Test

# Problem

As pointed out through doc-comments in this PR, `drop_old_connection` is
not cancellation-safe.

This means we can leave a `handle_walreceiver_connection` tokio task
dangling during Timeline shutdown.

More details described in the corresponding issue #7062.

# Solution

Don't cancel-by-drop the `connection_manager_loop_step` from the
`tokio::select!()` in the task_mgr task.
Instead, transform the code to use a `CancellationToken` ---
specifically, `task_mgr::shutdown_token()` --- and make code responsive
to it.

The `drop_old_connection()` is still not cancellation-safe and also
doesn't get a cancellation token, because there's no point inside the
function where we could return early if cancellation were requested
using a token.

We rely on the `handle_walreceiver_connection` to be sensitive to the
`TaskHandle`s cancellation token (argument name: `cancellation`).
Currently it checks for `cancellation` on each WAL message. It is
probably also sensitive to `Timeline::cancel` because ultimately all
that `handle_walreceiver_connection` does is interact with the
`Timeline`.

In summary, the above means that the following code (which is found in
`Timeline::shutdown`) now might **take longer**, but actually ensures
that all `handle_walreceiver_connection` tasks are finished:

```rust
task_mgr::shutdown_tasks(
    Some(TaskKind::WalReceiverManager),
    Some(self.tenant_shard_id),
    Some(self.timeline_id)
)
```

# Refs

refs #7062

.github/workflows/build_and_test.yml

@@ -1121,10 +1121,16 @@ jobs:
         run: |
           if [[ "$GITHUB_REF_NAME" == "main" ]]; then
             gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f deployPreprodRegion=false
-
-            # TODO: move deployPreprodRegion to release (`"$GITHUB_REF_NAME" == "release"` block), once Staging support different compute tag prefixes for different regions
-            gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f deployPreprodRegion=true
           elif [[ "$GITHUB_REF_NAME" == "release" ]]; then
+            gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main \
+              -f deployPgSniRouter=false \
+              -f deployProxy=false \
+              -f deployStorage=true \
+              -f deployStorageBroker=true \
+              -f branch=main \
+              -f dockerTag=${{needs.tag.outputs.build-tag}} \
+              -f deployPreprodRegion=true
+
             gh workflow --repo neondatabase/aws run deploy-prod.yml --ref main \
               -f deployPgSniRouter=false \
               -f deployProxy=false \
@@ -1133,6 +1139,15 @@ jobs:
               -f branch=main \
               -f dockerTag=${{needs.tag.outputs.build-tag}}
           elif [[ "$GITHUB_REF_NAME" == "release-proxy" ]]; then
+            gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main \
+              -f deployPgSniRouter=true \
+              -f deployProxy=true \
+              -f deployStorage=false \
+              -f deployStorageBroker=false \
+              -f branch=main \
+              -f dockerTag=${{needs.tag.outputs.build-tag}} \
+              -f deployPreprodRegion=true
+
             gh workflow --repo neondatabase/aws run deploy-proxy-prod.yml --ref main \
               -f deployPgSniRouter=true \
               -f deployProxy=true \

CODEOWNERS

@@ -1,12 +1,13 @@
 /compute_tools/ @neondatabase/control-plane @neondatabase/compute
 /control_plane/attachment_service @neondatabase/storage
 /libs/pageserver_api/ @neondatabase/storage
-/libs/postgres_ffi/ @neondatabase/compute
+/libs/postgres_ffi/ @neondatabase/compute @neondatabase/safekeepers
 /libs/remote_storage/ @neondatabase/storage
 /libs/safekeeper_api/ @neondatabase/safekeepers
 /libs/vm_monitor/ @neondatabase/autoscaling
 /pageserver/ @neondatabase/storage
 /pgxn/ @neondatabase/compute
+/pgxn/neon/ @neondatabase/compute @neondatabase/safekeepers
 /proxy/ @neondatabase/proxy
 /safekeeper/ @neondatabase/safekeepers
 /vendor/ @neondatabase/compute

Cargo.lock

@@ -276,7 +276,7 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "aws-config",
- "aws-sdk-secretsmanager",
+ "bytes",
  "camino",
  "clap",
  "control_plane",
@@ -288,6 +288,8 @@ dependencies = [
  "hex",
  "humantime",
  "hyper",
+ "lasso",
+ "measured",
  "metrics",
  "once_cell",
  "pageserver_api",
@@ -295,6 +297,7 @@ dependencies = [
  "postgres_connection",
  "r2d2",
  "reqwest",
+ "routerify",
  "serde",
  "serde_json",
  "thiserror",
@@ -343,9 +346,9 @@ dependencies = [
 
 [[package]]
 name = "aws-credential-types"
-version = "1.1.4"
+version = "1.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "33cc49dcdd31c8b6e79850a179af4c367669150c7ac0135f176c61bec81a70f7"
+checksum = "fa8587ae17c8e967e4b05a62d495be2fb7701bec52a97f7acfe8a29f938384c8"
 dependencies = [
  "aws-smithy-async",
  "aws-smithy-runtime-api",
@@ -355,9 +358,9 @@ dependencies = [
 
 [[package]]
 name = "aws-runtime"
-version = "1.1.4"
+version = "1.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eb031bff99877c26c28895766f7bb8484a05e24547e370768d6cc9db514662aa"
+checksum = "b13dc54b4b49f8288532334bba8f87386a40571c47c37b1304979b556dc613c8"
 dependencies = [
  "aws-credential-types",
  "aws-sigv4",
@@ -377,6 +380,29 @@ dependencies = [
  "uuid",
 ]
 
+[[package]]
+name = "aws-sdk-iam"
+version = "1.17.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8ae76026bfb1b80a6aed0bb400c1139cd9c0563e26bce1986cd021c6a968c7b"
+dependencies = [
+ "aws-credential-types",
+ "aws-runtime",
+ "aws-smithy-async",
+ "aws-smithy-http",
+ "aws-smithy-json",
+ "aws-smithy-query",
+ "aws-smithy-runtime",
+ "aws-smithy-runtime-api",
+ "aws-smithy-types",
+ "aws-smithy-xml",
+ "aws-types",
+ "http 0.2.9",
+ "once_cell",
+ "regex-lite",
+ "tracing",
+]
+
 [[package]]
 name = "aws-sdk-s3"
 version = "1.14.0"
@@ -406,29 +432,6 @@ dependencies = [
  "url",
 ]
 
-[[package]]
-name = "aws-sdk-secretsmanager"
-version = "1.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0a0b64e61e7d632d9df90a2e0f32630c68c24960cab1d27d848718180af883d3"
-dependencies = [
- "aws-credential-types",
- "aws-runtime",
- "aws-smithy-async",
- "aws-smithy-http",
- "aws-smithy-json",
- "aws-smithy-runtime",
- "aws-smithy-runtime-api",
- "aws-smithy-types",
- "aws-types",
- "bytes",
- "fastrand 2.0.0",
- "http 0.2.9",
- "once_cell",
- "regex-lite",
- "tracing",
-]
-
 [[package]]
 name = "aws-sdk-sso"
 version = "1.12.0"
@@ -498,9 +501,9 @@ dependencies = [
 
 [[package]]
 name = "aws-sigv4"
-version = "1.1.4"
+version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c371c6b0ac54d4605eb6f016624fb5c7c2925d315fdf600ac1bf21b19d5f1742"
+checksum = "11d6f29688a4be9895c0ba8bef861ad0c0dac5c15e9618b9b7a6c233990fc263"
 dependencies = [
  "aws-credential-types",
  "aws-smithy-eventstream",
@@ -513,7 +516,7 @@ dependencies = [
  "hex",
  "hmac",
  "http 0.2.9",
- "http 1.0.0",
+ "http 1.1.0",
  "once_cell",
  "p256",
  "percent-encoding",
@@ -527,9 +530,9 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-async"
-version = "1.1.4"
+version = "1.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "72ee2d09cce0ef3ae526679b522835d63e75fb427aca5413cd371e490d52dcc6"
+checksum = "d26ea8fa03025b2face2b3038a63525a10891e3d8829901d502e5384a0d8cd46"
 dependencies = [
  "futures-util",
  "pin-project-lite",
@@ -570,9 +573,9 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-http"
-version = "0.60.4"
+version = "0.60.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dab56aea3cd9e1101a0a999447fb346afb680ab1406cebc44b32346e25b4117d"
+checksum = "3f10fa66956f01540051b0aa7ad54574640f748f9839e843442d99b970d3aff9"
 dependencies = [
  "aws-smithy-eventstream",
  "aws-smithy-runtime-api",
@@ -591,18 +594,18 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-json"
-version = "0.60.4"
+version = "0.60.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fd3898ca6518f9215f62678870064398f00031912390efd03f1f6ef56d83aa8e"
+checksum = "4683df9469ef09468dad3473d129960119a0d3593617542b7d52086c8486f2d6"
 dependencies = [
  "aws-smithy-types",
 ]
 
 [[package]]
 name = "aws-smithy-query"
-version = "0.60.4"
+version = "0.60.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bda4b1dfc9810e35fba8a620e900522cd1bd4f9578c446e82f49d1ce41d2e9f9"
+checksum = "f2fbd61ceb3fe8a1cb7352e42689cec5335833cd9f94103a61e98f9bb61c64bb"
 dependencies = [
  "aws-smithy-types",
  "urlencoding",
@@ -610,9 +613,9 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-runtime"
-version = "1.1.4"
+version = "1.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fafdab38f40ad7816e7da5dec279400dd505160780083759f01441af1bbb10ea"
+checksum = "ec81002d883e5a7fd2bb063d6fb51c4999eb55d404f4fff3dd878bf4733b9f01"
 dependencies = [
  "aws-smithy-async",
  "aws-smithy-http",
@@ -635,14 +638,15 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-runtime-api"
-version = "1.1.4"
+version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c18276dd28852f34b3bf501f4f3719781f4999a51c7bff1a5c6dc8c4529adc29"
+checksum = "9acb931e0adaf5132de878f1398d83f8677f90ba70f01f65ff87f6d7244be1c5"
 dependencies = [
  "aws-smithy-async",
  "aws-smithy-types",
  "bytes",
  "http 0.2.9",
+ "http 1.1.0",
  "pin-project-lite",
  "tokio",
  "tracing",
@@ -651,9 +655,9 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-types"
-version = "1.1.4"
+version = "1.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bb3e134004170d3303718baa2a4eb4ca64ee0a1c0a7041dca31b38be0fb414f3"
+checksum = "abe14dceea1e70101d38fbf2a99e6a34159477c0fb95e68e05c66bd7ae4c3729"
 dependencies = [
  "base64-simd",
  "bytes",
@@ -674,18 +678,18 @@ dependencies = [
 
 [[package]]
 name = "aws-smithy-xml"
-version = "0.60.4"
+version = "0.60.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8604a11b25e9ecaf32f9aa56b9fe253c5e2f606a3477f0071e96d3155a5ed218"
+checksum = "872c68cf019c0e4afc5de7753c4f7288ce4b71663212771bf5e4542eb9346ca9"
 dependencies = [
  "xmlparser",
 ]
 
 [[package]]
 name = "aws-types"
-version = "1.1.4"
+version = "1.1.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "789bbe008e65636fe1b6dbbb374c40c8960d1232b96af5ff4aec349f9c4accf4"
+checksum = "0dbf2f3da841a8930f159163175cf6a3d16ddde517c1b0fba7aa776822800f40"
 dependencies = [
  "aws-credential-types",
  "aws-smithy-async",
@@ -2392,9 +2396,9 @@ dependencies = [
 
 [[package]]
 name = "http"
-version = "1.0.0"
+version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b32afd38673a8016f7c9ae69e5af41a58f81b1d31689040f2f1959594ce194ea"
+checksum = "21b9ddb458710bc376481b842f5da65cdf31522de232c1ca8146abce2a358258"
 dependencies = [
  "bytes",
  "fnv",
@@ -2494,7 +2498,7 @@ dependencies = [
  "hyper",
  "log",
  "rustls 0.21.9",
- "rustls-native-certs",
+ "rustls-native-certs 0.6.2",
  "tokio",
  "tokio-rustls 0.24.0",
 ]
@@ -2880,6 +2884,35 @@ version = "0.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "490cc448043f947bae3cbee9c203358d62dbee0db12107a74be5c30ccfd09771"
 
+[[package]]
+name = "measured"
+version = "0.0.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f246648d027839a34b420e27c7de1165ace96e19ef894985d0a6ff89a7840a9f"
+dependencies = [
+ "bytes",
+ "hashbrown 0.14.0",
+ "itoa",
+ "lasso",
+ "measured-derive",
+ "memchr",
+ "parking_lot 0.12.1",
+ "rustc-hash",
+ "ryu",
+]
+
+[[package]]
+name = "measured-derive"
+version = "0.0.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "edaa5cc22d99d5d6d7d99c3b5b5f7e7f8034c22f1b5d62a1adecd2ed005d9b80"
+dependencies = [
+ "heck",
+ "proc-macro2",
+ "quote",
+ "syn 2.0.52",
+]
+
 [[package]]
 name = "memchr"
 version = "2.6.4"
@@ -3548,6 +3581,7 @@ dependencies = [
  "strum_macros",
  "svg_fmt",
  "sync_wrapper",
+ "sysinfo",
  "tenant_size_model",
  "thiserror",
  "tokio",
@@ -3901,7 +3935,7 @@ dependencies = [
 [[package]]
 name = "postgres"
 version = "0.19.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#988d0ddb4184c408fa7fc1bd0ecca7993c02978f"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#20031d7a9ee1addeae6e0968e3899ae6bf01cee2"
 dependencies = [
  "bytes",
  "fallible-iterator",
@@ -3914,7 +3948,7 @@ dependencies = [
 [[package]]
 name = "postgres-native-tls"
 version = "0.5.0"
-source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#988d0ddb4184c408fa7fc1bd0ecca7993c02978f"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#20031d7a9ee1addeae6e0968e3899ae6bf01cee2"
 dependencies = [
  "native-tls",
  "tokio",
@@ -3925,7 +3959,7 @@ dependencies = [
 [[package]]
 name = "postgres-protocol"
 version = "0.6.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#988d0ddb4184c408fa7fc1bd0ecca7993c02978f"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#20031d7a9ee1addeae6e0968e3899ae6bf01cee2"
 dependencies = [
  "base64 0.20.0",
  "byteorder",
@@ -3938,12 +3972,13 @@ dependencies = [
  "rand 0.8.5",
  "sha2",
  "stringprep",
+ "tokio",
 ]
 
 [[package]]
 name = "postgres-types"
 version = "0.2.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#988d0ddb4184c408fa7fc1bd0ecca7993c02978f"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#20031d7a9ee1addeae6e0968e3899ae6bf01cee2"
 dependencies = [
  "bytes",
  "fallible-iterator",
@@ -4165,6 +4200,10 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "async-trait",
+ "aws-config",
+ "aws-sdk-iam",
+ "aws-sigv4",
+ "aws-types",
  "base64 0.13.1",
  "bstr",
  "bytes",
@@ -4175,6 +4214,7 @@ dependencies = [
  "consumption_metrics",
  "dashmap",
  "env_logger",
+ "fallible-iterator",
  "futures",
  "git-version",
  "hashbrown 0.13.2",
@@ -4182,6 +4222,7 @@ dependencies = [
  "hex",
  "hmac",
  "hostname",
+ "http 1.1.0",
  "humantime",
  "hyper",
  "hyper-tungstenite",
@@ -4225,6 +4266,7 @@ dependencies = [
  "smallvec",
  "smol_str",
  "socket2 0.5.5",
+ "subtle",
  "sync_wrapper",
  "task-local-extensions",
  "thiserror",
@@ -4396,9 +4438,9 @@ dependencies = [
 
 [[package]]
 name = "redis"
-version = "0.24.0"
+version = "0.25.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c580d9cbbe1d1b479e8d67cf9daf6a62c957e6846048408b80b43ac3f6af84cd"
+checksum = "71d64e978fd98a0e6b105d066ba4889a7301fca65aeac850a877d8797343feeb"
 dependencies = [
  "async-trait",
  "bytes",
@@ -4407,15 +4449,15 @@ dependencies = [
  "itoa",
  "percent-encoding",
  "pin-project-lite",
- "rustls 0.21.9",
- "rustls-native-certs",
- "rustls-pemfile 1.0.2",
- "rustls-webpki 0.101.7",
+ "rustls 0.22.2",
+ "rustls-native-certs 0.7.0",
+ "rustls-pemfile 2.1.1",
+ "rustls-pki-types",
  "ryu",
  "sha1_smol",
- "socket2 0.4.9",
+ "socket2 0.5.5",
  "tokio",
- "tokio-rustls 0.24.0",
+ "tokio-rustls 0.25.0",
  "tokio-util",
  "url",
 ]
@@ -4844,6 +4886,19 @@ dependencies = [
  "security-framework",
 ]
 
+[[package]]
+name = "rustls-native-certs"
+version = "0.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f1fb85efa936c42c6d5fc28d2629bb51e4b2f4b8a5211e297d599cc5a093792"
+dependencies = [
+ "openssl-probe",
+ "rustls-pemfile 2.1.1",
+ "rustls-pki-types",
+ "schannel",
+ "security-framework",
+]
+
 [[package]]
 name = "rustls-pemfile"
 version = "1.0.2"
@@ -5346,13 +5401,23 @@ checksum = "ae1a47186c03a32177042e55dbc5fd5aee900b8e0069a8d70fba96a9375cd012"
 
 [[package]]
 name = "sha2"
-version = "0.10.6"
+version = "0.10.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "82e6b795fe2e3b1e845bafcb27aa35405c4d47cdfc92af5fc8d3002f76cebdc0"
+checksum = "793db75ad2bcafc3ffa7c68b215fee268f537982cd901d132f89c6343f3a3dc8"
 dependencies = [
  "cfg-if",
  "cpufeatures",
  "digest",
+ "sha2-asm",
+]
+
+[[package]]
+name = "sha2-asm"
+version = "0.6.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f27ba7066011e3fb30d808b51affff34f0a66d3a03a58edd787c6e420e40e44e"
+dependencies = [
+ "cc",
 ]
 
 [[package]]
@@ -5935,7 +6000,7 @@ dependencies = [
 [[package]]
 name = "tokio-postgres"
 version = "0.7.7"
-source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#988d0ddb4184c408fa7fc1bd0ecca7993c02978f"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#20031d7a9ee1addeae6e0968e3899ae6bf01cee2"
 dependencies = [
  "async-trait",
  "byteorder",
@@ -6101,7 +6166,7 @@ dependencies = [
  "percent-encoding",
  "pin-project",
  "prost",
- "rustls-native-certs",
+ "rustls-native-certs 0.6.2",
  "rustls-pemfile 1.0.2",
  "tokio",
  "tokio-rustls 0.24.0",
@@ -6468,6 +6533,7 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "arc-swap",
+ "async-compression",
  "async-trait",
  "bincode",
  "byteorder",
@@ -6506,12 +6572,14 @@ dependencies = [
  "thiserror",
  "tokio",
  "tokio-stream",
+ "tokio-tar",
  "tokio-util",
  "tracing",
  "tracing-error",
  "tracing-subscriber",
  "url",
  "uuid",
+ "walkdir",
  "workspace_hack",
 ]
 
@@ -6983,7 +7051,6 @@ dependencies = [
  "aws-sigv4",
  "aws-smithy-async",
  "aws-smithy-http",
- "aws-smithy-runtime-api",
  "aws-smithy-types",
  "axum",
  "base64 0.21.1",
@@ -7029,6 +7096,7 @@ dependencies = [
  "scopeguard",
  "serde",
  "serde_json",
+ "sha2",
  "smallvec",
  "subtle",
  "syn 1.0.109",

Cargo.toml

@@ -52,10 +52,12 @@ async-stream = "0.3"
 async-trait = "0.1"
 aws-config = { version = "1.1.4", default-features = false, features=["rustls"] }
 aws-sdk-s3 = "1.14"
-aws-sdk-secretsmanager = { version = "1.14.0" }
+aws-sdk-iam = "1.15.0"
 aws-smithy-async = { version = "1.1.4", default-features = false, features=["rt-tokio"] }
 aws-smithy-types = "1.1.4"
 aws-credential-types = "1.1.4"
+aws-sigv4 = { version = "1.2.0", features = ["sign-http"] }
+aws-types = "1.1.7"
 axum = { version = "0.6.20", features = ["ws"] }
 base64 = "0.13.0"
 bincode = "1.3"
@@ -76,6 +78,7 @@ either = "1.8"
 enum-map = "2.4.2"
 enumset = "1.0.12"
 fail = "0.5.0"
+fallible-iterator = "0.2"
 fs2 = "0.4.3"
 futures = "0.3"
 futures-core = "0.3"
@@ -88,6 +91,7 @@ hex = "0.4"
 hex-literal = "0.4"
 hmac = "0.12.1"
 hostname = "0.3.1"
+http = {version = "1.1.0", features = ["std"]}
 http-types = { version = "2", default-features = false }
 humantime = "2.1"
 humantime-serde = "1.1.1"
@@ -101,6 +105,7 @@ lasso = "0.7"
 leaky-bucket = "1.0.1"
 libc = "0.2"
 md5 = "0.7.0"
+measured = { version = "0.0.13", features=["default", "lasso"] }
 memoffset = "0.8"
 native-tls = "0.2"
 nix = { version = "0.27", features = ["fs", "process", "socket", "signal", "poll"] }
@@ -120,7 +125,7 @@ procfs = "0.14"
 prometheus = {version = "0.13", default_features=false, features = ["process"]} # removes protobuf dependency
 prost = "0.11"
 rand = "0.8"
-redis = { version = "0.24.0", features = ["tokio-rustls-comp", "keep-alive"] }
+redis = { version = "0.25.2", features = ["tokio-rustls-comp", "keep-alive"] }
 regex = "1.10.2"
 reqwest = { version = "0.11", default-features = false, features = ["rustls-tls"] }
 reqwest-tracing = { version = "0.4.7", features = ["opentelemetry_0_20"] }
@@ -148,6 +153,7 @@ smol_str = { version = "0.2.0", features = ["serde"] }
 socket2 = "0.5"
 strum = "0.24"
 strum_macros = "0.24"
+"subtle"  = "2.5.0"
 svg_fmt = "0.4.1"
 sync_wrapper = "0.1.2"
 tar = "0.4"

