scenario(layer_name: &str, expected_l0: bool) {
@@ -903,49 +727,44 @@ mod tests {
            let downloaded = Arc::new(skeleton);

            let mut map = LayerMap::default();
+            let mut mapping = LayerFileManager::new();

            // two disjoint Arcs in different lifecycle phases. even if it seems they must be the
            // same layer, we use LayerMap::compare_arced_layers as the identity of layers.
-            assert!(!LayerMap::compare_arced_layers(&remote, &downloaded));
+            assert_eq!(remote.layer_desc(), downloaded.layer_desc());

            let expected_in_counts = (1, usize::from(expected_l0));

            map.batch_update()
-                .insert_historic(remote.get_persistent_layer_desc(), remote.clone());
-            assert_eq!(count_layer_in(&map, &remote), expected_in_counts);
-
-            let replaced = map
-                .batch_update()
-                .replace_historic(
-                    remote.get_persistent_layer_desc(),
-                    &remote,
-                    downloaded.get_persistent_layer_desc(),
-                    downloaded.clone(),
-                )
-                .expect("name derived attributes are the same");
-            assert!(
-                matches!(replaced, Replacement::Replaced { .. }),
-                "{replaced:?}"
+                .insert_historic(remote.layer_desc().clone());
+            mapping.insert(remote.clone());
+            assert_eq!(
+                count_layer_in(&map, remote.layer_desc()),
+                expected_in_counts
+            );
+
+            mapping
+                .replace_and_verify(remote, downloaded.clone())
+                .expect("name derived attributes are the same");
+            assert_eq!(
+                count_layer_in(&map, downloaded.layer_desc()),
+                expected_in_counts
            );
-            assert_eq!(count_layer_in(&map, &downloaded), expected_in_counts);

            map.batch_update()
-                .remove_historic(downloaded.get_persistent_layer_desc(), downloaded.clone());
-            assert_eq!(count_layer_in(&map, &downloaded), (0, 0));
+                .remove_historic(downloaded.layer_desc().clone());
+            assert_eq!(count_layer_in(&map, downloaded.layer_desc()), (0, 0));
        }

-        fn count_layer_in<L: Layer + ?Sized>(map: &LayerMap<L>, layer: &Arc<L>) -> (usize, usize) {
+        fn count_layer_in(map: &LayerMap, layer: &PersistentLayerDesc) -> (usize, usize) {
            let historic = map
                .iter_historic_layers()
-                .filter(|x| LayerMap::compare_arced_layers(x, layer))
+                .filter(|x| x.key() == layer.key())
                .count();
            let l0s = map
                .get_level0_deltas()
                .expect("why does this return a result");
-            let l0 = l0s
-                .iter()
-                .filter(|x| LayerMap::compare_arced_layers(x, layer))
-                .count();
+            let l0 = l0s.iter().filter(|x| x.key() == layer.key()).count();

            (historic, l0)
        }
--- a/pageserver/src/tenant/layer_map/historic_layer_coverage.rs
+++ b/pageserver/src/tenant/layer_map/historic_layer_coverage.rs
@@ -3,6 +3,8 @@ use std::ops::Range;

 use tracing::info;

+use crate::tenant::storage_layer::PersistentLayerDesc;
+
 use super::layer_coverage::LayerCoverageTuple;

 /// Layers in this module are identified and indexed by this data.
@@ -41,8 +43,8 @@ impl Ord for LayerKey {
    }
 }

-impl<'a, L: crate::tenant::storage_layer::Layer + ?Sized> From<&'a L> for LayerKey {
-    fn from(layer: &'a L) -> Self {
+impl From<&PersistentLayerDesc> for LayerKey {
+    fn from(layer: &PersistentLayerDesc) -> Self {
        let kr = layer.get_key_range();
        let lr = layer.get_lsn_range();
        LayerKey {
@@ -454,59 +456,6 @@ impl<Value: Clone> BufferedHistoricLayerCoverage<Value> {
        self.buffer.insert(layer_key, None);
    }

-    /// Replaces a previous layer with a new layer value.
-    ///
-    /// The replacement is conditional on:
-    /// - there is an existing `LayerKey` record
-    /// - there is no buffered removal for the given `LayerKey`
-    /// - the given closure returns true for the current `Value`
-    ///
-    /// The closure is used to compare the latest value (buffered insert, or existing layer)
-    /// against some expectation. This allows to use `Arc::ptr_eq` or similar which would be
-    /// inaccessible via `PartialEq` trait.
-    ///
-    /// Returns a `Replacement` value describing the outcome; only the case of
-    /// `Replacement::Replaced` modifies the map and requires a rebuild.
-    pub fn replace<F>(
-        &mut self,
-        layer_key: &LayerKey,
-        new: Value,
-        check_expected: F,
-    ) -> Replacement<Value>
-    where
-        F: FnOnce(&Value) -> bool,
-    {
-        let (slot, in_buffered) = match self.buffer.get(layer_key) {
-            Some(inner @ Some(_)) => {
-                // we compare against the buffered version, because there will be a later
-                // rebuild before querying
-                (inner.as_ref(), true)
-            }
-            Some(None) => {
-                // buffer has removal for this key; it will not be equivalent by any check_expected.
-                return Replacement::RemovalBuffered;
-            }
-            None => {
-                // no pending modification for the key, check layers
-                (self.layers.get(layer_key), false)
-            }
-        };
-
-        match slot {
-            Some(existing) if !check_expected(existing) => {
-                // unfortunate clone here, but otherwise the nll borrowck grows the region of
-                // 'a to cover the whole function, and we could not mutate in the other
-                // Some(existing) branch
-                Replacement::Unexpected(existing.clone())
-            }
-            None => Replacement::NotFound,
-            Some(_existing) => {
-                self.insert(layer_key.to_owned(), new);
-                Replacement::Replaced { in_buffered }
-            }
-        }
-    }
-
    pub fn rebuild(&mut self) {
        // Find the first LSN that needs to be rebuilt
        let rebuild_since: u64 = match self.buffer.iter().next() {
@@ -575,22 +524,6 @@ impl<Value: Clone> BufferedHistoricLayerCoverage<Value> {
    }
 }

-/// Outcome of the replace operation.
-#[derive(Debug)]
-pub enum Replacement<Value> {
-    /// Previous value was replaced with the new value.
-    Replaced {
-        /// Replacement happened for a scheduled insert.
-        in_buffered: bool,
-    },
-    /// Key was not found buffered updates or existing layers.
-    NotFound,
-    /// Key has been scheduled for removal, it was not replaced.
-    RemovalBuffered,
-    /// Previous value was rejected by the closure.
-    Unexpected(Value),
-}
-
 #[test]
 fn test_retroactive_regression_1() {
    let mut map = BufferedHistoricLayerCoverage::new();
@@ -699,139 +632,3 @@ fn test_retroactive_simple() {
        assert_eq!(version.image_coverage.query(8), Some("Image 4".to_string()));
    }
 }
-
-#[test]
-fn test_retroactive_replacement() {
-    let mut map = BufferedHistoricLayerCoverage::new();
-
-    let keys = [
-        LayerKey {
-            key: 0..5,
-            lsn: 100..101,
-            is_image: true,
-        },
-        LayerKey {
-            key: 3..9,
-            lsn: 110..111,
-            is_image: true,
-        },
-        LayerKey {
-            key: 4..6,
-            lsn: 120..121,
-            is_image: true,
-        },
-    ];
-
-    let layers = [
-        "Image 1".to_string(),
-        "Image 2".to_string(),
-        "Image 3".to_string(),
-    ];
-
-    for (key, layer) in keys.iter().zip(layers.iter()) {
-        map.insert(key.to_owned(), layer.to_owned());
-    }
-
-    // rebuild is not necessary here, because replace works for both buffered updates and existing
-    // layers.
-
-    for (key, orig_layer) in keys.iter().zip(layers.iter()) {
-        let replacement = format!("Remote {orig_layer}");
-
-        // evict
-        let ret = map.replace(key, replacement.clone(), |l| l == orig_layer);
-        assert!(
-            matches!(ret, Replacement::Replaced { .. }),
-            "replace {orig_layer}: {ret:?}"
-        );
-        map.rebuild();
-
-        let at = key.lsn.end + 1;
-
-        let version = map.get().expect("rebuilt").get_version(at).unwrap();
-        assert_eq!(
-            version.image_coverage.query(4).as_deref(),
-            Some(replacement.as_str()),
-            "query for 4 at version {at} after eviction",
-        );
-
-        // download
-        let ret = map.replace(key, orig_layer.clone(), |l| l == &replacement);
-        assert!(
-            matches!(ret, Replacement::Replaced { .. }),
-            "replace {orig_layer} back: {ret:?}"
-        );
-        map.rebuild();
-        let version = map.get().expect("rebuilt").get_version(at).unwrap();
-        assert_eq!(
-            version.image_coverage.query(4).as_deref(),
-            Some(orig_layer.as_str()),
-            "query for 4 at version {at} after download",
-        );
-    }
-}
-
-#[test]
-fn missing_key_is_not_inserted_with_replace() {
-    let mut map = BufferedHistoricLayerCoverage::new();
-    let key = LayerKey {
-        key: 0..5,
-        lsn: 100..101,
-        is_image: true,
-    };
-
-    let ret = map.replace(&key, "should not replace", |_| true);
-    assert!(matches!(ret, Replacement::NotFound), "{ret:?}");
-    map.rebuild();
-    assert!(map
-        .get()
-        .expect("no changes to rebuild")
-        .get_version(102)
-        .is_none());
-}
-
-#[test]
-fn replacing_buffered_insert_and_remove() {
-    let mut map = BufferedHistoricLayerCoverage::new();
-    let key = LayerKey {
-        key: 0..5,
-        lsn: 100..101,
-        is_image: true,
-    };
-
-    map.insert(key.clone(), "Image 1");
-    let ret = map.replace(&key, "Remote Image 1", |&l| l == "Image 1");
-    assert!(
-        matches!(ret, Replacement::Replaced { in_buffered: true }),
-        "{ret:?}"
-    );
-    map.rebuild();
-
-    assert_eq!(
-        map.get()
-            .expect("rebuilt")
-            .get_version(102)
-            .unwrap()
-            .image_coverage
-            .query(4),
-        Some("Remote Image 1")
-    );
-
-    map.remove(key.clone());
-    let ret = map.replace(&key, "should not replace", |_| true);
-    assert!(
-        matches!(ret, Replacement::RemovalBuffered),
-        "cannot replace after scheduled remove: {ret:?}"
-    );
-
-    map.rebuild();
-
-    let ret = map.replace(&key, "should not replace", |_| true);
-    assert!(
-        matches!(ret, Replacement::NotFound),
-        "cannot replace after remove + rebuild: {ret:?}"
-    );
-
-    let at_version = map.get().expect("rebuilt").get_version(102);
-    assert!(at_version.is_none());
-}
--- a/pageserver/src/tenant/manifest.rs
+++ b/pageserver/src/tenant/manifest.rs
@@ -0,0 +1,325 @@
+//! This module contains the encoding and decoding of the local manifest file.
+//!
+//! MANIFEST is a write-ahead log which is stored locally to each timeline. It
+//! records the state of the storage engine. It contains a snapshot of the
+//! state and all operations proceeding that snapshot. The file begins with a
+//! header recording MANIFEST version number. After that, it contains a snapshot.
+//! The snapshot is followed by a list of operations. Each operation is a list
+//! of records. Each record is either an addition or a removal of a layer.
+//!
+//! With MANIFEST, we can:
+//!
+//! 1. recover state quickly by reading the file, potentially boosting the
+//!    startup speed.
+//! 2. ensure all operations are atomic and avoid corruption, solving issues
+//!    like redundant image layer and preparing us for future compaction
+//!    strategies.
+//!
+//! There is also a format for storing all layer files on S3, called
+//! `index_part.json`. Compared with index_part, MANIFEST is an WAL which
+//! records all operations as logs, and therefore we can easily replay the
+//! operations when recovering from crash, while ensuring those operations
+//! are atomic upon restart.
+//!
+//! 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
+
+use std::io::{self, Read, Write};
+
+use crate::virtual_file::VirtualFile;
+use anyhow::Result;
+use bytes::{Buf, BufMut, Bytes, BytesMut};
+use crc32c::crc32c;
+use serde::{Deserialize, Serialize};
+use tracing::log::warn;
+use utils::lsn::Lsn;
+
+use super::storage_layer::PersistentLayerDesc;
+
+pub struct Manifest {
+    file: VirtualFile,
+}
+
+#[derive(Clone, Serialize, Deserialize, PartialEq, Eq, Debug)]
+pub struct Snapshot {
+    pub layers: Vec<PersistentLayerDesc>,
+}
+
+/// serde by default encode this in tagged enum, and therefore it will be something
+/// like `{ "AddLayer": { ... } }`.
+#[derive(Clone, Serialize, Deserialize, PartialEq, Eq, Debug)]
+pub enum Record {
+    AddLayer(PersistentLayerDesc),
+    RemoveLayer(PersistentLayerDesc),
+}
+
+/// `echo neon.manifest | sha1sum` and take the leading 8 bytes.
+const MANIFEST_MAGIC_NUMBER: u64 = 0xf5c44592b806109c;
+const MANIFEST_VERSION: u64 = 1;
+
+#[derive(Clone, Serialize, Deserialize, PartialEq, Eq, Debug)]
+pub struct ManifestHeader {
+    magic_number: u64,
+    version: u64,
+}
+
+const MANIFEST_HEADER_LEN: usize = 16;
+
+impl ManifestHeader {
+    fn encode(&self) -> BytesMut {
+        let mut buf = BytesMut::with_capacity(MANIFEST_HEADER_LEN);
+        buf.put_u64(self.magic_number);
+        buf.put_u64(self.version);
+        buf
+    }
+
+    fn decode(mut buf: &[u8]) -> Self {
+        assert!(buf.len() == MANIFEST_HEADER_LEN, "invalid header");
+        Self {
+            magic_number: buf.get_u64(),
+            version: buf.get_u64(),
+        }
+    }
+}
+
+#[derive(Clone, Serialize, Deserialize, PartialEq, Eq, Debug)]
+pub enum Operation {
+    /// A snapshot of the current state.
+    ///
+    /// Lsn field represents the LSN that is persisted to disk for this snapshot.
+    Snapshot(Snapshot, Lsn),
+    /// An atomic operation that changes the state.
+    ///
+    /// Lsn field represents the LSN that is persisted to disk after the operation is done.
+    /// This will only change when new L0 is flushed to the disk.
+    Operation(Vec<Record>, Lsn),
+}
+
+struct RecordHeader {
+    size: u32,
+    checksum: u32,
+}
+
+const RECORD_HEADER_LEN: usize = 8;
+
+impl RecordHeader {
+    fn encode(&self) -> BytesMut {
+        let mut buf = BytesMut::with_capacity(RECORD_HEADER_LEN);
+        buf.put_u32(self.size);
+        buf.put_u32(self.checksum);
+        buf
+    }
+
+    fn decode(mut buf: &[u8]) -> Self {
+        assert!(buf.len() == RECORD_HEADER_LEN, "invalid header");
+        Self {
+            size: buf.get_u32(),
+            checksum: buf.get_u32(),
+        }
+    }
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum ManifestLoadError {
+    #[error("manifest header is corrupted")]
+    CorruptedManifestHeader,
+    #[error("unsupported manifest version: got {0}, expected {1}")]
+    UnsupportedVersion(u64, u64),
+    #[error("error when decoding record: {0}")]
+    DecodeRecord(serde_json::Error),
+    #[error("I/O error: {0}")]
+    Io(io::Error),
+}
+
+#[must_use = "Should check if the manifest is partially corrupted"]
+pub struct ManifestPartiallyCorrupted(bool);
+
+impl Manifest {
+    /// Create a new manifest by writing the manifest header and a snapshot record to the given file.
+    pub fn init(file: VirtualFile, snapshot: Snapshot, lsn: Lsn) -> Result<Self> {
+        let mut manifest = Self { file };
+        manifest.append_manifest_header(ManifestHeader {
+            magic_number: MANIFEST_MAGIC_NUMBER,
+            version: MANIFEST_VERSION,
+        })?;
+        manifest.append_operation(Operation::Snapshot(snapshot, lsn))?;
+        Ok(manifest)
+    }
+
+    /// Load a manifest. Returns the manifest and a list of operations. If the manifest is corrupted,
+    /// the bool flag will be set to true and the user is responsible to reconstruct a new manifest and
+    /// backup the current one.
+    pub fn load(
+        mut file: VirtualFile,
+    ) -> Result<(Self, Vec<Operation>, ManifestPartiallyCorrupted), ManifestLoadError> {
+        let mut buf = vec![];
+        file.read_to_end(&mut buf).map_err(ManifestLoadError::Io)?;
+
+        // Read manifest header
+        let mut buf = Bytes::from(buf);
+        if buf.remaining() < MANIFEST_HEADER_LEN {
+            return Err(ManifestLoadError::CorruptedManifestHeader);
+        }
+        let header = ManifestHeader::decode(&buf[..MANIFEST_HEADER_LEN]);
+        buf.advance(MANIFEST_HEADER_LEN);
+        if header.version != MANIFEST_VERSION {
+            return Err(ManifestLoadError::UnsupportedVersion(
+                header.version,
+                MANIFEST_VERSION,
+            ));
+        }
+
+        // Read operations
+        let mut operations = Vec::new();
+        let corrupted = loop {
+            if buf.remaining() == 0 {
+                break false;
+            }
+            if buf.remaining() < RECORD_HEADER_LEN {
+                warn!("incomplete header when decoding manifest, could be corrupted");
+                break true;
+            }
+            let RecordHeader { size, checksum } = RecordHeader::decode(&buf[..RECORD_HEADER_LEN]);
+            let size = size as usize;
+            buf.advance(RECORD_HEADER_LEN);
+            if buf.remaining() < size {
+                warn!("incomplete data when decoding manifest, could be corrupted");
+                break true;
+            }
+            let data = &buf[..size];
+            if crc32c(data) != checksum {
+                warn!("checksum mismatch when decoding manifest, could be corrupted");
+                break true;
+            }
+            // if the following decode fails, we cannot use the manifest or safely ignore any record.
+            operations.push(serde_json::from_slice(data).map_err(ManifestLoadError::DecodeRecord)?);
+            buf.advance(size);
+        };
+        Ok((
+            Self { file },
+            operations,
+            ManifestPartiallyCorrupted(corrupted),
+        ))
+    }
+
+    fn append_data(&mut self, data: &[u8]) -> Result<()> {
+        if data.len() >= u32::MAX as usize {
+            panic!("data too large");
+        }
+        let header = RecordHeader {
+            size: data.len() as u32,
+            checksum: crc32c(data),
+        };
+        let header = header.encode();
+        self.file.write_all(&header)?;
+        self.file.write_all(data)?;
+        self.file.sync_all()?;
+        Ok(())
+    }
+
+    fn append_manifest_header(&mut self, header: ManifestHeader) -> Result<()> {
+        let encoded = header.encode();
+        self.file.write_all(&encoded)?;
+        Ok(())
+    }
+
+    /// Add an operation to the manifest. The operation will be appended to the end of the file,
+    /// and the file will fsync.
+    pub fn append_operation(&mut self, operation: Operation) -> Result<()> {
+        let encoded = Vec::from(serde_json::to_string(&operation)?);
+        self.append_data(&encoded)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::fs::OpenOptions;
+
+    use crate::repository::Key;
+
+    use super::*;
+
+    #[test]
+    fn test_read_manifest() {
+        let testdir = crate::config::PageServerConf::test_repo_dir("test_read_manifest");
+        std::fs::create_dir_all(&testdir).unwrap();
+        let file = VirtualFile::create(&testdir.join("MANIFEST")).unwrap();
+        let layer1 = PersistentLayerDesc::new_test(Key::from_i128(0)..Key::from_i128(233));
+        let layer2 = PersistentLayerDesc::new_test(Key::from_i128(233)..Key::from_i128(2333));
+        let layer3 = PersistentLayerDesc::new_test(Key::from_i128(2333)..Key::from_i128(23333));
+        let layer4 = PersistentLayerDesc::new_test(Key::from_i128(23333)..Key::from_i128(233333));
+
+        // Write a manifest with a snapshot and some operations
+        let snapshot = Snapshot {
+            layers: vec![layer1, layer2],
+        };
+        let mut manifest = Manifest::init(file, snapshot.clone(), Lsn::from(0)).unwrap();
+        manifest
+            .append_operation(Operation::Operation(
+                vec![Record::AddLayer(layer3.clone())],
+                Lsn::from(1),
+            ))
+            .unwrap();
+        drop(manifest);
+
+        // Open the second time and write
+        let file = VirtualFile::open_with_options(
+            &testdir.join("MANIFEST"),
+            OpenOptions::new()
+                .read(true)
+                .write(true)
+                .create_new(false)
+                .truncate(false),
+        )
+        .unwrap();
+        let (mut manifest, operations, corrupted) = Manifest::load(file).unwrap();
+        assert!(!corrupted.0);
+        assert_eq!(operations.len(), 2);
+        assert_eq!(
+            &operations[0],
+            &Operation::Snapshot(snapshot.clone(), Lsn::from(0))
+        );
+        assert_eq!(
+            &operations[1],
+            &Operation::Operation(vec![Record::AddLayer(layer3.clone())], Lsn::from(1))
+        );
+        manifest
+            .append_operation(Operation::Operation(
+                vec![
+                    Record::RemoveLayer(layer3.clone()),
+                    Record::AddLayer(layer4.clone()),
+                ],
+                Lsn::from(2),
+            ))
+            .unwrap();
+        drop(manifest);
+
+        // Open the third time and verify
+        let file = VirtualFile::open_with_options(
+            &testdir.join("MANIFEST"),
+            OpenOptions::new()
+                .read(true)
+                .write(true)
+                .create_new(false)
+                .truncate(false),
+        )
+        .unwrap();
+        let (_manifest, operations, corrupted) = Manifest::load(file).unwrap();
+        assert!(!corrupted.0);
+        assert_eq!(operations.len(), 3);
+        assert_eq!(&operations[0], &Operation::Snapshot(snapshot, Lsn::from(0)));
+        assert_eq!(
+            &operations[1],
+            &Operation::Operation(vec![Record::AddLayer(layer3.clone())], Lsn::from(1))
+        );
+        assert_eq!(
+            &operations[2],
+            &Operation::Operation(
+                vec![Record::RemoveLayer(layer3), Record::AddLayer(layer4)],
+                Lsn::from(2)
+            )
+        );
+    }
+}
--- a/pageserver/src/tenant/mgr.rs
+++ b/pageserver/src/tenant/mgr.rs
@@ -396,7 +396,9 @@ pub async fn delete_timeline(
    ctx: &RequestContext,
 ) -> Result<(), DeleteTimelineError> {
    let tenant = get_tenant(tenant_id, true).await?;
-    tenant.delete_timeline(timeline_id, ctx).await?;
+    tenant
+        .prepare_and_schedule_delete_timeline(timeline_id, ctx)
+        .await?;
    Ok(())
 }