Dockerfile.build-tools

@@ -135,7 +135,7 @@ WORKDIR /home/nonroot
 
 # Rust
 # Please keep the version of llvm (installed above) in sync with rust llvm (`rustc --version --verbose | grep LLVM`)
-ENV RUSTC_VERSION=1.76.0
+ENV RUSTC_VERSION=1.77.0
 ENV RUSTUP_HOME="/home/nonroot/.rustup"
 ENV PATH="/home/nonroot/.cargo/bin:${PATH}"
 RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && whoami && \
@@ -149,7 +149,7 @@ RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux
     cargo install --git https://github.com/paritytech/cachepot && \
     cargo install rustfilt && \
     cargo install cargo-hakari && \
-    cargo install cargo-deny && \
+    cargo install cargo-deny --locked && \
     cargo install cargo-hack && \
     cargo install cargo-nextest && \
     rm -rf /home/nonroot/.cargo/registry && \

compute_tools/README.md

@@ -32,6 +32,29 @@ compute_ctl -D /var/db/postgres/compute \
             -b /usr/local/bin/postgres
 ```
 
+## State Diagram
+
+Computes can be in various states. Below is a diagram that details how a
+compute moves between states.
+
+```mermaid
+%% https://mermaid.js.org/syntax/stateDiagram.html
+stateDiagram-v2
+  [*] --> Empty : Compute spawned
+  Empty --> ConfigurationPending : Waiting for compute spec
+  ConfigurationPending --> Configuration : Received compute spec
+  Configuration --> Failed : Failed to configure the compute
+  Configuration --> Running : Compute has been configured
+  Empty --> Init : Compute spec is immediately available
+  Empty --> TerminationPending : Requested termination
+  Init --> Failed : Failed to start Postgres
+  Init --> Running : Started Postgres
+  Running --> TerminationPending : Requested termination
+  TerminationPending --> Terminated : Terminated compute
+  Failed --> [*] : Compute exited
+  Terminated --> [*] : Compute exited
+```
+
 ## Tests
 
 Cargo formatter:

compute_tools/src/config.rs

@@ -17,6 +17,7 @@ pub fn line_in_file(path: &Path, line: &str) -> Result<bool> {
         .write(true)
         .create(true)
         .append(false)
+        .truncate(false)
         .open(path)?;
     let buf = io::BufReader::new(&file);
     let mut count: usize = 0;

control_plane/attachment_service/Cargo.toml

@@ -16,7 +16,7 @@ testing = []
 [dependencies]
 anyhow.workspace = true
 aws-config.workspace = true
-aws-sdk-secretsmanager.workspace = true
+bytes.workspace = true
 camino.workspace = true
 clap.workspace = true
 fail.workspace = true
@@ -25,17 +25,20 @@ git-version.workspace = true
 hex.workspace = true
 hyper.workspace = true
 humantime.workspace = true
+lasso.workspace = true
 once_cell.workspace = true
 pageserver_api.workspace = true
 pageserver_client.workspace = true
 postgres_connection.workspace = true
 reqwest.workspace = true
+routerify.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 thiserror.workspace = true
 tokio.workspace = true
 tokio-util.workspace = true
 tracing.workspace = true
+measured.workspace = true
 
 diesel = { version = "2.1.4", features = ["serde_json", "postgres", "r2d2"] }
 diesel_migrations = { version = "2.1.0" }

control_plane/attachment_service/migrations/2024-03-18-184429_rename_policy/down.sql

@@ -0,0 +1,3 @@
+
+UPDATE tenant_shards set placement_policy='{"Double": 1}' where placement_policy='{"Attached": 1}';
+UPDATE tenant_shards set placement_policy='"Single"' where placement_policy='{"Attached": 0}';

control_plane/attachment_service/migrations/2024-03-18-184429_rename_policy/up.sql

@@ -0,0 +1,3 @@
+
+UPDATE tenant_shards set placement_policy='{"Attached": 1}' where placement_policy='{"Double": 1}';
+UPDATE tenant_shards set placement_policy='{"Attached": 0}' where placement_policy='"Single"';

control_plane/attachment_service/src/heartbeater.rs

@@ -139,7 +139,7 @@ impl HeartbeaterTask {
                         .with_client_retries(
                             |client| async move { client.get_utilization().await },
                             &jwt_token,
-                            2,
+                            3,
                             3,
                             Duration::from_secs(1),
                             &cancel,

control_plane/attachment_service/src/http.rs

@@ -1,5 +1,11 @@
+use crate::metrics::{
+    HttpRequestLatencyLabelGroup, HttpRequestStatusLabelGroup, PageserverRequestLabelGroup,
+    METRICS_REGISTRY,
+};
 use crate::reconciler::ReconcileError;
 use crate::service::{Service, STARTUP_RECONCILE_TIMEOUT};
+use futures::Future;
+use hyper::header::CONTENT_TYPE;
 use hyper::{Body, Request, Response};
 use hyper::{StatusCode, Uri};
 use pageserver_api::models::{
@@ -34,6 +40,8 @@ use pageserver_api::upcall_api::{ReAttachRequest, ValidateRequest};
 
 use control_plane::storage_controller::{AttachHookRequest, InspectRequest};
 
+use routerify::Middleware;
+
 /// State available to HTTP request handlers
 #[derive(Clone)]
 pub struct HttpState {
@@ -313,7 +321,7 @@ async fn handle_tenant_timeline_passthrough(
     tracing::info!("Proxying request for tenant {} ({})", tenant_id, path);
 
     // Find the node that holds shard zero
-    let (base_url, tenant_shard_id) = service.tenant_shard0_baseurl(tenant_id)?;
+    let (node, tenant_shard_id) = service.tenant_shard0_node(tenant_id)?;
 
     // Callers will always pass an unsharded tenant ID.  Before proxying, we must
     // rewrite this to a shard-aware shard zero ID.
@@ -322,12 +330,39 @@ async fn handle_tenant_timeline_passthrough(
     let tenant_shard_str = format!("{}", tenant_shard_id);
     let path = path.replace(&tenant_str, &tenant_shard_str);
 
-    let client = mgmt_api::Client::new(base_url, service.get_config().jwt_token.as_deref());
+    let latency = &METRICS_REGISTRY
+        .metrics_group
+        .storage_controller_passthrough_request_latency;
+
+    // This is a bit awkward. We remove the param from the request
+    // and join the words by '_' to get a label for the request.
+    let just_path = path.replace(&tenant_shard_str, "");
+    let path_label = just_path
+        .split('/')
+        .filter(|token| !token.is_empty())
+        .collect::<Vec<_>>()
+        .join("_");
+    let labels = PageserverRequestLabelGroup {
+        pageserver_id: &node.get_id().to_string(),
+        path: &path_label,
+        method: crate::metrics::Method::Get,
+    };
+
+    let _timer = latency.start_timer(labels.clone());
+
+    let client = mgmt_api::Client::new(node.base_url(), service.get_config().jwt_token.as_deref());
     let resp = client.get_raw(path).await.map_err(|_e|
         // FIXME: give APiError a proper Unavailable variant.  We return 503 here because
         // if we can't successfully send a request to the pageserver, we aren't available.
         ApiError::ShuttingDown)?;
 
+    if !resp.status().is_success() {
+        let error_counter = &METRICS_REGISTRY
+            .metrics_group
+            .storage_controller_passthrough_request_error;
+        error_counter.inc(labels);
+    }
+
     // We have a reqest::Response, would like a http::Response
     let mut builder = hyper::Response::builder()
         .status(resp.status())
@@ -353,6 +388,16 @@ async fn handle_tenant_locate(
     json_response(StatusCode::OK, service.tenant_locate(tenant_id)?)
 }
 
+async fn handle_tenant_describe(
+    service: Arc<Service>,
+    req: Request<Body>,
+) -> Result<Response<Body>, ApiError> {
+    check_permissions(&req, Scope::Admin)?;
+
+    let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
+    json_response(StatusCode::OK, service.tenant_describe(tenant_id)?)
+}
+
 async fn handle_node_register(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
     check_permissions(&req, Scope::Admin)?;
 
@@ -488,7 +533,11 @@ impl From<ReconcileError> for ApiError {
 
 /// Common wrapper for request handlers that call into Service and will operate on tenants: they must only
 /// be allowed to run if Service has finished its initial reconciliation.
-async fn tenant_service_handler<R, H>(request: Request<Body>, handler: H) -> R::Output
+async fn tenant_service_handler<R, H>(
+    request: Request<Body>,
+    handler: H,
+    request_name: RequestName,
+) -> R::Output
 where
     R: std::future::Future<Output = Result<Response<Body>, ApiError>> + Send + 'static,
     H: FnOnce(Arc<Service>, Request<Body>) -> R + Send + Sync + 'static,
@@ -508,9 +557,10 @@ where
         ));
     }
 
-    request_span(
+    named_request_span(
         request,
         |request| async move { handler(service, request).await },
+        request_name,
     )
     .await
 }
@@ -521,11 +571,98 @@ fn check_permissions(request: &Request<Body>, required_scope: Scope) -> Result<(
     })
 }
 
+#[derive(Clone, Debug)]
+struct RequestMeta {
+    method: hyper::http::Method,
+    at: Instant,
+}
+
+fn prologue_metrics_middleware<B: hyper::body::HttpBody + Send + Sync + 'static>(
+) -> Middleware<B, ApiError> {
+    Middleware::pre(move |req| async move {
+        let meta = RequestMeta {
+            method: req.method().clone(),
+            at: Instant::now(),
+        };
+
+        req.set_context(meta);
+
+        Ok(req)
+    })
+}
+
+fn epilogue_metrics_middleware<B: hyper::body::HttpBody + Send + Sync + 'static>(
+) -> Middleware<B, ApiError> {
+    Middleware::post_with_info(move |resp, req_info| async move {
+        let request_name = match req_info.context::<RequestName>() {
+            Some(name) => name,
+            None => {
+                return Ok(resp);
+            }
+        };
+
+        if let Some(meta) = req_info.context::<RequestMeta>() {
+            let status = &crate::metrics::METRICS_REGISTRY
+                .metrics_group
+                .storage_controller_http_request_status;
+            let latency = &crate::metrics::METRICS_REGISTRY
+                .metrics_group
+                .storage_controller_http_request_latency;
+
+            status.inc(HttpRequestStatusLabelGroup {
+                path: request_name.0,
+                method: meta.method.clone().into(),
+                status: crate::metrics::StatusCode(resp.status()),
+            });
+
+            latency.observe(
+                HttpRequestLatencyLabelGroup {
+                    path: request_name.0,
+                    method: meta.method.into(),
+                },
+                meta.at.elapsed().as_secs_f64(),
+            );
+        }
+        Ok(resp)
+    })
+}
+
+pub async fn measured_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    pub const TEXT_FORMAT: &str = "text/plain; version=0.0.4";
+
+    let payload = crate::metrics::METRICS_REGISTRY.encode();
+    let response = Response::builder()
+        .status(200)
+        .header(CONTENT_TYPE, TEXT_FORMAT)
+        .body(payload.into())
+        .unwrap();
+
+    Ok(response)
+}
+
+#[derive(Clone)]
+struct RequestName(&'static str);
+
+async fn named_request_span<R, H>(
+    request: Request<Body>,
+    handler: H,
+    name: RequestName,
+) -> R::Output
+where
+    R: Future<Output = Result<Response<Body>, ApiError>> + Send + 'static,
+    H: FnOnce(Request<Body>) -> R + Send + Sync + 'static,
+{
+    request.set_context(name);
+    request_span(request, handler).await
+}
+
 pub fn make_router(
     service: Arc<Service>,
     auth: Option<Arc<SwappableJwtAuth>>,
 ) -> RouterBuilder<hyper::Body, ApiError> {
-    let mut router = endpoint::make_router();
+    let mut router = endpoint::make_router()
+        .middleware(prologue_metrics_middleware())
+        .middleware(epilogue_metrics_middleware());
     if auth.is_some() {
         router = router.middleware(auth_middleware(|request| {
             let state = get_state(request);
@@ -534,96 +671,166 @@ pub fn make_router(
             } else {
                 state.auth.as_deref()
             }
-        }))
+        }));
     }
 
     router
         .data(Arc::new(HttpState::new(service, auth)))
+        .get("/metrics", |r| {
+            named_request_span(r, measured_metrics_handler, RequestName("metrics"))
+        })
         // Non-prefixed generic endpoints (status, metrics)
-        .get("/status", |r| request_span(r, handle_status))
-        .get("/ready", |r| request_span(r, handle_ready))
+        .get("/status", |r| {
+            named_request_span(r, handle_status, RequestName("status"))
+        })
+        .get("/ready", |r| {
+            named_request_span(r, handle_ready, RequestName("ready"))
+        })
         // Upcalls for the pageserver: point the pageserver's `control_plane_api` config to this prefix
         .post("/upcall/v1/re-attach", |r| {
-            request_span(r, handle_re_attach)
+            named_request_span(r, handle_re_attach, RequestName("upcall_v1_reattach"))
+        })
+        .post("/upcall/v1/validate", |r| {
+            named_request_span(r, handle_validate, RequestName("upcall_v1_validate"))
         })
-        .post("/upcall/v1/validate", |r| request_span(r, handle_validate))
         // Test/dev/debug endpoints
         .post("/debug/v1/attach-hook", |r| {
-            request_span(r, handle_attach_hook)
+            named_request_span(r, handle_attach_hook, RequestName("debug_v1_attach_hook"))
+        })
+        .post("/debug/v1/inspect", |r| {
+            named_request_span(r, handle_inspect, RequestName("debug_v1_inspect"))
         })
-        .post("/debug/v1/inspect", |r| request_span(r, handle_inspect))
         .post("/debug/v1/tenant/:tenant_id/drop", |r| {
-            request_span(r, handle_tenant_drop)
+            named_request_span(r, handle_tenant_drop, RequestName("debug_v1_tenant_drop"))
         })
         .post("/debug/v1/node/:node_id/drop", |r| {
-            request_span(r, handle_node_drop)
+            named_request_span(r, handle_node_drop, RequestName("debug_v1_node_drop"))
+        })
+        .get("/debug/v1/tenant", |r| {
+            named_request_span(r, handle_tenants_dump, RequestName("debug_v1_tenant"))
+        })
+        .get("/debug/v1/tenant/:tenant_id/locate", |r| {
+            tenant_service_handler(
+                r,
+                handle_tenant_locate,
+                RequestName("debug_v1_tenant_locate"),
+            )
         })
-        .get("/debug/v1/tenant", |r| request_span(r, handle_tenants_dump))
         .get("/debug/v1/scheduler", |r| {
-            request_span(r, handle_scheduler_dump)
+            named_request_span(r, handle_scheduler_dump, RequestName("debug_v1_scheduler"))
         })
         .post("/debug/v1/consistency_check", |r| {
-            request_span(r, handle_consistency_check)
+            named_request_span(
+                r,
+                handle_consistency_check,
+                RequestName("debug_v1_consistency_check"),
+            )
         })
         .put("/debug/v1/failpoints", |r| {
             request_span(r, |r| failpoints_handler(r, CancellationToken::new()))
         })
-        .get("/control/v1/tenant/:tenant_id/locate", |r| {
-            tenant_service_handler(r, handle_tenant_locate)
-        })
         // Node operations
         .post("/control/v1/node", |r| {
-            request_span(r, handle_node_register)
+            named_request_span(r, handle_node_register, RequestName("control_v1_node"))
+        })
+        .get("/control/v1/node", |r| {
+            named_request_span(r, handle_node_list, RequestName("control_v1_node"))
         })
-        .get("/control/v1/node", |r| request_span(r, handle_node_list))
         .put("/control/v1/node/:node_id/config", |r| {
-            request_span(r, handle_node_configure)
+            named_request_span(
+                r,
+                handle_node_configure,
+                RequestName("control_v1_node_config"),
+            )
         })
         // Tenant Shard operations
         .put("/control/v1/tenant/:tenant_shard_id/migrate", |r| {
-            tenant_service_handler(r, handle_tenant_shard_migrate)
+            tenant_service_handler(
+                r,
+                handle_tenant_shard_migrate,
+                RequestName("control_v1_tenant_migrate"),
+            )
         })
         .put("/control/v1/tenant/:tenant_id/shard_split", |r| {
-            tenant_service_handler(r, handle_tenant_shard_split)
+            tenant_service_handler(
+                r,
+                handle_tenant_shard_split,
+                RequestName("control_v1_tenant_shard_split"),
+            )
+        })
+        .get("/control/v1/tenant/:tenant_id", |r| {
+            tenant_service_handler(
+                r,
+                handle_tenant_describe,
+                RequestName("control_v1_tenant_describe"),
+            )
         })
         // Tenant operations
         // The ^/v1/ endpoints act as a "Virtual Pageserver", enabling shard-naive clients to call into
         // this service to manage tenants that actually consist of many tenant shards, as if they are a single entity.
         .post("/v1/tenant", |r| {
-            tenant_service_handler(r, handle_tenant_create)
+            tenant_service_handler(r, handle_tenant_create, RequestName("v1_tenant"))
         })
         .delete("/v1/tenant/:tenant_id", |r| {
-            tenant_service_handler(r, handle_tenant_delete)
+            tenant_service_handler(r, handle_tenant_delete, RequestName("v1_tenant"))
         })
         .put("/v1/tenant/config", |r| {
-            tenant_service_handler(r, handle_tenant_config_set)
+            tenant_service_handler(r, handle_tenant_config_set, RequestName("v1_tenant_config"))
         })
         .get("/v1/tenant/:tenant_id/config", |r| {
-            tenant_service_handler(r, handle_tenant_config_get)
+            tenant_service_handler(r, handle_tenant_config_get, RequestName("v1_tenant_config"))
         })
         .put("/v1/tenant/:tenant_shard_id/location_config", |r| {
-            tenant_service_handler(r, handle_tenant_location_config)
+            tenant_service_handler(
+                r,
+                handle_tenant_location_config,
+                RequestName("v1_tenant_location_config"),
+            )
         })
         .put("/v1/tenant/:tenant_id/time_travel_remote_storage", |r| {
-            tenant_service_handler(r, handle_tenant_time_travel_remote_storage)
+            tenant_service_handler(
+                r,
+                handle_tenant_time_travel_remote_storage,
+                RequestName("v1_tenant_time_travel_remote_storage"),
+            )
         })
         .post("/v1/tenant/:tenant_id/secondary/download", |r| {
-            tenant_service_handler(r, handle_tenant_secondary_download)
+            tenant_service_handler(
+                r,
+                handle_tenant_secondary_download,
+                RequestName("v1_tenant_secondary_download"),
+            )
         })
         // Timeline operations
         .delete("/v1/tenant/:tenant_id/timeline/:timeline_id", |r| {
-            tenant_service_handler(r, handle_tenant_timeline_delete)
+            tenant_service_handler(
+                r,
+                handle_tenant_timeline_delete,
+                RequestName("v1_tenant_timeline"),
+            )
         })
         .post("/v1/tenant/:tenant_id/timeline", |r| {
-            tenant_service_handler(r, handle_tenant_timeline_create)
+            tenant_service_handler(
+                r,
+                handle_tenant_timeline_create,
+                RequestName("v1_tenant_timeline"),
+            )
         })
         // Tenant detail GET passthrough to shard zero
         .get("/v1/tenant/:tenant_id", |r| {
-            tenant_service_handler(r, handle_tenant_timeline_passthrough)
+            tenant_service_handler(
+                r,
+                handle_tenant_timeline_passthrough,
+                RequestName("v1_tenant_passthrough"),
+            )
         })
         // Timeline GET passthrough to shard zero.  Note that the `*` in the URL is a wildcard: any future
         // timeline GET APIs will be implicitly included.
         .get("/v1/tenant/:tenant_id/timeline*", |r| {
-            tenant_service_handler(r, handle_tenant_timeline_passthrough)
+            tenant_service_handler(
+                r,
+                handle_tenant_timeline_passthrough,
+                RequestName("v1_tenant_timeline_passthrough"),
+            )
         })
 }

control_plane/attachment_service/src/lib.rs

@@ -8,6 +8,7 @@ pub mod http;
 mod id_lock_map;
 pub mod metrics;
 mod node;
+mod pageserver_client;
 pub mod persistence;
 mod reconciler;
 mod scheduler;

control_plane/attachment_service/src/main.rs

@@ -3,7 +3,6 @@ use attachment_service::http::make_router;
 use attachment_service::metrics::preinitialize_metrics;
 use attachment_service::persistence::Persistence;
 use attachment_service::service::{Config, Service, MAX_UNAVAILABLE_INTERVAL_DEFAULT};
-use aws_config::{BehaviorVersion, Region};
 use camino::Utf8PathBuf;
 use clap::Parser;
 use diesel::Connection;
@@ -55,11 +54,31 @@ struct Cli {
     #[arg(long)]
     database_url: Option<String>,
 
+    /// Flag to enable dev mode, which permits running without auth
+    #[arg(long, default_value = "false")]
+    dev: bool,
+
     /// Grace period before marking unresponsive pageserver offline
     #[arg(long)]
     max_unavailable_interval: Option<humantime::Duration>,
 }
 
+enum StrictMode {
+    /// In strict mode, we will require that all secrets are loaded, i.e. security features
+    /// may not be implicitly turned off by omitting secrets in the environment.
+    Strict,
+    /// In dev mode, secrets are optional, and omitting a particular secret will implicitly
+    /// disable the auth related to it (e.g. no pageserver jwt key -> send unauthenticated
+    /// requests, no public key -> don't authenticate incoming requests).
+    Dev,
+}
+
+impl Default for StrictMode {
+    fn default() -> Self {
+        Self::Strict
+    }
+}
+
 /// Secrets may either be provided on the command line (for testing), or loaded from AWS SecretManager: this
 /// type encapsulates the logic to decide which and do the loading.
 struct Secrets {
@@ -70,13 +89,6 @@ struct Secrets {
 }
 
 impl Secrets {
-    const DATABASE_URL_SECRET: &'static str = "rds-neon-storage-controller-url";
-    const PAGESERVER_JWT_TOKEN_SECRET: &'static str =
-        "neon-storage-controller-pageserver-jwt-token";
-    const CONTROL_PLANE_JWT_TOKEN_SECRET: &'static str =
-        "neon-storage-controller-control-plane-jwt-token";
-    const PUBLIC_KEY_SECRET: &'static str = "neon-storage-controller-public-key";
-
     const DATABASE_URL_ENV: &'static str = "DATABASE_URL";
     const PAGESERVER_JWT_TOKEN_ENV: &'static str = "PAGESERVER_JWT_TOKEN";
     const CONTROL_PLANE_JWT_TOKEN_ENV: &'static str = "CONTROL_PLANE_JWT_TOKEN";
@@ -87,111 +99,41 @@ impl Secrets {
     /// - Environment variables if DATABASE_URL is set.
     /// - AWS Secrets Manager secrets
     async fn load(args: &Cli) -> anyhow::Result<Self> {
-        match &args.database_url {
-            Some(url) => Self::load_cli(url, args),
-            None => match std::env::var(Self::DATABASE_URL_ENV) {
-                Ok(database_url) => Self::load_env(database_url),
-                Err(_) => Self::load_aws_sm().await,
-            },
-        }
-    }
-
-    fn load_env(database_url: String) -> anyhow::Result<Self> {
-        let public_key = match std::env::var(Self::PUBLIC_KEY_ENV) {
-            Ok(public_key) => Some(JwtAuth::from_key(public_key).context("Loading public key")?),
-            Err(_) => None,
-        };
-        Ok(Self {
-            database_url,
-            public_key,
-            jwt_token: std::env::var(Self::PAGESERVER_JWT_TOKEN_ENV).ok(),
-            control_plane_jwt_token: std::env::var(Self::CONTROL_PLANE_JWT_TOKEN_ENV).ok(),
-        })
-    }
-
-    async fn load_aws_sm() -> anyhow::Result<Self> {
-        let Ok(region) = std::env::var("AWS_REGION") else {
-            anyhow::bail!("AWS_REGION is not set, cannot load secrets automatically: either set this, or use CLI args to supply secrets");
-        };
-        let config = aws_config::defaults(BehaviorVersion::v2023_11_09())
-            .region(Region::new(region.clone()))
-            .load()
-            .await;
-
-        let asm = aws_sdk_secretsmanager::Client::new(&config);
-
-        let Some(database_url) = asm