@@ -673,7 +675,7 @@ pub async fn immediate_gc(
        .get(&tenant_id)
        .map(Arc::clone)
        .with_context(|| format!("tenant {tenant_id}"))
-        .map_err(ApiError::NotFound)?;
+        .map_err(|e| ApiError::NotFound(e.into()))?;

    let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
    // Use tenant's pitr setting
@@ -722,11 +724,11 @@ pub async fn immediate_compact(
        .get(&tenant_id)
        .map(Arc::clone)
        .with_context(|| format!("tenant {tenant_id}"))
-        .map_err(ApiError::NotFound)?;
+        .map_err(|e| ApiError::NotFound(e.into()))?;

    let timeline = tenant
        .get_timeline(timeline_id, true)
-        .map_err(ApiError::NotFound)?;
+        .map_err(|e| ApiError::NotFound(e.into()))?;

    // Run in task_mgr to avoid race with tenant_detach operation
    let ctx = ctx.detached_child(TaskKind::Compaction, DownloadBehavior::Download);
--- a/pageserver/src/tenant/remote_timeline_client.rs
+++ b/pageserver/src/tenant/remote_timeline_client.rs
@@ -210,13 +210,15 @@ use chrono::{NaiveDateTime, Utc};
 pub use download::{is_temp_download_file, list_remote_timelines};
 use scopeguard::ScopeGuard;

+use std::collections::{HashMap, VecDeque};
+use std::path::Path;
 use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::{Arc, Mutex};

-use remote_storage::{DownloadError, GenericRemoteStorage};
+use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
 use std::ops::DerefMut;
 use tokio::runtime::Runtime;
-use tracing::{debug, error, info, warn};
+use tracing::{debug, error, info, instrument, warn};
 use tracing::{info_span, Instrument};
 use utils::lsn::Lsn;

@@ -225,7 +227,9 @@ use crate::metrics::{
    RemoteTimelineClientMetricsCallTrackSize, REMOTE_ONDEMAND_DOWNLOADED_BYTES,
    REMOTE_ONDEMAND_DOWNLOADED_LAYERS,
 };
+use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
+use crate::tenant::upload_queue::Delete;
 use crate::{
    config::PageServerConf,
    task_mgr,
@@ -259,7 +263,7 @@ const FAILED_UPLOAD_WARN_THRESHOLD: u32 = 3;

 pub enum MaybeDeletedIndexPart {
    IndexPart(IndexPart),
-    Deleted,
+    Deleted(IndexPart),
 }

 /// Errors that can arise when calling [`RemoteTimelineClient::stop`].
@@ -361,11 +365,42 @@ impl RemoteTimelineClient {
        Ok(())
    }

+    /// Initialize the queue in stopped state. Used in startup path
+    /// to continue deletion operation interrupted by pageserver crash or restart.
+    pub fn init_upload_queue_stopped_to_continue_deletion(
+        &self,
+        index_part: &IndexPart,
+    ) -> anyhow::Result<()> {
+        // FIXME: consider newtype for DeletedIndexPart.
+        let deleted_at = index_part.deleted_at.ok_or(anyhow::anyhow!(
+            "bug: it is responsibility of the caller to provide index part from MaybeDeletedIndexPart::Deleted"
+        ))?;
+
+        {
+            let mut upload_queue = self.upload_queue.lock().unwrap();
+            upload_queue.initialize_with_current_remote_index_part(index_part)?;
+            self.update_remote_physical_size_gauge(Some(index_part));
+        }
+        // also locks upload queue, without dropping the guard above it will be a deadlock
+        self.stop().expect("initialized line above");
+
+        let mut upload_queue = self.upload_queue.lock().unwrap();
+
+        upload_queue
+            .stopped_mut()
+            .expect("stopped above")
+            .deleted_at = SetDeletedFlagProgress::Successful(deleted_at);
+
+        Ok(())
+    }
+
    pub fn last_uploaded_consistent_lsn(&self) -> Option<Lsn> {
        match &*self.upload_queue.lock().unwrap() {
            UploadQueue::Uninitialized => None,
            UploadQueue::Initialized(q) => Some(q.last_uploaded_consistent_lsn),
-            UploadQueue::Stopped(q) => Some(q.last_uploaded_consistent_lsn),
+            UploadQueue::Stopped(q) => {
+                Some(q.upload_queue_for_deletion.last_uploaded_consistent_lsn)
+            }
        }
    }

@@ -420,7 +455,7 @@ impl RemoteTimelineClient {
        .await?;

        if index_part.deleted_at.is_some() {
-            Ok(MaybeDeletedIndexPart::Deleted)
+            Ok(MaybeDeletedIndexPart::Deleted(index_part))
        } else {
            Ok(MaybeDeletedIndexPart::IndexPart(index_part))
        }
@@ -622,7 +657,11 @@ impl RemoteTimelineClient {

            // schedule the actual deletions
            for name in names {
-                let op = UploadOp::Delete(RemoteOpFileKind::Layer, name.clone());
+                let op = UploadOp::Delete(Delete {
+                    file_kind: RemoteOpFileKind::Layer,
+                    layer_file_name: name.clone(),
+                    scheduled_from_timeline_delete: false,
+                });
                self.calls_unfinished_metric_begin(&op);
                upload_queue.queued_operations.push_back(op);
                info!("scheduled layer file deletion {}", name.file_name());
@@ -639,18 +678,11 @@ impl RemoteTimelineClient {
    /// Wait for all previously scheduled uploads/deletions to complete
    ///
    pub async fn wait_completion(self: &Arc<Self>) -> anyhow::Result<()> {
-        let (sender, mut receiver) = tokio::sync::watch::channel(());
-        let barrier_op = UploadOp::Barrier(sender);
-
-        {
+        let mut receiver = {
            let mut guard = self.upload_queue.lock().unwrap();
            let upload_queue = guard.initialized_mut()?;
-            upload_queue.queued_operations.push_back(barrier_op);
-            // Don't count this kind of operation!
-
-            // Launch the task immediately, if possible
-            self.launch_queued_tasks(upload_queue);
-        }
+            self.schedule_barrier(upload_queue)
+        };

        if receiver.changed().await.is_err() {
            anyhow::bail!("wait_completion aborted because upload queue was stopped");
@@ -658,6 +690,22 @@ impl RemoteTimelineClient {
        Ok(())
    }

+    fn schedule_barrier(
+        self: &Arc<Self>,
+        upload_queue: &mut UploadQueueInitialized,
+    ) -> tokio::sync::watch::Receiver<()> {
+        let (sender, receiver) = tokio::sync::watch::channel(());
+        let barrier_op = UploadOp::Barrier(sender);
+
+        upload_queue.queued_operations.push_back(barrier_op);
+        // Don't count this kind of operation!
+
+        // Launch the task immediately, if possible
+        self.launch_queued_tasks(upload_queue);
+
+        receiver
+    }
+
    /// Set the deleted_at field in the remote index file.
    ///
    /// This fails if the upload queue has not been `stop()`ed.
@@ -665,6 +713,7 @@ impl RemoteTimelineClient {
    /// The caller is responsible for calling `stop()` AND for waiting
    /// for any ongoing upload tasks to finish after `stop()` has succeeded.
    /// Check method [`RemoteTimelineClient::stop`] for details.
+    #[instrument(skip_all)]
    pub(crate) async fn persist_index_part_with_deleted_flag(
        self: &Arc<Self>,
    ) -> Result<(), PersistIndexPartWithDeletedFlagError> {
@@ -674,15 +723,7 @@ impl RemoteTimelineClient {
            // We must be in stopped state because otherwise
            // we can have inprogress index part upload that can overwrite the file
            // with missing is_deleted flag that we going to set below
-            let stopped = match &mut *locked {
-                UploadQueue::Uninitialized => {
-                    return Err(anyhow::anyhow!("is not Stopped but Uninitialized").into())
-                }
-                UploadQueue::Initialized(_) => {
-                    return Err(anyhow::anyhow!("is not Stopped but Initialized").into())
-                }
-                UploadQueue::Stopped(stopped) => stopped,
-            };
+            let stopped = locked.stopped_mut()?;

            match stopped.deleted_at {
                SetDeletedFlagProgress::NotRunning => (), // proceed
@@ -696,48 +737,34 @@ impl RemoteTimelineClient {
            let deleted_at = Utc::now().naive_utc();
            stopped.deleted_at = SetDeletedFlagProgress::InProgress(deleted_at);

-            let mut index_part = IndexPart::new(
-                stopped.latest_files.clone(),
-                stopped.last_uploaded_consistent_lsn,
-                stopped
-                    .latest_metadata
-                    .to_bytes()
-                    .context("serialize metadata")?,
-            );
+            let mut index_part = IndexPart::try_from(&stopped.upload_queue_for_deletion)
+                .context("IndexPart serialize")?;
            index_part.deleted_at = Some(deleted_at);
            index_part
        };

        let undo_deleted_at = scopeguard::guard(Arc::clone(self), |self_clone| {
            let mut locked = self_clone.upload_queue.lock().unwrap();
-            let stopped = match &mut *locked {
-                UploadQueue::Uninitialized | UploadQueue::Initialized(_) => unreachable!(
-                    "there's no way out of Stopping, and we checked it's Stopping above: {:?}",
-                    locked.as_str(),
-                ),
-                UploadQueue::Stopped(stopped) => stopped,
-            };
+            let stopped = locked
+                .stopped_mut()
+                .expect("there's no way out of Stopping, and we checked it's Stopping above");
            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.
-        #[cfg(feature = "testing")]
-        tokio::task::spawn_blocking({
-            let current = tracing::Span::current();
-            move || {
-                let _entered = current.entered();
-                tracing::info!(
-                    "at failpoint persist_index_part_with_deleted_flag_after_set_before_upload_pause"
-                );
-                fail::fail_point!(
-                    "persist_index_part_with_deleted_flag_after_set_before_upload_pause"
-                );
-            }
-        })
-        .await
-        .expect("spawn_blocking");
-
+        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");
+        }
        upload::upload_index_part(
            self.conf,
            &self.storage_impl,
@@ -751,13 +778,10 @@ impl RemoteTimelineClient {
        ScopeGuard::into_inner(undo_deleted_at);
        {
            let mut locked = self.upload_queue.lock().unwrap();
-            let stopped = match &mut *locked {
-                UploadQueue::Uninitialized | UploadQueue::Initialized(_) => unreachable!(
-                    "there's no way out of Stopping, and we checked it's Stopping above: {:?}",
-                    locked.as_str(),
-                ),
-                UploadQueue::Stopped(stopped) => stopped,
-            };
+
+            let stopped = locked
+                .stopped_mut()
+                .expect("there's no way out of Stopping, and we checked it's Stopping above");
            stopped.deleted_at = SetDeletedFlagProgress::Successful(
                index_part_with_deleted_at
                    .deleted_at
@@ -768,6 +792,90 @@ impl RemoteTimelineClient {
        Ok(())
    }

+    /// Prerequisites: UploadQueue should be in stopped state and deleted_at should be successfuly set.
+    /// The function deletes layer files one by one, then lists the prefix to see if we leaked something
+    /// deletes leaked files if any and proceeds with deletion of index file at the end.
+    pub(crate) async fn delete_all(self: &Arc<Self>) -> anyhow::Result<()> {
+        debug_assert_current_span_has_tenant_and_timeline_id();
+
+        let (mut receiver, deletions_queued) = {
+            let mut deletions_queued = 0;
+
+            let mut locked = self.upload_queue.lock().unwrap();
+            let stopped = locked.stopped_mut()?;
+
+            if !matches!(stopped.deleted_at, SetDeletedFlagProgress::Successful(_)) {
+                anyhow::bail!("deleted_at is not set")
+            }
+
+            debug_assert!(stopped.upload_queue_for_deletion.no_pending_work());
+
+            stopped
+                .upload_queue_for_deletion
+                .queued_operations
+                .reserve(stopped.upload_queue_for_deletion.latest_files.len());
+
+            // schedule the actual deletions
+            for name in stopped.upload_queue_for_deletion.latest_files.keys() {
+                let op = UploadOp::Delete(Delete {
+                    file_kind: RemoteOpFileKind::Layer,
+                    layer_file_name: name.clone(),
+                    scheduled_from_timeline_delete: true,
+                });
+                self.calls_unfinished_metric_begin(&op);
+                stopped
+                    .upload_queue_for_deletion
+                    .queued_operations
+                    .push_back(op);
+
+                info!("scheduled layer file deletion {}", name.file_name());
+                deletions_queued += 1;
+            }
+
+            self.launch_queued_tasks(&mut stopped.upload_queue_for_deletion);
+
+            (
+                self.schedule_barrier(&mut stopped.upload_queue_for_deletion),
+                deletions_queued,
+            )
+        };
+
+        receiver.changed().await?;
+
+        // Do not delete index part yet, it is needed for possible retry. If we remove it first
+        // and retry will arrive to different pageserver there wont be any traces of it on remote storage
+        let timeline_path = self.conf.timeline_path(&self.timeline_id, &self.tenant_id);
+        let timeline_storage_path = self.conf.remote_path(&timeline_path)?;
+
+        let remaining = self
+            .storage_impl
+            .list_prefixes(Some(&timeline_storage_path))
+            .await?;
+
+        let remaining: Vec<RemotePath> = remaining
+            .into_iter()
+            .filter(|p| p.object_name() != Some(IndexPart::FILE_NAME))
+            .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?;
+        }
+
+        let index_file_path = timeline_storage_path.join(Path::new(IndexPart::FILE_NAME));
+
+        debug!("deleting index part");
+        self.storage_impl.delete(&index_file_path).await?;
+
+        info!(deletions_queued, "done deleting, including index_part.json");
+
+        Ok(())
+    }
+
    ///
    /// Pick next tasks from the queue, and start as many of them as possible without violating
    /// the ordering constraints.
@@ -786,7 +894,7 @@ impl RemoteTimelineClient {
                    // have finished.
                    upload_queue.inprogress_tasks.is_empty()
                }
-                UploadOp::Delete(_, _) => {
+                UploadOp::Delete(_) => {
                    // Wait for preceding uploads to finish. Concurrent deletions are OK, though.
                    upload_queue.num_inprogress_deletions == upload_queue.inprogress_tasks.len()
                }
@@ -817,7 +925,7 @@ impl RemoteTimelineClient {
                UploadOp::UploadMetadata(_, _) => {
                    upload_queue.num_inprogress_metadata_uploads += 1;
                }
-                UploadOp::Delete(_, _) => {
+                UploadOp::Delete(_) => {
                    upload_queue.num_inprogress_deletions += 1;
                }
                UploadOp::Barrier(sender) => {
@@ -891,7 +999,6 @@ impl RemoteTimelineClient {
                        unreachable!("we never launch an upload task if the queue is uninitialized, and once it is initialized, we never go back")
                    }
                }
-                self.calls_unfinished_metric_end(&task.op);
                return;
            }

@@ -937,16 +1044,16 @@ impl RemoteTimelineClient {
                    }
                    res
                }
-                UploadOp::Delete(metric_file_kind, ref layer_file_name) => {
+                UploadOp::Delete(delete) => {
                    let path = &self
                        .conf
                        .timeline_path(&self.timeline_id, &self.tenant_id)
-                        .join(layer_file_name.file_name());
+                        .join(delete.layer_file_name.file_name());
                    delete::delete_layer(self.conf, &self.storage_impl, path)
                        .measure_remote_op(
                            self.tenant_id,
                            self.timeline_id,
-                            *metric_file_kind,
+                            delete.file_kind,
                            RemoteOpKind::Delete,
                            Arc::clone(&self.metrics),
                        )
@@ -1012,11 +1119,24 @@ impl RemoteTimelineClient {
            let mut upload_queue_guard = self.upload_queue.lock().unwrap();
            let upload_queue = match upload_queue_guard.deref_mut() {
                UploadQueue::Uninitialized => panic!("callers are responsible for ensuring this is only called on an initialized queue"),
-                UploadQueue::Stopped(_) => {
+                UploadQueue::Stopped(stopped) => {
+                    // Special care is needed for deletions, if it was an earlier deletion (not scheduled from deletion)
+                    // then stop() took care of it so we just return.
+                    // For deletions that come from delete_all we still want to maintain metrics, launch following tasks, etc.
+                    match &task.op {
+                        UploadOp::Delete(delete) if delete.scheduled_from_timeline_delete => Some(&mut stopped.upload_queue_for_deletion),
+                        _ => None
+                    }
+                },
+                UploadQueue::Initialized(qi) => { Some(qi) }
+            };
+
+            let upload_queue = match upload_queue {
+                Some(upload_queue) => upload_queue,
+                None => {
                    info!("another concurrent task already stopped the queue");
                    return;
-                }, // nothing to do
-                UploadQueue::Initialized(qi) => { qi }
+                }
            };

            upload_queue.inprogress_tasks.remove(&task.task_id);
@@ -1029,7 +1149,7 @@ impl RemoteTimelineClient {
                    upload_queue.num_inprogress_metadata_uploads -= 1;
                    upload_queue.last_uploaded_consistent_lsn = lsn; // XXX monotonicity check?
                }
-                UploadOp::Delete(_, _) => {
+                UploadOp::Delete(_) => {
                    upload_queue.num_inprogress_deletions -= 1;
                }
                UploadOp::Barrier(_) => unreachable!(),
@@ -1063,8 +1183,8 @@ impl RemoteTimelineClient {
                    reason: "metadata uploads are tiny",
                },
            ),
-            UploadOp::Delete(file_kind, _) => (
-                *file_kind,
+            UploadOp::Delete(delete) => (
+                delete.file_kind,
                RemoteOpKind::Delete,
                DontTrackSize {
                    reason: "should we track deletes? positive or negative sign?",
@@ -1111,32 +1231,36 @@ impl RemoteTimelineClient {
                info!("another concurrent task already shut down the queue");
                Ok(())
            }
-            UploadQueue::Initialized(UploadQueueInitialized {
-                latest_files,
-                latest_metadata,
-                last_uploaded_consistent_lsn,
-                ..
-            }) => {
+            UploadQueue::Initialized(initialized) => {
                info!("shutting down upload queue");

                // Replace the queue with the Stopped state, taking ownership of the old
                // Initialized queue. We will do some checks on it, and then drop it.
                let qi = {
-                    // take or clone what we need
-                    let latest_files = std::mem::take(latest_files);
-                    let last_uploaded_consistent_lsn = *last_uploaded_consistent_lsn;
-                    // this could be Copy
-                    let latest_metadata = latest_metadata.clone();
-
-                    let stopped = UploadQueueStopped {
-                        latest_files,
-                        last_uploaded_consistent_lsn,
-                        latest_metadata,
-                        deleted_at: SetDeletedFlagProgress::NotRunning,
+                    // Here we preserve working version of the upload queue for possible use during deletions.
+                    // In-place replace of Initialized to Stopped can be done with the help of https://github.com/Sgeo/take_mut
+                    // but for this use case it doesnt really makes sense to bring unsafe code only for this usage point.
+                    // Deletion is not really perf sensitive so there shouldnt be any problems with cloning a fraction of it.
+                    let upload_queue_for_deletion = UploadQueueInitialized {
+                        task_counter: 0,
+                        latest_files: initialized.latest_files.clone(),
+                        latest_files_changes_since_metadata_upload_scheduled: 0,
+                        latest_metadata: initialized.latest_metadata.clone(),
+                        last_uploaded_consistent_lsn: initialized.last_uploaded_consistent_lsn,
+                        num_inprogress_layer_uploads: 0,
+                        num_inprogress_metadata_uploads: 0,
+                        num_inprogress_deletions: 0,
+                        inprogress_tasks: HashMap::default(),
+                        queued_operations: VecDeque::default(),
                    };

-                    let upload_queue =
-                        std::mem::replace(&mut *guard, UploadQueue::Stopped(stopped));
+                    let upload_queue = std::mem::replace(
+                        &mut *guard,
+                        UploadQueue::Stopped(UploadQueueStopped {
+                            upload_queue_for_deletion,
+                            deleted_at: SetDeletedFlagProgress::NotRunning,
+                        }),
+                    );
                    if let UploadQueue::Initialized(qi) = upload_queue {
                        qi
                    } else {
@@ -1144,8 +1268,6 @@ impl RemoteTimelineClient {
                    }
                };