-            .get_secret_value()
-            .secret_id(Self::DATABASE_URL_SECRET)
-            .send()
-            .await?
-            .secret_string()
-            .map(str::to_string)
+        let Some(database_url) =
+            Self::load_secret(&args.database_url, Self::DATABASE_URL_ENV).await
         else {
             anyhow::bail!(
-                "Database URL secret not found at {region}/{}",
-                Self::DATABASE_URL_SECRET
+                "Database URL is not set (set `--database-url`, or `DATABASE_URL` environment)"
             )
         };
 
-        let jwt_token = asm
-            .get_secret_value()
-            .secret_id(Self::PAGESERVER_JWT_TOKEN_SECRET)
-            .send()
-            .await?
-            .secret_string()
-            .map(str::to_string);
-        if jwt_token.is_none() {
-            tracing::warn!("No pageserver JWT token set: this will only work if authentication is disabled on the pageserver");
-        }
-
-        let control_plane_jwt_token = asm
-            .get_secret_value()
-            .secret_id(Self::CONTROL_PLANE_JWT_TOKEN_SECRET)
-            .send()
-            .await?
-            .secret_string()
-            .map(str::to_string);
-        if jwt_token.is_none() {
-            tracing::warn!("No control plane JWT token set: this will only work if authentication is disabled on the pageserver");
-        }
-
-        let public_key = asm
-            .get_secret_value()
-            .secret_id(Self::PUBLIC_KEY_SECRET)
-            .send()
-            .await?
-            .secret_string()
-            .map(str::to_string);
-        let public_key = match public_key {
-            Some(key) => Some(JwtAuth::from_key(key)?),
-            None => {
-                tracing::warn!(
-                    "No public key set: inccoming HTTP requests will not be authenticated"
-                );
-                None
-            }
+        let public_key = match Self::load_secret(&args.public_key, Self::PUBLIC_KEY_ENV).await {
+            Some(v) => Some(JwtAuth::from_key(v).context("Loading public key")?),
+            None => None,
         };
 
-        Ok(Self {
+        let this = Self {
             database_url,
             public_key,
-            jwt_token,
-            control_plane_jwt_token,
-        })
+            jwt_token: Self::load_secret(&args.jwt_token, Self::PAGESERVER_JWT_TOKEN_ENV).await,
+            control_plane_jwt_token: Self::load_secret(
+                &args.control_plane_jwt_token,
+                Self::CONTROL_PLANE_JWT_TOKEN_ENV,
+            )
+            .await,
+        };
+
+        Ok(this)
     }
 
-    fn load_cli(database_url: &str, args: &Cli) -> anyhow::Result<Self> {
-        let public_key = match &args.public_key {
-            None => None,
-            Some(key) => Some(JwtAuth::from_key(key.clone()).context("Loading public key")?),
-        };
-        Ok(Self {
-            database_url: database_url.to_owned(),
-            public_key,
-            jwt_token: args.jwt_token.clone(),
-            control_plane_jwt_token: args.control_plane_jwt_token.clone(),
-        })
+    async fn load_secret(cli: &Option<String>, env_name: &str) -> Option<String> {
+        if let Some(v) = cli {
+            Some(v.clone())
+        } else if let Ok(v) = std::env::var(env_name) {
+            Some(v)
+        } else {
+            None
+        }
     }
 }
 
@@ -247,8 +189,42 @@ async fn async_main() -> anyhow::Result<()> {
         args.listen
     );
 
+    let strict_mode = if args.dev {
+        StrictMode::Dev
+    } else {
+        StrictMode::Strict
+    };
+
     let secrets = Secrets::load(&args).await?;
 
+    // Validate required secrets and arguments are provided in strict mode
+    match strict_mode {
+        StrictMode::Strict
+            if (secrets.public_key.is_none()
+                || secrets.jwt_token.is_none()
+                || secrets.control_plane_jwt_token.is_none()) =>
+        {
+            // Production systems should always have secrets configured: if public_key was not set
+            // then we would implicitly disable auth.
+            anyhow::bail!(
+                    "Insecure config!  One or more secrets is not set.  This is only permitted in `--dev` mode"
+                );
+        }
+        StrictMode::Strict if args.compute_hook_url.is_none() => {
+            // Production systems should always have a compute hook set, to prevent falling
+            // back to trying to use neon_local.
+            anyhow::bail!(
+                "`--compute-hook-url` is not set: this is only permitted in `--dev` mode"
+            );
+        }
+        StrictMode::Strict => {
+            tracing::info!("Starting in strict mode: configuration is OK.")
+        }
+        StrictMode::Dev => {
+            tracing::warn!("Starting in dev mode: this may be an insecure configuration.")
+        }
+    }
+
     let config = Config {
         jwt_token: secrets.jwt_token,
         control_plane_jwt_token: secrets.control_plane_jwt_token,

control_plane/attachment_service/src/metrics.rs

@@ -1,32 +1,284 @@
-use metrics::{register_int_counter, register_int_counter_vec, IntCounter, IntCounterVec};
+//!
+//! This module provides metric definitions for the storage controller.
+//!
+//! All metrics are grouped in [`StorageControllerMetricGroup`]. [`StorageControllerMetrics`] holds
+//! the mentioned metrics and their encoder. It's globally available via the [`METRICS_REGISTRY`]
+//! constant.
+//!
+//! The rest of the code defines label group types and deals with converting outer types to labels.
+//!
+use bytes::Bytes;
+use measured::{
+    label::{LabelValue, StaticLabelSet},
+    FixedCardinalityLabel, MetricGroup,
+};
 use once_cell::sync::Lazy;
+use std::sync::Mutex;
 
-pub(crate) struct ReconcilerMetrics {
-    pub(crate) spawned: IntCounter,
-    pub(crate) complete: IntCounterVec,
-}
+use crate::persistence::{DatabaseError, DatabaseOperation};
 
-impl ReconcilerMetrics {
-    // Labels used on [`Self::complete`]
-    pub(crate) const SUCCESS: &'static str = "ok";
-    pub(crate) const ERROR: &'static str = "success";
-    pub(crate) const CANCEL: &'static str = "cancel";
-}
-
-pub(crate) static RECONCILER: Lazy<ReconcilerMetrics> = Lazy::new(|| ReconcilerMetrics {
-    spawned: register_int_counter!(
-        "storage_controller_reconcile_spawn",
-        "Count of how many times we spawn a reconcile task",
-    )
-    .expect("failed to define a metric"),
-    complete: register_int_counter_vec!(
-        "storage_controller_reconcile_complete",
-        "Reconciler tasks completed, broken down by success/failure/cancelled",
-        &["status"],
-    )
-    .expect("failed to define a metric"),
-});
+pub(crate) static METRICS_REGISTRY: Lazy<StorageControllerMetrics> =
+    Lazy::new(StorageControllerMetrics::default);
 
 pub fn preinitialize_metrics() {
-    Lazy::force(&RECONCILER);
+    Lazy::force(&METRICS_REGISTRY);
+}
+
+pub(crate) struct StorageControllerMetrics {
+    pub(crate) metrics_group: StorageControllerMetricGroup,
+    encoder: Mutex<measured::text::TextEncoder>,
+}
+
+#[derive(measured::MetricGroup)]
+pub(crate) struct StorageControllerMetricGroup {
+    /// Count of how many times we spawn a reconcile task
+    pub(crate) storage_controller_reconcile_spawn: measured::Counter,
+    /// Reconciler tasks completed, broken down by success/failure/cancelled
+    pub(crate) storage_controller_reconcile_complete:
+        measured::CounterVec<ReconcileCompleteLabelGroupSet>,
+
+    /// HTTP request status counters for handled requests
+    pub(crate) storage_controller_http_request_status:
+        measured::CounterVec<HttpRequestStatusLabelGroupSet>,
+    /// HTTP request handler latency across all status codes
+    pub(crate) storage_controller_http_request_latency:
+        measured::HistogramVec<HttpRequestLatencyLabelGroupSet, 5>,
+
+    /// Count of HTTP requests to the pageserver that resulted in an error,
+    /// broken down by the pageserver node id, request name and method
+    pub(crate) storage_controller_pageserver_request_error:
+        measured::CounterVec<PageserverRequestLabelGroupSet>,
+
+    /// Latency of HTTP requests to the pageserver, broken down by pageserver
+    /// node id, request name and method. This include both successful and unsuccessful
+    /// requests.
+    pub(crate) storage_controller_pageserver_request_latency:
+        measured::HistogramVec<PageserverRequestLabelGroupSet, 5>,
+
+    /// Count of pass-through HTTP requests to the pageserver that resulted in an error,
+    /// broken down by the pageserver node id, request name and method
+    pub(crate) storage_controller_passthrough_request_error:
+        measured::CounterVec<PageserverRequestLabelGroupSet>,
+
+    /// Latency of pass-through HTTP requests to the pageserver, broken down by pageserver
+    /// node id, request name and method. This include both successful and unsuccessful
+    /// requests.
+    pub(crate) storage_controller_passthrough_request_latency:
+        measured::HistogramVec<PageserverRequestLabelGroupSet, 5>,
+
+    /// Count of errors in database queries, broken down by error type and operation.
+    pub(crate) storage_controller_database_query_error:
+        measured::CounterVec<DatabaseQueryErrorLabelGroupSet>,
+
+    /// Latency of database queries, broken down by operation.
+    pub(crate) storage_controller_database_query_latency:
+        measured::HistogramVec<DatabaseQueryLatencyLabelGroupSet, 5>,
+}
+
+impl StorageControllerMetrics {
+    pub(crate) fn encode(&self) -> Bytes {
+        let mut encoder = self.encoder.lock().unwrap();
+        self.metrics_group.collect_into(&mut *encoder);
+        encoder.finish()
+    }
+}
+
+impl Default for StorageControllerMetrics {
+    fn default() -> Self {
+        Self {
+            metrics_group: StorageControllerMetricGroup::new(),
+            encoder: Mutex::new(measured::text::TextEncoder::new()),
+        }
+    }
+}
+
+impl StorageControllerMetricGroup {
+    pub(crate) fn new() -> Self {
+        Self {
+            storage_controller_reconcile_spawn: measured::Counter::new(),
+            storage_controller_reconcile_complete: measured::CounterVec::new(
+                ReconcileCompleteLabelGroupSet {
+                    status: StaticLabelSet::new(),
+                },
+            ),
+            storage_controller_http_request_status: measured::CounterVec::new(
+                HttpRequestStatusLabelGroupSet {
+                    path: lasso::ThreadedRodeo::new(),
+                    method: StaticLabelSet::new(),
+                    status: StaticLabelSet::new(),
+                },
+            ),
+            storage_controller_http_request_latency: measured::HistogramVec::new(
+                measured::metric::histogram::Thresholds::exponential_buckets(0.1, 2.0),
+            ),
+            storage_controller_pageserver_request_error: measured::CounterVec::new(
+                PageserverRequestLabelGroupSet {
+                    pageserver_id: lasso::ThreadedRodeo::new(),
+                    path: lasso::ThreadedRodeo::new(),
+                    method: StaticLabelSet::new(),
+                },
+            ),
+            storage_controller_pageserver_request_latency: measured::HistogramVec::new(
+                measured::metric::histogram::Thresholds::exponential_buckets(0.1, 2.0),
+            ),
+            storage_controller_passthrough_request_error: measured::CounterVec::new(
+                PageserverRequestLabelGroupSet {
+                    pageserver_id: lasso::ThreadedRodeo::new(),
+                    path: lasso::ThreadedRodeo::new(),
+                    method: StaticLabelSet::new(),
+                },
+            ),
+            storage_controller_passthrough_request_latency: measured::HistogramVec::new(
+                measured::metric::histogram::Thresholds::exponential_buckets(0.1, 2.0),
+            ),
+            storage_controller_database_query_error: measured::CounterVec::new(
+                DatabaseQueryErrorLabelGroupSet {
+                    operation: StaticLabelSet::new(),
+                    error_type: StaticLabelSet::new(),
+                },
+            ),
+            storage_controller_database_query_latency: measured::HistogramVec::new(
+                measured::metric::histogram::Thresholds::exponential_buckets(0.1, 2.0),
+            ),
+        }
+    }
+}
+
+#[derive(measured::LabelGroup)]
+#[label(set = ReconcileCompleteLabelGroupSet)]
+pub(crate) struct ReconcileCompleteLabelGroup {
+    pub(crate) status: ReconcileOutcome,
+}
+
+#[derive(measured::LabelGroup)]
+#[label(set = HttpRequestStatusLabelGroupSet)]
+pub(crate) struct HttpRequestStatusLabelGroup<'a> {
+    #[label(dynamic_with = lasso::ThreadedRodeo)]
+    pub(crate) path: &'a str,
+    pub(crate) method: Method,
+    pub(crate) status: StatusCode,
+}
+
+#[derive(measured::LabelGroup)]
+#[label(set = HttpRequestLatencyLabelGroupSet)]
+pub(crate) struct HttpRequestLatencyLabelGroup<'a> {
+    #[label(dynamic_with = lasso::ThreadedRodeo)]
+    pub(crate) path: &'a str,
+    pub(crate) method: Method,
+}
+
+impl Default for HttpRequestLatencyLabelGroupSet {
+    fn default() -> Self {
+        Self {
+            path: lasso::ThreadedRodeo::new(),
+            method: StaticLabelSet::new(),
+        }
+    }
+}
+
+#[derive(measured::LabelGroup, Clone)]
+#[label(set = PageserverRequestLabelGroupSet)]
+pub(crate) struct PageserverRequestLabelGroup<'a> {
+    #[label(dynamic_with = lasso::ThreadedRodeo)]
+    pub(crate) pageserver_id: &'a str,
+    #[label(dynamic_with = lasso::ThreadedRodeo)]
+    pub(crate) path: &'a str,
+    pub(crate) method: Method,
+}
+
+impl Default for PageserverRequestLabelGroupSet {
+    fn default() -> Self {
+        Self {
+            pageserver_id: lasso::ThreadedRodeo::new(),
+            path: lasso::ThreadedRodeo::new(),
+            method: StaticLabelSet::new(),
+        }
+    }
+}
+
+#[derive(measured::LabelGroup)]
+#[label(set = DatabaseQueryErrorLabelGroupSet)]
+pub(crate) struct DatabaseQueryErrorLabelGroup {
+    pub(crate) error_type: DatabaseErrorLabel,
+    pub(crate) operation: DatabaseOperation,
+}
+
+#[derive(measured::LabelGroup)]
+#[label(set = DatabaseQueryLatencyLabelGroupSet)]
+pub(crate) struct DatabaseQueryLatencyLabelGroup {
+    pub(crate) operation: DatabaseOperation,
+}
+
+#[derive(FixedCardinalityLabel)]
+pub(crate) enum ReconcileOutcome {
+    #[label(rename = "ok")]
+    Success,
+    Error,
+    Cancel,
+}
+
+#[derive(FixedCardinalityLabel, Clone)]
+pub(crate) enum Method {
+    Get,
+    Put,
+    Post,
+    Delete,
+    Other,
+}
+
+impl From<hyper::Method> for Method {
+    fn from(value: hyper::Method) -> Self {
+        if value == hyper::Method::GET {
+            Method::Get
+        } else if value == hyper::Method::PUT {
+            Method::Put
+        } else if value == hyper::Method::POST {
+            Method::Post
+        } else if value == hyper::Method::DELETE {
+            Method::Delete
+        } else {
+            Method::Other
+        }
+    }
+}
+
+pub(crate) struct StatusCode(pub(crate) hyper::http::StatusCode);
+
+impl LabelValue for StatusCode {
+    fn visit<V: measured::label::LabelVisitor>(&self, v: V) -> V::Output {
+        v.write_int(self.0.as_u16() as u64)
+    }
+}
+
+impl FixedCardinalityLabel for StatusCode {
+    fn cardinality() -> usize {
+        (100..1000).len()
+    }
+
+    fn encode(&self) -> usize {
+        self.0.as_u16() as usize
+    }
+
+    fn decode(value: usize) -> Self {
+        Self(hyper::http::StatusCode::from_u16(u16::try_from(value).unwrap()).unwrap())
+    }
+}
+
+#[derive(FixedCardinalityLabel)]
+pub(crate) enum DatabaseErrorLabel {
+    Query,
+    Connection,
+    ConnectionPool,
+    Logical,
+}
+
+impl DatabaseError {
+    pub(crate) fn error_label(&self) -> DatabaseErrorLabel {
+        match self {
+            Self::Query(_) => DatabaseErrorLabel::Query,
+            Self::Connection(_) => DatabaseErrorLabel::Connection,
+            Self::ConnectionPool(_) => DatabaseErrorLabel::ConnectionPool,
+            Self::Logical(_) => DatabaseErrorLabel::Logical,
+        }
+    }
 }

control_plane/attachment_service/src/node.rs

@@ -12,7 +12,9 @@ use serde::Serialize;
 use tokio_util::sync::CancellationToken;
 use utils::{backoff, id::NodeId};
 
-use crate::{persistence::NodePersistence, scheduler::MaySchedule};
+use crate::{
+    pageserver_client::PageserverClient, persistence::NodePersistence, scheduler::MaySchedule,
+};
 
 /// Represents the in-memory description of a Node.
 ///
@@ -202,7 +204,7 @@ impl Node {
         cancel: &CancellationToken,
     ) -> Option<mgmt_api::Result<T>>
     where
-        O: FnMut(mgmt_api::Client) -> F,
+        O: FnMut(PageserverClient) -> F,
         F: std::future::Future<Output = mgmt_api::Result<T>>,
     {
         fn is_fatal(e: &mgmt_api::Error) -> bool {
@@ -224,8 +226,12 @@ impl Node {
                     .build()
                     .expect("Failed to construct HTTP client");
 
-                let client =
-                    mgmt_api::Client::from_client(http_client, self.base_url(), jwt.as_deref());
+                let client = PageserverClient::from_client(
+                    self.get_id(),
+                    http_client,
+                    self.base_url(),
+                    jwt.as_deref(),
+                );
 
                 let node_cancel_fut = self.cancel.cancelled();
 

control_plane/attachment_service/src/pageserver_client.rs

@@ -0,0 +1,203 @@
+use pageserver_api::{
+    models::{
+        LocationConfig, LocationConfigListResponse, PageserverUtilization, SecondaryProgress,
+        TenantShardSplitRequest, TenantShardSplitResponse, TimelineCreateRequest, TimelineInfo,
+    },
+    shard::TenantShardId,
+};
+use pageserver_client::mgmt_api::{Client, Result};
+use reqwest::StatusCode;
+use utils::id::{NodeId, TimelineId};
+
+/// Thin wrapper around [`pageserver_client::mgmt_api::Client`]. It allows the storage
+/// controller to collect metrics in a non-intrusive manner.
+#[derive(Debug, Clone)]
+pub(crate) struct PageserverClient {
+    inner: Client,
+    node_id_label: String,
+}
+
+macro_rules! measured_request {
+    ($name:literal, $method:expr, $node_id: expr, $invoke:expr) => {{
+        let labels = crate::metrics::PageserverRequestLabelGroup {
+            pageserver_id: $node_id,
+            path: $name,
+            method: $method,
+        };
+
+        let latency = &crate::metrics::METRICS_REGISTRY
+            .metrics_group
+            .storage_controller_pageserver_request_latency;
+        let _timer_guard = latency.start_timer(labels.clone());
+
+        let res = $invoke;
+
+        if res.is_err() {
+            let error_counters = &crate::metrics::METRICS_REGISTRY
+                .metrics_group
+                .storage_controller_pageserver_request_error;
+            error_counters.inc(labels)
+        }
+
+        res
+    }};
+}
+
+impl PageserverClient {
+    pub(crate) fn new(node_id: NodeId, mgmt_api_endpoint: String, jwt: Option<&str>) -> Self {
+        Self {
+            inner: Client::from_client(reqwest::Client::new(), mgmt_api_endpoint, jwt),
+            node_id_label: node_id.0.to_string(),
+        }
+    }
+
+    pub(crate) fn from_client(
+        node_id: NodeId,
+        raw_client: reqwest::Client,
+        mgmt_api_endpoint: String,
+        jwt: Option<&str>,
+    ) -> Self {
+        Self {
+            inner: Client::from_client(raw_client, mgmt_api_endpoint, jwt),
+            node_id_label: node_id.0.to_string(),
+        }
+    }
+
+    pub(crate) async fn tenant_delete(&self, tenant_shard_id: TenantShardId) -> Result<StatusCode> {
+        measured_request!(
+            "tenant",
+            crate::metrics::Method::Delete,
+            &self.node_id_label,
+            self.inner.tenant_delete(tenant_shard_id).await
+        )
+    }
+
+    pub(crate) async fn tenant_time_travel_remote_storage(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timestamp: &str,
+        done_if_after: &str,
+    ) -> Result<()> {
+        measured_request!(
+            "tenant_time_travel_remote_storage",
+            crate::metrics::Method::Put,
+            &self.node_id_label,
+            self.inner
+                .tenant_time_travel_remote_storage(tenant_shard_id, timestamp, done_if_after)
+                .await
+        )
+    }
+
+    pub(crate) async fn tenant_secondary_download(
+        &self,
+        tenant_id: TenantShardId,
+        wait: Option<std::time::Duration>,
+    ) -> Result<(StatusCode, SecondaryProgress)> {
+        measured_request!(
+            "tenant_secondary_download",
+            crate::metrics::Method::Post,
+            &self.node_id_label,
+            self.inner.tenant_secondary_download(tenant_id, wait).await
+        )
+    }
+
+    pub(crate) async fn location_config(
+        &self,
+        tenant_shard_id: TenantShardId,
+        config: LocationConfig,
+        flush_ms: Option<std::time::Duration>,
+        lazy: bool,
+    ) -> Result<()> {
+        measured_request!(
+            "location_config",
+            crate::metrics::Method::Put,
+            &self.node_id_label,
+            self.inner
+                .location_config(tenant_shard_id, config, flush_ms, lazy)
+                .await
+        )
+    }
+
+    pub(crate) async fn list_location_config(&self) -> Result<LocationConfigListResponse> {
+        measured_request!(
+            "location_configs",
+            crate::metrics::Method::Get,
+            &self.node_id_label,
+            self.inner.list_location_config().await
+        )
+    }
+
+    pub(crate) async fn get_location_config(
+        &self,
+        tenant_shard_id: TenantShardId,
+    ) -> Result<Option<LocationConfig>> {
+        measured_request!(
+            "location_config",
+            crate::metrics::Method::Get,
+            &self.node_id_label,
+            self.inner.get_location_config(tenant_shard_id).await
+        )
+    }
+
+    pub(crate) async fn timeline_create(
+        &self,
+        tenant_shard_id: TenantShardId,
+        req: &TimelineCreateRequest,
+    ) -> Result<TimelineInfo> {
+        measured_request!(
+            "timeline",
+            crate::metrics::Method::Post,
+            &self.node_id_label,
+            self.inner.timeline_create(tenant_shard_id, req).await
+        )
+    }
+
+    pub(crate) async fn timeline_delete(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+    ) -> Result<StatusCode> {
+        measured_request!(
+            "timeline",
+            crate::metrics::Method::Delete,
+            &self.node_id_label,
+            self.inner
+                .timeline_delete(tenant_shard_id, timeline_id)
+                .await
+        )
+    }
+
+    pub(crate) async fn tenant_shard_split(
+        &self,
+        tenant_shard_id: TenantShardId,
+        req: TenantShardSplitRequest,
+    ) -> Result<TenantShardSplitResponse> {
+        measured_request!(
+            "tenant_shard_split",
+            crate::metrics::Method::Put,
+            &self.node_id_label,
+            self.inner.tenant_shard_split(tenant_shard_id, req).await
+        )
+    }
+
+    pub(crate) async fn timeline_list(
+        &self,
+        tenant_shard_id: &TenantShardId,
+    ) -> Result<Vec<TimelineInfo>> {
+        measured_request!(
+            "timelines",
+            crate::metrics::Method::Get,
+            &self.node_id_label,
+            self.inner.timeline_list(tenant_shard_id).await
+        )
+    }
+
+    pub(crate) async fn get_utilization(&self) -> Result<PageserverUtilization> {
+        measured_request!(
+            "utilization",
+            crate::metrics::Method::Get,
+            &self.node_id_label,
+            self.inner.get_utilization().await
+        )
+    }
+}

control_plane/attachment_service/src/persistence.rs

@@ -19,6 +19,9 @@ use serde::{Deserialize, Serialize};
 use utils::generation::Generation;
 use utils::id::{NodeId, TenantId};
 
+use crate::metrics::{
+    DatabaseQueryErrorLabelGroup, DatabaseQueryLatencyLabelGroup, METRICS_REGISTRY,
+};
 use crate::node::Node;
 