-                assert!(qi.latest_files.is_empty(), "do not use this anymore");
-
                // consistency check
                assert_eq!(
                    qi.num_inprogress_layer_uploads
@@ -1243,7 +1365,7 @@ mod tests {
    struct TestSetup {
        runtime: &'static tokio::runtime::Runtime,
        entered_runtime: EnterGuard<'static>,
-        harness: TenantHarness<'static>,
+        harness: TenantHarness,
        tenant: Arc<Tenant>,
        tenant_ctx: RequestContext,
        remote_fs_dir: PathBuf,
@@ -1264,7 +1386,12 @@ mod tests {
            let harness = TenantHarness::create(test_name)?;
            let (tenant, ctx) = runtime.block_on(harness.load());
            // create an empty timeline directory
-            let _ = tenant.create_test_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)?;
+            let _ = runtime.block_on(tenant.create_test_timeline(
+                TIMELINE_ID,
+                Lsn(8),
+                DEFAULT_PG_VERSION,
+                &ctx,
+            ))?;

            let remote_fs_dir = harness.conf.workdir.join("remote_fs");
            std::fs::create_dir_all(remote_fs_dir)?;
@@ -1408,7 +1535,7 @@ mod tests {
        // Download back the index.json, and check that the list of files is correct
        let index_part = match runtime.block_on(client.download_index_file())? {
            MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
-            MaybeDeletedIndexPart::Deleted => panic!("unexpectedly got deleted index part"),
+            MaybeDeletedIndexPart::Deleted(_) => panic!("unexpectedly got deleted index part"),
        };

        assert_file_list(
--- a/pageserver/src/tenant/remote_timeline_client/index.rs
+++ b/pageserver/src/tenant/remote_timeline_client/index.rs
@@ -7,9 +7,11 @@ use std::collections::{HashMap, HashSet};
 use chrono::NaiveDateTime;
 use serde::{Deserialize, Serialize};
 use serde_with::{serde_as, DisplayFromStr};
+use utils::bin_ser::SerializeError;

 use crate::tenant::metadata::TimelineMetadata;
 use crate::tenant::storage_layer::LayerFileName;
+use crate::tenant::upload_queue::UploadQueueInitialized;

 use utils::lsn::Lsn;

@@ -115,6 +117,21 @@ impl IndexPart {
    }
 }

+impl TryFrom<&UploadQueueInitialized> for IndexPart {
+    type Error = SerializeError;
+
+    fn try_from(upload_queue: &UploadQueueInitialized) -> Result<Self, Self::Error> {
+        let disk_consistent_lsn = upload_queue.latest_metadata.disk_consistent_lsn();
+        let metadata_bytes = upload_queue.latest_metadata.to_bytes()?;
+
+        Ok(Self::new(
+            upload_queue.latest_files.clone(),
+            disk_consistent_lsn,
+            metadata_bytes,
+        ))
+    }
+}
+
 /// Serialized form of [`LayerFileMetadata`].
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize, Default)]
 pub struct IndexLayerMetadata {
--- a/pageserver/src/tenant/storage_layer.rs
+++ b/pageserver/src/tenant/storage_layer.rs
@@ -176,13 +176,10 @@ 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.
-    pub(crate) fn for_loading_layer<L>(
-        layer_map_lock_held_witness: &BatchedUpdates<'_, L>,
+    pub(crate) fn for_loading_layer(
+        layer_map_lock_held_witness: &BatchedUpdates<'_>,
        status: LayerResidenceStatus,
-    ) -> Self
-    where
-        L: ?Sized + Layer,
-    {
+    ) -> Self {
        let new = LayerAccessStats(Mutex::new(LayerAccessStatsLocked::default()));
        new.record_residence_event(
            layer_map_lock_held_witness,
@@ -197,14 +194,11 @@ 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.
-    pub(crate) fn clone_for_residence_change<L>(
+    pub(crate) fn clone_for_residence_change(
        &self,
-        layer_map_lock_held_witness: &BatchedUpdates<'_, L>,
+        layer_map_lock_held_witness: &BatchedUpdates<'_>,
        new_status: LayerResidenceStatus,
-    ) -> LayerAccessStats
-    where
-        L: ?Sized + Layer,
-    {
+    ) -> LayerAccessStats {
        let clone = {
            let inner = self.0.lock().unwrap();
            inner.clone()
@@ -232,14 +226,12 @@ impl LayerAccessStats {
    /// - Compact: Grab layer map lock, add the new L1 to layer map and remove the L0s, release layer map lock.
    /// - Eviction: observes the new L1 layer whose only activity timestamp is the LayerCreate event.
    ///
-    pub(crate) fn record_residence_event<L>(
+    pub(crate) fn record_residence_event(
        &self,
-        _layer_map_lock_held_witness: &BatchedUpdates<'_, L>,
+        _layer_map_lock_held_witness: &BatchedUpdates<'_>,
        status: LayerResidenceStatus,
        reason: LayerResidenceEventReason,
-    ) where
-        L: ?Sized + Layer,
-    {
+    ) {
        let mut locked = self.0.lock().unwrap();
        locked.iter_mut().for_each(|inner| {
            inner
@@ -389,10 +381,10 @@ pub trait Layer: std::fmt::Debug + Send + Sync {
 }

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

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

 /// A Layer contains all data in a "rectangle" consisting of a range of keys and
 /// range of LSNs.
@@ -473,94 +465,125 @@ pub fn downcast_remote_layer(
    }
 }

-/// 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 {
-    pub key: Range<Key>,
-    pub lsn: Range<Lsn>,
-    pub is_incremental: bool,
-    pub short_id: String,
-}
+pub mod tests {
+    use super::*;

-impl LayerDescriptor {
-    /// `LayerDescriptor` is only used for testing purpose so it does not matter whether it is image / delta,
-    /// and the tenant / timeline id does not matter.
-    pub fn get_persistent_layer_desc(&self) -> PersistentLayerDesc {
-        PersistentLayerDesc::new_delta(
-            TenantId::from_array([0; 16]),
-            TimelineId::from_array([0; 16]),
-            self.key.clone(),
-            self.lsn.clone(),
-            233,
-        )
-    }
-}
-
-impl Layer for LayerDescriptor {
-    fn get_key_range(&self) -> Range<Key> {
-        self.key.clone()
+    /// 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,
    }

-    fn get_lsn_range(&self) -> Range<Lsn> {
-        self.lsn.clone()
-    }
-
-    fn is_incremental(&self) -> bool {
-        self.is_incremental
-    }
-
-    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 short_id(&self) -> String {
-        self.short_id.clone()
-    }
-
-    fn dump(&self, _verbose: bool, _ctx: &RequestContext) -> Result<()> {
-        todo!()
-    }
-}
-
-impl From<DeltaFileName> for LayerDescriptor {
-    fn from(value: DeltaFileName) -> Self {
-        let short_id = value.to_string();
-        LayerDescriptor {
-            key: value.key_range,
-            lsn: value.lsn_range,
-            is_incremental: true,
-            short_id,
+    impl From<PersistentLayerDesc> for LayerDescriptor {
+        fn from(base: PersistentLayerDesc) -> Self {
+            Self { base }
        }
    }
-}

-impl From<ImageFileName> for LayerDescriptor {
-    fn from(value: ImageFileName) -> Self {
-        let short_id = value.to_string();
-        let lsn = value.lsn_as_range();
-        LayerDescriptor {
-            key: value.key_range,
-            lsn,
-            is_incremental: false,
-            short_id,
+    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.
+        fn short_id(&self) -> String {
+            self.layer_desc().short_id()
        }
    }
-}

-impl From<LayerFileName> for LayerDescriptor {
-    fn from(value: LayerFileName) -> Self {
-        match value {
-            LayerFileName::Delta(d) => Self::from(d),
-            LayerFileName::Image(i) => Self::from(i),
+    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 {
+        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,
+                ),
+            }
+        }
+    }
+
+    impl From<ImageFileName> for LayerDescriptor {
+        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,
+                ),
+            }
+        }
+    }
+
+    impl From<LayerFileName> for LayerDescriptor {
+        fn from(value: LayerFileName) -> Self {
+            match value {
+                LayerFileName::Delta(d) => Self::from(d),
+                LayerFileName::Image(i) => Self::from(i),
+            }
        }
    }
 }
--- a/pageserver/src/tenant/storage_layer/delta_layer.rs
+++ b/pageserver/src/tenant/storage_layer/delta_layer.rs
@@ -37,6 +37,7 @@ use crate::virtual_file::VirtualFile;
 use crate::{walrecord, TEMP_FILE_SUFFIX};
 use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
 use anyhow::{bail, ensure, Context, Result};
+use once_cell::sync::OnceCell;
 use pageserver_api::models::{HistoricLayerInfo, LayerAccessKind};
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
@@ -46,7 +47,6 @@ use std::io::{Seek, SeekFrom};
 use std::ops::Range;
 use std::os::unix::fs::FileExt;
 use std::path::{Path, PathBuf};
-use std::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
 use tracing::*;

 use utils::{
@@ -184,7 +184,7 @@ pub struct DeltaLayer {

    access_stats: LayerAccessStats,

-    inner: RwLock<DeltaLayerInner>,
+    inner: OnceCell<DeltaLayerInner>,
 }

 impl std::fmt::Debug for DeltaLayer {
@@ -201,21 +201,17 @@ impl std::fmt::Debug for DeltaLayer {
 }

 pub struct DeltaLayerInner {
-    /// If false, the fields below have not been loaded into memory yet.
-    loaded: bool,
-
    // values copied from summary
    index_start_blk: u32,
    index_root_blk: u32,

-    /// Reader object for reading blocks from the file. (None if not loaded yet)
-    file: Option<FileBlockReader<VirtualFile>>,
+    /// Reader object for reading blocks from the file.
+    file: FileBlockReader<VirtualFile>,
 }

 impl std::fmt::Debug for DeltaLayerInner {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("DeltaLayerInner")
-            .field("loaded", &self.loaded)
            .field("index_start_blk", &self.index_start_blk)
            .field("index_root_blk", &self.index_root_blk)
            .finish()
@@ -246,7 +242,7 @@ impl Layer for DeltaLayer {
            inner.index_start_blk, inner.index_root_blk
        );

-        let file = inner.file.as_ref().unwrap();
+        let file = &inner.file;
        let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
            inner.index_start_blk,
            inner.index_root_blk,
@@ -315,7 +311,7 @@ impl Layer for DeltaLayer {
            let inner = self.load(LayerAccessKind::GetValueReconstructData, ctx)?;

            // Scan the page versions backwards, starting from `lsn`.
-            let file = inner.file.as_ref().unwrap();
+            let file = &inner.file;
            let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
                inner.index_start_blk,
                inner.index_root_blk,
@@ -500,51 +496,22 @@ impl DeltaLayer {
    /// Open the underlying file and read the metadata into memory, if it's
    /// not loaded already.
    ///
-    fn load(
-        &self,
-        access_kind: LayerAccessKind,
-        ctx: &RequestContext,
-    ) -> Result<RwLockReadGuard<DeltaLayerInner>> {
+    fn load(&self, access_kind: LayerAccessKind, ctx: &RequestContext) -> Result<&DeltaLayerInner> {
        self.access_stats
            .record_access(access_kind, ctx.task_kind());
-        loop {
-            // Quick exit if already loaded
-            let inner = self.inner.read().unwrap();
-            if inner.loaded {
-                return Ok(inner);
-            }
-
-            // Need to open the file and load the metadata. Upgrade our lock to
-            // a write lock. (Or rather, release and re-lock in write mode.)
-            drop(inner);
-            let inner = self.inner.write().unwrap();
-            if !inner.loaded {
-                self.load_inner(inner).with_context(|| {
-                    format!("Failed to load delta layer {}", self.path().display())
-                })?;
-            } else {
-                // Another thread loaded it while we were not holding the lock.
-            }
-
-            // We now have the file open and loaded. There's no function to do
-            // that in the std library RwLock, so we have to release and re-lock
-            // in read mode. (To be precise, the lock guard was moved in the
-            // above call to `load_inner`, so it's already been released). And
-            // while we do that, another thread could unload again, so we have
-            // to re-check and retry if that happens.
-        }
+        // Quick exit if already loaded
+        self.inner
+            .get_or_try_init(|| self.load_inner())
+            .with_context(|| format!("Failed to load delta layer {}", self.path().display()))
    }

-    fn load_inner(&self, mut inner: RwLockWriteGuard<DeltaLayerInner>) -> Result<()> {
+    fn load_inner(&self) -> Result<DeltaLayerInner> {
        let path = self.path();

-        // Open the file if it's not open already.
-        if inner.file.is_none() {
-            let file = VirtualFile::open(&path)
-                .with_context(|| format!("Failed to open file '{}'", path.display()))?;
-            inner.file = Some(FileBlockReader::new(file));
-        }
-        let file = inner.file.as_mut().unwrap();
+        let file = VirtualFile::open(&path)
+            .with_context(|| format!("Failed to open file '{}'", path.display()))?;
+        let file = FileBlockReader::new(file);
+
        let summary_blk = file.read_blk(0)?;
        let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;

@@ -571,13 +538,13 @@ impl DeltaLayer {
            }
        }

-        inner.index_start_blk = actual_summary.index_start_blk;
-        inner.index_root_blk = actual_summary.index_root_blk;
-
        debug!("loaded from {}", &path.display());

-        inner.loaded = true;
-        Ok(())
+        Ok(DeltaLayerInner {
+            file,
+            index_start_blk: actual_summary.index_start_blk,
+            index_root_blk: actual_summary.index_root_blk,
+        })
    }

    /// Create a DeltaLayer struct representing an existing file on disk.
@@ -599,12 +566,7 @@ impl DeltaLayer {
                file_size,
            ),
            access_stats,
-            inner: RwLock::new(DeltaLayerInner {
-                loaded: false,
-                file: None,
-                index_start_blk: 0,
-                index_root_blk: 0,
-            }),
+            inner: once_cell::sync::OnceCell::new(),
        }
    }

@@ -631,12 +593,7 @@ impl DeltaLayer {
                metadata.len(),
            ),
            access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
-            inner: RwLock::new(DeltaLayerInner {
-                loaded: false,
-                file: None,
-                index_start_blk: 0,
-                index_root_blk: 0,
-            }),
+            inner: once_cell::sync::OnceCell::new(),
        })
    }

@@ -800,12 +757,7 @@ impl DeltaLayerWriterInner {
                metadata.len(),
            ),
            access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
-            inner: RwLock::new(DeltaLayerInner {
-                loaded: false,
-                file: None,
-                index_start_blk,
-                index_root_blk,
-            }),
+            inner: once_cell::sync::OnceCell::new(),
        };

        // fsync the file
@@ -940,13 +892,13 @@ struct DeltaValueIter<'a> {
    reader: BlockCursor<Adapter<'a>>,
 }

-struct Adapter<'a>(RwLockReadGuard<'a, DeltaLayerInner>);
+struct Adapter<'a>(&'a DeltaLayerInner);

 impl<'a> BlockReader for Adapter<'a> {
    type BlockLease = PageReadGuard<'static>;

    fn read_blk(&self, blknum: u32) -> Result<Self::BlockLease, std::io::Error> {
-        self.0.file.as_ref().unwrap().read_blk(blknum)
+        self.0.file.read_blk(blknum)
    }
 }

@@ -959,8 +911,8 @@ impl<'a> Iterator for DeltaValueIter<'a> {
 }

 impl<'a> DeltaValueIter<'a> {
-    fn new(inner: RwLockReadGuard<'a, DeltaLayerInner>) -> Result<Self> {
-        let file = inner.file.as_ref().unwrap();
+    fn new(inner: &'a DeltaLayerInner) -> Result<Self> {
+        let file = &inner.file;
        let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
            inner.index_start_blk,
            inner.index_root_blk,
@@ -1033,8 +985,8 @@ impl Iterator for DeltaKeyIter {
 }

 impl<'a> DeltaKeyIter {
-    fn new(inner: RwLockReadGuard<'a, DeltaLayerInner>) -> Result<Self> {
-        let file = inner.file.as_ref().unwrap();
+    fn new(inner: &'a DeltaLayerInner) -> Result<Self> {
+        let file = &inner.file;
        let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
            inner.index_start_blk,
            inner.index_root_blk,
@@ -1074,3 +1026,21 @@ impl<'a> DeltaKeyIter {
        Ok(iter)
    }
 }
+
+#[cfg(test)]
+mod test {
+    use super::DeltaKeyIter;
+    use super::DeltaLayer;
+    use super::DeltaValueIter;
+
+    // We will soon need the iters to be send in the compaction code.
+    // Cf https://github.com/neondatabase/neon/pull/4462#issuecomment-1587398883
+    // Cf https://github.com/neondatabase/neon/issues/4471
+    #[test]
+    fn is_send() {
+        fn assert_send<T: Send>() {}
+        assert_send::<DeltaLayer>();
+        assert_send::<DeltaValueIter>();
+        assert_send::<DeltaKeyIter>();
+    }
+}
--- a/pageserver/src/tenant/storage_layer/inmemory_layer.rs
+++ b/pageserver/src/tenant/storage_layer/inmemory_layer.rs
@@ -304,7 +304,7 @@ impl InMemoryLayer {
        Ok(())
    }

-    pub fn put_tombstone(&self, _key_range: Range<Key>, _lsn: Lsn) -> Result<()> {
+    pub async fn put_tombstone(&self, _key_range: Range<Key>, _lsn: Lsn) -> Result<()> {
        // TODO: Currently, we just leak the storage for any deleted keys

        Ok(())
--- a/pageserver/src/tenant/storage_layer/layer_desc.rs
+++ b/pageserver/src/tenant/storage_layer/layer_desc.rs
@@ -9,10 +9,12 @@ use crate::{context::RequestContext, repository::Key};

 use super::{DeltaFileName, ImageFileName, LayerFileName};

+use serde::{Deserialize, Serialize};
+
 /// A unique identifier of a persistent layer. This is different from `LayerDescriptor`, which is only used in the
 /// benchmarks. This struct contains all necessary information to find the image / delta layer. It also provides
 /// a unified way to generate layer information like file name.
-#[derive(Debug, PartialEq, Eq, Clone)]
+#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
 pub struct PersistentLayerDesc {
    pub tenant_id: TenantId,
    pub timeline_id: TimelineId,
@@ -50,6 +52,19 @@ impl PersistentLayerDesc {
        self.filename().file_name()
    }

+    #[cfg(test)]
+    pub fn new_test(key_range: Range<Key>) -> Self {
+        Self {
+            tenant_id: TenantId::generate(),
+            timeline_id: TimelineId::generate(),
+            key_range,
+            lsn_range: Lsn(0)..Lsn(1),
+            is_delta: false,
+            is_incremental: false,
+            file_size: 0,
+        }
+    }
+
    pub fn new_img(
        tenant_id: TenantId,
        timeline_id: TimelineId,
--- a/pageserver/src/tenant/storage_layer/remote_layer.rs
+++ b/pageserver/src/tenant/storage_layer/remote_layer.rs
@@ -218,15 +218,12 @@ impl RemoteLayer {
    }

    /// Create a Layer struct representing this layer, after it has been downloaded.
-    pub fn create_downloaded_layer<L>(
+    pub fn create_downloaded_layer(
        &self,
-        layer_map_lock_held_witness: &BatchedUpdates<'_, L>,
+        layer_map_lock_held_witness: &BatchedUpdates<'_>,
        conf: &'static PageServerConf,
        file_size: u64,
-    ) -> Arc<dyn PersistentLayer>
-    where
-        L: ?Sized + Layer,
-    {
+    ) -> Arc<dyn PersistentLayer> {
        if self.desc.is_delta {
            let fname = self.desc.delta_file_name();
            Arc::new(DeltaLayer::new(
--- 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
@@ -197,9 +197,11 @@ impl Timeline {
        // We don't want to hold the layer map lock during eviction.
        // So, we just need to deal with this.
        let candidates: Vec<Arc<dyn PersistentLayer>> = {
-            let layers = self.layers.read().unwrap();
+            let guard = self.layers.read().await;
+            let (layers, mapping) = &*guard;
            let mut candidates = Vec::new();
            for hist_layer in layers.iter_historic_layers() {
+                let hist_layer = mapping.get_from_desc(&hist_layer);
                if hist_layer.is_remote_layer() {
                    continue;
                }
--- a/pageserver/src/tenant/timeline/walreceiver/connection_manager.rs
+++ b/pageserver/src/tenant/timeline/walreceiver/connection_manager.rs
@@ -153,7 +153,7 @@ pub(super) async fn connection_manager_loop_step(
                            match new_state {
                                // we're already active as walreceiver, no need to reactivate
                                TimelineState::Active => continue,
-                                TimelineState::Broken | TimelineState::Stopping => {
+                                TimelineState::Broken { .. } | TimelineState::Stopping => {
                                    debug!("timeline entered terminal state {new_state:?}, stopping wal connection manager loop");
                                    return ControlFlow::Break(());
                                }
@@ -1321,10 +1321,11 @@ mod tests {

    const DUMMY_SAFEKEEPER_HOST: &str = "safekeeper_connstr";

-    async fn dummy_state(harness: &TenantHarness<'_>) -> ConnectionManagerState {
+    async fn dummy_state(harness: &TenantHarness) -> ConnectionManagerState {
        let (tenant, ctx) = harness.load().await;
        let timeline = tenant
-            .create_test_timeline(TIMELINE_ID, Lsn(0), crate::DEFAULT_PG_VERSION, &ctx)
+            .create_test_timeline(TIMELINE_ID, Lsn(0x8), crate::DEFAULT_PG_VERSION, &ctx)
+            .await
            .expect("Failed to create an empty timeline for dummy wal connection manager");

        ConnectionManagerState {
--- a/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs
+++ b/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs
@@ -304,12 +304,15 @@ pub(super) async fn handle_walreceiver_connection(
            }
        }

-        timeline.check_checkpoint_distance().with_context(|| {
-            format!(
-                "Failed to check checkpoint distance for timeline {}",
-                timeline.timeline_id
-            )
-        })?;
+        timeline
+            .check_checkpoint_distance()
+            .await
+            .with_context(|| {
+                format!(
+                    "Failed to check checkpoint distance for timeline {}",
+                    timeline.timeline_id
+                )
+            })?;

        if let Some(last_lsn) = status_update {
            let timeline_remote_consistent_lsn =
--- a/pageserver/src/tenant/upload_queue.rs
+++ b/pageserver/src/tenant/upload_queue.rs
@@ -76,6 +76,12 @@ pub(crate) struct UploadQueueInitialized {
    pub(crate) queued_operations: VecDeque<UploadOp>,
 }

+impl UploadQueueInitialized {
+    pub(super) fn no_pending_work(&self) -> bool {
+        self.inprogress_tasks.is_empty() && self.queued_operations.is_empty()
+    }
+}
+
 #[derive(Clone, Copy)]
 pub(super) enum SetDeletedFlagProgress {
    NotRunning,
@@ -84,9 +90,7 @@ pub(super) enum SetDeletedFlagProgress {
 }

 pub(super) struct UploadQueueStopped {
-    pub(super) latest_files: HashMap<LayerFileName, LayerFileMetadata>,
-    pub(super) last_uploaded_consistent_lsn: Lsn,
-    pub(super) latest_metadata: TimelineMetadata,
+    pub(super) upload_queue_for_deletion: UploadQueueInitialized,
    pub(super) deleted_at: SetDeletedFlagProgress,
 }

@@ -187,6 +191,15 @@ impl UploadQueue {
            UploadQueue::Initialized(x) => Ok(x),
        }
    }
+
+    pub(crate) fn stopped_mut(&mut self) -> anyhow::Result<&mut UploadQueueStopped> {
+        match self {
+            UploadQueue::Initialized(_) | UploadQueue::Uninitialized => {
+                anyhow::bail!("queue is in state {}", self.as_str())
+            }
+            UploadQueue::Stopped(stopped) => Ok(stopped),
+        }
+    }
 }

 /// An in-progress upload or delete task.
@@ -199,6 +212,13 @@ pub(crate) struct UploadTask {
    pub(crate) op: UploadOp,
 }

+#[derive(Debug)]
+pub(crate) struct Delete {
+    pub(crate) file_kind: RemoteOpFileKind,
+    pub(crate) layer_file_name: LayerFileName,
+    pub(crate) scheduled_from_timeline_delete: bool,
+}
+
 #[derive(Debug)]
 pub(crate) enum UploadOp {
    /// Upload a layer file
@@ -207,8 +227,8 @@ pub(crate) enum UploadOp {
    /// Upload the metadata file
    UploadMetadata(IndexPart, Lsn),

-    /// Delete a file.
-    Delete(RemoteOpFileKind, LayerFileName),
+    /// Delete a layer file
+    Delete(Delete),

    /// Barrier. When the barrier operation is reached,
    Barrier(tokio::sync::watch::Sender<()>),
@@ -226,7 +246,12 @@ impl std::fmt::Display for UploadOp {
                )
            }
            UploadOp::UploadMetadata(_, lsn) => write!(f, "UploadMetadata(lsn: {})", lsn),
-            UploadOp::Delete(_, path) => write!(f, "Delete({})", path.file_name()),
+            UploadOp::Delete(delete) => write!(
+                f,
+                "Delete(path: {}, scheduled_from_timeline_delete: {})",
+                delete.layer_file_name.file_name(),
+                delete.scheduled_from_timeline_delete
+            ),
            UploadOp::Barrier(_) => write!(f, "Barrier"),
        }
    }
--- a/pageserver/src/walingest.rs
+++ b/pageserver/src/walingest.rs
@@ -25,7 +25,7 @@ use postgres_ffi::v14::nonrelfile_utils::clogpage_precedes;
 use postgres_ffi::v14::nonrelfile_utils::slru_may_delete_clogsegment;
 use postgres_ffi::{fsm_logical_to_physical, page_is_new, page_set_lsn};

-use anyhow::Result;
+use anyhow::{Context, Result};
 use bytes::{Buf, Bytes, BytesMut};
 use tracing::*;

@@ -333,7 +333,7 @@ impl<'a> WalIngest<'a> {

        // Now that this record has been fully handled, including updating the
        // checkpoint data, let the repository know that it is up-to-date to this LSN
-        modification.commit()?;
+        modification.commit().await?;

        Ok(())
    }
@@ -1082,7 +1082,10 @@ impl<'a> WalIngest<'a> {
            .await?
        {
            // create it with 0 size initially, the logic below will extend it
-            modification.put_rel_creation(rel, 0, ctx).await?;
+            modification
+                .put_rel_creation(rel, 0, ctx)
+                .await
+                .context("Relation Error")?;
            0
        } else {
            self.timeline.get_rel_size(rel, last_lsn, true, ctx).await?
@@ -1171,7 +1174,6 @@ impl<'a> WalIngest<'a> {
 #[cfg(test)]
 mod tests {
    use super::*;
-    use crate::pgdatadir_mapping::create_test_timeline;
    use crate::tenant::harness::*;
    use crate::tenant::Timeline;
    use postgres_ffi::v14::xlog_utils::SIZEOF_CHECKPOINT;
@@ -1200,7 +1202,7 @@ mod tests {
        let mut m = tline.begin_modification(Lsn(0x10));
        m.put_checkpoint(ZERO_CHECKPOINT.clone())?;
        m.put_relmap_file(0, 111, Bytes::from(""), ctx).await?; // dummy relmapper file
-        m.commit()?;
+        m.commit().await?;
        let walingest = WalIngest::new(tline, Lsn(0x10), ctx).await?;

        Ok(walingest)
@@ -1209,7 +1211,9 @@ mod tests {
    #[tokio::test]
    async fn test_relsize() -> Result<()> {
        let (tenant, ctx) = TenantHarness::create("test_relsize")?.load().await;
-        let tline = create_test_timeline(&tenant, TIMELINE_ID, DEFAULT_PG_VERSION, &ctx)?;
+        let tline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
+            .await?;
        let mut walingest = init_walingest_test(&tline, &ctx).await?;

        let mut m = tline.begin_modification(Lsn(0x20));
@@ -1217,22 +1221,22 @@ mod tests {
        walingest
            .put_rel_page_image(&mut m, TESTREL_A, 0, TEST_IMG("foo blk 0 at 2"), &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;
        let mut m = tline.begin_modification(Lsn(0x30));
        walingest
            .put_rel_page_image(&mut m, TESTREL_A, 0, TEST_IMG("foo blk 0 at 3"), &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;
        let mut m = tline.begin_modification(Lsn(0x40));
        walingest
            .put_rel_page_image(&mut m, TESTREL_A, 1, TEST_IMG("foo blk 1 at 4"), &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;
        let mut m = tline.begin_modification(Lsn(0x50));
        walingest
            .put_rel_page_image(&mut m, TESTREL_A, 2, TEST_IMG("foo blk 2 at 5"), &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;

        assert_current_logical_size(&tline, Lsn(0x50));

@@ -1318,7 +1322,7 @@ mod tests {
        walingest
            .put_rel_truncation(&mut m, TESTREL_A, 2, &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;
        assert_current_logical_size(&tline, Lsn(0x60));

        // Check reported size and contents after truncation
@@ -1360,7 +1364,7 @@ mod tests {
        walingest
            .put_rel_truncation(&mut m, TESTREL_A, 0, &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;
        assert_eq!(
            tline
                .get_rel_size(TESTREL_A, Lsn(0x68), false, &ctx)
@@ -1373,7 +1377,7 @@ mod tests {
        walingest
            .put_rel_page_image(&mut m, TESTREL_A, 1, TEST_IMG("foo blk 1"), &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;
        assert_eq!(
            tline
                .get_rel_size(TESTREL_A, Lsn(0x70), false, &ctx)
@@ -1398,7 +1402,7 @@ mod tests {
        walingest
            .put_rel_page_image(&mut m, TESTREL_A, 1500, TEST_IMG("foo blk 1500"), &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;
        assert_eq!(
            tline
                .get_rel_size(TESTREL_A, Lsn(0x80), false, &ctx)
@@ -1428,14 +1432,16 @@ mod tests {
    #[tokio::test]
    async fn test_drop_extend() -> Result<()> {
        let (tenant, ctx) = TenantHarness::create("test_drop_extend")?.load().await;
-        let tline = create_test_timeline(&tenant, TIMELINE_ID, DEFAULT_PG_VERSION, &ctx)?;
+        let tline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
+            .await?;
        let mut walingest = init_walingest_test(&tline, &ctx).await?;

        let mut m = tline.begin_modification(Lsn(0x20));
        walingest
            .put_rel_page_image(&mut m, TESTREL_A, 0, TEST_IMG("foo blk 0 at 2"), &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;

        // Check that rel exists and size is correct
        assert_eq!(
@@ -1454,7 +1460,7 @@ mod tests {
        // Drop rel
        let mut m = tline.begin_modification(Lsn(0x30));
        walingest.put_rel_drop(&mut m, TESTREL_A, &ctx).await?;
-        m.commit()?;
+        m.commit().await?;

        // Check that rel is not visible anymore
        assert_eq!(
@@ -1472,7 +1478,7 @@ mod tests {
        walingest
            .put_rel_page_image(&mut m, TESTREL_A, 0, TEST_IMG("foo blk 0 at 4"), &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;

        // Check that rel exists and size is correct
        assert_eq!(
@@ -1497,7 +1503,9 @@ mod tests {
    #[tokio::test]
    async fn test_truncate_extend() -> Result<()> {
        let (tenant, ctx) = TenantHarness::create("test_truncate_extend")?.load().await;
-        let tline = create_test_timeline(&tenant, TIMELINE_ID, DEFAULT_PG_VERSION, &ctx)?;
+        let tline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
+            .await?;
        let mut walingest = init_walingest_test(&tline, &ctx).await?;

        // Create a 20 MB relation (the size is arbitrary)
@@ -1509,7 +1517,7 @@ mod tests {
                .put_rel_page_image(&mut m, TESTREL_A, blkno, TEST_IMG(&data), &ctx)
                .await?;
        }
-        m.commit()?;
+        m.commit().await?;

        // The relation was created at LSN 20, not visible at LSN 1 yet.
        assert_eq!(
@@ -1554,7 +1562,7 @@ mod tests {
        walingest
            .put_rel_truncation(&mut m, TESTREL_A, 1, &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;

        // Check reported size and contents after truncation
        assert_eq!(
@@ -1603,7 +1611,7 @@ mod tests {
                .put_rel_page_image(&mut m, TESTREL_A, blkno, TEST_IMG(&data), &ctx)
                .await?;
        }
-        m.commit()?;
+        m.commit().await?;

        assert_eq!(
            tline
@@ -1637,7 +1645,9 @@ mod tests {
    #[tokio::test]
    async fn test_large_rel() -> Result<()> {
        let (tenant, ctx) = TenantHarness::create("test_large_rel")?.load().await;
-        let tline = create_test_timeline(&tenant, TIMELINE_ID, DEFAULT_PG_VERSION, &ctx)?;
+        let tline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
+            .await?;
        let mut walingest = init_walingest_test(&tline, &ctx).await?;

        let mut lsn = 0x10;
@@ -1648,7 +1658,7 @@ mod tests {
            walingest
                .put_rel_page_image(&mut m, TESTREL_A, blknum as BlockNumber, img, &ctx)
                .await?;
-            m.commit()?;
+            m.commit().await?;
        }

        assert_current_logical_size(&tline, Lsn(lsn));
@@ -1664,7 +1674,7 @@ mod tests {
        walingest
            .put_rel_truncation(&mut m, TESTREL_A, RELSEG_SIZE, &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;
        assert_eq!(
            tline.get_rel_size(TESTREL_A, Lsn(lsn), false, &ctx).await?,
            RELSEG_SIZE
@@ -1677,7 +1687,7 @@ mod tests {
        walingest
            .put_rel_truncation(&mut m, TESTREL_A, RELSEG_SIZE - 1, &ctx)
            .await?;
-        m.commit()?;
+        m.commit().await?;
        assert_eq!(
            tline.get_rel_size(TESTREL_A, Lsn(lsn), false, &ctx).await?,
            RELSEG_SIZE - 1
@@ -1693,7 +1703,7 @@ mod tests {
            walingest
                .put_rel_truncation(&mut m, TESTREL_A, size as BlockNumber, &ctx)
                .await?;
-            m.commit()?;
+            m.commit().await?;
            assert_eq!(
                tline.get_rel_size(TESTREL_A, Lsn(lsn), false, &ctx).await?,
                size as BlockNumber
--- a/pgxn/neon/Makefile
+++ b/pgxn/neon/Makefile
@@ -4,6 +4,7 @@
 MODULE_big = neon
 OBJS = \
 	$(WIN32RES) \
+	extension_server.o \
 	file_cache.o \
 	libpagestore.o \
 	libpqwalproposer.o \
--- a/pgxn/neon/control_plane_connector.c
+++ b/pgxn/neon/control_plane_connector.c
@@ -32,6 +32,7 @@
 #include "port.h"
 #include <curl/curl.h>
 #include "utils/jsonb.h"
+#include "libpq/crypt.h"

 static ProcessUtility_hook_type PreviousProcessUtilityHook = NULL;

@@ -161,7 +162,22 @@ ConstructDeltaMessage()
 			PushKeyValue(&state, "name", entry->name);
 			if (entry->password)
 			{
+#if PG_MAJORVERSION_NUM == 14
+				char	   *logdetail;
+#else
+				const char *logdetail;
+#endif
 				PushKeyValue(&state, "password", (char *) entry->password);
+				char	   *encrypted_password = get_role_password(entry->name, &logdetail);
+
+				if (encrypted_password)
+				{
+					PushKeyValue(&state, "encrypted_password", encrypted_password);
+				}
+				else
+				{
+					elog(ERROR, "Failed to get encrypted password: %s", logdetail);
+				}
 			}
 			if (entry->old_name[0] != '\0')
 			{
--- a/pgxn/neon/extension_server.c
+++ b/pgxn/neon/extension_server.c
@@ -0,0 +1,104 @@
+
+/*-------------------------------------------------------------------------
+ *
+ * extension_server.c
+ *	  Request compute_ctl to download extension files.
+ *
+ * IDENTIFICATION
+ *	 contrib/neon/extension_server.c
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "postgres.h"
+#include "tcop/pquery.h"
+#include "tcop/utility.h"
+#include "access/xact.h"
+#include "utils/hsearch.h"
+#include "utils/memutils.h"
+#include "commands/defrem.h"
+#include "miscadmin.h"
+#include "utils/acl.h"
+#include "fmgr.h"
+#include "utils/guc.h"
+#include "port.h"
+#include "fmgr.h"
+
+#include <curl/curl.h>
+
+static int extension_server_port = 0;
+
+static download_extension_file_hook_type prev_download_extension_file_hook = NULL;
+
+// to download all SQL (and data) files for an extension:
+// curl -X POST http://localhost:8080/extension_server/postgis
+// it covers two possible extension files layouts:
+// 1. extension_name--version--platform.sql
+// 2. extension_name/extension_name--version.sql
+//    extension_name/extra_files.csv
+//
+// to download specific library file:
+// curl -X POST http://localhost:8080/extension_server/postgis-3.so?is_library=true
+static bool
+neon_download_extension_file_http(const char *filename, bool is_library)
+{
+    CURL *curl;
+    CURLcode res;
+    char *compute_ctl_url;
+    char *postdata;
+    bool ret = false;
+
+    if ((curl = curl_easy_init()) == NULL)
+    {
+        elog(ERROR, "Failed to initialize curl handle");
+    }
+
+    compute_ctl_url = psprintf("http://localhost:%d/extension_server/%s%s",
+                               extension_server_port, filename, is_library ? "?is_library=true" : "");
+
+    elog(LOG, "Sending request to compute_ctl: %s", compute_ctl_url);
+
+    curl_easy_setopt(curl, CURLOPT_CUSTOMREQUEST, "POST");
+    curl_easy_setopt(curl, CURLOPT_URL, compute_ctl_url);
+    // NOTE: 15L may be insufficient time for large extensions like postgis
+    curl_easy_setopt(curl, CURLOPT_TIMEOUT, 15L /* seconds */);
+
+    if (curl)
+    {
+        /* Perform the request, res will get the return code */
+        res = curl_easy_perform(curl);
+        /* Check for errors */
+        if (res == CURLE_OK)
+        {
+            ret = true;
+        }
+        else
+        {
+            // Don't error here because postgres will try to find the file
+            // and will fail with some proper error message if it's not found.
+            elog(WARNING, "neon_download_extension_file_http failed: %s\n", curl_easy_strerror(res));
+        }
+
+        /* always cleanup */
+        curl_easy_cleanup(curl);
+    }
+
+    return ret;
+}
+
+void pg_init_extension_server()
+{
+    // Port to connect to compute_ctl on localhost
+    // to request extension files.
+    DefineCustomIntVariable("neon.extension_server_port",
+                            "connection string to the compute_ctl",
+                            NULL,
+                            &extension_server_port,
+                            0, 0, INT_MAX,
+                            PGC_POSTMASTER,
+                            0, /* no flags required */
+                            NULL, NULL, NULL);
+
+    // set download_extension_file_hook
+    prev_download_extension_file_hook = download_extension_file_hook;
+    download_extension_file_hook = neon_download_extension_file_http;
+}
--- a/pgxn/neon/extension_server.h
+++ b/pgxn/neon/extension_server.h
@@ -0,0 +1 @@
+
--- a/pgxn/neon/file_cache.c
+++ b/pgxn/neon/file_cache.c
@@ -190,7 +190,7 @@ lfc_change_limit_hook(int newval, void *extra)
 		hash_search(lfc_hash, &victim->key, HASH_REMOVE, NULL);
 		lfc_ctl->used -= 1;
 	}
-	elog(LOG, "set local file cache limit to %d", new_size);
+	elog(DEBUG1, "set local file cache limit to %d", new_size);
 	LWLockRelease(lfc_lock);
 }

--- a/pgxn/neon/neon.c
+++ b/pgxn/neon/neon.c
@@ -35,8 +35,11 @@ _PG_init(void)
 {
 	pg_init_libpagestore();
 	pg_init_walproposer();
+
 	InitControlPlaneConnector();

+	pg_init_extension_server();
+
        // Important: This must happen after other parts of the extension
        // are loaded, otherwise any settings to GUCs that were set before
        // the extension was loaded will be removed.
--- a/pgxn/neon/neon.h
+++ b/pgxn/neon/neon.h
@@ -21,6 +21,8 @@ extern char *neon_tenant;
 extern void pg_init_libpagestore(void);
 extern void pg_init_walproposer(void);

+extern void pg_init_extension_server(void);
+
 /*
 * Returns true if we shouldn't do REDO on that block in record indicated by
 * block_id; false otherwise.
--- a/pgxn/neon/pagestore_smgr.c
+++ b/pgxn/neon/pagestore_smgr.c
@@ -2675,7 +2675,6 @@ bool
 neon_redo_read_buffer_filter(XLogReaderState *record, uint8 block_id)
 {
 	XLogRecPtr	end_recptr = record->EndRecPtr;
-	XLogRecPtr	prev_end_recptr = record->ReadRecPtr - 1;
 	RelFileNode	rnode;
 	ForkNumber	forknum;
 	BlockNumber	blkno;
@@ -2719,16 +2718,15 @@ neon_redo_read_buffer_filter(XLogReaderState *record, uint8 block_id)

 	no_redo_needed = buffer < 0;

-	/* we don't have the buffer in memory, update lwLsn past this record */
+	/* In both cases st lwlsn past this WAL record */
+	SetLastWrittenLSNForBlock(end_recptr, rnode, forknum, blkno);
+
+	/* we don't have the buffer in memory, update lwLsn past this record,
+	 * also evict page fro file cache
+	 */
 	if (no_redo_needed)
-	{
-		SetLastWrittenLSNForBlock(end_recptr, rnode, forknum, blkno);
 		lfc_evict(rnode, forknum, blkno);
-	}
-	else
-	{
-		SetLastWrittenLSNForBlock(prev_end_recptr, rnode, forknum, blkno);
-	}
+

 	LWLockRelease(partitionLock);

@@ -2736,7 +2734,10 @@ neon_redo_read_buffer_filter(XLogReaderState *record, uint8 block_id)
 	if (get_cached_relsize(rnode, forknum, &relsize))
 	{
 		if (relsize < blkno + 1)
+		{
 			update_cached_relsize(rnode, forknum, blkno + 1);
+			SetLastWrittenLSNForRelation(end_recptr, rnode, forknum);
+		}
 	}
 	else
 	{
@@ -2768,6 +2769,7 @@ neon_redo_read_buffer_filter(XLogReaderState *record, uint8 block_id)
 		Assert(nbresponse->n_blocks > blkno);

 		set_cached_relsize(rnode, forknum, nbresponse->n_blocks);
+		SetLastWrittenLSNForRelation(end_recptr, rnode, forknum);

 		elog(SmgrTrace, "Set length to %d", nbresponse->n_blocks);
 	}
--- a/pgxn/neon/walproposer.c
+++ b/pgxn/neon/walproposer.c
@@ -257,7 +257,7 @@ nwp_register_gucs(void)
 							"Walproposer reconnects to offline safekeepers once in this interval.",
 							NULL,
 							&wal_acceptor_reconnect_timeout,
-							5000, 0, INT_MAX,	/* default, min, max */
+							1000, 0, INT_MAX,	/* default, min, max */
 							PGC_SIGHUP, /* context */
 							GUC_UNIT_MS,	/* flags */
 							NULL, NULL, NULL);
--- a/poetry.lock
+++ b/poetry.lock
@@ -1,4 +1,4 @@
-# This file is automatically @generated by Poetry 1.4.2 and should not be changed by hand.
+# This file is automatically @generated by Poetry and should not be changed by hand.