 /// ## What do we store?
@@ -75,6 +78,25 @@ pub(crate) enum DatabaseError {
     Logical(String),
 }
 
+#[derive(measured::FixedCardinalityLabel, Clone)]
+pub(crate) enum DatabaseOperation {
+    InsertNode,
+    UpdateNode,
+    DeleteNode,
+    ListNodes,
+    BeginShardSplit,
+    CompleteShardSplit,
+    AbortShardSplit,
+    Detach,
+    ReAttach,
+    IncrementGeneration,
+    ListTenantShards,
+    InsertTenantShards,
+    UpdateTenantShard,
+    DeleteTenant,
+    UpdateTenantConfig,
+}
+
 #[must_use]
 pub(crate) enum AbortShardSplitStatus {
     /// We aborted the split in the database by reverting to the parent shards
@@ -115,6 +137,34 @@ impl Persistence {
         }
     }
 
+    /// Wraps `with_conn` in order to collect latency and error metrics
+    async fn with_measured_conn<F, R>(&self, op: DatabaseOperation, func: F) -> DatabaseResult<R>
+    where
+        F: Fn(&mut PgConnection) -> DatabaseResult<R> + Send + 'static,
+        R: Send + 'static,
+    {
+        let latency = &METRICS_REGISTRY
+            .metrics_group
+            .storage_controller_database_query_latency;
+        let _timer = latency.start_timer(DatabaseQueryLatencyLabelGroup {
+            operation: op.clone(),
+        });
+
+        let res = self.with_conn(func).await;
+
+        if let Err(err) = &res {
+            let error_counter = &METRICS_REGISTRY
+                .metrics_group
+                .storage_controller_database_query_error;
+            error_counter.inc(DatabaseQueryErrorLabelGroup {
+                error_type: err.error_label(),
+                operation: op,
+            })
+        }
+
+        res
+    }
+
     /// Call the provided function in a tokio blocking thread, with a Diesel database connection.
     async fn with_conn<F, R>(&self, func: F) -> DatabaseResult<R>
     where
@@ -130,21 +180,27 @@ impl Persistence {
     /// When a node is first registered, persist it before using it for anything
     pub(crate) async fn insert_node(&self, node: &Node) -> DatabaseResult<()> {
         let np = node.to_persistent();
-        self.with_conn(move |conn| -> DatabaseResult<()> {
-            diesel::insert_into(crate::schema::nodes::table)
-                .values(&np)
-                .execute(conn)?;
-            Ok(())
-        })
+        self.with_measured_conn(
+            DatabaseOperation::InsertNode,
+            move |conn| -> DatabaseResult<()> {
+                diesel::insert_into(crate::schema::nodes::table)
+                    .values(&np)
+                    .execute(conn)?;
+                Ok(())
+            },
+        )
         .await
     }
 
     /// At startup, populate the list of nodes which our shards may be placed on
     pub(crate) async fn list_nodes(&self) -> DatabaseResult<Vec<NodePersistence>> {
         let nodes: Vec<NodePersistence> = self
-            .with_conn(move |conn| -> DatabaseResult<_> {
-                Ok(crate::schema::nodes::table.load::<NodePersistence>(conn)?)
-            })
+            .with_measured_conn(
+                DatabaseOperation::ListNodes,
+                move |conn| -> DatabaseResult<_> {
+                    Ok(crate::schema::nodes::table.load::<NodePersistence>(conn)?)
+                },
+            )
             .await?;
 
         tracing::info!("list_nodes: loaded {} nodes", nodes.len());
@@ -159,7 +215,7 @@ impl Persistence {
     ) -> DatabaseResult<()> {
         use crate::schema::nodes::dsl::*;
         let updated = self
-            .with_conn(move |conn| {
+            .with_measured_conn(DatabaseOperation::UpdateNode, move |conn| {
                 let updated = diesel::update(nodes)
                     .filter(node_id.eq(input_node_id.0 as i64))
                     .set((scheduling_policy.eq(String::from(input_scheduling)),))
@@ -181,9 +237,12 @@ impl Persistence {
     /// be enriched at runtime with state discovered on pageservers.
     pub(crate) async fn list_tenant_shards(&self) -> DatabaseResult<Vec<TenantShardPersistence>> {
         let loaded = self
-            .with_conn(move |conn| -> DatabaseResult<_> {
-                Ok(crate::schema::tenant_shards::table.load::<TenantShardPersistence>(conn)?)
-            })
+            .with_measured_conn(
+                DatabaseOperation::ListTenantShards,
+                move |conn| -> DatabaseResult<_> {
+                    Ok(crate::schema::tenant_shards::table.load::<TenantShardPersistence>(conn)?)
+                },
+            )
             .await?;
 
         if loaded.is_empty() {
@@ -211,15 +270,10 @@ impl Persistence {
 
         let mut decoded = serde_json::from_slice::<JsonPersistence>(&bytes)
             .map_err(|e| DatabaseError::Logical(format!("Deserialization error: {e}")))?;
-        for (tenant_id, tenant) in &mut decoded.tenants {
-            // Backward compat: an old attachments.json from before PR #6251, replace
-            // empty strings with proper defaults.
-            if tenant.tenant_id.is_empty() {
-                tenant.tenant_id = tenant_id.to_string();
-                tenant.config = serde_json::to_string(&TenantConfig::default())
-                    .map_err(|e| DatabaseError::Logical(format!("Serialization error: {e}")))?;
-                tenant.placement_policy = serde_json::to_string(&PlacementPolicy::Single)
-                    .map_err(|e| DatabaseError::Logical(format!("Serialization error: {e}")))?;
+        for shard in decoded.tenants.values_mut() {
+            if shard.placement_policy == "\"Single\"" {
+                // Backward compat for test data after PR https://github.com/neondatabase/neon/pull/7165
+                shard.placement_policy = "{\"Attached\":0}".to_string();
             }
         }
 
@@ -265,17 +319,20 @@ impl Persistence {
         shards: Vec<TenantShardPersistence>,
     ) -> DatabaseResult<()> {
         use crate::schema::tenant_shards::dsl::*;
-        self.with_conn(move |conn| -> DatabaseResult<()> {
-            conn.transaction(|conn| -> QueryResult<()> {
-                for tenant in &shards {
-                    diesel::insert_into(tenant_shards)
-                        .values(tenant)
-                        .execute(conn)?;
-                }
+        self.with_measured_conn(
+            DatabaseOperation::InsertTenantShards,
+            move |conn| -> DatabaseResult<()> {
+                conn.transaction(|conn| -> QueryResult<()> {
+                    for tenant in &shards {
+                        diesel::insert_into(tenant_shards)
+                            .values(tenant)
+                            .execute(conn)?;
+                    }
+                    Ok(())
+                })?;
                 Ok(())
-            })?;
-            Ok(())
-        })
+            },
+        )
         .await
     }
 
@@ -283,25 +340,31 @@ impl Persistence {
     /// the tenant from memory on this server.
     pub(crate) async fn delete_tenant(&self, del_tenant_id: TenantId) -> DatabaseResult<()> {
         use crate::schema::tenant_shards::dsl::*;
-        self.with_conn(move |conn| -> DatabaseResult<()> {
-            diesel::delete(tenant_shards)
-                .filter(tenant_id.eq(del_tenant_id.to_string()))
-                .execute(conn)?;
+        self.with_measured_conn(
+            DatabaseOperation::DeleteTenant,
+            move |conn| -> DatabaseResult<()> {
+                diesel::delete(tenant_shards)
+                    .filter(tenant_id.eq(del_tenant_id.to_string()))
+                    .execute(conn)?;
 
-            Ok(())
-        })
+                Ok(())
+            },
+        )
         .await
     }
 
     pub(crate) async fn delete_node(&self, del_node_id: NodeId) -> DatabaseResult<()> {
         use crate::schema::nodes::dsl::*;
-        self.with_conn(move |conn| -> DatabaseResult<()> {
-            diesel::delete(nodes)
-                .filter(node_id.eq(del_node_id.0 as i64))
-                .execute(conn)?;
+        self.with_measured_conn(
+            DatabaseOperation::DeleteNode,
+            move |conn| -> DatabaseResult<()> {
+                diesel::delete(nodes)
+                    .filter(node_id.eq(del_node_id.0 as i64))
+                    .execute(conn)?;
 
-            Ok(())
-        })
+                Ok(())
+            },
+        )
         .await
     }
 
@@ -315,7 +378,7 @@ impl Persistence {
     ) -> DatabaseResult<HashMap<TenantShardId, Generation>> {
         use crate::schema::tenant_shards::dsl::*;
         let updated = self
-            .with_conn(move |conn| {
+            .with_measured_conn(DatabaseOperation::ReAttach, move |conn| {
                 let rows_updated = diesel::update(tenant_shards)
                     .filter(generation_pageserver.eq(node_id.0 as i64))
                     .set(generation.eq(generation + 1))
@@ -365,7 +428,7 @@ impl Persistence {
     ) -> anyhow::Result<Generation> {
         use crate::schema::tenant_shards::dsl::*;
         let updated = self
-            .with_conn(move |conn| {
+            .with_measured_conn(DatabaseOperation::IncrementGeneration, move |conn| {
                 let updated = diesel::update(tenant_shards)
                     .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
                     .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
@@ -409,7 +472,7 @@ impl Persistence {
     ) -> DatabaseResult<()> {
         use crate::schema::tenant_shards::dsl::*;
 
-        self.with_conn(move |conn| {
+        self.with_measured_conn(DatabaseOperation::UpdateTenantShard, move |conn| {
             let query = diesel::update(tenant_shards)
                 .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
                 .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
@@ -450,7 +513,7 @@ impl Persistence {
     ) -> DatabaseResult<()> {
         use crate::schema::tenant_shards::dsl::*;
 
-        self.with_conn(move |conn| {
+        self.with_measured_conn(DatabaseOperation::UpdateTenantConfig, move |conn| {
             diesel::update(tenant_shards)
                 .filter(tenant_id.eq(input_tenant_id.to_string()))
                 .set((config.eq(serde_json::to_string(&input_config).unwrap()),))
@@ -465,7 +528,7 @@ impl Persistence {
 
     pub(crate) async fn detach(&self, tenant_shard_id: TenantShardId) -> anyhow::Result<()> {
         use crate::schema::tenant_shards::dsl::*;
-        self.with_conn(move |conn| {
+        self.with_measured_conn(DatabaseOperation::Detach, move |conn| {
             let updated = diesel::update(tenant_shards)
                 .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
                 .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
@@ -495,7 +558,7 @@ impl Persistence {
         parent_to_children: Vec<(TenantShardId, Vec<TenantShardPersistence>)>,
     ) -> DatabaseResult<()> {
         use crate::schema::tenant_shards::dsl::*;
-        self.with_conn(move |conn| -> DatabaseResult<()> {
+        self.with_measured_conn(DatabaseOperation::BeginShardSplit, move |conn| -> DatabaseResult<()> {
             conn.transaction(|conn| -> DatabaseResult<()> {
                 // Mark parent shards as splitting
 
@@ -559,26 +622,29 @@ impl Persistence {
         old_shard_count: ShardCount,
     ) -> DatabaseResult<()> {
         use crate::schema::tenant_shards::dsl::*;
-        self.with_conn(move |conn| -> DatabaseResult<()> {
-            conn.transaction(|conn| -> QueryResult<()> {
-                // Drop parent shards
-                diesel::delete(tenant_shards)
-                    .filter(tenant_id.eq(split_tenant_id.to_string()))
-                    .filter(shard_count.eq(old_shard_count.literal() as i32))
-                    .execute(conn)?;
+        self.with_measured_conn(
+            DatabaseOperation::CompleteShardSplit,
+            move |conn| -> DatabaseResult<()> {
+                conn.transaction(|conn| -> QueryResult<()> {
+                    // Drop parent shards
+                    diesel::delete(tenant_shards)
+                        .filter(tenant_id.eq(split_tenant_id.to_string()))
+                        .filter(shard_count.eq(old_shard_count.literal() as i32))
+                        .execute(conn)?;
 
-                // Clear sharding flag
-                let updated = diesel::update(tenant_shards)
-                    .filter(tenant_id.eq(split_tenant_id.to_string()))
-                    .set((splitting.eq(0),))
-                    .execute(conn)?;
-                debug_assert!(updated > 0);
+                    // Clear sharding flag
+                    let updated = diesel::update(tenant_shards)
+                        .filter(tenant_id.eq(split_tenant_id.to_string()))
+                        .set((splitting.eq(0),))
+                        .execute(conn)?;
+                    debug_assert!(updated > 0);
+
+                    Ok(())
+                })?;
 
                 Ok(())
-            })?;
-
-            Ok(())
-        })
+            },
+        )
         .await
     }
 
@@ -590,40 +656,44 @@ impl Persistence {
         new_shard_count: ShardCount,
     ) -> DatabaseResult<AbortShardSplitStatus> {
         use crate::schema::tenant_shards::dsl::*;
-        self.with_conn(move |conn| -> DatabaseResult<AbortShardSplitStatus> {
-            let aborted = conn.transaction(|conn| -> DatabaseResult<AbortShardSplitStatus> {
-                // Clear the splitting state on parent shards
-                let updated = diesel::update(tenant_shards)
-                    .filter(tenant_id.eq(split_tenant_id.to_string()))
-                    .filter(shard_count.ne(new_shard_count.literal() as i32))
-                    .set((splitting.eq(0),))
-                    .execute(conn)?;
+        self.with_measured_conn(
+            DatabaseOperation::AbortShardSplit,
+            move |conn| -> DatabaseResult<AbortShardSplitStatus> {
+                let aborted =
+                    conn.transaction(|conn| -> DatabaseResult<AbortShardSplitStatus> {
+                        // Clear the splitting state on parent shards
+                        let updated = diesel::update(tenant_shards)
+                            .filter(tenant_id.eq(split_tenant_id.to_string()))
+                            .filter(shard_count.ne(new_shard_count.literal() as i32))
+                            .set((splitting.eq(0),))
+                            .execute(conn)?;
 
-                // Parent shards are already gone: we cannot abort.
-                if updated == 0 {
-                    return Ok(AbortShardSplitStatus::Complete);
-                }
+                        // Parent shards are already gone: we cannot abort.
+                        if updated == 0 {
+                            return Ok(AbortShardSplitStatus::Complete);
+                        }
 
-                // Sanity check: if parent shards were present, their cardinality should
-                // be less than the number of child shards.
-                if updated >= new_shard_count.count() as usize {
-                    return Err(DatabaseError::Logical(format!(
-                        "Unexpected parent shard count {updated} while aborting split to \
+                        // Sanity check: if parent shards were present, their cardinality should
+                        // be less than the number of child shards.
+                        if updated >= new_shard_count.count() as usize {
+                            return Err(DatabaseError::Logical(format!(
+                                "Unexpected parent shard count {updated} while aborting split to \
                             count {new_shard_count:?} on tenant {split_tenant_id}"
-                    )));
-                }
+                            )));
+                        }
 
-                // Erase child shards
-                diesel::delete(tenant_shards)
-                    .filter(tenant_id.eq(split_tenant_id.to_string()))
-                    .filter(shard_count.eq(new_shard_count.literal() as i32))
-                    .execute(conn)?;
+                        // Erase child shards
+                        diesel::delete(tenant_shards)
+                            .filter(tenant_id.eq(split_tenant_id.to_string()))
+                            .filter(shard_count.eq(new_shard_count.literal() as i32))
+                            .execute(conn)?;
 
-                Ok(AbortShardSplitStatus::Aborted)
-            })?;
+                        Ok(AbortShardSplitStatus::Aborted)
+                    })?;
 
-            Ok(aborted)
-        })
+                Ok(aborted)
+            },
+        )
         .await
     }
 }

control_plane/attachment_service/src/reconciler.rs

@@ -1,3 +1,4 @@
+use crate::pageserver_client::PageserverClient;
 use crate::persistence::Persistence;
 use crate::service;
 use hyper::StatusCode;
@@ -117,6 +118,15 @@ impl Reconciler {
         flush_ms: Option<Duration>,
         lazy: bool,
     ) -> Result<(), ReconcileError> {
+        if !node.is_available() && config.mode == LocationConfigMode::Detached {
+            // Attempts to detach from offline nodes may be imitated without doing I/O: a node which is offline
+            // will get fully reconciled wrt the shard's intent state when it is reactivated, irrespective of
+            // what we put into `observed`, in [`crate::service::Service::node_activate_reconcile`]
+            tracing::info!("Node {node} is unavailable during detach: proceeding anyway, it will be detached on next activation");
+            self.observed.locations.remove(&node.get_id());
+            return Ok(());
+        }
+
         self.observed
             .locations
             .insert(node.get_id(), ObservedStateLocation { conf: None });
@@ -149,9 +159,16 @@ impl Reconciler {
         };
         tracing::info!("location_config({node}) complete: {:?}", config);
 
-        self.observed
-            .locations
-            .insert(node.get_id(), ObservedStateLocation { conf: Some(config) });
+        match config.mode {
+            LocationConfigMode::Detached => {
+                self.observed.locations.remove(&node.get_id());
+            }
+            _ => {
+                self.observed
+                    .locations
+                    .insert(node.get_id(), ObservedStateLocation { conf: Some(config) });
+            }
+        }
 
         Ok(())
     }
@@ -243,8 +260,11 @@ impl Reconciler {
         tenant_shard_id: TenantShardId,
         node: &Node,
     ) -> anyhow::Result<HashMap<TimelineId, Lsn>> {
-        let client =
-            mgmt_api::Client::new(node.base_url(), self.service_config.jwt_token.as_deref());
+        let client = PageserverClient::new(
+            node.get_id(),
+            node.base_url(),
+            self.service_config.jwt_token.as_deref(),
+        );
 
         let timelines = client.timeline_list(&tenant_shard_id).await?;
         Ok(timelines
@@ -475,7 +495,7 @@ impl Reconciler {
             }
         }
 
-        // Downgrade the origin to secondary.  If the tenant's policy is PlacementPolicy::Single, then
+        // Downgrade the origin to secondary.  If the tenant's policy is PlacementPolicy::Attached(0), then
         // this location will be deleted in the general case reconciliation that runs after this.
         let origin_secondary_conf = build_location_config(
             &self.shard,

control_plane/attachment_service/src/service.rs

@@ -7,7 +7,9 @@ use std::{
     time::{Duration, Instant},
 };
 
-use crate::{id_lock_map::IdLockMap, persistence::AbortShardSplitStatus};
+use crate::{
+    id_lock_map::IdLockMap, persistence::AbortShardSplitStatus, reconciler::ReconcileError,
+};
 use anyhow::Context;
 use control_plane::storage_controller::{
     AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
@@ -18,12 +20,14 @@ use hyper::StatusCode;
 use pageserver_api::{
     controller_api::{
         NodeAvailability, NodeRegisterRequest, NodeSchedulingPolicy, PlacementPolicy,
-        TenantCreateResponse, TenantCreateResponseShard, TenantLocateResponse,
-        TenantShardMigrateRequest, TenantShardMigrateResponse, UtilizationScore,
+        TenantCreateResponse, TenantCreateResponseShard, TenantDescribeResponse,
+        TenantDescribeResponseShard, TenantLocateResponse, TenantShardMigrateRequest,
+        TenantShardMigrateResponse, UtilizationScore,
     },
     models::{SecondaryProgress, TenantConfigRequest},
 };
 
+use crate::pageserver_client::PageserverClient;
 use pageserver_api::{
     models::{
         self, LocationConfig, LocationConfigListResponse, LocationConfigMode,
@@ -200,6 +204,30 @@ enum TenantCreateOrUpdate {
     Update(Vec<ShardUpdate>),
 }
 
+struct ShardSplitParams {
+    old_shard_count: ShardCount,
+    new_shard_count: ShardCount,
+    new_stripe_size: Option<ShardStripeSize>,
+    targets: Vec<ShardSplitTarget>,
+    policy: PlacementPolicy,
+    config: TenantConfig,
+    shard_ident: ShardIdentity,
+}
+
+// When preparing for a shard split, we may either choose to proceed with the split,
+// or find that the work is already done and return NoOp.
+enum ShardSplitAction {
+    Split(ShardSplitParams),
+    NoOp(TenantShardSplitResponse),
+}
+
+// A parent shard which will be split
+struct ShardSplitTarget {
+    parent_id: TenantShardId,
+    node: Node,
+    child_ids: Vec<TenantShardId>,
+}
+
 /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
 /// might not be available.  We therefore use a queue of abort operations processed in the background.
 struct TenantShardSplitAbort {
@@ -525,7 +553,11 @@ impl Service {
                 break;
             }
 
-            let client = mgmt_api::Client::new(node.base_url(), self.config.jwt_token.as_deref());
+            let client = PageserverClient::new(
+                node.get_id(),
+                node.base_url(),
+                self.config.jwt_token.as_deref(),
+            );
             match client
                 .location_config(
                     tenant_shard_id,
@@ -733,7 +765,19 @@ impl Service {
                 tenant.waiter.advance(result.sequence);
             }
             Err(e) => {
-                tracing::warn!("Reconcile error: {}", e);
+                match e {
+                    ReconcileError::Cancel => {
+                        tracing::info!("Reconciler was cancelled");
+                    }
+                    ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
+                        // This might be due to the reconciler getting cancelled, or it might
+                        // be due to the `Node` being marked offline.
+                        tracing::info!("Reconciler cancelled during pageserver API call");
+                    }
+                    _ => {
+                        tracing::warn!("Reconcile error: {}", e);
+                    }
+                }
 
                 // Ordering: populate last_error before advancing error_seq,
                 // so that waiters will see the correct error after waiting.
@@ -1057,7 +1101,7 @@ impl Service {
                 shard_stripe_size: 0,
                 generation: Some(0),
                 generation_pageserver: None,
-                placement_policy: serde_json::to_string(&PlacementPolicy::Single).unwrap(),
+                placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
                 config: serde_json::to_string(&TenantConfig::default()).unwrap(),
                 splitting: SplitState::default(),
             };
@@ -1084,7 +1128,7 @@ impl Service {
                         TenantState::new(
                             attach_req.tenant_shard_id,
                             ShardIdentity::unsharded(),
-                            PlacementPolicy::Single,
+                            PlacementPolicy::Attached(0),
                         ),
                     );
                     tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
@@ -1113,7 +1157,7 @@ impl Service {
                     self.persistence
                         .update_tenant_shard(
                             attach_req.tenant_shard_id,
-                            PlacementPolicy::Single,
+                            PlacementPolicy::Attached(0),