 [[package]]
 name = "aiohttp"
@@ -855,35 +855,31 @@ files = [

 [[package]]
 name = "cryptography"
-version = "39.0.1"
+version = "41.0.0"
 description = "cryptography is a package which provides cryptographic recipes and primitives to Python developers."
 category = "main"
 optional = false
-python-versions = ">=3.6"
+python-versions = ">=3.7"
 files = [
-    {file = "cryptography-39.0.1-cp36-abi3-macosx_10_12_universal2.whl", hash = "sha256:6687ef6d0a6497e2b58e7c5b852b53f62142cfa7cd1555795758934da363a965"},
-    {file = "cryptography-39.0.1-cp36-abi3-macosx_10_12_x86_64.whl", hash = "sha256:706843b48f9a3f9b9911979761c91541e3d90db1ca905fd63fee540a217698bc"},
-    {file = "cryptography-39.0.1-cp36-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.manylinux_2_24_aarch64.whl", hash = "sha256:5d2d8b87a490bfcd407ed9d49093793d0f75198a35e6eb1a923ce1ee86c62b41"},
-    {file = "cryptography-39.0.1-cp36-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:83e17b26de248c33f3acffb922748151d71827d6021d98c70e6c1a25ddd78505"},
-    {file = "cryptography-39.0.1-cp36-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e124352fd3db36a9d4a21c1aa27fd5d051e621845cb87fb851c08f4f75ce8be6"},
-    {file = "cryptography-39.0.1-cp36-abi3-manylinux_2_24_x86_64.whl", hash = "sha256:5aa67414fcdfa22cf052e640cb5ddc461924a045cacf325cd164e65312d99502"},
-    {file = "cryptography-39.0.1-cp36-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:35f7c7d015d474f4011e859e93e789c87d21f6f4880ebdc29896a60403328f1f"},
-    {file = "cryptography-39.0.1-cp36-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:f24077a3b5298a5a06a8e0536e3ea9ec60e4c7ac486755e5fb6e6ea9b3500106"},
-    {file = "cryptography-39.0.1-cp36-abi3-musllinux_1_1_aarch64.whl", hash = "sha256:f0c64d1bd842ca2633e74a1a28033d139368ad959872533b1bab8c80e8240a0c"},
-    {file = "cryptography-39.0.1-cp36-abi3-musllinux_1_1_x86_64.whl", hash = "sha256:0f8da300b5c8af9f98111ffd512910bc792b4c77392a9523624680f7956a99d4"},
-    {file = "cryptography-39.0.1-cp36-abi3-win32.whl", hash = "sha256:fe913f20024eb2cb2f323e42a64bdf2911bb9738a15dba7d3cce48151034e3a8"},
-    {file = "cryptography-39.0.1-cp36-abi3-win_amd64.whl", hash = "sha256:ced4e447ae29ca194449a3f1ce132ded8fcab06971ef5f618605aacaa612beac"},
-    {file = "cryptography-39.0.1-pp38-pypy38_pp73-macosx_10_12_x86_64.whl", hash = "sha256:807ce09d4434881ca3a7594733669bd834f5b2c6d5c7e36f8c00f691887042ad"},
-    {file = "cryptography-39.0.1-pp38-pypy38_pp73-manylinux_2_28_aarch64.whl", hash = "sha256:c5caeb8188c24888c90b5108a441c106f7faa4c4c075a2bcae438c6e8ca73cef"},
-    {file = "cryptography-39.0.1-pp38-pypy38_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:4789d1e3e257965e960232345002262ede4d094d1a19f4d3b52e48d4d8f3b885"},
-    {file = "cryptography-39.0.1-pp38-pypy38_pp73-win_amd64.whl", hash = "sha256:96f1157a7c08b5b189b16b47bc9db2332269d6680a196341bf30046330d15388"},
-    {file = "cryptography-39.0.1-pp39-pypy39_pp73-macosx_10_12_x86_64.whl", hash = "sha256:e422abdec8b5fa8462aa016786680720d78bdce7a30c652b7fadf83a4ba35336"},
-    {file = "cryptography-39.0.1-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.manylinux_2_24_aarch64.whl", hash = "sha256:b0afd054cd42f3d213bf82c629efb1ee5f22eba35bf0eec88ea9ea7304f511a2"},
-    {file = "cryptography-39.0.1-pp39-pypy39_pp73-manylinux_2_24_x86_64.whl", hash = "sha256:6f8ba7f0328b79f08bdacc3e4e66fb4d7aab0c3584e0bd41328dce5262e26b2e"},
-    {file = "cryptography-39.0.1-pp39-pypy39_pp73-manylinux_2_28_aarch64.whl", hash = "sha256:ef8b72fa70b348724ff1218267e7f7375b8de4e8194d1636ee60510aae104cd0"},
-    {file = "cryptography-39.0.1-pp39-pypy39_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:aec5a6c9864be7df2240c382740fcf3b96928c46604eaa7f3091f58b878c0bb6"},
-    {file = "cryptography-39.0.1-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:fdd188c8a6ef8769f148f88f859884507b954cc64db6b52f66ef199bb9ad660a"},
-    {file = "cryptography-39.0.1.tar.gz", hash = "sha256:d1f6198ee6d9148405e49887803907fe8962a23e6c6f83ea7d98f1c0de375695"},
+    {file = "cryptography-41.0.0-cp37-abi3-macosx_10_12_universal2.whl", hash = "sha256:3c5ef25d060c80d6d9f7f9892e1d41bb1c79b78ce74805b8cb4aa373cb7d5ec8"},
+    {file = "cryptography-41.0.0-cp37-abi3-macosx_10_12_x86_64.whl", hash = "sha256:8362565b3835ceacf4dc8f3b56471a2289cf51ac80946f9087e66dc283a810e0"},
+    {file = "cryptography-41.0.0-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3680248309d340fda9611498a5319b0193a8dbdb73586a1acf8109d06f25b92d"},
+    {file = "cryptography-41.0.0-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:84a165379cb9d411d58ed739e4af3396e544eac190805a54ba2e0322feb55c46"},
+    {file = "cryptography-41.0.0-cp37-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:4ab14d567f7bbe7f1cdff1c53d5324ed4d3fc8bd17c481b395db224fb405c237"},
+    {file = "cryptography-41.0.0-cp37-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:9f65e842cb02550fac96536edb1d17f24c0a338fd84eaf582be25926e993dde4"},
+    {file = "cryptography-41.0.0-cp37-abi3-musllinux_1_1_aarch64.whl", hash = "sha256:b7f2f5c525a642cecad24ee8670443ba27ac1fab81bba4cc24c7b6b41f2d0c75"},
+    {file = "cryptography-41.0.0-cp37-abi3-musllinux_1_1_x86_64.whl", hash = "sha256:7d92f0248d38faa411d17f4107fc0bce0c42cae0b0ba5415505df72d751bf62d"},
+    {file = "cryptography-41.0.0-cp37-abi3-win32.whl", hash = "sha256:34d405ea69a8b34566ba3dfb0521379b210ea5d560fafedf9f800a9a94a41928"},
+    {file = "cryptography-41.0.0-cp37-abi3-win_amd64.whl", hash = "sha256:344c6de9f8bda3c425b3a41b319522ba3208551b70c2ae00099c205f0d9fd3be"},
+    {file = "cryptography-41.0.0-pp38-pypy38_pp73-macosx_10_12_x86_64.whl", hash = "sha256:88ff107f211ea696455ea8d911389f6d2b276aabf3231bf72c8853d22db755c5"},
+    {file = "cryptography-41.0.0-pp38-pypy38_pp73-manylinux_2_28_aarch64.whl", hash = "sha256:b846d59a8d5a9ba87e2c3d757ca019fa576793e8758174d3868aecb88d6fc8eb"},
+    {file = "cryptography-41.0.0-pp38-pypy38_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:f5d0bf9b252f30a31664b6f64432b4730bb7038339bd18b1fafe129cfc2be9be"},
+    {file = "cryptography-41.0.0-pp38-pypy38_pp73-win_amd64.whl", hash = "sha256:5c1f7293c31ebc72163a9a0df246f890d65f66b4a40d9ec80081969ba8c78cc9"},
+    {file = "cryptography-41.0.0-pp39-pypy39_pp73-macosx_10_12_x86_64.whl", hash = "sha256:bf8fc66012ca857d62f6a347007e166ed59c0bc150cefa49f28376ebe7d992a2"},
+    {file = "cryptography-41.0.0-pp39-pypy39_pp73-manylinux_2_28_aarch64.whl", hash = "sha256:a4fc68d1c5b951cfb72dfd54702afdbbf0fb7acdc9b7dc4301bbf2225a27714d"},
+    {file = "cryptography-41.0.0-pp39-pypy39_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:14754bcdae909d66ff24b7b5f166d69340ccc6cb15731670435efd5719294895"},
+    {file = "cryptography-41.0.0-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:0ddaee209d1cf1f180f1efa338a68c4621154de0afaef92b89486f5f96047c55"},
+    {file = "cryptography-41.0.0.tar.gz", hash = "sha256:6b71f64beeea341c9b4f963b48ee3b62d62d57ba93eb120e1196b31dc1025e78"},
 ]

 [package.dependencies]
@@ -892,12 +888,12 @@ cffi = ">=1.12"
 [package.extras]
 docs = ["sphinx (>=5.3.0)", "sphinx-rtd-theme (>=1.1.1)"]
 docstest = ["pyenchant (>=1.6.11)", "sphinxcontrib-spelling (>=4.0.1)", "twine (>=1.12.0)"]
-pep8test = ["black", "check-manifest", "mypy", "ruff", "types-pytz", "types-requests"]
-sdist = ["setuptools-rust (>=0.11.4)"]
+nox = ["nox"]
+pep8test = ["black", "check-sdist", "mypy", "ruff"]
+sdist = ["build"]
 ssh = ["bcrypt (>=3.1.5)"]
-test = ["hypothesis (>=1.11.4,!=3.79.2)", "iso8601", "pretend", "pytest (>=6.2.0)", "pytest-benchmark", "pytest-cov", "pytest-shard (>=0.1.2)", "pytest-subtests", "pytest-xdist", "pytz"]
+test = ["pretend", "pytest (>=6.2.0)", "pytest-benchmark", "pytest-cov", "pytest-xdist"]
 test-randomorder = ["pytest-randomly"]
-tox = ["tox"]

 [[package]]
 name = "docker"
@@ -1658,71 +1654,74 @@ test = ["enum34", "ipaddress", "mock", "pywin32", "wmi"]

 [[package]]
 name = "psycopg2-binary"
-version = "2.9.3"
+version = "2.9.6"
 description = "psycopg2 - Python-PostgreSQL Database Adapter"
 category = "main"
 optional = false
 python-versions = ">=3.6"
 files = [
-    {file = "psycopg2-binary-2.9.3.tar.gz", hash = "sha256:761df5313dc15da1502b21453642d7599d26be88bff659382f8f9747c7ebea4e"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-macosx_10_14_x86_64.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl", hash = "sha256:539b28661b71da7c0e428692438efbcd048ca21ea81af618d845e06ebfd29478"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2f2534ab7dc7e776a263b463a16e189eb30e85ec9bbe1bff9e78dae802608932"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6e82d38390a03da28c7985b394ec3f56873174e2c88130e6966cb1c946508e65"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:57804fc02ca3ce0dbfbef35c4b3a4a774da66d66ea20f4bda601294ad2ea6092"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-manylinux_2_24_aarch64.whl", hash = "sha256:083a55275f09a62b8ca4902dd11f4b33075b743cf0d360419e2051a8a5d5ff76"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-manylinux_2_24_ppc64le.whl", hash = "sha256:0a29729145aaaf1ad8bafe663131890e2111f13416b60e460dae0a96af5905c9"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:3a79d622f5206d695d7824cbf609a4f5b88ea6d6dab5f7c147fc6d333a8787e4"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:090f3348c0ab2cceb6dfbe6bf721ef61262ddf518cd6cc6ecc7d334996d64efa"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:a9e1f75f96ea388fbcef36c70640c4efbe4650658f3d6a2967b4cc70e907352e"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:c3ae8e75eb7160851e59adc77b3a19a976e50622e44fd4fd47b8b18208189d42"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-win32.whl", hash = "sha256:7b1e9b80afca7b7a386ef087db614faebbf8839b7f4db5eb107d0f1a53225029"},
-    {file = "psycopg2_binary-2.9.3-cp310-cp310-win_amd64.whl", hash = "sha256:8b344adbb9a862de0c635f4f0425b7958bf5a4b927c8594e6e8d261775796d53"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-macosx_10_14_x86_64.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl", hash = "sha256:e847774f8ffd5b398a75bc1c18fbb56564cda3d629fe68fd81971fece2d3c67e"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:68641a34023d306be959101b345732360fc2ea4938982309b786f7be1b43a4a1"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3303f8807f342641851578ee7ed1f3efc9802d00a6f83c101d21c608cb864460"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-manylinux_2_24_aarch64.whl", hash = "sha256:e3699852e22aa68c10de06524a3721ade969abf382da95884e6a10ff798f9281"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-manylinux_2_24_ppc64le.whl", hash = "sha256:526ea0378246d9b080148f2d6681229f4b5964543c170dd10bf4faaab6e0d27f"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-musllinux_1_1_aarch64.whl", hash = "sha256:b1c8068513f5b158cf7e29c43a77eb34b407db29aca749d3eb9293ee0d3103ca"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-musllinux_1_1_i686.whl", hash = "sha256:15803fa813ea05bef089fa78835118b5434204f3a17cb9f1e5dbfd0b9deea5af"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-musllinux_1_1_ppc64le.whl", hash = "sha256:152f09f57417b831418304c7f30d727dc83a12761627bb826951692cc6491e57"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-musllinux_1_1_x86_64.whl", hash = "sha256:404224e5fef3b193f892abdbf8961ce20e0b6642886cfe1fe1923f41aaa75c9d"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-win32.whl", hash = "sha256:1f6b813106a3abdf7b03640d36e24669234120c72e91d5cbaeb87c5f7c36c65b"},
-    {file = "psycopg2_binary-2.9.3-cp36-cp36m-win_amd64.whl", hash = "sha256:2d872e3c9d5d075a2e104540965a1cf898b52274a5923936e5bfddb58c59c7c2"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-macosx_10_14_x86_64.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl", hash = "sha256:10bb90fb4d523a2aa67773d4ff2b833ec00857f5912bafcfd5f5414e45280fb1"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:874a52ecab70af13e899f7847b3e074eeb16ebac5615665db33bce8a1009cf33"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a29b3ca4ec9defec6d42bf5feb36bb5817ba3c0230dd83b4edf4bf02684cd0ae"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-manylinux_2_24_aarch64.whl", hash = "sha256:12b11322ea00ad8db8c46f18b7dfc47ae215e4df55b46c67a94b4effbaec7094"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-manylinux_2_24_ppc64le.whl", hash = "sha256:53293533fcbb94c202b7c800a12c873cfe24599656b341f56e71dd2b557be063"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:c381bda330ddf2fccbafab789d83ebc6c53db126e4383e73794c74eedce855ef"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:9d29409b625a143649d03d0fd7b57e4b92e0ecad9726ba682244b73be91d2fdb"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:183a517a3a63503f70f808b58bfbf962f23d73b6dccddae5aa56152ef2bcb232"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:15c4e4cfa45f5a60599d9cec5f46cd7b1b29d86a6390ec23e8eebaae84e64554"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-win32.whl", hash = "sha256:adf20d9a67e0b6393eac162eb81fb10bc9130a80540f4df7e7355c2dd4af9fba"},
-    {file = "psycopg2_binary-2.9.3-cp37-cp37m-win_amd64.whl", hash = "sha256:2f9ffd643bc7349eeb664eba8864d9e01f057880f510e4681ba40a6532f93c71"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-macosx_10_14_x86_64.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl", hash = "sha256:def68d7c21984b0f8218e8a15d514f714d96904265164f75f8d3a70f9c295667"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:e6aa71ae45f952a2205377773e76f4e3f27951df38e69a4c95440c779e013560"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:dffc08ca91c9ac09008870c9eb77b00a46b3378719584059c034b8945e26b272"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:280b0bb5cbfe8039205c7981cceb006156a675362a00fe29b16fbc264e242834"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-manylinux_2_24_aarch64.whl", hash = "sha256:af9813db73395fb1fc211bac696faea4ca9ef53f32dc0cfa27e4e7cf766dcf24"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-manylinux_2_24_ppc64le.whl", hash = "sha256:63638d875be8c2784cfc952c9ac34e2b50e43f9f0a0660b65e2a87d656b3116c"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:ffb7a888a047696e7f8240d649b43fb3644f14f0ee229077e7f6b9f9081635bd"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:0c9d5450c566c80c396b7402895c4369a410cab5a82707b11aee1e624da7d004"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:d1c1b569ecafe3a69380a94e6ae09a4789bbb23666f3d3a08d06bbd2451f5ef1"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:8fc53f9af09426a61db9ba357865c77f26076d48669f2e1bb24d85a22fb52307"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-win32.whl", hash = "sha256:6472a178e291b59e7f16ab49ec8b4f3bdada0a879c68d3817ff0963e722a82ce"},
-    {file = "psycopg2_binary-2.9.3-cp38-cp38-win_amd64.whl", hash = "sha256:35168209c9d51b145e459e05c31a9eaeffa9a6b0fd61689b48e07464ffd1a83e"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-macosx_10_14_x86_64.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl", hash = "sha256:47133f3f872faf28c1e87d4357220e809dfd3fa7c64295a4a148bcd1e6e34ec9"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:b3a24a1982ae56461cc24f6680604fffa2c1b818e9dc55680da038792e004d18"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:91920527dea30175cc02a1099f331aa8c1ba39bf8b7762b7b56cbf54bc5cce42"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:887dd9aac71765ac0d0bac1d0d4b4f2c99d5f5c1382d8b770404f0f3d0ce8a39"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-manylinux_2_24_aarch64.whl", hash = "sha256:1f14c8b0942714eb3c74e1e71700cbbcb415acbc311c730370e70c578a44a25c"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-manylinux_2_24_ppc64le.whl", hash = "sha256:7af0dd86ddb2f8af5da57a976d27cd2cd15510518d582b478fbb2292428710b4"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:93cd1967a18aa0edd4b95b1dfd554cf15af657cb606280996d393dadc88c3c35"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:bda845b664bb6c91446ca9609fc69f7db6c334ec5e4adc87571c34e4f47b7ddb"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:01310cf4cf26db9aea5158c217caa92d291f0500051a6469ac52166e1a16f5b7"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:99485cab9ba0fa9b84f1f9e1fef106f44a46ef6afdeec8885e0b88d0772b49e8"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-win32.whl", hash = "sha256:46f0e0a6b5fa5851bbd9ab1bc805eef362d3a230fbdfbc209f4a236d0a7a990d"},
-    {file = "psycopg2_binary-2.9.3-cp39-cp39-win_amd64.whl", hash = "sha256:accfe7e982411da3178ec690baaceaad3c278652998b2c45828aaac66cd8285f"},
+    {file = "psycopg2-binary-2.9.6.tar.gz", hash = "sha256:1f64dcfb8f6e0c014c7f55e51c9759f024f70ea572fbdef123f85318c297947c"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:d26e0342183c762de3276cca7a530d574d4e25121ca7d6e4a98e4f05cb8e4df7"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:c48d8f2db17f27d41fb0e2ecd703ea41984ee19362cbce52c097963b3a1b4365"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ffe9dc0a884a8848075e576c1de0290d85a533a9f6e9c4e564f19adf8f6e54a7"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:8a76e027f87753f9bd1ab5f7c9cb8c7628d1077ef927f5e2446477153a602f2c"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:6460c7a99fc939b849431f1e73e013d54aa54293f30f1109019c56a0b2b2ec2f"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ae102a98c547ee2288637af07393dd33f440c25e5cd79556b04e3fca13325e5f"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:9972aad21f965599ed0106f65334230ce826e5ae69fda7cbd688d24fa922415e"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:7a40c00dbe17c0af5bdd55aafd6ff6679f94a9be9513a4c7e071baf3d7d22a70"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:cacbdc5839bdff804dfebc058fe25684cae322987f7a38b0168bc1b2df703fb1"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:7f0438fa20fb6c7e202863e0d5ab02c246d35efb1d164e052f2f3bfe2b152bd0"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-win32.whl", hash = "sha256:b6c8288bb8a84b47e07013bb4850f50538aa913d487579e1921724631d02ea1b"},
+    {file = "psycopg2_binary-2.9.6-cp310-cp310-win_amd64.whl", hash = "sha256:61b047a0537bbc3afae10f134dc6393823882eb263088c271331602b672e52e9"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:964b4dfb7c1c1965ac4c1978b0f755cc4bd698e8aa2b7667c575fb5f04ebe06b"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:afe64e9b8ea66866a771996f6ff14447e8082ea26e675a295ad3bdbffdd72afb"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:15e2ee79e7cf29582ef770de7dab3d286431b01c3bb598f8e05e09601b890081"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:dfa74c903a3c1f0d9b1c7e7b53ed2d929a4910e272add6700c38f365a6002820"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:b83456c2d4979e08ff56180a76429263ea254c3f6552cd14ada95cff1dec9bb8"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0645376d399bfd64da57148694d78e1f431b1e1ee1054872a5713125681cf1be"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:e99e34c82309dd78959ba3c1590975b5d3c862d6f279f843d47d26ff89d7d7e1"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:4ea29fc3ad9d91162c52b578f211ff1c931d8a38e1f58e684c45aa470adf19e2"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-musllinux_1_1_ppc64le.whl", hash = "sha256:4ac30da8b4f57187dbf449294d23b808f8f53cad6b1fc3623fa8a6c11d176dd0"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:e78e6e2a00c223e164c417628572a90093c031ed724492c763721c2e0bc2a8df"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-win32.whl", hash = "sha256:1876843d8e31c89c399e31b97d4b9725a3575bb9c2af92038464231ec40f9edb"},
+    {file = "psycopg2_binary-2.9.6-cp311-cp311-win_amd64.whl", hash = "sha256:b4b24f75d16a89cc6b4cdff0eb6a910a966ecd476d1e73f7ce5985ff1328e9a6"},
+    {file = "psycopg2_binary-2.9.6-cp36-cp36m-win32.whl", hash = "sha256:498807b927ca2510baea1b05cc91d7da4718a0f53cb766c154c417a39f1820a0"},
+    {file = "psycopg2_binary-2.9.6-cp36-cp36m-win_amd64.whl", hash = "sha256:0d236c2825fa656a2d98bbb0e52370a2e852e5a0ec45fc4f402977313329174d"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:34b9ccdf210cbbb1303c7c4db2905fa0319391bd5904d32689e6dd5c963d2ea8"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:84d2222e61f313c4848ff05353653bf5f5cf6ce34df540e4274516880d9c3763"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:30637a20623e2a2eacc420059be11527f4458ef54352d870b8181a4c3020ae6b"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8122cfc7cae0da9a3077216528b8bb3629c43b25053284cc868744bfe71eb141"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:38601cbbfe600362c43714482f43b7c110b20cb0f8172422c616b09b85a750c5"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:c7e62ab8b332147a7593a385d4f368874d5fe4ad4e341770d4983442d89603e3"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:2ab652e729ff4ad76d400df2624d223d6e265ef81bb8aa17fbd63607878ecbee"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:c83a74b68270028dc8ee74d38ecfaf9c90eed23c8959fca95bd703d25b82c88e"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:d4e6036decf4b72d6425d5b29bbd3e8f0ff1059cda7ac7b96d6ac5ed34ffbacd"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-win32.whl", hash = "sha256:a8c28fd40a4226b4a84bdf2d2b5b37d2c7bd49486b5adcc200e8c7ec991dfa7e"},
+    {file = "psycopg2_binary-2.9.6-cp37-cp37m-win_amd64.whl", hash = "sha256:51537e3d299be0db9137b321dfb6a5022caaab275775680e0c3d281feefaca6b"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:cf4499e0a83b7b7edcb8dabecbd8501d0d3a5ef66457200f77bde3d210d5debb"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:7e13a5a2c01151f1208d5207e42f33ba86d561b7a89fca67c700b9486a06d0e2"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0e0f754d27fddcfd74006455b6e04e6705d6c31a612ec69ddc040a5468e44b4e"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:d57c3fd55d9058645d26ae37d76e61156a27722097229d32a9e73ed54819982a"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:71f14375d6f73b62800530b581aed3ada394039877818b2d5f7fc77e3bb6894d"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:441cc2f8869a4f0f4bb408475e5ae0ee1f3b55b33f350406150277f7f35384fc"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:65bee1e49fa6f9cf327ce0e01c4c10f39165ee76d35c846ade7cb0ec6683e303"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:af335bac6b666cc6aea16f11d486c3b794029d9df029967f9938a4bed59b6a19"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:cfec476887aa231b8548ece2e06d28edc87c1397ebd83922299af2e051cf2827"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:65c07febd1936d63bfde78948b76cd4c2a411572a44ac50719ead41947d0f26b"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-win32.whl", hash = "sha256:4dfb4be774c4436a4526d0c554af0cc2e02082c38303852a36f6456ece7b3503"},
+    {file = "psycopg2_binary-2.9.6-cp38-cp38-win_amd64.whl", hash = "sha256:02c6e3cf3439e213e4ee930308dc122d6fb4d4bea9aef4a12535fbd605d1a2fe"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:e9182eb20f41417ea1dd8e8f7888c4d7c6e805f8a7c98c1081778a3da2bee3e4"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:8a6979cf527e2603d349a91060f428bcb135aea2be3201dff794813256c274f1"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8338a271cb71d8da40b023a35d9c1e919eba6cbd8fa20a54b748a332c355d896"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:e3ed340d2b858d6e6fb5083f87c09996506af483227735de6964a6100b4e6a54"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f81e65376e52f03422e1fb475c9514185669943798ed019ac50410fb4c4df232"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bfb13af3c5dd3a9588000910178de17010ebcccd37b4f9794b00595e3a8ddad3"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:4c727b597c6444a16e9119386b59388f8a424223302d0c06c676ec8b4bc1f963"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:4d67fbdaf177da06374473ef6f7ed8cc0a9dc640b01abfe9e8a2ccb1b1402c1f"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:0892ef645c2fabb0c75ec32d79f4252542d0caec1d5d949630e7d242ca4681a3"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:02c0f3757a4300cf379eb49f543fb7ac527fb00144d39246ee40e1df684ab514"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-win32.whl", hash = "sha256:c3dba7dab16709a33a847e5cd756767271697041fbe3fe97c215b1fc1f5c9848"},
+    {file = "psycopg2_binary-2.9.6-cp39-cp39-win_amd64.whl", hash = "sha256:f6a88f384335bb27812293fdb11ac6aee2ca3f51d3c7820fe03de0a304ab6249"},
 ]

 [[package]]
--- a/proxy/Cargo.toml
+++ b/proxy/Cargo.toml
@@ -7,7 +7,6 @@ license.workspace = true
 [dependencies]
 anyhow.workspace = true
 async-trait.workspace = true
-atty.workspace = true
 base64.workspace = true
 bstr.workspace = true
 bytes = { workspace = true, features = ["serde"] }
@@ -38,6 +37,7 @@ rand.workspace = true
 regex.workspace = true
 reqwest = { workspace = true, features = ["json"] }
 reqwest-middleware.workspace = true
+reqwest-retry.workspace = true
 reqwest-tracing.workspace = true
 routerify.workspace = true
 rustls-pemfile.workspace = true
--- a/proxy/README.md
+++ b/proxy/README.md
@@ -93,6 +93,15 @@ With the current approach we made the following design decisions:
   and column oids. Command tag capturing was added to the rust-postgres
   functionality as part of this change.

+### Output options
+
+User can pass several optional headers that will affect resulting json.
+
+1. `Neon-Raw-Text-Output: true`. Return postgres values as text, without parsing them. So numbers, objects, booleans, nulls and arrays will be returned as text. That can be useful in cases when client code wants to implement it's own parsing or reuse parsing libraries from e.g. node-postgres.
+2. `Neon-Array-Mode: true`. Return postgres rows as arrays instead of objects. That is more compact representation and also helps in some edge
+cases where it is hard to use rows represented as objects (e.g. when several fields have the same name).
+
+
 ## Using SNI-based routing on localhost

 Now proxy determines project name from the subdomain, request to the `round-rice-566201.somedomain.tld` will be routed to the project named `round-rice-566201`. Unfortunately, `/etc/hosts` does not support domain wildcards, so I usually use `*.localtest.me` which resolves to `127.0.0.1`. Now we can create self-signed certificate and play with proxy:
--- a/proxy/src/console/mgmt.rs
+++ b/proxy/src/console/mgmt.rs
@@ -8,7 +8,7 @@ use postgres_backend::{self, AuthType, PostgresBackend, PostgresBackendTCP, Quer
 use pq_proto::{BeMessage, SINGLE_COL_ROWDESC};
 use std::future;
 use tokio::net::{TcpListener, TcpStream};
-use tracing::{error, info, info_span};
+use tracing::{error, info, info_span, Instrument};

 static CPLANE_WAITERS: Lazy<Waiters<ComputeReady>> = Lazy::new(Default::default);

@@ -44,19 +44,30 @@ pub async fn task_main(listener: TcpListener) -> anyhow::Result<()> {
            .set_nodelay(true)
            .context("failed to set client socket option")?;

-        tokio::task::spawn(async move {
-            let span = info_span!("mgmt", peer = %peer_addr);
-            let _enter = span.enter();
+        let span = info_span!("mgmt", peer = %peer_addr);

-            info!("started a new console management API thread");
-            scopeguard::defer! {
-                info!("console management API thread is about to finish");
-            }
+        tokio::task::spawn(
+            async move {
+                info!("serving a new console management API connection");

-            if let Err(e) = handle_connection(socket).await {
-                error!("thread failed with an error: {e}");
+                // these might be long running connections, have a separate logging for cancelling
+                // on shutdown and other ways of stopping.
+                let cancelled = scopeguard::guard(tracing::Span::current(), |span| {
+                    let _e = span.entered();
+                    info!("console management API task cancelled");
+                });
+
+                if let Err(e) = handle_connection(socket).await {
+                    error!("serving failed with an error: {e}");
+                } else {
+                    info!("serving completed");
+                }
+
+                // we can no longer get dropped
+                scopeguard::ScopeGuard::into_inner(cancelled);
            }
-        });
+            .instrument(span),
+        );
    }
 }

@@ -77,14 +88,14 @@ impl postgres_backend::Handler<tokio::net::TcpStream> for MgmtHandler {
        pgb: &mut PostgresBackendTCP,
        query: &str,
    ) -> Result<(), QueryError> {
-        try_process_query(pgb, query).await.map_err(|e| {
+        try_process_query(pgb, query).map_err(|e| {
            error!("failed to process response: {e:?}");
            e
        })
    }
 }

-async fn try_process_query(pgb: &mut PostgresBackendTCP, query: &str) -> Result<(), QueryError> {
+fn try_process_query(pgb: &mut PostgresBackendTCP, query: &str) -> Result<(), QueryError> {
    let resp: KickSession = serde_json::from_str(query).context("Failed to parse query as json")?;

    let span = info_span!("event", session_id = resp.session_id);
--- a/proxy/src/http.rs
+++ b/proxy/src/http.rs
@@ -6,8 +6,11 @@ pub mod server;
 pub mod sql_over_http;
 pub mod websocket;

+use std::time::Duration;
+
 pub use reqwest::{Request, Response, StatusCode};
 pub use reqwest_middleware::{ClientWithMiddleware, Error};
+pub use reqwest_retry::{policies::ExponentialBackoff, RetryTransientMiddleware};

 use crate::url::ApiUrl;
 use reqwest_middleware::RequestBuilder;
@@ -21,6 +24,24 @@ pub fn new_client() -> ClientWithMiddleware {
        .build()
 }

+pub fn new_client_with_timeout(default_timout: Duration) -> ClientWithMiddleware {
+    let timeout_client = reqwest::ClientBuilder::new()
+        .timeout(default_timout)
+        .build()
+        .expect("Failed to create http client with timeout");
+
+    let retry_policy =
+        ExponentialBackoff::builder().build_with_total_retry_duration(default_timout);
+
+    reqwest_middleware::ClientBuilder::new(timeout_client)
+        .with(reqwest_tracing::TracingMiddleware::default())
+        // As per docs, "This middleware always errors when given requests with streaming bodies".
+        // That's all right because we only use this client to send `serde_json::RawValue`, which
+        // is not a stream.
+        .with(RetryTransientMiddleware::new_with_policy(retry_policy))
+        .build()
+}
+
 /// Thin convenience wrapper for an API provided by an http endpoint.
 #[derive(Debug, Clone)]
 pub struct Endpoint {
--- a/proxy/src/http/sql_over_http.rs
+++ b/proxy/src/http/sql_over_http.rs
@@ -1,6 +1,9 @@
 use futures::pin_mut;
 use futures::StreamExt;
+use futures::TryFutureExt;
 use hyper::body::HttpBody;
+use hyper::http::HeaderName;
+use hyper::http::HeaderValue;
 use hyper::{Body, HeaderMap, Request};
 use pq_proto::StartupMessageParams;
 use serde_json::json;
@@ -9,8 +12,13 @@ use serde_json::Value;
 use tokio_postgres::types::Kind;
 use tokio_postgres::types::Type;
 use tokio_postgres::Row;
+use tracing::error;
+use tracing::info;
+use tracing::instrument;
 use url::Url;

+use crate::proxy::invalidate_cache;
+use crate::proxy::NUM_RETRIES_WAKE_COMPUTE;
 use crate::{auth, config::ProxyConfig, console};

 #[derive(serde::Deserialize)]
@@ -23,21 +31,28 @@ const APP_NAME: &str = "sql_over_http";
 const MAX_RESPONSE_SIZE: usize = 1024 * 1024; // 1 MB
 const MAX_REQUEST_SIZE: u64 = 1024 * 1024; // 1 MB

+static RAW_TEXT_OUTPUT: HeaderName = HeaderName::from_static("neon-raw-text-output");
+static ARRAY_MODE: HeaderName = HeaderName::from_static("neon-array-mode");
+static HEADER_VALUE_TRUE: HeaderValue = HeaderValue::from_static("true");
+
 //
 // Convert json non-string types to strings, so that they can be passed to Postgres
 // as parameters.
 //
-fn json_to_pg_text(json: Vec<Value>) -> Result<Vec<String>, serde_json::Error> {
+fn json_to_pg_text(json: Vec<Value>) -> Result<Vec<Option<String>>, serde_json::Error> {
    json.iter()
        .map(|value| {
            match value {
-                Value::Null => serde_json::to_string(value),
-                Value::Bool(_) => serde_json::to_string(value),
-                Value::Number(_) => serde_json::to_string(value),
-                Value::Object(_) => serde_json::to_string(value),
+                // special care for nulls
+                Value::Null => Ok(None),

-                // no need to escape
-                Value::String(s) => Ok(s.to_string()),
+                // convert to text with escaping
+                Value::Bool(_) => serde_json::to_string(value).map(Some),
+                Value::Number(_) => serde_json::to_string(value).map(Some),
+                Value::Object(_) => serde_json::to_string(value).map(Some),
+
+                // avoid escaping here, as we pass this as a parameter
+                Value::String(s) => Ok(Some(s.to_string())),

                // special care for arrays
                Value::Array(_) => json_array_to_pg_array(value),
@@ -54,33 +69,44 @@ fn json_to_pg_text(json: Vec<Value>) -> Result<Vec<String>, serde_json::Error> {
 //
 // Example of the same escaping in node-postgres: packages/pg/lib/utils.js
 //
-fn json_array_to_pg_array(value: &Value) -> Result<String, serde_json::Error> {
+fn json_array_to_pg_array(value: &Value) -> Result<Option<String>, serde_json::Error> {
    match value {
-        // same
-        Value::Null => serde_json::to_string(value),
-        Value::Bool(_) => serde_json::to_string(value),
-        Value::Number(_) => serde_json::to_string(value),
-        Value::Object(_) => serde_json::to_string(value),
+        // special care for nulls
+        Value::Null => Ok(None),

-        // now needs to be escaped, as it is part of the array
-        Value::String(_) => serde_json::to_string(value),
+        // convert to text with escaping
+        Value::Bool(_) => serde_json::to_string(value).map(Some),
+        Value::Number(_) => serde_json::to_string(value).map(Some),
+        Value::Object(_) => serde_json::to_string(value).map(Some),
+
+        // here string needs to be escaped, as it is part of the array
+        Value::String(_) => serde_json::to_string(value).map(Some),

        // recurse into array
        Value::Array(arr) => {
            let vals = arr
                .iter()
                .map(json_array_to_pg_array)
+                .map(|r| r.map(|v| v.unwrap_or_else(|| "NULL".to_string())))
                .collect::<Result<Vec<_>, _>>()?
                .join(",");
-            Ok(format!("{{{}}}", vals))
+
+            Ok(Some(format!("{{{}}}", vals)))
        }
    }
 }

+struct ConnInfo {
+    username: String,
+    dbname: String,
+    hostname: String,
+    password: String,
+}
+
 fn get_conn_info(
    headers: &HeaderMap,
    sni_hostname: Option<String>,
-) -> Result<(String, String, String, String), anyhow::Error> {
+) -> Result<ConnInfo, anyhow::Error> {
    let connection_string = headers
        .get("Neon-Connection-String")
        .ok_or(anyhow::anyhow!("missing connection string"))?
@@ -133,12 +159,12 @@ fn get_conn_info(
        }
    }

-    Ok((
-        username.to_owned(),
-        dbname.to_owned(),
-        hostname.to_owned(),
-        password.to_owned(),
-    ))
+    Ok(ConnInfo {
+        username: username.to_owned(),
+        dbname: dbname.to_owned(),
+        hostname: hostname.to_owned(),
+        password: password.to_owned(),
+    })
 }

 // TODO: return different http error codes
@@ -151,13 +177,18 @@ pub async fn handle(
    // Determine the destination and connection params
    //
    let headers = request.headers();
-    let (username, dbname, hostname, password) = get_conn_info(headers, sni_hostname)?;
+    let conn_info = get_conn_info(headers, sni_hostname)?;
    let credential_params = StartupMessageParams::new([
-        ("user", &username),
-        ("database", &dbname),
+        ("user", &conn_info.username),
+        ("database", &conn_info.dbname),
        ("application_name", APP_NAME),
    ]);

+    // Determine the output options. Default behaviour is 'false'. Anything that is not
+    // strictly 'true' assumed to be false.
+    let raw_output = headers.get(&RAW_TEXT_OUTPUT) == Some(&HEADER_VALUE_TRUE);
+    let array_mode = headers.get(&ARRAY_MODE) == Some(&HEADER_VALUE_TRUE);
+
    //
    // Wake up the destination if needed. Code here is a bit involved because
    // we reuse the code from the usual proxy and we need to prepare few structures
@@ -168,21 +199,20 @@ pub async fn handle(
    let creds = config
        .auth_backend
        .as_ref()
-        .map(|_| auth::ClientCredentials::parse(&credential_params, Some(&hostname), common_names))
+        .map(|_| {
+            auth::ClientCredentials::parse(
+                &credential_params,
+                Some(&conn_info.hostname),
+                common_names,
+            )
+        })
        .transpose()?;
    let extra = console::ConsoleReqExtra {
        session_id: uuid::Uuid::new_v4(),
        application_name: Some(APP_NAME),
    };
-    let node = creds.wake_compute(&extra).await?.expect("msg");
-    let conf = node.value.config;
-    let port = *conf.get_ports().first().expect("no port");
-    let host = match conf.get_hosts().first().expect("no host") {
-        tokio_postgres::config::Host::Tcp(host) => host,
-        tokio_postgres::config::Host::Unix(_) => {
-            return Err(anyhow::anyhow!("unix socket is not supported"));
-        }
-    };
+
+    let mut node_info = creds.wake_compute(&extra).await?.expect("msg");

    let request_content_length = match request.body().size_hint().upper() {
        Some(v) => v,
@@ -202,28 +232,10 @@ pub async fn handle(
    let QueryData { query, params } = serde_json::from_slice(&body)?;
    let query_params = json_to_pg_text(params)?;

-    //
-    // Connenct to the destination
-    //
-    let (client, connection) = tokio_postgres::Config::new()
-        .host(host)
-        .port(port)
-        .user(&username)
-        .password(&password)
-        .dbname(&dbname)
-        .max_backend_message_size(MAX_RESPONSE_SIZE)
-        .connect(tokio_postgres::NoTls)
-        .await?;
-
-    tokio::spawn(async move {
-        if let Err(e) = connection.await {
-            eprintln!("connection error: {}", e);
-        }
-    });
-
    //
    // Now execute the query and return the result
    //
+    let client = connect_to_compute(&mut node_info, &extra, &creds, &conn_info).await?;
    let row_stream = client.query_raw_txt(query, query_params).await?;

    // Manually drain the stream into a vector to leave row_stream hanging
@@ -262,6 +274,11 @@ pub async fn handle(
                json!({
                    "name": Value::String(c.name().to_owned()),
                    "dataTypeID": Value::Number(c.type_().oid().into()),
+                    "tableID": c.table_oid(),
+                    "columnID": c.column_id(),
+                    "dataTypeSize": c.type_size(),
+                    "dataTypeModifier": c.type_modifier(),
+                    "format": "text",
                })
            })
            .collect::<Vec<_>>()
@@ -272,7 +289,7 @@ pub async fn handle(
    // convert rows to JSON
    let rows = rows
        .iter()
-        .map(pg_text_row_to_json)
+        .map(|row| pg_text_row_to_json(row, raw_output, array_mode))
        .collect::<Result<Vec<_>, _>>()?;

    // resulting JSON format is based on the format of node-postgres result
@@ -281,26 +298,106 @@ pub async fn handle(
        "rowCount": command_tag_count,
        "rows": rows,
        "fields": fields,
+        "rowAsArray": array_mode,
    }))
 }