                             conf,
                             None,
                         )
@@ -1138,7 +1182,7 @@ impl Service {
 
         if let Some(new_generation) = new_generation {
             tenant_state.generation = Some(new_generation);
-            tenant_state.policy = PlacementPolicy::Single;
+            tenant_state.policy = PlacementPolicy::Attached(0);
         } else {
             // This is a detach notification.  We must update placement policy to avoid re-attaching
             // during background scheduling/reconciliation, or during storage controller restart.
@@ -1350,7 +1394,8 @@ impl Service {
             incremented_generations.len()
         );
 
-        // Apply the updated generation to our in-memory state
+        // Apply the updated generation to our in-memory state, and
+        // gather discover secondary locations.
         let mut locked = self.inner.write().unwrap();
         let (nodes, tenants, scheduler) = locked.parts_mut();
 
@@ -1358,62 +1403,65 @@ impl Service {
             tenants: Vec::new(),
         };
 
-        for (tenant_shard_id, new_gen) in incremented_generations {
-            response.tenants.push(ReAttachResponseTenant {
-                id: tenant_shard_id,
-                gen: new_gen.into().unwrap(),
-            });
-            // Apply the new generation number to our in-memory state
-            let shard_state = tenants.get_mut(&tenant_shard_id);
-            let Some(shard_state) = shard_state else {
-                // Not fatal.  This edge case requires a re-attach to happen
-                // between inserting a new tenant shard in to the database, and updating our in-memory
-                // state to know about the shard, _and_ that the state inserted to the database referenced
-                // a pageserver.  Should never happen, but handle it rather than panicking, since it should
-                // be harmless.
-                tracing::error!(
-                    "Shard {} is in database for node {} but not in-memory state",
-                    tenant_shard_id,
-                    reattach_req.node_id
-                );
-                continue;
-            };
+        // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
+        // to call location_conf API with an old generation.  Wait for cancellation to complete
+        // before responding to this request.  Requires well implemented CancellationToken logic
+        // all the way to where we call location_conf.  Even then, there can still be a location_conf
+        // request in flight over the network: TODO handle that by making location_conf API refuse
+        // to go backward in generations.
 
-            // If [`Persistence::re_attach`] selected this shard, it must have alread
-            // had a generation set.
-            debug_assert!(shard_state.generation.is_some());
-            let Some(old_gen) = shard_state.generation else {
-                // Should never happen:  would only return incremented generation
-                // for a tenant that already had a non-null generation.
-                return Err(ApiError::InternalServerError(anyhow::anyhow!(
-                    "Generation must be set while re-attaching"
-                )));
-            };
-            shard_state.generation = Some(std::cmp::max(old_gen, new_gen));
-            if let Some(observed) = shard_state
-                .observed
-                .locations
-                .get_mut(&reattach_req.node_id)
-            {
-                if let Some(conf) = observed.conf.as_mut() {
-                    conf.generation = new_gen.into();
+        // Scan through all shards, applying updates for ones where we updated generation
+        // and identifying shards that intend to have a secondary location on this node.
+        for (tenant_shard_id, shard) in tenants {
+            if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
+                let new_gen = *new_gen;
+                response.tenants.push(ReAttachResponseTenant {
+                    id: *tenant_shard_id,
+                    gen: Some(new_gen.into().unwrap()),
+                    // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
+                    // execution.  If a pageserver is restarted during that process, then the reconcile pass will
+                    // fail, and start from scratch, so it doesn't make sense for us to try and preserve
+                    // the stale/multi states at this point.
+                    mode: LocationConfigMode::AttachedSingle,
+                });
+
+                shard.generation = std::cmp::max(shard.generation, Some(new_gen));
+                if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
+                    // Why can we update `observed` even though we're not sure our response will be received
+                    // by the pageserver?  Because the pageserver will not proceed with startup until
+                    // it has processed response: if it loses it, we'll see another request and increment
+                    // generation again, avoiding any uncertainty about dirtiness of tenant's state.
+                    if let Some(conf) = observed.conf.as_mut() {
+                        conf.generation = new_gen.into();
+                    }
+                } else {
+                    // This node has no observed state for the shard: perhaps it was offline
+                    // when the pageserver restarted.  Insert a None, so that the Reconciler
+                    // will be prompted to learn the location's state before it makes changes.
+                    shard
+                        .observed
+                        .locations
+                        .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
                 }
-            } else {
-                // This node has no observed state for the shard: perhaps it was offline
-                // when the pageserver restarted.  Insert a None, so that the Reconciler
-                // will be prompted to learn the location's state before it makes changes.
-                shard_state
-                    .observed
-                    .locations
-                    .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
-            }
+            } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
+                // Ordering: pageserver will not accept /location_config requests until it has
+                // finished processing the response from re-attach.  So we can update our in-memory state
+                // now, and be confident that we are not stamping on the result of some later location config.
+                // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
+                // so we might update observed state here, and then get over-written by some racing
+                // ReconcileResult.  The impact is low however, since we have set state on pageserver something
+                // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
 
-            // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
-            // to call location_conf API with an old generation.  Wait for cancellation to complete
-            // before responding to this request.  Requires well implemented CancellationToken logic
-            // all the way to where we call location_conf.  Even then, there can still be a location_conf
-            // request in flight over the network: TODO handle that by making location_conf API refuse
-            // to go backward in generations.
+                response.tenants.push(ReAttachResponseTenant {
+                    id: *tenant_shard_id,
+                    gen: None,
+                    mode: LocationConfigMode::Secondary,
+                });
+
+                // We must not update observed, because we have no guarantee that our
+                // response will be received by the pageserver. This could leave it
+                // falsely dirty, but the resulting reconcile should be idempotent.
+            }
         }
 
         // We consider a node Active once we have composed a re-attach response, but we
@@ -1475,6 +1523,8 @@ impl Service {
         &self,
         create_req: TenantCreateRequest,
     ) -> Result<TenantCreateResponse, ApiError> {
+        let tenant_id = create_req.new_tenant_id.tenant_id;
+
         // Exclude any concurrent attempts to create/access the same tenant ID
         let _tenant_lock = self
             .tenant_op_locks
@@ -1483,7 +1533,12 @@ impl Service {
 
         let (response, waiters) = self.do_tenant_create(create_req).await?;
 
-        self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await?;
+        if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
+            // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
+            // accept compute notifications while it is in the process of creating.  Reconciliation will
+            // be retried in the background.
+            tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
+        }
         Ok(response)
     }
 
@@ -1491,11 +1546,11 @@ impl Service {
         &self,
         create_req: TenantCreateRequest,
     ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
-        // As a default, single is convenient for tests that don't choose a policy.
         let placement_policy = create_req
             .placement_policy
             .clone()
-            .unwrap_or(PlacementPolicy::Single);
+            // As a default, zero secondaries is convenient for tests that don't choose a policy.
+            .unwrap_or(PlacementPolicy::Attached(0));
 
         // This service expects to handle sharding itself: it is an error to try and directly create
         // a particular shard here.
@@ -1562,13 +1617,25 @@ impl Service {
                 splitting: SplitState::default(),
             })
             .collect();
-        self.persistence
+
+        match self
+            .persistence
             .insert_tenant_shards(persist_tenant_shards)
             .await
-            .map_err(|e| {
-                // TODO: distinguish primary key constraint (idempotent, OK), from other errors
-                ApiError::InternalServerError(anyhow::anyhow!(e))
-            })?;
+        {
+            Ok(_) => {}
+            Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
+                DatabaseErrorKind::UniqueViolation,
+                _,
+            ))) => {
+                // Unique key violation: this is probably a retry.  Because the shard count is part of the unique key,
+                // if we see a unique key violation it means that the creation request's shard count matches the previous
+                // creation's shard count.
+                tracing::info!("Tenant shards already present in database, proceeding with idempotent creation...");
+            }
+            // Any other database error is unexpected and a bug.
+            Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
+        };
 
         let (waiters, response_shards) = {
             let mut locked = self.inner.write().unwrap();
@@ -1705,11 +1772,11 @@ impl Service {
             | LocationConfigMode::AttachedSingle
             | LocationConfigMode::AttachedStale => {
                 if nodes.len() > 1 {
-                    PlacementPolicy::Double(1)
+                    PlacementPolicy::Attached(1)
                 } else {
                     // Convenience for dev/test: if we just have one pageserver, import
-                    // tenants into Single mode so that scheduling will succeed.
-                    PlacementPolicy::Single
+                    // tenants into non-HA mode so that scheduling will succeed.
+                    PlacementPolicy::Attached(0)
                 }
             }
         };
@@ -2058,8 +2125,11 @@ impl Service {
                 })
                 .collect::<Vec<_>>();
             for tenant_shard_id in shard_ids {
-                let client =
-                    mgmt_api::Client::new(node.base_url(), self.config.jwt_token.as_deref());
+                let client = PageserverClient::new(
+                    node.get_id(),
+                    node.base_url(),
+                    self.config.jwt_token.as_deref(),
+                );
 
                 tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
 
@@ -2111,7 +2181,11 @@ impl Service {
         // Issue concurrent requests to all shards' locations
         let mut futs = FuturesUnordered::new();
         for (tenant_shard_id, node) in targets {
-            let client = mgmt_api::Client::new(node.base_url(), self.config.jwt_token.as_deref());
+            let client = PageserverClient::new(
+                node.get_id(),
+                node.base_url(),
+                self.config.jwt_token.as_deref(),
+            );
             futs.push(async move {
                 let result = client
                     .tenant_secondary_download(tenant_shard_id, wait)
@@ -2204,7 +2278,11 @@ impl Service {
         // Phase 1: delete on the pageservers
         let mut any_pending = false;
         for (tenant_shard_id, node) in targets {
-            let client = mgmt_api::Client::new(node.base_url(), self.config.jwt_token.as_deref());
+            let client = PageserverClient::new(
+                node.get_id(),
+                node.base_url(),
+                self.config.jwt_token.as_deref(),
+            );
             // TODO: this, like many other places, requires proper retry handling for 503, timeout: those should not
             // surface immediately as an error to our caller.
             let status = client.tenant_delete(tenant_shard_id).await.map_err(|e| {
@@ -2316,7 +2394,7 @@ impl Service {
                 tenant_shard_id,
                 create_req.new_timeline_id,
             );
-            let client = mgmt_api::Client::new(node.base_url(), jwt.as_deref());
+            let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
 
             client
                 .timeline_create(tenant_shard_id, &create_req)
@@ -2440,7 +2518,7 @@ impl Service {
                 "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
             );
 
-            let client = mgmt_api::Client::new(node.base_url(), jwt.as_deref());
+            let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
             client
                 .timeline_delete(tenant_shard_id, timeline_id)
                 .await
@@ -2481,11 +2559,11 @@ impl Service {
     }
 
     /// When you need to send an HTTP request to the pageserver that holds shard0 of a tenant, this
-    /// function looks it up and returns the url.  If the tenant isn't found, returns Err(ApiError::NotFound)
-    pub(crate) fn tenant_shard0_baseurl(
+    /// function looks up and returns node. If the tenant isn't found, returns Err(ApiError::NotFound)
+    pub(crate) fn tenant_shard0_node(
         &self,
         tenant_id: TenantId,
-    ) -> Result<(String, TenantShardId), ApiError> {
+    ) -> Result<(Node, TenantShardId), ApiError> {
         let locked = self.inner.read().unwrap();
         let Some((tenant_shard_id, shard)) = locked
             .tenants
@@ -2517,7 +2595,7 @@ impl Service {
             )));
         };
 
-        Ok((node.base_url(), *tenant_shard_id))
+        Ok((node.clone(), *tenant_shard_id))
     }
 
     pub(crate) fn tenant_locate(
@@ -2527,9 +2605,6 @@ impl Service {
         let locked = self.inner.read().unwrap();
         tracing::info!("Locating shards for tenant {tenant_id}");
 
-        // Take a snapshot of pageservers
-        let pageservers = locked.nodes.clone();
-
         let mut result = Vec::new();
         let mut shard_params: Option<ShardParameters> = None;
 
@@ -2543,7 +2618,8 @@ impl Service {
                         "Cannot locate a tenant that is not attached"
                     )))?;
 
-            let node = pageservers
+            let node = locked
+                .nodes
                 .get(&node_id)
                 .expect("Pageservers may not be deleted while referenced");
 
@@ -2591,6 +2667,47 @@ impl Service {
         })
     }
 
+    pub(crate) fn tenant_describe(
+        &self,
+        tenant_id: TenantId,
+    ) -> Result<TenantDescribeResponse, ApiError> {
+        let locked = self.inner.read().unwrap();
+
+        let mut shard_zero = None;
+        let mut shards = Vec::new();
+
+        for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
+        {
+            if tenant_shard_id.is_zero() {
+                shard_zero = Some(shard);
+            }
+
+            let response_shard = TenantDescribeResponseShard {
+                tenant_shard_id: *tenant_shard_id,
+                node_attached: *shard.intent.get_attached(),
+                node_secondary: shard.intent.get_secondary().to_vec(),
+                last_error: shard.last_error.lock().unwrap().clone(),
+                is_reconciling: shard.reconciler.is_some(),
+                is_pending_compute_notification: shard.pending_compute_notification,
+                is_splitting: matches!(shard.splitting, SplitState::Splitting),
+            };
+            shards.push(response_shard);
+        }
+
+        let Some(shard_zero) = shard_zero else {
+            return Err(ApiError::NotFound(
+                anyhow::anyhow!("Tenant {tenant_id} not found").into(),
+            ));
+        };
+
+        Ok(TenantDescribeResponse {
+            shards,
+            stripe_size: shard_zero.shard.stripe_size,
+            policy: shard_zero.policy.clone(),
+            config: shard_zero.config.clone(),
+        })
+    }
+
     #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
     async fn abort_tenant_shard_split(
         &self,
@@ -2648,7 +2765,7 @@ impl Service {
         let detach_locations: Vec<(Node, TenantShardId)> = {
             let mut detach_locations = Vec::new();
             let mut locked = self.inner.write().unwrap();
-            let (nodes, tenants, _scheduler) = locked.parts_mut();
+            let (nodes, tenants, scheduler) = locked.parts_mut();
 
             for (tenant_shard_id, shard) in
                 tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
@@ -2681,6 +2798,13 @@ impl Service {
 
                 tracing::info!("Restoring parent shard {tenant_shard_id}");
                 shard.splitting = SplitState::Idle;
+                if let Err(e) = shard.schedule(scheduler) {
+                    // If this shard can't be scheduled now (perhaps due to offline nodes or
+                    // capacity issues), that must not prevent us rolling back a split.  In this
+                    // case it should be eventually scheduled in the background.
+                    tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
+                }
+
                 self.maybe_reconcile_shard(shard, nodes);
             }
 
@@ -2772,7 +2896,7 @@ impl Service {
                 .map(|(shard_id, _)| *shard_id)
                 .collect::<Vec<_>>();
 
-            let (_nodes, tenants, scheduler) = locked.parts_mut();
+            let (nodes, tenants, scheduler) = locked.parts_mut();
             for parent_id in parent_ids {
                 let child_ids = parent_id.split(new_shard_count);
 
@@ -2814,7 +2938,7 @@ impl Service {
                                 generation,
                                 &child_shard,
                                 &config,
-                                matches!(policy, PlacementPolicy::Double(n) if n > 0),
+                                matches!(policy, PlacementPolicy::Attached(n) if n > 0),
                             )),
                         },
                     );
@@ -2839,6 +2963,8 @@ impl Service {
                         // find a secondary (e.g. because cluster is overloaded).
                         tracing::warn!("Failed to schedule child shard {child}: {e}");
                     }
+                    // In the background, attach secondary locations for the new shards
+                    self.maybe_reconcile_shard(&mut child_state, nodes);
 
                     tenants.insert(child, child_state);
                     response.new_shards.push(child);
@@ -2861,17 +2987,23 @@ impl Service {
         let new_shard_count = ShardCount::new(split_req.new_shard_count);
         let new_stripe_size = split_req.new_stripe_size;
 
-        let r = self.do_tenant_shard_split(tenant_id, split_req).await;
+        // Validate the request and construct parameters.  This phase is fallible, but does not require
+        // rollback on errors, as it does no I/O and mutates no state.
+        let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
+            ShardSplitAction::NoOp(resp) => return Ok(resp),
+            ShardSplitAction::Split(params) => params,
+        };
+
+        // Execute this split: this phase mutates state and does remote I/O on pageservers.  If it fails,
+        // we must roll back.
+        let r = self
+            .do_tenant_shard_split(tenant_id, shard_split_params)
+            .await;
 
         match r {
             Ok(r) => Ok(r),
-            Err(ApiError::BadRequest(_)) => {
-                // A request validation error does not require rollback: we rejected it before we started making any changes: just
-                // return the error
-                r
-            }
             Err(e) => {
-                // General case error handling: split might be part-done, we must do work to abort it.
+                // Split might be part-done, we must do work to abort it.
                 tracing::warn!("Enqueuing background abort of split on {tenant_id}");
                 self.abort_tx
                     .send(TenantShardSplitAbort {
@@ -2887,25 +3019,18 @@ impl Service {
         }
     }
 
-    pub(crate) async fn do_tenant_shard_split(
+    fn prepare_tenant_shard_split(
         &self,
         tenant_id: TenantId,
         split_req: TenantShardSplitRequest,
-    ) -> Result<TenantShardSplitResponse, ApiError> {
-        let mut policy = None;
-        let mut shard_ident = None;
-
-        // A parent shard which will be split
-        struct SplitTarget {
-            parent_id: TenantShardId,
-            node: Node,
-            child_ids: Vec<TenantShardId>,
-        }
-
+    ) -> Result<ShardSplitAction, ApiError> {
         fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
             anyhow::anyhow!("failpoint")
         )));
 
+        let mut policy = None;
+        let mut config = None;
+        let mut shard_ident = None;
         // Validate input, and calculate which shards we will create
         let (old_shard_count, targets) =
             {
@@ -2961,6 +3086,9 @@ impl Service {
                     if shard_ident.is_none() {
                         shard_ident = Some(shard.shard);
                     }
+                    if config.is_none() {
+                        config = Some(shard.config.clone());
+                    }
 
                     if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
                         tracing::info!(
@@ -2979,9 +3107,7 @@ impl Service {
                         .get(&node_id)
                         .expect("Pageservers may not be deleted while referenced");
 
-                    // TODO: if any reconciliation is currently in progress for this shard, wait for it.
-
-                    targets.push(SplitTarget {
+                    targets.push(ShardSplitTarget {
                         parent_id: *tenant_shard_id,
                         node: node.clone(),
                         child_ids: tenant_shard_id
@@ -2991,9 +3117,9 @@ impl Service {
 
                 if targets.is_empty() {
                     if children_found.len() == split_req.new_shard_count as usize {
-                        return Ok(TenantShardSplitResponse {
+                        return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
                             new_shards: children_found,
-                        });
+                        }));
                     } else {
                         // No shards found to split, and no existing children found: the
                         // tenant doesn't exist at all.
@@ -3023,13 +3149,77 @@ impl Service {
             shard_ident.unwrap()
         };
         let policy = policy.unwrap();
+        let config = config.unwrap();
 
+        Ok(ShardSplitAction::Split(ShardSplitParams {
+            old_shard_count,
+            new_shard_count: ShardCount::new(split_req.new_shard_count),
+            new_stripe_size: split_req.new_stripe_size,
+            targets,
+            policy,
+            config,
+            shard_ident,
+        }))
+    }
+
+    async fn do_tenant_shard_split(
+        &self,
+        tenant_id: TenantId,
+        params: ShardSplitParams,
+    ) -> Result<TenantShardSplitResponse, ApiError> {
         // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
         // request could occur here, deleting or mutating the tenant.  begin_shard_split checks that the
         // parent shards exist as expected, but it would be neater to do the above pre-checks within the
         // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
         // (https://github.com/neondatabase/neon/issues/6676)
 
+        let ShardSplitParams {
+            old_shard_count,
+            new_shard_count,
+            new_stripe_size,
+            mut targets,
+            policy,
+            config,
+            shard_ident,
+        } = params;
+
+        // Drop any secondary locations: pageservers do not support splitting these, and in any case the
+        // end-state for a split tenant will usually be to have secondary locations on different nodes.
+        // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
+        // at the time of split.
+        let waiters = {
+            let mut locked = self.inner.write().unwrap();
+            let mut waiters = Vec::new();
+            let (nodes, tenants, scheduler) = locked.parts_mut();
+            for target in &mut targets {
+                let Some(shard) = tenants.get_mut(&target.parent_id) else {
+                    // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
+                    return Err(ApiError::InternalServerError(anyhow::anyhow!(
+                        "Shard {} not found",
+                        target.parent_id
+                    )));
+                };
+
+                if shard.intent.get_attached() != &Some(target.node.get_id()) {
+                    // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
+                    return Err(ApiError::Conflict(format!(
+                        "Shard {} unexpectedly rescheduled during split",
+                        target.parent_id
+                    )));
+                }
+
+                // Irrespective of PlacementPolicy, clear secondary locations from intent
+                shard.intent.clear_secondary(scheduler);
+
+                // Run Reconciler to execute detach fo secondary locations.
+                if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
+                    waiters.push(waiter);
+                }
+            }
+            waiters
+        };
+        self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
+
         // Before creating any new child shards in memory or on the pageservers, persist them: this
         // enables us to ensure that we will always be able to clean up if something goes wrong.  This also
         // acts as the protection against two concurrent attempts to split: one of them will get a database
@@ -3058,8 +3248,7 @@ impl Service {
                     generation: None,
                     generation_pageserver: Some(target.node.get_id().0 as i64),
                     placement_policy: serde_json::to_string(&policy).unwrap(),
-                    // TODO: get the config out of the map
-                    config: serde_json::to_string(&TenantConfig::default()).unwrap(),
+                    config: serde_json::to_string(&config).unwrap(),
                     splitting: SplitState::Splitting,
                 });
             }
@@ -3111,18 +3300,22 @@ impl Service {
         // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
 
         for target in &targets {
-            let SplitTarget {
+            let ShardSplitTarget {
                 parent_id,
                 node,
                 child_ids,
             } = target;
-            let client = mgmt_api::Client::new(node.base_url(), self.config.jwt_token.as_deref());
+            let client = PageserverClient::new(
+                node.get_id(),
+                node.base_url(),
+                self.config.jwt_token.as_deref(),
+            );
             let response = client
                 .tenant_shard_split(
                     *parent_id,
                     TenantShardSplitRequest {
-                        new_shard_count: split_req.new_shard_count,
-                        new_stripe_size: split_req.new_stripe_size,
+                        new_shard_count: new_shard_count.literal(),
+                        new_stripe_size,
                     },
                 )
                 .await
@@ -3171,11 +3364,8 @@ impl Service {
         ));
 
         // Replace all the shards we just split with their children: this phase is infallible.
-        let (response, child_locations) = self.tenant_shard_split_commit_inmem(
-            tenant_id,
-            ShardCount::new(split_req.new_shard_count),
-            split_req.new_stripe_size,
-        );
+        let (response, child_locations) =
+            self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
 
         // Send compute notifications for all the new shards
         let mut failed_notifications = Vec::new();
@@ -3240,17 +3430,20 @@ impl Service {
                 let old_attached = *shard.intent.get_attached();
 
                 match shard.policy {
-                    PlacementPolicy::Single => {
-                        shard.intent.clear_secondary(scheduler);
-                        shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
-                    }
-                    PlacementPolicy::Double(_n) => {
+                    PlacementPolicy::Attached(n) => {
                         // If our new attached node was a secondary, it no longer should be.
                         shard.intent.remove_secondary(scheduler, migrate_req.node_id);
 
                         // If we were already attached to something, demote that to a secondary
                         if let Some(old_attached) = old_attached {
-                            shard.intent.push_secondary(scheduler, old_attached);
+                            if n > 0 {
+                                // Remove other secondaries to make room for the location we'll demote
+                                while shard.intent.get_secondary().len() >= n {
+                                    shard.intent.pop_secondary(scheduler);
+                                }
+
+                                shard.intent.push_secondary(scheduler, old_attached);
+                            }
                         }
 
                         shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
@@ -3276,7 +3469,7 @@ impl Service {
         if let Some(waiter) = waiter {
             waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
         } else {
-            tracing::warn!("Migration is a no-op");
+            tracing::info!("Migration is a no-op");
         }
 
         Ok(TenantShardMigrateResponse {})
@@ -3631,6 +3824,13 @@ impl Service {
                         observed_loc.conf = None;
                     }
 
+                    if new_nodes.len() == 1 {
+                        // Special case for single-node cluster: there is no point trying to reschedule
+                        // any tenant shards: avoid doing so, in order to avoid spewing warnings about
+                        // failures to schedule them.
+                        continue;
+                    }
+
                     if tenant_state.intent.demote_attached(node_id) {
                         tenant_state.sequence = tenant_state.sequence.next();
                         match tenant_state.schedule(scheduler) {

control_plane/attachment_service/src/tenant_state.rs

@@ -4,7 +4,10 @@ use std::{
     time::Duration,
 };
 
-use crate::{metrics, persistence::TenantShardPersistence};
+use crate::{
+    metrics::{self, ReconcileCompleteLabelGroup, ReconcileOutcome},
+    persistence::TenantShardPersistence,
+};
 use pageserver_api::controller_api::PlacementPolicy;
 use pageserver_api::{
     models::{LocationConfig, LocationConfigMode, TenantConfig},
@@ -457,22 +460,7 @@ impl TenantState {
         // Add/remove nodes to fulfil policy
         use PlacementPolicy::*;
         match self.policy {
-            Single => {
-                // Should have exactly one attached, and zero secondaries
-                if !self.intent.secondary.is_empty() {
-                    self.intent.clear_secondary(scheduler);
-                    modified = true;
-                }
-
-                let (modified_attached, _attached_node_id) = self.schedule_attached(scheduler)?;
-                modified |= modified_attached;
-
-                if !self.intent.secondary.is_empty() {
-                    self.intent.clear_secondary(scheduler);
-                    modified = true;
-                }
-            }
-            Double(secondary_count) => {
+            Attached(secondary_count) => {
                 let retain_secondaries = if self.intent.attached.is_none()
                     && scheduler.node_preferred(&self.intent.secondary).is_some()
                 {
@@ -733,7 +721,10 @@ impl TenantState {
         let reconciler_span = tracing::info_span!(parent: None, "reconciler", seq=%reconcile_seq,
                                                         tenant_id=%reconciler.tenant_shard_id.tenant_id,
                                                         shard_id=%reconciler.tenant_shard_id.shard_slug());
-        metrics::RECONCILER.spawned.inc();
+        metrics::METRICS_REGISTRY
+            .metrics_group
+            .storage_controller_reconcile_spawn
+            .inc();
         let result_tx = result_tx.clone();
         let join_handle = tokio::task::spawn(
             async move {
@@ -751,10 +742,12 @@ impl TenantState {
                 // TODO: wrap all remote API operations in cancellation check
                 // as well.
                 if reconciler.cancel.is_cancelled() {
-                    metrics::RECONCILER
-                        .complete
-                        .with_label_values(&[metrics::ReconcilerMetrics::CANCEL])
-                        .inc();
+                    metrics::METRICS_REGISTRY
+                        .metrics_group
+                        .storage_controller_reconcile_complete
+                        .inc(ReconcileCompleteLabelGroup {
+                            status: ReconcileOutcome::Cancel,
+                        });
                     return;
                 }
 
@@ -769,18 +762,18 @@ impl TenantState {
                 }
 
                 // Update result counter
-                match &result {
-                    Ok(_) => metrics::RECONCILER
-                        .complete
-                        .with_label_values(&[metrics::ReconcilerMetrics::SUCCESS]),
-                    Err(ReconcileError::Cancel) => metrics::RECONCILER
-                        .complete
-                        .with_label_values(&[metrics::ReconcilerMetrics::CANCEL]),
-                    Err(_) => metrics::RECONCILER
-                        .complete
-                        .with_label_values(&[metrics::ReconcilerMetrics::ERROR]),
-                }
-                .inc();
+                let outcome_label = match &result {
+                    Ok(_) => ReconcileOutcome::Success,
+                    Err(ReconcileError::Cancel) => ReconcileOutcome::Cancel,
+                    Err(_) => ReconcileOutcome::Error,
+                };
+
+                metrics::METRICS_REGISTRY
+                    .metrics_group
+                    .storage_controller_reconcile_complete
+                    .inc(ReconcileCompleteLabelGroup {
+                        status: outcome_label,
+                    });
 
                 result_tx
                     .send(ReconcileResult {
@@ -895,7 +888,7 @@ pub(crate) mod tests {
 
         let mut scheduler = Scheduler::new(nodes.values());
 
-        let mut tenant_state = make_test_tenant_shard(PlacementPolicy::Double(1));
+        let mut tenant_state = make_test_tenant_shard(PlacementPolicy::Attached(1));
         tenant_state
             .schedule(&mut scheduler)
             .expect("we have enough nodes, scheduling should work");
@@ -943,7 +936,7 @@ pub(crate) mod tests {
         let nodes = make_test_nodes(3);
         let mut scheduler = Scheduler::new(nodes.values());
 
-        let mut tenant_state = make_test_tenant_shard(PlacementPolicy::Double(1));
+        let mut tenant_state = make_test_tenant_shard(PlacementPolicy::Attached(1));
 
         tenant_state.observed.locations.insert(
             NodeId(3),

control_plane/src/background_process.rs

@@ -294,7 +294,7 @@ where
     //      is in state 'taken' but the thread that would unlock it is
     //      not there.
     //   2. A rust object that represented some external resource in the
-    //      parent now got implicitly copied by the the fork, even though
+    //      parent now got implicitly copied by the fork, even though
     //      the object's type is not `Copy`. The parent program may use
     //      non-copyability as way to enforce unique ownership of an
     //      external resource in the typesystem. The fork breaks that

control_plane/src/bin/neon_local.rs

@@ -437,7 +437,7 @@ async fn handle_tenant(
 
             let placement_policy = match create_match.get_one::<String>("placement-policy") {
                 Some(s) if !s.is_empty() => serde_json::from_str::<PlacementPolicy>(s)?,
-                _ => PlacementPolicy::Single,
+                _ => PlacementPolicy::Attached(0),
             };
 
             let tenant_conf = PageServerNode::parse_config(tenant_conf)?;
@@ -523,88 +523,6 @@ async fn handle_tenant(
                 .with_context(|| format!("Tenant config failed for tenant with id {tenant_id}"))?;
             println!("tenant {tenant_id} successfully configured on the pageserver");
         }
-        Some(("migrate", matches)) => {
-            let tenant_shard_id = get_tenant_shard_id(matches, env)?;
-            let new_pageserver = get_pageserver(env, matches)?;
-            let new_pageserver_id = new_pageserver.conf.id;
-
-            let storage_controller = StorageController::from_env(env);
-            storage_controller
-                .tenant_migrate(tenant_shard_id, new_pageserver_id)
-                .await?;
-
-            println!("tenant {tenant_shard_id} migrated to {}", new_pageserver_id);
-        }
-        Some(("status", matches)) => {
-            let tenant_id = get_tenant_id(matches, env)?;
-
-            let mut shard_table = comfy_table::Table::new();
-            shard_table.set_header(["Shard", "Pageserver", "Physical Size"]);
-
-            let mut tenant_synthetic_size = None;
-
-            let storage_controller = StorageController::from_env(env);
-            for shard in storage_controller.tenant_locate(tenant_id).await?.shards {
-                let pageserver =
-                    PageServerNode::from_env(env, env.get_pageserver_conf(shard.node_id)?);
-
-                let size = pageserver
-                    .http_client
-                    .tenant_details(shard.shard_id)
-                    .await?
-                    .tenant_info
-                    .current_physical_size
-                    .unwrap();
-
-                shard_table.add_row([
-                    format!("{}", shard.shard_id.shard_slug()),
-                    format!("{}", shard.node_id.0),
-                    format!("{} MiB", size / (1024 * 1024)),
-                ]);
-
-                if shard.shard_id.is_zero() {
-                    tenant_synthetic_size =
-                        Some(pageserver.tenant_synthetic_size(shard.shard_id).await?);
-                }
-            }
-
-            let Some(synthetic_size) = tenant_synthetic_size else {
-                bail!("Shard 0 not found")
-            };
-
-            let mut tenant_table = comfy_table::Table::new();
-            tenant_table.add_row(["Tenant ID".to_string(), tenant_id.to_string()]);
-            tenant_table.add_row([
-                "Synthetic size".to_string(),
-                format!("{} MiB", synthetic_size.size.unwrap_or(0) / (1024 * 1024)),
-            ]);
-
-            println!("{tenant_table}");
-            println!("{shard_table}");
-        }
-        Some(("shard-split", matches)) => {
-            let tenant_id = get_tenant_id(matches, env)?;
-            let shard_count: u8 = matches.get_one::<u8>("shard-count").cloned().unwrap_or(0);
-            let shard_stripe_size: Option<ShardStripeSize> = matches
-                .get_one::<Option<ShardStripeSize>>("shard-stripe-size")
-                .cloned()
-                .unwrap();
-
-            let storage_controller = StorageController::from_env(env);
-            let result = storage_controller
-                .tenant_split(tenant_id, shard_count, shard_stripe_size)
-                .await?;
-            println!(
-                "Split tenant {} into shards {}",
-                tenant_id,
-                result
-                    .new_shards
-                    .iter()
-                    .map(|s| format!("{:?}", s))
-                    .collect::<Vec<_>>()
-                    .join(",")
-            );
-        }
 
         Some((sub_name, _)) => bail!("Unexpected tenant subcommand '{}'", sub_name),
         None => bail!("no tenant subcommand provided"),
@@ -1578,19 +1496,6 @@ fn cli() -> Command {
             .subcommand(Command::new("config")
                 .arg(tenant_id_arg.clone())
                 .arg(Arg::new("config").short('c').num_args(1).action(ArgAction::Append).required(false)))
-            .subcommand(Command::new("migrate")
-                .about("Migrate a tenant from one pageserver to another")
-                .arg(tenant_id_arg.clone())
-                .arg(pageserver_id_arg.clone()))
-            .subcommand(Command::new("status")
-                .about("Human readable summary of the tenant's shards and attachment locations")
-                .arg(tenant_id_arg.clone()))
-            .subcommand(Command::new("shard-split")
-                .about("Increase the number of shards in the tenant")
-                .arg(tenant_id_arg.clone())
-                .arg(Arg::new("shard-count").value_parser(value_parser!(u8)).long("shard-count").action(ArgAction::Set).help("Number of shards in the new tenant (default 1)"))
-                .arg(Arg::new("shard-stripe-size").value_parser(value_parser!(u32)).long("shard-stripe-size").action(ArgAction::Set).help("Sharding stripe size in pages"))
-                )
         )
         .subcommand(
             Command::new("pageserver")

control_plane/src/endpoint.rs

@@ -12,7 +12,7 @@
 //!
 //! The endpoint is managed by the `compute_ctl` binary. When an endpoint is
 //! started, we launch `compute_ctl` It synchronizes the safekeepers, downloads
-//! the basebackup from the pageserver to initialize the the data directory, and
+//! the basebackup from the pageserver to initialize the data directory, and
 //! finally launches the PostgreSQL process. It watches the PostgreSQL process
 //! until it exits.
 //!

control_plane/src/local_env.rs

@@ -127,8 +127,8 @@ pub struct PageServerConf {
     pub pg_auth_type: AuthType,
     pub http_auth_type: AuthType,
 
-    pub(crate) virtual_file_io_engine: String,
-    pub(crate) get_vectored_impl: String,
+    pub(crate) virtual_file_io_engine: Option<String>,
+    pub(crate) get_vectored_impl: Option<String>,
 }
 
 impl Default for PageServerConf {
@@ -139,9 +139,8 @@ impl Default for PageServerConf {
             listen_http_addr: String::new(),
             pg_auth_type: AuthType::Trust,
             http_auth_type: AuthType::Trust,
-            // FIXME: use the ones exposed by pageserver crate
-            virtual_file_io_engine: "tokio-epoll-uring".to_owned(),
-            get_vectored_impl: "sequential".to_owned(),
+            virtual_file_io_engine: None,
+            get_vectored_impl: None,
         }
     }
 }

control_plane/src/pageserver.rs

@@ -101,8 +101,16 @@ impl PageServerNode {
 
         let pg_auth_type_param = format!("pg_auth_type='{}'", pg_auth_type);
         let listen_pg_addr_param = format!("listen_pg_addr='{}'", listen_pg_addr);
-        let virtual_file_io_engine = format!("virtual_file_io_engine='{virtual_file_io_engine}'");
-        let get_vectored_impl = format!("get_vectored_impl='{get_vectored_impl}'");
+        let virtual_file_io_engine = if let Some(virtual_file_io_engine) = virtual_file_io_engine {
+            format!("virtual_file_io_engine='{virtual_file_io_engine}'")
+        } else {
+            String::new()
+        };
+        let get_vectored_impl = if let Some(get_vectored_impl) = get_vectored_impl {
+            format!("get_vectored_impl='{get_vectored_impl}'")
+        } else {
+            String::new()
+        };
 
         let broker_endpoint_param = format!("broker_endpoint='{}'", self.env.broker.client_url());
 

control_plane/src/storage_controller.rs

@@ -279,6 +279,7 @@ impl StorageController {
             &self.listen,
             "-p",
             self.path.as_ref(),
+            "--dev",
             "--database-url",
             &database_url,
             "--max-unavailable-interval",
@@ -475,7 +476,7 @@ impl StorageController {
     pub async fn tenant_locate(&self, tenant_id: TenantId) -> anyhow::Result<TenantLocateResponse> {
         self.dispatch::<(), _>(
             Method::GET,
-            format!("control/v1/tenant/{tenant_id}/locate"),
+            format!("debug/v1/tenant/{tenant_id}/locate"),
             None,
         )
         .await

libs/metrics/src/hll.rs

@@ -40,7 +40,7 @@ macro_rules! register_hll {
     }};
 
     ($N:literal, $NAME:expr, $HELP:expr $(,)?) => {{
-        $crate::register_hll!($N, $crate::opts!($NAME, $HELP), $LABELS_NAMES)
+        $crate::register_hll!($N, $crate::opts!($NAME, $HELP))
     }};
 }
 

libs/pageserver_api/src/controller_api.rs

@@ -6,7 +6,10 @@ use std::str::FromStr;
 use serde::{Deserialize, Serialize};
 use utils::id::NodeId;
 
-use crate::{models::ShardParameters, shard::TenantShardId};
+use crate::{
+    models::{ShardParameters, TenantConfig},
+    shard::{ShardStripeSize, TenantShardId},
+};
 
 #[derive(Serialize, Deserialize)]
 pub struct TenantCreateResponseShard {
@@ -57,6 +60,31 @@ pub struct TenantLocateResponse {
     pub shard_params: ShardParameters,
 }
 
+#[derive(Serialize, Deserialize)]
+pub struct TenantDescribeResponse {
+    pub shards: Vec<TenantDescribeResponseShard>,
+    pub stripe_size: ShardStripeSize,
+    pub policy: PlacementPolicy,
+    pub config: TenantConfig,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct TenantDescribeResponseShard {
+    pub tenant_shard_id: TenantShardId,
+
+    pub node_attached: Option<NodeId>,
+    pub node_secondary: Vec<NodeId>,
+
+    pub last_error: String,
+
+    /// A task is currently running to reconcile this tenant's intent state with the state on pageservers
+    pub is_reconciling: bool,
+    /// This shard failed in sending a compute notification to the cloud control plane, and a retry is pending.
+    pub is_pending_compute_notification: bool,
+    /// A shard split is currently underway
+    pub is_splitting: bool,
+}
+
 /// Explicitly migrating a particular shard is a low level operation
 /// TODO: higher level "Reschedule tenant" operation where the request
 /// specifies some constraints, e.g. asking it to get off particular node(s)
@@ -181,11 +209,8 @@ impl From<NodeSchedulingPolicy> for String {
 /// to create secondary locations.
 #[derive(Clone, Serialize, Deserialize, Debug, PartialEq, Eq)]
 pub enum PlacementPolicy {
-    /// Cheapest way to attach a tenant: just one pageserver, no secondary
-    Single,