+/// This function is a copy of `connect_to_compute` from `src/proxy.rs` with
+/// the difference that it uses `tokio_postgres` for the connection.
+#[instrument(skip_all)]
+async fn connect_to_compute(
+    node_info: &mut console::CachedNodeInfo,
+    extra: &console::ConsoleReqExtra<'_>,
+    creds: &auth::BackendType<'_, auth::ClientCredentials<'_>>,
+    conn_info: &ConnInfo,
+) -> anyhow::Result<tokio_postgres::Client> {
+    let mut num_retries: usize = NUM_RETRIES_WAKE_COMPUTE;
+
+    loop {
+        match connect_to_compute_once(node_info, conn_info).await {
+            Err(e) if num_retries > 0 => {
+                info!("compute node's state has changed; requesting a wake-up");
+                match creds.wake_compute(extra).await? {
+                    // Update `node_info` and try one more time.
+                    Some(new) => {
+                        *node_info = new;
+                    }
+                    // Link auth doesn't work that way, so we just exit.
+                    None => return Err(e),
+                }
+            }
+            other => return other,
+        }
+
+        num_retries -= 1;
+        info!("retrying after wake-up ({num_retries} attempts left)");
+    }
+}
+
+async fn connect_to_compute_once(
+    node_info: &console::CachedNodeInfo,
+    conn_info: &ConnInfo,
+) -> anyhow::Result<tokio_postgres::Client> {
+    let mut config = (*node_info.config).clone();
+
+    let (client, connection) = config
+        .user(&conn_info.username)
+        .password(&conn_info.password)
+        .dbname(&conn_info.dbname)
+        .max_backend_message_size(MAX_RESPONSE_SIZE)
+        .connect(tokio_postgres::NoTls)
+        .inspect_err(|e: &tokio_postgres::Error| {
+            error!(
+                "failed to connect to compute node hosts={:?} ports={:?}: {}",
+                node_info.config.get_hosts(),
+                node_info.config.get_ports(),
+                e
+            );
+            invalidate_cache(node_info)
+        })
+        .await?;
+
+    tokio::spawn(async move {
+        if let Err(e) = connection.await {
+            error!("connection error: {}", e);
+        }
+    });
+
+    Ok(client)
+}
+
 //
 // Convert postgres row with text-encoded values to JSON object
 //
-pub fn pg_text_row_to_json(row: &Row) -> Result<Value, anyhow::Error> {
-    let res = row
-        .columns()
-        .iter()
-        .enumerate()
-        .map(|(i, column)| {
-            let name = column.name();
-            let pg_value = row.as_text(i)?;
-            let json_value = pg_text_to_json(pg_value, column.type_())?;
-            Ok((name.to_string(), json_value))
-        })
-        .collect::<Result<Map<String, Value>, anyhow::Error>>()?;
+pub fn pg_text_row_to_json(
+    row: &Row,
+    raw_output: bool,
+    array_mode: bool,
+) -> Result<Value, anyhow::Error> {
+    let iter = row.columns().iter().enumerate().map(|(i, column)| {
+        let name = column.name();
+        let pg_value = row.as_text(i)?;
+        let json_value = if raw_output {
+            match pg_value {
+                Some(v) => Value::String(v.to_string()),
+                None => Value::Null,
+            }
+        } else {
+            pg_text_to_json(pg_value, column.type_())?
+        };
+        Ok((name.to_string(), json_value))
+    });

-    Ok(Value::Object(res))
+    if array_mode {
+        // drop keys and aggregate into array
+        let arr = iter
+            .map(|r| r.map(|(_key, val)| val))
+            .collect::<Result<Vec<Value>, anyhow::Error>>()?;
+        Ok(Value::Array(arr))
+    } else {
+        let obj = iter.collect::<Result<Map<String, Value>, anyhow::Error>>()?;
+        Ok(Value::Object(obj))
+    }
 }

 //
@@ -308,10 +405,6 @@ pub fn pg_text_row_to_json(row: &Row) -> Result<Value, anyhow::Error> {
 //
 pub fn pg_text_to_json(pg_value: Option<&str>, pg_type: &Type) -> Result<Value, anyhow::Error> {
    if let Some(val) = pg_value {
-        if val == "NULL" {
-            return Ok(Value::Null);
-        }
-
        if let Kind::Array(elem_type) = pg_type.kind() {
            return pg_array_parse(val, elem_type);
        }
@@ -373,6 +466,27 @@ fn _pg_array_parse(
        }
    }

+    fn push_checked(
+        entry: &mut String,
+        entries: &mut Vec<Value>,
+        elem_type: &Type,
+    ) -> Result<(), anyhow::Error> {
+        if !entry.is_empty() {
+            // While in usual postgres response we get nulls as None and everything else
+            // as Some(&str), in arrays we get NULL as unquoted 'NULL' string (while
+            // string with value 'NULL' will be represented by '"NULL"'). So catch NULLs
+            // here while we have quotation info and convert them to None.
+            if entry == "NULL" {
+                entries.push(pg_text_to_json(None, elem_type)?);
+            } else {
+                entries.push(pg_text_to_json(Some(entry), elem_type)?);
+            }
+            entry.clear();
+        }
+
+        Ok(())
+    }
+
    while let Some((mut i, mut c)) = pg_array_chr.next() {
        let mut escaped = false;

@@ -395,9 +509,7 @@ fn _pg_array_parse(
            '}' => {
                level -= 1;
                if level == 0 {
-                    if !entry.is_empty() {
-                        entries.push(pg_text_to_json(Some(&entry), elem_type)?);
-                    }
+                    push_checked(&mut entry, &mut entries, elem_type)?;
                    if nested {
                        return Ok((Value::Array(entries), i));
                    }
@@ -405,17 +517,15 @@ fn _pg_array_parse(
            }
            '"' if !escaped => {
                if quote {
-                    // push even if empty
+                    // end of quoted string, so push it manually without any checks
+                    // for emptiness or nulls
                    entries.push(pg_text_to_json(Some(&entry), elem_type)?);
-                    entry = String::new();
+                    entry.clear();
                }
                quote = !quote;
            }
            ',' if !quote => {
-                if !entry.is_empty() {
-                    entries.push(pg_text_to_json(Some(&entry), elem_type)?);
-                    entry = String::new();
-                }
+                push_checked(&mut entry, &mut entries, elem_type)?;
            }
            _ => {
                entry.push(c);
@@ -439,30 +549,35 @@ mod tests {
    fn test_atomic_types_to_pg_params() {
        let json = vec![Value::Bool(true), Value::Bool(false)];
        let pg_params = json_to_pg_text(json).unwrap();
-        assert_eq!(pg_params, vec!["true", "false"]);
+        assert_eq!(
+            pg_params,
+            vec![Some("true".to_owned()), Some("false".to_owned())]
+        );

        let json = vec![Value::Number(serde_json::Number::from(42))];
        let pg_params = json_to_pg_text(json).unwrap();
-        assert_eq!(pg_params, vec!["42"]);
+        assert_eq!(pg_params, vec![Some("42".to_owned())]);

        let json = vec![Value::String("foo\"".to_string())];
        let pg_params = json_to_pg_text(json).unwrap();
-        assert_eq!(pg_params, vec!["foo\""]);
+        assert_eq!(pg_params, vec![Some("foo\"".to_owned())]);

        let json = vec![Value::Null];
        let pg_params = json_to_pg_text(json).unwrap();
-        assert_eq!(pg_params, vec!["null"]);
+        assert_eq!(pg_params, vec![None]);
    }

    #[test]
    fn test_json_array_to_pg_array() {
        // atoms and escaping
-        let json = "[true, false, null, 42, \"foo\", \"bar\\\"-\\\\\"]";
+        let json = "[true, false, null, \"NULL\", 42, \"foo\", \"bar\\\"-\\\\\"]";
        let json: Value = serde_json::from_str(json).unwrap();
        let pg_params = json_to_pg_text(vec![json]).unwrap();
        assert_eq!(
            pg_params,
-            vec!["{true,false,null,42,\"foo\",\"bar\\\"-\\\\\"}"]
+            vec![Some(
+                "{true,false,NULL,\"NULL\",42,\"foo\",\"bar\\\"-\\\\\"}".to_owned()
+            )]
        );

        // nested arrays
@@ -471,7 +586,9 @@ mod tests {
        let pg_params = json_to_pg_text(vec![json]).unwrap();
        assert_eq!(
            pg_params,
-            vec!["{{true,false},{null,42},{\"foo\",\"bar\\\"-\\\\\"}}"]
+            vec![Some(
+                "{{true,false},{NULL,42},{\"foo\",\"bar\\\"-\\\\\"}}".to_owned()
+            )]
        );
    }

--- a/proxy/src/http/websocket.rs
+++ b/proxy/src/http/websocket.rs
@@ -26,7 +26,6 @@ use tls_listener::TlsListener;
 use tokio::{
    io::{self, AsyncBufRead, AsyncRead, AsyncWrite, ReadBuf},
    net::TcpListener,
-    select,
 };
 use tokio_util::sync::CancellationToken;
 use tracing::{error, info, info_span, warn, Instrument};
@@ -193,14 +192,9 @@ async fn ws_handler(
    // TODO: that deserves a refactor as now this function also handles http json client besides websockets.
    // Right now I don't want to blow up sql-over-http patch with file renames and do that as a follow up instead.
    } else if request.uri().path() == "/sql" && request.method() == Method::POST {
-        let result = select! {
-            _ = tokio::time::sleep(std::time::Duration::from_secs(10)) => {
-                Err(anyhow::anyhow!("Query timed out"))
-            }
-            response = sql_over_http::handle(config, request, sni_hostname) => {
-                response
-            }
-        };
+        let result = sql_over_http::handle(config, request, sni_hostname)
+            .instrument(info_span!("sql-over-http"))
+            .await;
        let status_code = match result {
            Ok(_) => StatusCode::OK,
            Err(_) => StatusCode::BAD_REQUEST,
--- a/proxy/src/logging.rs
+++ b/proxy/src/logging.rs
@@ -18,7 +18,7 @@ pub async fn init() -> anyhow::Result<LoggingGuard> {
        .from_env_lossy();

    let fmt_layer = tracing_subscriber::fmt::layer()
-        .with_ansi(atty::is(atty::Stream::Stderr))
+        .with_ansi(false)
        .with_writer(std::io::stderr)
        .with_target(false);

--- a/proxy/src/metrics.rs
+++ b/proxy/src/metrics.rs
@@ -4,11 +4,13 @@ use crate::{config::MetricCollectionConfig, http};
 use chrono::{DateTime, Utc};
 use consumption_metrics::{idempotency_key, Event, EventChunk, EventType, CHUNK_SIZE};
 use serde::Serialize;
-use std::collections::HashMap;
+use std::{collections::HashMap, time::Duration};
 use tracing::{error, info, instrument, trace, warn};

 const PROXY_IO_BYTES_PER_CLIENT: &str = "proxy_io_bytes_per_client";

+const DEFAULT_HTTP_REPORTING_TIMEOUT: Duration = Duration::from_secs(60);
+
 ///
 /// Key that uniquely identifies the object, this metric describes.
 /// Currently, endpoint_id is enough, but this may change later,
@@ -30,7 +32,7 @@ pub async fn task_main(config: &MetricCollectionConfig) -> anyhow::Result<()> {
        info!("metrics collector has shut down");
    }

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

@@ -182,36 +184,36 @@ async fn collect_metrics_iteration(
            }
        };

-        if res.status().is_success() {
-            // update cached metrics after they were sent successfully
-            for send_metric in chunk {
-                let stop_time = match send_metric.kind {
-                    EventType::Incremental { stop_time, .. } => stop_time,
-                    _ => unreachable!(),
-                };
-
-                cached_metrics
-                    .entry(Ids {
-                        endpoint_id: send_metric.extra.endpoint_id.clone(),
-                        branch_id: send_metric.extra.branch_id.clone(),
-                    })
-                    // update cached value (add delta) and time
-                    .and_modify(|e| {
-                        e.0 = e.0.saturating_add(send_metric.value);
-                        e.1 = stop_time
-                    })
-                    // cache new metric
-                    .or_insert((send_metric.value, stop_time));
-            }
-        } else {
+        if !res.status().is_success() {
            error!("metrics endpoint refused the sent metrics: {:?}", res);
-            for metric in chunk.iter() {
+            for metric in chunk.iter().filter(|metric| metric.value > (1u64 << 40)) {
                // Report if the metric value is suspiciously large
-                if metric.value > (1u64 << 40) {
-                    error!("potentially abnormal metric value: {:?}", metric);
-                }
+                error!("potentially abnormal metric value: {:?}", metric);
            }
        }
+        // update cached metrics after they were sent
+        // (to avoid sending the same metrics twice)
+        // see the relevant discussion on why to do so even if the status is not success:
+        // https://github.com/neondatabase/neon/pull/4563#discussion_r1246710956
+        for send_metric in chunk {
+            let stop_time = match send_metric.kind {
+                EventType::Incremental { stop_time, .. } => stop_time,
+                _ => unreachable!(),
+            };
+
+            cached_metrics
+                .entry(Ids {
+                    endpoint_id: send_metric.extra.endpoint_id.clone(),
+                    branch_id: send_metric.extra.branch_id.clone(),
+                })
+                // update cached value (add delta) and time
+                .and_modify(|e| {
+                    e.0 = e.0.saturating_add(send_metric.value);
+                    e.1 = stop_time
+                })
+                // cache new metric
+                .or_insert((send_metric.value, stop_time));
+        }
    }
    Ok(())
 }
--- a/proxy/src/proxy.rs
+++ b/proxy/src/proxy.rs
@@ -22,7 +22,7 @@ use tracing::{error, info, warn};
 use utils::measured_stream::MeasuredStream;

 /// Number of times we should retry the `/proxy_wake_compute` http request.
-const NUM_RETRIES_WAKE_COMPUTE: usize = 1;
+pub const NUM_RETRIES_WAKE_COMPUTE: usize = 1;

 const ERR_INSECURE_CONNECTION: &str = "connection is insecure (try using `sslmode=require`)";
 const ERR_PROTO_VIOLATION: &str = "protocol violation";
@@ -283,34 +283,35 @@ async fn handshake<S: AsyncRead + AsyncWrite + Unpin>(
    }
 }

+/// If we couldn't connect, a cached connection info might be to blame
+/// (e.g. the compute node's address might've changed at the wrong time).
+/// Invalidate the cache entry (if any) to prevent subsequent errors.
+#[tracing::instrument(name = "invalidate_cache", skip_all)]
+pub fn invalidate_cache(node_info: &console::CachedNodeInfo) {
+    let is_cached = node_info.cached();
+    if is_cached {
+        warn!("invalidating stalled compute node info cache entry");
+        node_info.invalidate();
+    }
+
+    let label = match is_cached {
+        true => "compute_cached",
+        false => "compute_uncached",
+    };
+    NUM_CONNECTION_FAILURES.with_label_values(&[label]).inc();
+}
+
 /// Try to connect to the compute node once.
 #[tracing::instrument(name = "connect_once", skip_all)]
 async fn connect_to_compute_once(
    node_info: &console::CachedNodeInfo,
 ) -> Result<PostgresConnection, compute::ConnectionError> {
-    // If we couldn't connect, a cached connection info might be to blame
-    // (e.g. the compute node's address might've changed at the wrong time).
-    // Invalidate the cache entry (if any) to prevent subsequent errors.
-    let invalidate_cache = |_: &compute::ConnectionError| {
-        let is_cached = node_info.cached();
-        if is_cached {
-            warn!("invalidating stalled compute node info cache entry");
-            node_info.invalidate();
-        }
-
-        let label = match is_cached {
-            true => "compute_cached",
-            false => "compute_uncached",
-        };
-        NUM_CONNECTION_FAILURES.with_label_values(&[label]).inc();
-    };
-
    let allow_self_signed_compute = node_info.allow_self_signed_compute;

    node_info
        .config
        .connect(allow_self_signed_compute)
-        .inspect_err(invalidate_cache)
+        .inspect_err(|_: &compute::ConnectionError| invalidate_cache(node_info))
        .await
 }

--- a/pyproject.toml
+++ b/pyproject.toml
@@ -7,7 +7,7 @@ authors = []
 [tool.poetry.dependencies]
 python = "^3.9"
 pytest = "^7.3.1"
-psycopg2-binary = "^2.9.1"
+psycopg2-binary = "^2.9.6"
 typing-extensions = "^4.6.1"
 PyJWT = {version = "^2.1.0", extras = ["crypto"]}
 requests = "^2.31.0"
--- a/rust-toolchain.toml
+++ b/rust-toolchain.toml
@@ -1,5 +1,5 @@
 [toolchain]
-channel = "1.68.2"
+channel = "1.70.0"
 profile = "default"
 # The default profile includes rustc, rust-std, cargo, rust-docs, rustfmt and clippy.
 # https://rust-lang.github.io/rustup/concepts/profiles.html
--- a/safekeeper/src/bin/safekeeper.rs
+++ b/safekeeper/src/bin/safekeeper.rs
@@ -3,15 +3,19 @@
 //
 use anyhow::{bail, Context, Result};
 use clap::Parser;
+use futures::future::BoxFuture;
+use futures::stream::FuturesUnordered;
+use futures::{FutureExt, StreamExt};
 use remote_storage::RemoteStorageConfig;
+use tokio::runtime::Handle;
+use tokio::signal::unix::{signal, SignalKind};
+use tokio::task::JoinError;
 use toml_edit::Document;
-use utils::signals::ShutdownSignals;

 use std::fs::{self, File};
 use std::io::{ErrorKind, Write};
 use std::path::{Path, PathBuf};
 use std::sync::Arc;
-use std::thread;
 use std::time::Duration;
 use storage_broker::Uri;
 use tokio::sync::mpsc;
@@ -20,22 +24,21 @@ use tracing::*;
 use utils::pid_file;

 use metrics::set_build_info_metric;
-use safekeeper::broker;
-use safekeeper::control_file;
 use safekeeper::defaults::{
    DEFAULT_HEARTBEAT_TIMEOUT, DEFAULT_HTTP_LISTEN_ADDR, DEFAULT_MAX_OFFLOADER_LAG_BYTES,
    DEFAULT_PG_LISTEN_ADDR,
 };
-use safekeeper::http;
-use safekeeper::remove_wal;
-use safekeeper::wal_backup;
 use safekeeper::wal_service;
 use safekeeper::GlobalTimelines;
 use safekeeper::SafeKeeperConf;
+use safekeeper::{broker, WAL_SERVICE_RUNTIME};
+use safekeeper::{control_file, BROKER_RUNTIME};
+use safekeeper::{http, WAL_REMOVER_RUNTIME};
+use safekeeper::{remove_wal, WAL_BACKUP_RUNTIME};
+use safekeeper::{wal_backup, HTTP_RUNTIME};
 use storage_broker::DEFAULT_ENDPOINT;
 use utils::auth::JwtAuth;
 use utils::{
-    http::endpoint,
    id::NodeId,
    logging::{self, LogFormat},
    project_git_version,
@@ -104,10 +107,6 @@ struct Args {
    /// Safekeeper won't be elected for WAL offloading if it is lagging for more than this value in bytes
    #[arg(long, default_value_t = DEFAULT_MAX_OFFLOADER_LAG_BYTES)]
    max_offloader_lag: u64,
-    /// Number of threads for wal backup runtime, by default number of cores
-    /// available to the system.
-    #[arg(long)]
-    wal_backup_threads: Option<usize>,
    /// Number of max parallel WAL segments to be offloaded to remote storage.
    #[arg(long, default_value = "5")]
    wal_backup_parallel_jobs: usize,
@@ -121,9 +120,14 @@ struct Args {
    /// Format for logging, either 'plain' or 'json'.
    #[arg(long, default_value = "plain")]
    log_format: String,
+    /// Run everything in single threaded current thread runtime, might be
+    /// useful for debugging.
+    #[arg(long)]
+    current_thread_runtime: bool,
 }

-fn main() -> anyhow::Result<()> {
+#[tokio::main(flavor = "current_thread")]
+async fn main() -> anyhow::Result<()> {
    let args = Args::parse();

    if let Some(addr) = args.dump_control_file {
@@ -183,10 +187,10 @@ fn main() -> anyhow::Result<()> {
        heartbeat_timeout: args.heartbeat_timeout,
        remote_storage: args.remote_storage,
        max_offloader_lag_bytes: args.max_offloader_lag,
-        backup_runtime_threads: args.wal_backup_threads,
        wal_backup_enabled: !args.disable_wal_backup,
        backup_parallel_jobs: args.wal_backup_parallel_jobs,
        auth,
+        current_thread_runtime: args.current_thread_runtime,
    };

    // initialize sentry if SENTRY_DSN is provided
@@ -194,10 +198,14 @@ fn main() -> anyhow::Result<()> {
        Some(GIT_VERSION.into()),
        &[("node_id", &conf.my_id.to_string())],
    );
-    start_safekeeper(conf)
+    start_safekeeper(conf).await
 }

-fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
+/// Result of joining any of main tasks: upper error means task failed to
+/// complete, e.g. panicked, inner is error produced by task itself.
+type JoinTaskRes = Result<anyhow::Result<()>, JoinError>;
+
+async fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
    // Prevent running multiple safekeepers on the same directory
    let lock_file_path = conf.workdir.join(PID_FILE_NAME);
    let lock_file =
@@ -208,14 +216,18 @@ fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
    // we need to release the lock file only when the current process is gone
    std::mem::forget(lock_file);

-    let http_listener = tcp_listener::bind(conf.listen_http_addr.clone()).map_err(|e| {
-        error!("failed to bind to address {}: {}", conf.listen_http_addr, e);
+    info!("starting safekeeper WAL service on {}", conf.listen_pg_addr);
+    let pg_listener = tcp_listener::bind(conf.listen_pg_addr.clone()).map_err(|e| {
+        error!("failed to bind to address {}: {}", conf.listen_pg_addr, e);
        e
    })?;