-    /// Production-ready way to attach a tenant: one attached pageserver and
-    /// some number of secondaries.
-    Double(usize),
+    /// Normal live state: one attached pageserver and zero or more secondaries.
+    Attached(usize),
     /// Create one secondary mode locations. This is useful when onboarding
     /// a tenant, or for an idle tenant that we might want to bring online quickly.
     Secondary,
@@ -207,14 +232,14 @@ mod test {
     /// Check stability of PlacementPolicy's serialization
     #[test]
     fn placement_policy_encoding() -> anyhow::Result<()> {
-        let v = PlacementPolicy::Double(1);
+        let v = PlacementPolicy::Attached(1);
         let encoded = serde_json::to_string(&v)?;
-        assert_eq!(encoded, "{\"Double\":1}");
+        assert_eq!(encoded, "{\"Attached\":1}");
         assert_eq!(serde_json::from_str::<PlacementPolicy>(&encoded)?, v);
 
-        let v = PlacementPolicy::Single;
+        let v = PlacementPolicy::Detached;
         let encoded = serde_json::to_string(&v)?;
-        assert_eq!(encoded, "\"Single\"");
+        assert_eq!(encoded, "\"Detached\"");
         assert_eq!(serde_json::from_str::<PlacementPolicy>(&encoded)?, v);
         Ok(())
     }

libs/pageserver_api/src/upcall_api.rs

@@ -6,7 +6,9 @@
 use serde::{Deserialize, Serialize};
 use utils::id::NodeId;
 
-use crate::{controller_api::NodeRegisterRequest, shard::TenantShardId};
+use crate::{
+    controller_api::NodeRegisterRequest, models::LocationConfigMode, shard::TenantShardId,
+};
 
 /// Upcall message sent by the pageserver to the configured `control_plane_api` on
 /// startup.
@@ -20,12 +22,20 @@ pub struct ReAttachRequest {
     pub register: Option<NodeRegisterRequest>,
 }
 
-#[derive(Serialize, Deserialize)]
-pub struct ReAttachResponseTenant {
-    pub id: TenantShardId,
-    pub gen: u32,
+fn default_mode() -> LocationConfigMode {
+    LocationConfigMode::AttachedSingle
 }
 
+#[derive(Serialize, Deserialize, Debug)]
+pub struct ReAttachResponseTenant {
+    pub id: TenantShardId,
+    /// Mandatory if LocationConfigMode is None or set to an Attached* mode
+    pub gen: Option<u32>,
+
+    /// Default value only for backward compat: this field should be set
+    #[serde(default = "default_mode")]
+    pub mode: LocationConfigMode,
+}
 #[derive(Serialize, Deserialize)]
 pub struct ReAttachResponse {
     pub tenants: Vec<ReAttachResponseTenant>,

libs/postgres_ffi/wal_craft/src/bin/wal_craft.rs

@@ -1,5 +1,6 @@
 use anyhow::*;
 use clap::{value_parser, Arg, ArgMatches, Command};
+use postgres::Client;
 use std::{path::PathBuf, str::FromStr};
 use wal_craft::*;
 
@@ -8,8 +9,8 @@ fn main() -> Result<()> {
         .init();
     let arg_matches = cli().get_matches();
 
-    let wal_craft = |arg_matches: &ArgMatches, client| {
-        let (intermediate_lsns, end_of_wal_lsn) = match arg_matches
+    let wal_craft = |arg_matches: &ArgMatches, client: &mut Client| {
+        let intermediate_lsns = match arg_matches
             .get_one::<String>("type")
             .map(|s| s.as_str())
             .context("'type' is required")?
@@ -25,6 +26,7 @@ fn main() -> Result<()> {
             LastWalRecordCrossingSegment::NAME => LastWalRecordCrossingSegment::craft(client)?,
             a => panic!("Unknown --type argument: {a}"),
         };
+        let end_of_wal_lsn = client.pg_current_wal_insert_lsn()?;
         for lsn in intermediate_lsns {
             println!("intermediate_lsn = {lsn}");
         }

libs/postgres_ffi/wal_craft/src/lib.rs

@@ -5,7 +5,6 @@ use postgres::types::PgLsn;
 use postgres::Client;
 use postgres_ffi::{WAL_SEGMENT_SIZE, XLOG_BLCKSZ};
 use postgres_ffi::{XLOG_SIZE_OF_XLOG_RECORD, XLOG_SIZE_OF_XLOG_SHORT_PHD};
-use std::cmp::Ordering;
 use std::path::{Path, PathBuf};
 use std::process::Command;
 use std::time::{Duration, Instant};
@@ -232,59 +231,52 @@ pub fn ensure_server_config(client: &mut impl postgres::GenericClient) -> anyhow
 pub trait Crafter {
     const NAME: &'static str;
 
-    /// Generates WAL using the client `client`. Returns a pair of:
-    /// * A vector of some valid "interesting" intermediate LSNs which one may start reading from.
-    ///   May include or exclude Lsn(0) and the end-of-wal.
-    /// * The expected end-of-wal LSN.
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)>;
+    /// Generates WAL using the client `client`. Returns a vector of some valid
+    /// "interesting" intermediate LSNs which one may start reading from.
+    /// test_end_of_wal uses this to check various starting points.
+    ///
+    /// Note that postgres is generally keen about writing some WAL. While we
+    /// try to disable it (autovacuum, big wal_writer_delay, etc) it is always
+    /// possible, e.g. xl_running_xacts are dumped each 15s. So checks about
+    /// stable WAL end would be flaky unless postgres is shut down. For this
+    /// reason returning potential end of WAL here is pointless. Most of the
+    /// time this doesn't happen though, so it is reasonable to create needed
+    /// WAL structure and immediately kill postgres like test_end_of_wal does.
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>>;
 }
 
+/// Wraps some WAL craft function, providing current LSN to it before the
+/// insertion and flushing WAL afterwards. Also pushes initial LSN to the
+/// result.
 fn craft_internal<C: postgres::GenericClient>(
     client: &mut C,
-    f: impl Fn(&mut C, PgLsn) -> anyhow::Result<(Vec<PgLsn>, Option<PgLsn>)>,
-) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
+    f: impl Fn(&mut C, PgLsn) -> anyhow::Result<Vec<PgLsn>>,
+) -> anyhow::Result<Vec<PgLsn>> {
     ensure_server_config(client)?;
 
     let initial_lsn = client.pg_current_wal_insert_lsn()?;
     info!("LSN initial = {}", initial_lsn);
 
-    let (mut intermediate_lsns, last_lsn) = f(client, initial_lsn)?;
-    let last_lsn = match last_lsn {
-        None => client.pg_current_wal_insert_lsn()?,
-        Some(last_lsn) => {
-            let insert_lsn = client.pg_current_wal_insert_lsn()?;
-            match last_lsn.cmp(&insert_lsn) {
-                Ordering::Less => bail!(
-                    "Some records were inserted after the crafted WAL: {} vs {}",
-                    last_lsn,
-                    insert_lsn
-                ),
-                Ordering::Equal => last_lsn,
-                Ordering::Greater => bail!("Reported LSN is greater than insert_lsn"),
-            }
-        }
-    };
+    let mut intermediate_lsns = f(client, initial_lsn)?;
     if !intermediate_lsns.starts_with(&[initial_lsn]) {
         intermediate_lsns.insert(0, initial_lsn);
     }
 
     // Some records may be not flushed, e.g. non-transactional logical messages.
+    //
+    // Note: this is broken if pg_current_wal_insert_lsn is at page boundary
+    // because pg_current_wal_insert_lsn skips page headers.
     client.execute("select neon_xlogflush(pg_current_wal_insert_lsn())", &[])?;
-    match last_lsn.cmp(&client.pg_current_wal_flush_lsn()?) {
-        Ordering::Less => bail!("Some records were flushed after the crafted WAL"),
-        Ordering::Equal => {}
-        Ordering::Greater => bail!("Reported LSN is greater than flush_lsn"),
-    }
-    Ok((intermediate_lsns, last_lsn))
+    Ok(intermediate_lsns)
 }
 
 pub struct Simple;
 impl Crafter for Simple {
     const NAME: &'static str = "simple";
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>> {
         craft_internal(client, |client, _| {
             client.execute("CREATE table t(x int)", &[])?;
-            Ok((Vec::new(), None))
+            Ok(Vec::new())
         })
     }
 }
@@ -292,29 +284,36 @@ impl Crafter for Simple {
 pub struct LastWalRecordXlogSwitch;
 impl Crafter for LastWalRecordXlogSwitch {
     const NAME: &'static str = "last_wal_record_xlog_switch";
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
-        // Do not use generate_internal because here we end up with flush_lsn exactly on
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>> {
+        // Do not use craft_internal because here we end up with flush_lsn exactly on
         // the segment boundary and insert_lsn after the initial page header, which is unusual.
         ensure_server_config(client)?;
 
         client.execute("CREATE table t(x int)", &[])?;
         let before_xlog_switch = client.pg_current_wal_insert_lsn()?;
-        let after_xlog_switch: PgLsn = client.query_one("SELECT pg_switch_wal()", &[])?.get(0);
-        let next_segment = PgLsn::from(0x0200_0000);
+        // pg_switch_wal returns end of last record of the switched segment,
+        // i.e. end of SWITCH itself.
+        let xlog_switch_record_end: PgLsn = client.query_one("SELECT pg_switch_wal()", &[])?.get(0);
+        let before_xlog_switch_u64 = u64::from(before_xlog_switch);
+        let next_segment = PgLsn::from(
+            before_xlog_switch_u64 - (before_xlog_switch_u64 % WAL_SEGMENT_SIZE as u64)
+                + WAL_SEGMENT_SIZE as u64,
+        );
         ensure!(
-            after_xlog_switch <= next_segment,
-            "XLOG_SWITCH message ended after the expected segment boundary: {} > {}",
-            after_xlog_switch,
+            xlog_switch_record_end <= next_segment,
+            "XLOG_SWITCH record ended after the expected segment boundary: {} > {}",
+            xlog_switch_record_end,
             next_segment
         );
-        Ok((vec![before_xlog_switch, after_xlog_switch], next_segment))
+        Ok(vec![before_xlog_switch, xlog_switch_record_end])
     }
 }
 
 pub struct LastWalRecordXlogSwitchEndsOnPageBoundary;
+/// Craft xlog SWITCH record ending at page boundary.
 impl Crafter for LastWalRecordXlogSwitchEndsOnPageBoundary {
     const NAME: &'static str = "last_wal_record_xlog_switch_ends_on_page_boundary";
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>> {
         // Do not use generate_internal because here we end up with flush_lsn exactly on
         // the segment boundary and insert_lsn after the initial page header, which is unusual.
         ensure_server_config(client)?;
@@ -361,28 +360,29 @@ impl Crafter for LastWalRecordXlogSwitchEndsOnPageBoundary {
 
         // Emit the XLOG_SWITCH
         let before_xlog_switch = client.pg_current_wal_insert_lsn()?;
-        let after_xlog_switch: PgLsn = client.query_one("SELECT pg_switch_wal()", &[])?.get(0);
+        let xlog_switch_record_end: PgLsn = client.query_one("SELECT pg_switch_wal()", &[])?.get(0);
         let next_segment = PgLsn::from(0x0200_0000);
         ensure!(
-            after_xlog_switch < next_segment,
-            "XLOG_SWITCH message ended on or after the expected segment boundary: {} > {}",
-            after_xlog_switch,
+            xlog_switch_record_end < next_segment,
+            "XLOG_SWITCH record ended on or after the expected segment boundary: {} > {}",
+            xlog_switch_record_end,
             next_segment
         );
         ensure!(
-            u64::from(after_xlog_switch) as usize % XLOG_BLCKSZ == XLOG_SIZE_OF_XLOG_SHORT_PHD,
+            u64::from(xlog_switch_record_end) as usize % XLOG_BLCKSZ == XLOG_SIZE_OF_XLOG_SHORT_PHD,
             "XLOG_SWITCH message ended not on page boundary: {}, offset = {}",
-            after_xlog_switch,
-            u64::from(after_xlog_switch) as usize % XLOG_BLCKSZ
+            xlog_switch_record_end,
+            u64::from(xlog_switch_record_end) as usize % XLOG_BLCKSZ
         );
-        Ok((vec![before_xlog_switch, after_xlog_switch], next_segment))
+        Ok(vec![before_xlog_switch, xlog_switch_record_end])
     }
 }
 
-fn craft_single_logical_message(
+/// Write ~16MB logical message; it should cross WAL segment.
+fn craft_seg_size_logical_message(
     client: &mut impl postgres::GenericClient,
     transactional: bool,
-) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
+) -> anyhow::Result<Vec<PgLsn>> {
     craft_internal(client, |client, initial_lsn| {
         ensure!(
             initial_lsn < PgLsn::from(0x0200_0000 - 1024 * 1024),
@@ -405,34 +405,24 @@ fn craft_single_logical_message(
             "Logical message crossed two segments"
         );
 
-        if transactional {
-            // Transactional logical messages are part of a transaction, so the one above is
-            // followed by a small COMMIT record.
-
-            let after_message_lsn = client.pg_current_wal_insert_lsn()?;
-            ensure!(
-                message_lsn < after_message_lsn,
-                "No record found after the emitted message"
-            );
-            Ok((vec![message_lsn], Some(after_message_lsn)))
-        } else {
-            Ok((Vec::new(), Some(message_lsn)))
-        }
+        Ok(vec![message_lsn])
     })
 }
 
 pub struct WalRecordCrossingSegmentFollowedBySmallOne;
 impl Crafter for WalRecordCrossingSegmentFollowedBySmallOne {
     const NAME: &'static str = "wal_record_crossing_segment_followed_by_small_one";
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
-        craft_single_logical_message(client, true)
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>> {
+        // Transactional message crossing WAL segment will be followed by small
+        // commit record.
+        craft_seg_size_logical_message(client, true)
     }
 }
 
 pub struct LastWalRecordCrossingSegment;
 impl Crafter for LastWalRecordCrossingSegment {
     const NAME: &'static str = "last_wal_record_crossing_segment";
-    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<(Vec<PgLsn>, PgLsn)> {
-        craft_single_logical_message(client, false)
+    fn craft(client: &mut impl postgres::GenericClient) -> anyhow::Result<Vec<PgLsn>> {
+        craft_seg_size_logical_message(client, false)
     }
 }

libs/postgres_ffi/wal_craft/src/xlog_utils_test.rs

@@ -11,13 +11,15 @@ use utils::const_assert;
 use utils::lsn::Lsn;
 
 fn init_logging() {
-    let _ = env_logger::Builder::from_env(env_logger::Env::default().default_filter_or(
-        format!("crate=info,postgres_ffi::{PG_MAJORVERSION}::xlog_utils=trace"),
-    ))
+    let _ = env_logger::Builder::from_env(env_logger::Env::default().default_filter_or(format!(
+        "crate=info,postgres_ffi::{PG_MAJORVERSION}::xlog_utils=trace"
+    )))
     .is_test(true)
     .try_init();
 }
 
+/// Test that find_end_of_wal returns the same results as pg_dump on various
+/// WALs created by Crafter.
 fn test_end_of_wal<C: crate::Crafter>(test_name: &str) {
     use crate::*;
 
@@ -38,13 +40,13 @@ fn test_end_of_wal<C: crate::Crafter>(test_name: &str) {
     }
     cfg.initdb().unwrap();
     let srv = cfg.start_server().unwrap();
-    let (intermediate_lsns, expected_end_of_wal_partial) =
-        C::craft(&mut srv.connect_with_timeout().unwrap()).unwrap();
+    let intermediate_lsns = C::craft(&mut srv.connect_with_timeout().unwrap()).unwrap();
     let intermediate_lsns: Vec<Lsn> = intermediate_lsns
         .iter()
         .map(|&lsn| u64::from(lsn).into())
         .collect();
-    let expected_end_of_wal: Lsn = u64::from(expected_end_of_wal_partial).into();
+    // Kill postgres. Note that it might have inserted to WAL something after
+    // 'craft' did its job.
     srv.kill();
 
     // Check find_end_of_wal on the initial WAL
@@ -56,7 +58,7 @@ fn test_end_of_wal<C: crate::Crafter>(test_name: &str) {
         .filter(|fname| IsXLogFileName(fname))
         .max()
         .unwrap();
-    check_pg_waldump_end_of_wal(&cfg, &last_segment, expected_end_of_wal);
+    let expected_end_of_wal = find_pg_waldump_end_of_wal(&cfg, &last_segment);
     for start_lsn in intermediate_lsns
         .iter()
         .chain(std::iter::once(&expected_end_of_wal))
@@ -91,11 +93,7 @@ fn test_end_of_wal<C: crate::Crafter>(test_name: &str) {
     }
 }
 
-fn check_pg_waldump_end_of_wal(
-    cfg: &crate::Conf,
-    last_segment: &str,
-    expected_end_of_wal: Lsn,
-) {
+fn find_pg_waldump_end_of_wal(cfg: &crate::Conf, last_segment: &str) -> Lsn {
     // Get the actual end of WAL by pg_waldump
     let waldump_output = cfg
         .pg_waldump("000000010000000000000001", last_segment)
@@ -113,11 +111,8 @@ fn check_pg_waldump_end_of_wal(
         }
     };
     let waldump_wal_end = Lsn::from_str(caps.get(1).unwrap().as_str()).unwrap();
-    info!(
-        "waldump erred on {}, expected wal end at {}",
-        waldump_wal_end, expected_end_of_wal
-    );
-    assert_eq!(waldump_wal_end, expected_end_of_wal);
+    info!("waldump erred on {}", waldump_wal_end);
+    waldump_wal_end
 }
 
 fn check_end_of_wal(
@@ -210,9 +205,9 @@ pub fn test_update_next_xid() {
 #[test]
 pub fn test_encode_logical_message() {
     let expected = [
-        64, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 21, 0, 0, 170, 34, 166, 227, 255,
-        38, 0, 0, 0, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 112, 114,
-        101, 102, 105, 120, 0, 109, 101, 115, 115, 97, 103, 101,
+        64, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 21, 0, 0, 170, 34, 166, 227, 255, 38,
+        0, 0, 0, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 112, 114, 101, 102,
+        105, 120, 0, 109, 101, 115, 115, 97, 103, 101,
     ];
     let actual = encode_logical_message("prefix", "message");
     assert_eq!(expected, actual[..]);

libs/remote_storage/src/local_fs.rs

@@ -198,6 +198,7 @@ impl LocalFs {
             fs::OpenOptions::new()
                 .write(true)
                 .create(true)
+                .truncate(true)
                 .open(&temp_file_path)
                 .await
                 .with_context(|| {

libs/tenant_size_model/tests/tests.rs

@@ -247,7 +247,7 @@ fn scenario_4() {
     //
     // This is in total 5000 + 1000 + 5000 + 1000 = 12000
     //
-    // (If we used the the method from the previous scenario, and
+    // (If we used the method from the previous scenario, and
     // kept only snapshot at the branch point, we'd need to keep
     // all the WAL between 10000-18000 on the main branch, so
     // the total size would be 5000 + 1000 + 8000 = 14000. The

libs/utils/Cargo.toml

@@ -13,6 +13,7 @@ testing = ["fail/failpoints"]
 [dependencies]
 arc-swap.workspace = true
 sentry.workspace = true
+async-compression.workspace = true
 async-trait.workspace = true
 anyhow.workspace = true
 bincode.workspace = true
@@ -36,6 +37,7 @@ serde_json.workspace = true
 signal-hook.workspace = true
 thiserror.workspace = true
 tokio.workspace = true
+tokio-tar.workspace = true
 tokio-util.workspace = true
 tracing.workspace = true
 tracing-error.workspace = true
@@ -46,6 +48,7 @@ strum.workspace = true
 strum_macros.workspace = true
 url.workspace = true
 uuid.workspace = true
+walkdir.workspace = true
 
 pq_proto.workspace = true
 postgres_connection.workspace = true

libs/utils/src/http/endpoint.rs

@@ -245,7 +245,7 @@ impl std::io::Write for ChannelWriter {
     }
 }
 
-async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
     SERVE_METRICS_COUNT.inc();
 
     let started_at = std::time::Instant::now();
@@ -367,7 +367,6 @@ pub fn make_router() -> RouterBuilder<hyper::Body, ApiError> {
         .middleware(Middleware::post_with_info(
             add_request_id_header_to_response,
         ))
-        .get("/metrics", |r| request_span(r, prometheus_metrics_handler))
         .err_handler(route_error_handler)
 }
 

libs/utils/src/lib.rs

@@ -87,6 +87,8 @@ pub mod failpoint_support;
 
 pub mod yielding_loop;
 
+pub mod zstd;
+
 /// This is a shortcut to embed git sha into binaries and avoid copying the same build script to all packages
 ///
 /// we have several cases:

libs/utils/src/lock_file.rs

@@ -63,6 +63,7 @@ impl UnwrittenLockFile {
 pub fn create_exclusive(lock_file_path: &Utf8Path) -> anyhow::Result<UnwrittenLockFile> {
     let lock_file = fs::OpenOptions::new()
         .create(true) // O_CREAT
+        .truncate(true)
         .write(true)
         .open(lock_file_path)
         .context("open lock file")?;

libs/utils/src/sync/heavier_once_cell.rs

@@ -245,7 +245,7 @@ impl<'a, T> Guard<'a, T> {
     ///
     /// The permit will be on a semaphore part of the new internal value, and any following
     /// [`OnceCell::get_or_init`] will wait on it to complete.
-    pub fn take_and_deinit(&mut self) -> (T, InitPermit) {
+    pub fn take_and_deinit(mut self) -> (T, InitPermit) {
         let mut swapped = Inner::default();
         let sem = swapped.init_semaphore.clone();
         // acquire and forget right away, moving the control over to InitPermit
@@ -543,7 +543,7 @@ mod tests {
         target.set(42, permit);
 
         let (_answer, permit) = {
-            let mut guard = target
+            let guard = target
                 .get_or_init(|permit| async { Ok::<_, Infallible>((11, permit)) })
                 .await
                 .unwrap();

libs/utils/src/vec_map.rs

@@ -1,27 +1,60 @@
 use std::{alloc::Layout, cmp::Ordering, ops::RangeBounds};
 
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum VecMapOrdering {
+    Greater,
+    GreaterOrEqual,
+}
+
 /// Ordered map datastructure implemented in a Vec.
 /// Append only - can only add keys that are larger than the
 /// current max key.
+/// Ordering can be adjusted using [`VecMapOrdering`]
+/// during `VecMap` construction.
 #[derive(Clone, Debug)]
-pub struct VecMap<K, V>(Vec<(K, V)>);
+pub struct VecMap<K, V> {
+    data: Vec<(K, V)>,
+    ordering: VecMapOrdering,
+}
 
 impl<K, V> Default for VecMap<K, V> {
     fn default() -> Self {
-        VecMap(Default::default())
+        VecMap {
+            data: Default::default(),
+            ordering: VecMapOrdering::Greater,
+        }
     }
 }
 
-#[derive(Debug)]
-pub struct InvalidKey;
+#[derive(thiserror::Error, Debug)]
+pub enum VecMapError {
+    #[error("Key violates ordering constraint")]
+    InvalidKey,
+    #[error("Mismatched ordering constraints")]
+    ExtendOrderingError,
+}
 
 impl<K: Ord, V> VecMap<K, V> {
+    pub fn new(ordering: VecMapOrdering) -> Self {
+        Self {
+            data: Vec::new(),
+            ordering,
+        }
+    }
+
+    pub fn with_capacity(capacity: usize, ordering: VecMapOrdering) -> Self {
+        Self {
+            data: Vec::with_capacity(capacity),
+            ordering,
+        }
+    }
+
     pub fn is_empty(&self) -> bool {
-        self.0.is_empty()
+        self.data.is_empty()
     }
 
     pub fn as_slice(&self) -> &[(K, V)] {
-        self.0.as_slice()
+        self.data.as_slice()
     }
 
     /// This function may panic if given a range where the lower bound is
@@ -29,7 +62,7 @@ impl<K: Ord, V> VecMap<K, V> {
     pub fn slice_range<R: RangeBounds<K>>(&self, range: R) -> &[(K, V)] {
         use std::ops::Bound::*;
 
-        let binary_search = |k: &K| self.0.binary_search_by_key(&k, extract_key);
+        let binary_search = |k: &K| self.data.binary_search_by_key(&k, extract_key);
 
         let start_idx = match range.start_bound() {
             Unbounded => 0,
@@ -41,7 +74,7 @@ impl<K: Ord, V> VecMap<K, V> {
         };
 
         let end_idx = match range.end_bound() {
-            Unbounded => self.0.len(),
+            Unbounded => self.data.len(),
             Included(k) => match binary_search(k) {
                 Ok(idx) => idx + 1,
                 Err(idx) => idx,
@@ -49,34 +82,30 @@ impl<K: Ord, V> VecMap<K, V> {
             Excluded(k) => binary_search(k).unwrap_or_else(std::convert::identity),
         };
 
-        &self.0[start_idx..end_idx]
+        &self.data[start_idx..end_idx]
     }
 
     /// Add a key value pair to the map.
-    /// If `key` is less than or equal to the current maximum key
-    /// the pair will not be added and InvalidKey error will be returned.
-    pub fn append(&mut self, key: K, value: V) -> Result<usize, InvalidKey> {
-        if let Some((last_key, _last_value)) = self.0.last() {
-            if &key <= last_key {
-                return Err(InvalidKey);
-            }
-        }
+    /// If `key` is not respective of the `self` ordering the
+    /// pair will not be added and `InvalidKey` error will be returned.
+    pub fn append(&mut self, key: K, value: V) -> Result<usize, VecMapError> {
+        self.validate_key_order(&key)?;
 
         let delta_size = self.instrument_vec_op(|vec| vec.push((key, value)));
         Ok(delta_size)
     }
 
     /// Update the maximum key value pair or add a new key value pair to the map.