-    info!("starting safekeeper on {}", conf.listen_pg_addr);
-    let pg_listener = tcp_listener::bind(conf.listen_pg_addr.clone()).map_err(|e| {
-        error!("failed to bind to address {}: {}", conf.listen_pg_addr, e);
+    info!(
+        "starting safekeeper HTTP service on {}",
+        conf.listen_http_addr
+    );
+    let http_listener = tcp_listener::bind(conf.listen_http_addr.clone()).map_err(|e| {
+        error!("failed to bind to address {}: {}", conf.listen_http_addr, e);
        e
    })?;

@@ -224,71 +236,88 @@ fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
    let timeline_collector = safekeeper::metrics::TimelineCollector::new();
    metrics::register_internal(Box::new(timeline_collector))?;

-    let mut threads = vec![];
    let (wal_backup_launcher_tx, wal_backup_launcher_rx) = mpsc::channel(100);

    // Load all timelines from disk to memory.
    GlobalTimelines::init(conf.clone(), wal_backup_launcher_tx)?;

-    let conf_ = conf.clone();
-    threads.push(
-        thread::Builder::new()
-            .name("http_endpoint_thread".into())
-            .spawn(|| {
-                let router = http::make_router(conf_);
-                endpoint::serve_thread_main(
-                    router,
-                    http_listener,
-                    std::future::pending(), // never shut down
-                )
-                .unwrap();
-            })?,
-    );
-
-    let conf_cloned = conf.clone();
-    let safekeeper_thread = thread::Builder::new()
-        .name("WAL service thread".into())
-        .spawn(|| wal_service::thread_main(conf_cloned, pg_listener))
-        .unwrap();
-
-    threads.push(safekeeper_thread);
+    // Keep handles to main tasks to die if any of them disappears.
+    let mut tasks_handles: FuturesUnordered<BoxFuture<(String, JoinTaskRes)>> =
+        FuturesUnordered::new();

    let conf_ = conf.clone();
-    threads.push(
-        thread::Builder::new()
-            .name("broker thread".into())
-            .spawn(|| {
-                broker::thread_main(conf_);
-            })?,
-    );
+    // Run everything in current thread rt, if asked.
+    if conf.current_thread_runtime {
+        info!("running in current thread runtime");
+    }
+    let current_thread_rt = conf
+        .current_thread_runtime
+        .then(|| Handle::try_current().expect("no runtime in main"));
+    let wal_service_handle = current_thread_rt
+        .as_ref()
+        .unwrap_or_else(|| WAL_SERVICE_RUNTIME.handle())
+        .spawn(wal_service::task_main(conf_, pg_listener))
+        // wrap with task name for error reporting
+        .map(|res| ("WAL service main".to_owned(), res));
+    tasks_handles.push(Box::pin(wal_service_handle));

    let conf_ = conf.clone();
-    threads.push(
-        thread::Builder::new()
-            .name("WAL removal thread".into())
-            .spawn(|| {
-                remove_wal::thread_main(conf_);
-            })?,
-    );
+    let http_handle = current_thread_rt
+        .as_ref()
+        .unwrap_or_else(|| HTTP_RUNTIME.handle())
+        .spawn(http::task_main(conf_, http_listener))
+        .map(|res| ("HTTP service main".to_owned(), res));
+    tasks_handles.push(Box::pin(http_handle));

-    threads.push(
-        thread::Builder::new()
-            .name("WAL backup launcher thread".into())
-            .spawn(move || {
-                wal_backup::wal_backup_launcher_thread_main(conf, wal_backup_launcher_rx);
-            })?,
-    );
+    let conf_ = conf.clone();
+    let broker_task_handle = current_thread_rt
+        .as_ref()
+        .unwrap_or_else(|| BROKER_RUNTIME.handle())
+        .spawn(broker::task_main(conf_).instrument(info_span!("broker")))
+        .map(|res| ("broker main".to_owned(), res));
+    tasks_handles.push(Box::pin(broker_task_handle));
+
+    let conf_ = conf.clone();
+    let wal_remover_handle = current_thread_rt
+        .as_ref()
+        .unwrap_or_else(|| WAL_REMOVER_RUNTIME.handle())
+        .spawn(remove_wal::task_main(conf_))
+        .map(|res| ("WAL remover".to_owned(), res));
+    tasks_handles.push(Box::pin(wal_remover_handle));
+
+    let conf_ = conf.clone();
+    let wal_backup_handle = current_thread_rt
+        .as_ref()
+        .unwrap_or_else(|| WAL_BACKUP_RUNTIME.handle())
+        .spawn(wal_backup::wal_backup_launcher_task_main(
+            conf_,
+            wal_backup_launcher_rx,
+        ))
+        .map(|res| ("WAL backup launcher".to_owned(), res));
+    tasks_handles.push(Box::pin(wal_backup_handle));

    set_build_info_metric(GIT_VERSION);
-    // TODO: put more thoughts into handling of failed threads
-    // We should catch & die if they are in trouble.

-    // On any shutdown signal, log receival and exit. Additionally, handling
-    // SIGQUIT prevents coredump.
-    ShutdownSignals::handle(|signal| {
-        info!("received {}, terminating", signal.name());
-        std::process::exit(0);
-    })
+    // TODO: update tokio-stream, convert to real async Stream with
+    // SignalStream, map it to obtain missing signal name, combine streams into
+    // single stream we can easily sit on.
+    let mut sigquit_stream = signal(SignalKind::quit())?;
+    let mut sigint_stream = signal(SignalKind::interrupt())?;
+    let mut sigterm_stream = signal(SignalKind::terminate())?;
+
+    tokio::select! {
+        Some((task_name, res)) = tasks_handles.next()=> {
+            error!("{} task failed: {:?}, exiting", task_name, res);
+            std::process::exit(1);
+        }
+        // On any shutdown signal, log receival and exit. Additionally, handling
+        // SIGQUIT prevents coredump.
+        _ = sigquit_stream.recv() => info!("received SIGQUIT, terminating"),
+        _ = sigint_stream.recv() => info!("received SIGINT, terminating"),
+        _ = sigterm_stream.recv() => info!("received SIGTERM, terminating")
+
+    };
+    std::process::exit(0);
 }

 /// Determine safekeeper id.
--- a/safekeeper/src/broker.rs
+++ b/safekeeper/src/broker.rs
@@ -8,7 +8,7 @@ use anyhow::Error;
 use anyhow::Result;

 use storage_broker::parse_proto_ttid;
-use storage_broker::proto::broker_service_client::BrokerServiceClient;
+
 use storage_broker::proto::subscribe_safekeeper_info_request::SubscriptionKey as ProtoSubscriptionKey;
 use storage_broker::proto::SubscribeSafekeeperInfoRequest;
 use storage_broker::Request;
@@ -16,7 +16,7 @@ use storage_broker::Request;
 use std::time::Duration;
 use std::time::Instant;
 use tokio::task::JoinHandle;
-use tokio::{runtime, time::sleep};
+use tokio::time::sleep;
 use tracing::*;

 use crate::metrics::BROKER_ITERATION_TIMELINES;
@@ -29,23 +29,10 @@ use crate::SafeKeeperConf;
 const RETRY_INTERVAL_MSEC: u64 = 1000;
 const PUSH_INTERVAL_MSEC: u64 = 1000;

-pub fn thread_main(conf: SafeKeeperConf) {
-    let runtime = runtime::Builder::new_current_thread()
-        .enable_all()
-        .build()
-        .unwrap();
-
-    let _enter = info_span!("broker").entered();
-    info!("started, broker endpoint {:?}", conf.broker_endpoint);
-
-    runtime.block_on(async {
-        main_loop(conf).await;
-    });
-}
-
 /// Push once in a while data about all active timelines to the broker.
 async fn push_loop(conf: SafeKeeperConf) -> anyhow::Result<()> {
-    let mut client = BrokerServiceClient::connect(conf.broker_endpoint.clone()).await?;
+    let mut client =
+        storage_broker::connect(conf.broker_endpoint.clone(), conf.broker_keepalive_interval)?;
    let push_interval = Duration::from_millis(PUSH_INTERVAL_MSEC);

    let outbound = async_stream::stream! {
@@ -55,20 +42,27 @@ async fn push_loop(conf: SafeKeeperConf) -> anyhow::Result<()> {
            // sensitive and there is no risk of deadlock as we don't await while
            // lock is held.
            let now = Instant::now();
-            let mut active_tlis = GlobalTimelines::get_all();
-            active_tlis.retain(|tli| tli.is_active());
-            for tli in &active_tlis {
-                let sk_info = tli.get_safekeeper_info(&conf);
+            let all_tlis = GlobalTimelines::get_all();
+            let mut n_pushed_tlis = 0;
+            for tli in &all_tlis {
+                // filtering alternative futures::stream::iter(all_tlis)
+                //   .filter(|tli| {let tli = tli.clone(); async move { tli.is_active().await}}).collect::<Vec<_>>().await;
+                // doesn't look better, and I'm not sure how to do that without collect.
+                if !tli.is_active().await {
+                    continue;
+                }
+                let sk_info = tli.get_safekeeper_info(&conf).await;
                yield sk_info;
                BROKER_PUSHED_UPDATES.inc();
+                n_pushed_tlis += 1;
            }
            let elapsed = now.elapsed();

            BROKER_PUSH_ALL_UPDATES_SECONDS.observe(elapsed.as_secs_f64());
-            BROKER_ITERATION_TIMELINES.observe(active_tlis.len() as f64);
+            BROKER_ITERATION_TIMELINES.observe(n_pushed_tlis as f64);

            if elapsed > push_interval / 2 {
-                info!("broker push is too long, pushed {} timeline updates to broker in {:?}", active_tlis.len(), elapsed);
+                info!("broker push is too long, pushed {} timeline updates to broker in {:?}", n_pushed_tlis, elapsed);
            }

            sleep(push_interval).await;
@@ -125,10 +119,13 @@ async fn pull_loop(conf: SafeKeeperConf) -> Result<()> {
    bail!("end of stream");
 }

-async fn main_loop(conf: SafeKeeperConf) {
+pub async fn task_main(conf: SafeKeeperConf) -> anyhow::Result<()> {
+    info!("started, broker endpoint {:?}", conf.broker_endpoint);
+
    let mut ticker = tokio::time::interval(Duration::from_millis(RETRY_INTERVAL_MSEC));
    let mut push_handle: Option<JoinHandle<Result<(), Error>>> = None;
    let mut pull_handle: Option<JoinHandle<Result<(), Error>>> = None;
+
    // Selecting on JoinHandles requires some squats; is there a better way to
    // reap tasks individually?

--- a/safekeeper/src/control_file.rs
+++ b/safekeeper/src/control_file.rs
@@ -2,9 +2,10 @@

 use anyhow::{bail, ensure, Context, Result};
 use byteorder::{LittleEndian, ReadBytesExt, WriteBytesExt};
+use tokio::fs::{self, File};
+use tokio::io::AsyncWriteExt;

-use std::fs::{self, File, OpenOptions};
-use std::io::{Read, Write};
+use std::io::Read;
 use std::ops::Deref;
 use std::path::{Path, PathBuf};
 use std::time::Instant;
@@ -26,9 +27,10 @@ pub const CHECKSUM_SIZE: usize = std::mem::size_of::<u32>();

 /// Storage should keep actual state inside of it. It should implement Deref
 /// trait to access state fields and have persist method for updating that state.
+#[async_trait::async_trait]
 pub trait Storage: Deref<Target = SafeKeeperState> {
    /// Persist safekeeper state on disk and update internal state.
-    fn persist(&mut self, s: &SafeKeeperState) -> Result<()>;
+    async fn persist(&mut self, s: &SafeKeeperState) -> Result<()>;

    /// Timestamp of last persist.
    fn last_persist_at(&self) -> Instant;
@@ -82,7 +84,7 @@ impl FileStorage {
    /// Check the magic/version in the on-disk data and deserialize it, if possible.
    fn deser_sk_state(buf: &mut &[u8]) -> Result<SafeKeeperState> {
        // Read the version independent part
-        let magic = buf.read_u32::<LittleEndian>()?;
+        let magic = ReadBytesExt::read_u32::<LittleEndian>(buf)?;
        if magic != SK_MAGIC {
            bail!(
                "bad control file magic: {:X}, expected {:X}",
@@ -90,7 +92,7 @@ impl FileStorage {
                SK_MAGIC
            );
        }
-        let version = buf.read_u32::<LittleEndian>()?;
+        let version = ReadBytesExt::read_u32::<LittleEndian>(buf)?;
        if version == SK_FORMAT_VERSION {
            let res = SafeKeeperState::des(buf)?;
            return Ok(res);
@@ -110,7 +112,7 @@ impl FileStorage {

    /// Read in the control file.
    pub fn load_control_file<P: AsRef<Path>>(control_file_path: P) -> Result<SafeKeeperState> {
-        let mut control_file = OpenOptions::new()
+        let mut control_file = std::fs::OpenOptions::new()
            .read(true)
            .write(true)
            .open(&control_file_path)
@@ -159,30 +161,31 @@ impl Deref for FileStorage {
    }
 }

+#[async_trait::async_trait]
 impl Storage for FileStorage {
    /// persists state durably to underlying storage
    /// for description see https://lwn.net/Articles/457667/
-    fn persist(&mut self, s: &SafeKeeperState) -> Result<()> {
+    async fn persist(&mut self, s: &SafeKeeperState) -> Result<()> {
        let _timer = PERSIST_CONTROL_FILE_SECONDS.start_timer();

        // write data to safekeeper.control.partial
        let control_partial_path = self.timeline_dir.join(CONTROL_FILE_NAME_PARTIAL);
-        let mut control_partial = File::create(&control_partial_path).with_context(|| {
+        let mut control_partial = File::create(&control_partial_path).await.with_context(|| {
            format!(
                "failed to create partial control file at: {}",
                &control_partial_path.display()
            )
        })?;
        let mut buf: Vec<u8> = Vec::new();
-        buf.write_u32::<LittleEndian>(SK_MAGIC)?;
-        buf.write_u32::<LittleEndian>(SK_FORMAT_VERSION)?;
+        WriteBytesExt::write_u32::<LittleEndian>(&mut buf, SK_MAGIC)?;
+        WriteBytesExt::write_u32::<LittleEndian>(&mut buf, SK_FORMAT_VERSION)?;
        s.ser_into(&mut buf)?;

        // calculate checksum before resize
        let checksum = crc32c::crc32c(&buf);
        buf.extend_from_slice(&checksum.to_le_bytes());

-        control_partial.write_all(&buf).with_context(|| {
+        control_partial.write_all(&buf).await.with_context(|| {
            format!(
                "failed to write safekeeper state into control file at: {}",
                control_partial_path.display()
@@ -191,7 +194,7 @@ impl Storage for FileStorage {

        // fsync the file
        if !self.conf.no_sync {
-            control_partial.sync_all().with_context(|| {
+            control_partial.sync_all().await.with_context(|| {
                format!(
                    "failed to sync partial control file at {}",
                    control_partial_path.display()
@@ -202,21 +205,22 @@ impl Storage for FileStorage {
        let control_path = self.timeline_dir.join(CONTROL_FILE_NAME);

        // rename should be atomic
-        fs::rename(&control_partial_path, &control_path)?;
+        fs::rename(&control_partial_path, &control_path).await?;
        // this sync is not required by any standard but postgres does this (see durable_rename)
        if !self.conf.no_sync {
-            File::open(&control_path)
-                .and_then(|f| f.sync_all())
-                .with_context(|| {
-                    format!(
-                        "failed to sync control file at: {}",
-                        &control_path.display()
-                    )
-                })?;
+            let new_f = File::open(&control_path).await?;
+            new_f.sync_all().await.with_context(|| {
+                format!(
+                    "failed to sync control file at: {}",
+                    &control_path.display()
+                )
+            })?;

            // fsync the directory (linux specific)
-            File::open(&self.timeline_dir)
-                .and_then(|f| f.sync_all())
+            let tli_dir = File::open(&self.timeline_dir).await?;
+            tli_dir
+                .sync_all()
+                .await
                .context("failed to sync control file directory")?;
        }

@@ -236,7 +240,6 @@ mod test {
    use super::*;
    use crate::{safekeeper::SafeKeeperState, SafeKeeperConf};
    use anyhow::Result;
-    use std::fs;
    use utils::{id::TenantTimelineId, lsn::Lsn};

    fn stub_conf() -> SafeKeeperConf {
@@ -247,59 +250,75 @@ mod test {
        }
    }

-    fn load_from_control_file(
+    async fn load_from_control_file(
        conf: &SafeKeeperConf,
        ttid: &TenantTimelineId,
    ) -> Result<(FileStorage, SafeKeeperState)> {
-        fs::create_dir_all(conf.timeline_dir(ttid)).expect("failed to create timeline dir");
+        fs::create_dir_all(conf.timeline_dir(ttid))
+            .await
+            .expect("failed to create timeline dir");
        Ok((
            FileStorage::restore_new(ttid, conf)?,
            FileStorage::load_control_file_conf(conf, ttid)?,
        ))
    }

-    fn create(
+    async fn create(
        conf: &SafeKeeperConf,
        ttid: &TenantTimelineId,
    ) -> Result<(FileStorage, SafeKeeperState)> {
-        fs::create_dir_all(conf.timeline_dir(ttid)).expect("failed to create timeline dir");
+        fs::create_dir_all(conf.timeline_dir(ttid))
+            .await
+            .expect("failed to create timeline dir");
        let state = SafeKeeperState::empty();
        let storage = FileStorage::create_new(ttid, conf, state.clone())?;
        Ok((storage, state))
    }

-    #[test]
-    fn test_read_write_safekeeper_state() {
+    #[tokio::test]
+    async fn test_read_write_safekeeper_state() {
        let conf = stub_conf();
        let ttid = TenantTimelineId::generate();
        {
-            let (mut storage, mut state) = create(&conf, &ttid).expect("failed to create state");
+            let (mut storage, mut state) =
+                create(&conf, &ttid).await.expect("failed to create state");
            // change something
            state.commit_lsn = Lsn(42);
-            storage.persist(&state).expect("failed to persist state");
+            storage
+                .persist(&state)
+                .await
+                .expect("failed to persist state");
        }

-        let (_, state) = load_from_control_file(&conf, &ttid).expect("failed to read state");
+        let (_, state) = load_from_control_file(&conf, &ttid)
+            .await
+            .expect("failed to read state");
        assert_eq!(state.commit_lsn, Lsn(42));
    }

-    #[test]
-    fn test_safekeeper_state_checksum_mismatch() {
+    #[tokio::test]
+    async fn test_safekeeper_state_checksum_mismatch() {
        let conf = stub_conf();
        let ttid = TenantTimelineId::generate();
        {
-            let (mut storage, mut state) = create(&conf, &ttid).expect("failed to read state");
+            let (mut storage, mut state) =
+                create(&conf, &ttid).await.expect("failed to read state");

            // change something
            state.commit_lsn = Lsn(42);
-            storage.persist(&state).expect("failed to persist state");
+            storage
+                .persist(&state)
+                .await
+                .expect("failed to persist state");
        }
        let control_path = conf.timeline_dir(&ttid).join(CONTROL_FILE_NAME);
-        let mut data = fs::read(&control_path).unwrap();
+        let mut data = fs::read(&control_path).await.unwrap();
        data[0] += 1; // change the first byte of the file to fail checksum validation
-        fs::write(&control_path, &data).expect("failed to write control file");
+        fs::write(&control_path, &data)
+            .await
+            .expect("failed to write control file");

-        match load_from_control_file(&conf, &ttid) {
+        match load_from_control_file(&conf, &ttid).await {
            Err(err) => assert!(err
                .to_string()
                .contains("safekeeper control file checksum mismatch")),
--- a/safekeeper/src/debug_dump.rs
+++ b/safekeeper/src/debug_dump.rs
@@ -121,7 +121,7 @@ pub struct FileInfo {
 }

 /// Build debug dump response, using the provided [`Args`] filters.
-pub fn build(args: Args) -> Result<Response> {
+pub async fn build(args: Args) -> Result<Response> {
    let start_time = Utc::now();
    let timelines_count = GlobalTimelines::timelines_count();

@@ -155,7 +155,7 @@ pub fn build(args: Args) -> Result<Response> {
        }

        let control_file = if args.dump_control_file {
-            let mut state = tli.get_state().1;
+            let mut state = tli.get_state().await.1;
            if !args.dump_term_history {
                state.acceptor_state.term_history = TermHistory(vec![]);
            }
@@ -165,7 +165,7 @@ pub fn build(args: Args) -> Result<Response> {
        };

        let memory = if args.dump_memory {
-            Some(tli.memory_dump())
+            Some(tli.memory_dump().await)
        } else {
            None
        };
--- a/safekeeper/src/handler.rs
+++ b/safekeeper/src/handler.rs
@@ -256,14 +256,14 @@ impl SafekeeperPostgresHandler {

        let lsn = if self.is_walproposer_recovery() {
            // walproposer should get all local WAL until flush_lsn
-            tli.get_flush_lsn()
+            tli.get_flush_lsn().await
        } else {
            // other clients shouldn't get any uncommitted WAL
-            tli.get_state().0.commit_lsn
+            tli.get_state().await.0.commit_lsn
        }
        .to_string();

-        let sysid = tli.get_state().1.server.system_id.to_string();
+        let sysid = tli.get_state().await.1.server.system_id.to_string();
        let lsn_bytes = lsn.as_bytes();
        let tli = PG_TLI.to_string();
        let tli_bytes = tli.as_bytes();
--- a/Show More
+++ b/Show More