-    /// If `key` is less than the current maximum key no updates or additions
-    /// will occur and InvalidKey error will be returned.
+    /// If `key` is not respective of the `self` ordering no updates or additions
+    /// will occur and `InvalidKey` error will be returned.
     pub fn append_or_update_last(
         &mut self,
         key: K,
         mut value: V,
-    ) -> Result<(Option<V>, usize), InvalidKey> {
-        if let Some((last_key, last_value)) = self.0.last_mut() {
+    ) -> Result<(Option<V>, usize), VecMapError> {
+        if let Some((last_key, last_value)) = self.data.last_mut() {
             match key.cmp(last_key) {
-                Ordering::Less => return Err(InvalidKey),
+                Ordering::Less => return Err(VecMapError::InvalidKey),
                 Ordering::Equal => {
                     std::mem::swap(last_value, &mut value);
                     const DELTA_SIZE: usize = 0;
@@ -100,40 +129,67 @@ impl<K: Ord, V> VecMap<K, V> {
         V: Clone,
     {
         let split_idx = self
-            .0
+            .data
             .binary_search_by_key(&cutoff, extract_key)
             .unwrap_or_else(std::convert::identity);
 
         (
-            VecMap(self.0[..split_idx].to_vec()),
-            VecMap(self.0[split_idx..].to_vec()),
+            VecMap {
+                data: self.data[..split_idx].to_vec(),
+                ordering: self.ordering,
+            },
+            VecMap {
+                data: self.data[split_idx..].to_vec(),
+                ordering: self.ordering,
+            },
         )
     }
 
     /// Move items from `other` to the end of `self`, leaving `other` empty.
-    /// If any keys in `other` is less than or equal to any key in `self`,
-    /// `InvalidKey` error will be returned and no mutation will occur.
-    pub fn extend(&mut self, other: &mut Self) -> Result<usize, InvalidKey> {
-        let self_last_opt = self.0.last().map(extract_key);
-        let other_first_opt = other.0.last().map(extract_key);
+    /// If the `other` ordering is different from `self` ordering
+    /// `ExtendOrderingError` error will be returned.
+    /// If any keys in `other` is not respective of the ordering defined in
+    /// `self`, `InvalidKey` error will be returned and no mutation will occur.
+    pub fn extend(&mut self, other: &mut Self) -> Result<usize, VecMapError> {
+        if self.ordering != other.ordering {
+            return Err(VecMapError::ExtendOrderingError);
+        }
 
-        if let (Some(self_last), Some(other_first)) = (self_last_opt, other_first_opt) {
-            if self_last >= other_first {
-                return Err(InvalidKey);
+        let other_first_opt = other.data.last().map(extract_key);
+        if let Some(other_first) = other_first_opt {
+            self.validate_key_order(other_first)?;
+        }
+
+        let delta_size = self.instrument_vec_op(|vec| vec.append(&mut other.data));
+        Ok(delta_size)
+    }
+
+    /// Validate the current last key in `self` and key being
+    /// inserted against the order defined in `self`.
+    fn validate_key_order(&self, key: &K) -> Result<(), VecMapError> {
+        if let Some(last_key) = self.data.last().map(extract_key) {
+            match (&self.ordering, &key.cmp(last_key)) {
+                (VecMapOrdering::Greater, Ordering::Less | Ordering::Equal) => {
+                    return Err(VecMapError::InvalidKey);
+                }
+                (VecMapOrdering::Greater, Ordering::Greater) => {}
+                (VecMapOrdering::GreaterOrEqual, Ordering::Less) => {
+                    return Err(VecMapError::InvalidKey);
+                }
+                (VecMapOrdering::GreaterOrEqual, Ordering::Equal | Ordering::Greater) => {}
             }
         }
 
-        let delta_size = self.instrument_vec_op(|vec| vec.append(&mut other.0));
-        Ok(delta_size)
+        Ok(())
     }
 
     /// Instrument an operation on the underlying [`Vec`].
     /// Will panic if the operation decreases capacity.
     /// Returns the increase in memory usage caused by the op.
     fn instrument_vec_op(&mut self, op: impl FnOnce(&mut Vec<(K, V)>)) -> usize {
-        let old_cap = self.0.capacity();
-        op(&mut self.0);
-        let new_cap = self.0.capacity();
+        let old_cap = self.data.capacity();
+        op(&mut self.data);
+        let new_cap = self.data.capacity();
 
         match old_cap.cmp(&new_cap) {
             Ordering::Less => {
@@ -145,6 +201,36 @@ impl<K: Ord, V> VecMap<K, V> {
             Ordering::Greater => panic!("VecMap capacity shouldn't ever decrease"),
         }
     }
+
+    /// Similar to `from_iter` defined in `FromIter` trait except
+    /// that it accepts an [`VecMapOrdering`]
+    pub fn from_iter<I: IntoIterator<Item = (K, V)>>(iter: I, ordering: VecMapOrdering) -> Self {
+        let iter = iter.into_iter();
+        let initial_capacity = {
+            match iter.size_hint() {
+                (lower_bound, None) => lower_bound,
+                (_, Some(upper_bound)) => upper_bound,
+            }
+        };
+
+        let mut vec_map = VecMap::with_capacity(initial_capacity, ordering);
+        for (key, value) in iter {
+            vec_map
+                .append(key, value)
+                .expect("The passed collection needs to be sorted!");
+        }
+
+        vec_map
+    }
+}
+
+impl<K: Ord, V> IntoIterator for VecMap<K, V> {
+    type Item = (K, V);
+    type IntoIter = std::vec::IntoIter<(K, V)>;
+
+    fn into_iter(self) -> Self::IntoIter {
+        self.data.into_iter()
+    }
 }
 
 fn extract_key<K, V>(entry: &(K, V)) -> &K {
@@ -155,7 +241,7 @@ fn extract_key<K, V>(entry: &(K, V)) -> &K {
 mod tests {
     use std::{collections::BTreeMap, ops::Bound};
 
-    use super::VecMap;
+    use super::{VecMap, VecMapOrdering};
 
     #[test]
     fn unbounded_range() {
@@ -310,5 +396,59 @@ mod tests {
         left.extend(&mut one_map).unwrap_err();
         assert_eq!(left.as_slice(), &[(0, ()), (1, ())]);
         assert_eq!(one_map.as_slice(), &[(1, ())]);
+
+        let mut map_greater_or_equal = VecMap::new(VecMapOrdering::GreaterOrEqual);
+        map_greater_or_equal.append(2, ()).unwrap();
+        map_greater_or_equal.append(2, ()).unwrap();
+
+        left.extend(&mut map_greater_or_equal).unwrap_err();
+        assert_eq!(left.as_slice(), &[(0, ()), (1, ())]);
+        assert_eq!(map_greater_or_equal.as_slice(), &[(2, ()), (2, ())]);
+    }
+
+    #[test]
+    fn extend_with_ordering() {
+        let mut left = VecMap::new(VecMapOrdering::GreaterOrEqual);
+        left.append(0, ()).unwrap();
+        assert_eq!(left.as_slice(), &[(0, ())]);
+
+        let mut greater_right = VecMap::new(VecMapOrdering::Greater);
+        greater_right.append(0, ()).unwrap();
+        left.extend(&mut greater_right).unwrap_err();
+        assert_eq!(left.as_slice(), &[(0, ())]);
+
+        let mut greater_or_equal_right = VecMap::new(VecMapOrdering::GreaterOrEqual);
+        greater_or_equal_right.append(2, ()).unwrap();
+        greater_or_equal_right.append(2, ()).unwrap();
+        left.extend(&mut greater_or_equal_right).unwrap();
+        assert_eq!(left.as_slice(), &[(0, ()), (2, ()), (2, ())]);
+    }
+
+    #[test]
+    fn vec_map_from_sorted() {
+        let vec = vec![(1, ()), (2, ()), (3, ()), (6, ())];
+        let vec_map = VecMap::from_iter(vec, VecMapOrdering::Greater);
+        assert_eq!(vec_map.as_slice(), &[(1, ()), (2, ()), (3, ()), (6, ())]);
+
+        let vec = vec![(1, ()), (2, ()), (3, ()), (3, ()), (6, ()), (6, ())];
+        let vec_map = VecMap::from_iter(vec, VecMapOrdering::GreaterOrEqual);
+        assert_eq!(
+            vec_map.as_slice(),
+            &[(1, ()), (2, ()), (3, ()), (3, ()), (6, ()), (6, ())]
+        );
+    }
+
+    #[test]
+    #[should_panic]
+    fn vec_map_from_unsorted_greater() {
+        let vec = vec![(1, ()), (2, ()), (2, ()), (3, ()), (6, ())];
+        let _ = VecMap::from_iter(vec, VecMapOrdering::Greater);
+    }
+
+    #[test]
+    #[should_panic]
+    fn vec_map_from_unsorted_greater_or_equal() {
+        let vec = vec![(1, ()), (2, ()), (3, ()), (6, ()), (5, ())];
+        let _ = VecMap::from_iter(vec, VecMapOrdering::GreaterOrEqual);
     }
 }

libs/utils/src/zstd.rs

@@ -0,0 +1,78 @@
+use std::io::SeekFrom;
+
+use anyhow::{Context, Result};
+use async_compression::{
+    tokio::{bufread::ZstdDecoder, write::ZstdEncoder},
+    zstd::CParameter,
+    Level,
+};
+use camino::Utf8Path;
+use nix::NixPath;
+use tokio::{
+    fs::{File, OpenOptions},
+    io::AsyncBufRead,
+    io::AsyncSeekExt,
+    io::AsyncWriteExt,
+};
+use tokio_tar::{Archive, Builder, HeaderMode};
+use walkdir::WalkDir;
+
+/// Creates a Zstandard tarball.
+pub async fn create_zst_tarball(path: &Utf8Path, tarball: &Utf8Path) -> Result<(File, u64)> {
+    let file = OpenOptions::new()
+        .create(true)
+        .truncate(true)
+        .read(true)
+        .write(true)
+        .open(&tarball)
+        .await
+        .with_context(|| format!("tempfile creation {tarball}"))?;
+
+    let mut paths = Vec::new();
+    for entry in WalkDir::new(path) {
+        let entry = entry?;
+        let metadata = entry.metadata().expect("error getting dir entry metadata");
+        // Also allow directories so that we also get empty directories
+        if !(metadata.is_file() || metadata.is_dir()) {
+            continue;
+        }
+        let path = entry.into_path();
+        paths.push(path);
+    }
+    // Do a sort to get a more consistent listing
+    paths.sort_unstable();
+    let zstd = ZstdEncoder::with_quality_and_params(
+        file,
+        Level::Default,
+        &[CParameter::enable_long_distance_matching(true)],
+    );
+    let mut builder = Builder::new(zstd);
+    // Use reproducible header mode
+    builder.mode(HeaderMode::Deterministic);
+    for p in paths {
+        let rel_path = p.strip_prefix(path)?;
+        if rel_path.is_empty() {
+            // The top directory should not be compressed,
+            // the tar crate doesn't like that
+            continue;
+        }
+        builder.append_path_with_name(&p, rel_path).await?;
+    }
+    let mut zstd = builder.into_inner().await?;
+    zstd.shutdown().await?;
+    let mut compressed = zstd.into_inner();
+    let compressed_len = compressed.metadata().await?.len();
+    compressed.seek(SeekFrom::Start(0)).await?;
+    Ok((compressed, compressed_len))
+}
+
+/// Creates a Zstandard tarball.
+pub async fn extract_zst_tarball(
+    path: &Utf8Path,
+    tarball: impl AsyncBufRead + Unpin,
+) -> Result<()> {
+    let decoder = Box::pin(ZstdDecoder::new(tarball));
+    let mut archive = Archive::new(decoder);
+    archive.unpack(path).await?;
+    Ok(())
+}

libs/vm_monitor/src/runner.rs

@@ -69,7 +69,7 @@ pub struct Config {
     /// should be removed once we have a better solution there.
     sys_buffer_bytes: u64,
 
-    /// Minimum fraction of total system memory reserved *before* the the cgroup threshold; in
+    /// Minimum fraction of total system memory reserved *before* the cgroup threshold; in
     /// other words, providing a ceiling for the highest value of the threshold by enforcing that
     /// there's at least `cgroup_min_overhead_fraction` of the total memory remaining beyond the
     /// threshold.

pageserver/Cargo.toml

@@ -59,6 +59,7 @@ signal-hook.workspace = true
 smallvec = { workspace = true, features = ["write"] }
 svg_fmt.workspace = true
 sync_wrapper.workspace = true
+sysinfo.workspace = true
 tokio-tar.workspace = true
 thiserror.workspace = true
 tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time"] }

pageserver/benches/bench_walredo.rs

@@ -1,160 +1,156 @@
-//! Simple benchmarking around walredo.
+//! Quantify a single walredo manager's throughput under N concurrent callers.
 //!
-//! Right now they hope to just set a baseline. Later we can try to expand into latency and
-//! throughput after figuring out the coordinated omission problems below.
+//! The benchmark implementation ([`bench_impl`]) is parametrized by
+//! - `redo_work` => [`Request::short_request`] or [`Request::medium_request`]
+//! - `n_redos` => number of times the benchmark shell execute the `redo_work`
+//! - `nclients` => number of clients (more on this shortly).
 //!
-//! There are two sets of inputs; `short` and `medium`. They were collected on postgres v14 by
-//! logging what happens when a sequential scan is requested on a small table, then picking out two
-//! suitable from logs.
+//! The benchmark impl sets up a multi-threaded tokio runtime with default parameters.
+//! It spawns `nclients` times [`client`] tokio tasks.
+//! Each task executes the `redo_work` `n_redos/nclients` times.
 //!
+//! We exercise the following combinations:
+//! - `redo_work = short / medium``
+//! - `nclients = [1, 2, 4, 8, 16, 32, 64, 128]`
 //!
-//! Reference data (git blame to see commit) on an i3en.3xlarge
-// ```text
-//! short/short/1           time:   [39.175 µs 39.348 µs 39.536 µs]
-//! short/short/2           time:   [51.227 µs 51.487 µs 51.755 µs]
-//! short/short/4           time:   [76.048 µs 76.362 µs 76.674 µs]
-//! short/short/8           time:   [128.94 µs 129.82 µs 130.74 µs]
-//! short/short/16          time:   [227.84 µs 229.00 µs 230.28 µs]
-//! short/short/32          time:   [455.97 µs 457.81 µs 459.90 µs]
-//! short/short/64          time:   [902.46 µs 904.84 µs 907.32 µs]
-//! short/short/128         time:   [1.7416 ms 1.7487 ms 1.7561 ms]
-//! ``
-
-use std::sync::Arc;
+//! We let `criterion` determine the `n_redos` using `iter_custom`.
+//! The idea is that for each `(redo_work, nclients)` combination,
+//! criterion will run the `bench_impl` multiple times with different `n_redos`.
+//! The `bench_impl` reports the aggregate wall clock time from the clients' perspective.
+//! Criterion will divide that by `n_redos` to compute the "time per iteration".
+//! In our case, "time per iteration" means "time per redo_work execution".
+//!
+//! NB: the way by which `iter_custom` determines the "number of iterations"
+//! is called sampling. Apparently the idea here is to detect outliers.
+//! We're not sure whether the current choice of sampling method makes sense.
+//! See https://bheisler.github.io/criterion.rs/book/user_guide/command_line_output.html#collecting-samples
+//!
+//! # Reference Numbers
+//!
+//! 2024-03-20 on i3en.3xlarge
+//!
+//! ```text
+//! short/1                 time:   [26.483 µs 26.614 µs 26.767 µs]
+//! short/2                 time:   [32.223 µs 32.465 µs 32.767 µs]
+//! short/4                 time:   [47.203 µs 47.583 µs 47.984 µs]
+//! short/8                 time:   [89.135 µs 89.612 µs 90.139 µs]
+//! short/16                time:   [190.12 µs 191.52 µs 192.88 µs]
+//! short/32                time:   [380.96 µs 382.63 µs 384.20 µs]
+//! short/64                time:   [736.86 µs 741.07 µs 745.03 µs]
+//! short/128               time:   [1.4106 ms 1.4206 ms 1.4294 ms]
+//! medium/1                time:   [111.81 µs 112.25 µs 112.79 µs]
+//! medium/2                time:   [158.26 µs 159.13 µs 160.21 µs]
+//! medium/4                time:   [334.65 µs 337.14 µs 340.07 µs]
+//! medium/8                time:   [675.32 µs 679.91 µs 685.25 µs]
+//! medium/16               time:   [1.2929 ms 1.2996 ms 1.3067 ms]
+//! medium/32               time:   [2.4295 ms 2.4461 ms 2.4623 ms]
+//! medium/64               time:   [4.3973 ms 4.4458 ms 4.4875 ms]
+//! medium/128              time:   [7.5955 ms 7.7847 ms 7.9481 ms]
+//! ```
 
 use bytes::{Buf, Bytes};
-use pageserver::{
-    config::PageServerConf, repository::Key, walrecord::NeonWalRecord, walredo::PostgresRedoManager,
+use criterion::{BenchmarkId, Criterion};
+use pageserver::{config::PageServerConf, walrecord::NeonWalRecord, walredo::PostgresRedoManager};
+use pageserver_api::{key::Key, shard::TenantShardId};
+use std::{
+    sync::Arc,
+    time::{Duration, Instant},
 };
-use pageserver_api::shard::TenantShardId;
-use tokio::task::JoinSet;
+use tokio::{sync::Barrier, task::JoinSet};
 use utils::{id::TenantId, lsn::Lsn};
 
-use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion};
+fn bench(c: &mut Criterion) {
+    {
+        let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
+        for nclients in nclients {
+            let mut group = c.benchmark_group("short");
+            group.bench_with_input(
+                BenchmarkId::from_parameter(nclients),
+                &nclients,
+                |b, nclients| {
+                    let redo_work = Arc::new(Request::short_input());
+                    b.iter_custom(|iters| bench_impl(Arc::clone(&redo_work), iters, *nclients));
+                },
+            );
+        }
+    }
 
-fn redo_scenarios(c: &mut Criterion) {
-    // logging should be enabled when adding more inputs, since walredo will only report malformed
-    // input to the stderr.
-    // utils::logging::init(utils::logging::LogFormat::Plain).unwrap();
+    {
+        let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
+        for nclients in nclients {
+            let mut group = c.benchmark_group("medium");
+            group.bench_with_input(
+                BenchmarkId::from_parameter(nclients),
+                &nclients,
+                |b, nclients| {
+                    let redo_work = Arc::new(Request::medium_input());
+                    b.iter_custom(|iters| bench_impl(Arc::clone(&redo_work), iters, *nclients));
+                },
+            );
+        }
+    }
+}
+criterion::criterion_group!(benches, bench);
+criterion::criterion_main!(benches);
 
+// Returns the sum of each client's wall-clock time spent executing their share of the n_redos.
+fn bench_impl(redo_work: Arc<Request>, n_redos: u64, nclients: u64) -> Duration {
     let repo_dir = camino_tempfile::tempdir_in(env!("CARGO_TARGET_TMPDIR")).unwrap();
 
     let conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
     let conf = Box::leak(Box::new(conf));
     let tenant_shard_id = TenantShardId::unsharded(TenantId::generate());
 
-    let manager = PostgresRedoManager::new(conf, tenant_shard_id);
-
-    let manager = Arc::new(manager);
-
-    {
-        let rt = tokio::runtime::Builder::new_current_thread()
-            .enable_all()
-            .build()
-            .unwrap();
-        tracing::info!("executing first");
-        rt.block_on(short().execute(&manager)).unwrap();
-        tracing::info!("first executed");
-    }
-
-    let thread_counts = [1, 2, 4, 8, 16, 32, 64, 128];
-
-    let mut group = c.benchmark_group("short");
-    group.sampling_mode(criterion::SamplingMode::Flat);
-
-    for thread_count in thread_counts {
-        group.bench_with_input(
-            BenchmarkId::new("short", thread_count),
-            &thread_count,
-            |b, thread_count| {
-                add_multithreaded_walredo_requesters(b, *thread_count, &manager, short);
-            },
-        );
-    }
-    drop(group);
-
-    let mut group = c.benchmark_group("medium");
-    group.sampling_mode(criterion::SamplingMode::Flat);
-
-    for thread_count in thread_counts {
-        group.bench_with_input(
-            BenchmarkId::new("medium", thread_count),
-            &thread_count,
-            |b, thread_count| {
-                add_multithreaded_walredo_requesters(b, *thread_count, &manager, medium);
-            },
-        );
-    }
-    drop(group);
-}
-
-/// Sets up a multi-threaded tokio runtime with default worker thread count,
-/// then, spawn `requesters` tasks that repeatedly:
-/// - get input from `input_factor()`
-/// - call `manager.request_redo()` with their input
-///
-/// This stress-tests the scalability of a single walredo manager at high tokio-level concurrency.
-///
-/// Using tokio's default worker thread count means the results will differ on machines
-/// with different core countrs. We don't care about that, the performance will always
-/// be different on different hardware. To compare performance of different software versions,
-/// use the same hardware.
-fn add_multithreaded_walredo_requesters(
-    b: &mut criterion::Bencher,
-    nrequesters: usize,
-    manager: &Arc<PostgresRedoManager>,
-    input_factory: fn() -> Request,
-) {
-    assert_ne!(nrequesters, 0);
-
     let rt = tokio::runtime::Builder::new_multi_thread()
         .enable_all()
         .build()
         .unwrap();
 
-    let barrier = Arc::new(tokio::sync::Barrier::new(nrequesters + 1));
+    let start = Arc::new(Barrier::new(nclients as usize));
 
-    let mut requesters = JoinSet::new();
-    for _ in 0..nrequesters {
-        let _entered = rt.enter();
-        let manager = manager.clone();
-        let barrier = barrier.clone();
-        requesters.spawn(async move {
-            loop {
-                let input = input_factory();
-                barrier.wait().await;
-                let page = input.execute(&manager).await.unwrap();
-                assert_eq!(page.remaining(), 8192);
-                barrier.wait().await;
-            }
+    let mut tasks = JoinSet::new();
+
+    let manager = PostgresRedoManager::new(conf, tenant_shard_id);
+    let manager = Arc::new(manager);
+
+    for _ in 0..nclients {
+        rt.block_on(async {
+            tasks.spawn(client(
+                Arc::clone(&manager),
+                Arc::clone(&start),
+                Arc::clone(&redo_work),
+                // divide the amount of work equally among the clients
+                n_redos / nclients,
+            ))
         });
     }
 
-    let do_one_iteration = || {
-        rt.block_on(async {
-            barrier.wait().await;
-            // wait for work to complete
-            barrier.wait().await;
-        })
-    };
-
-    b.iter_batched(
-        || {
-            // warmup
-            do_one_iteration();
-        },
-        |()| {
-            // work loop
-            do_one_iteration();
-        },
-        criterion::BatchSize::PerIteration,
-    );
-
-    rt.block_on(requesters.shutdown());
+    rt.block_on(async move {
+        let mut total_wallclock_time = std::time::Duration::from_millis(0);
+        while let Some(res) = tasks.join_next().await {
+            total_wallclock_time += res.unwrap();
+        }
+        total_wallclock_time
+    })
 }
 
-criterion_group!(benches, redo_scenarios);
-criterion_main!(benches);
+async fn client(
+    mgr: Arc<PostgresRedoManager>,
+    start: Arc<Barrier>,
+    redo_work: Arc<Request>,
+    n_redos: u64,
+) -> Duration {
+    start.wait().await;
+    let start = Instant::now();
+    for _ in 0..n_redos {
+        let page = redo_work.execute(&mgr).await.unwrap();
+        assert_eq!(page.remaining(), 8192);
+        // The real pageserver will rarely if ever do 2 walredos in a row without
+        // yielding to the executor.
+        tokio::task::yield_now().await;
+    }
+    start.elapsed()
+}
 
 macro_rules! lsn {
     ($input:expr) => {{
@@ -166,12 +162,46 @@ macro_rules! lsn {
     }};
 }
 
-/// Short payload, 1132 bytes.
-// pg_records are copypasted from log, where they are put with Debug impl of Bytes, which uses \0
-// for null bytes.
-#[allow(clippy::octal_escapes)]
-fn short() -> Request {
-    Request {
+/// Simple wrapper around `WalRedoManager::request_redo`.
+///
+/// In benchmarks this is cloned around.
+#[derive(Clone)]
+struct Request {
+    key: Key,
+    lsn: Lsn,
+    base_img: Option<(Lsn, Bytes)>,
+    records: Vec<(Lsn, NeonWalRecord)>,
+    pg_version: u32,
+}
+
+impl Request {
+    async fn execute(&self, manager: &PostgresRedoManager) -> anyhow::Result<Bytes> {
+        let Request {
+            key,
+            lsn,
+            base_img,
+            records,
+            pg_version,
+        } = self;
+
+        // TODO: avoid these clones
+        manager
+            .request_redo(*key, *lsn, base_img.clone(), records.clone(), *pg_version)
+            .await
+    }
+
+    fn pg_record(will_init: bool, bytes: &'static [u8]) -> NeonWalRecord {
+        let rec = Bytes::from_static(bytes);
+        NeonWalRecord::Postgres { will_init, rec }
+    }
+
+    /// Short payload, 1132 bytes.
+    // pg_records are copypasted from log, where they are put with Debug impl of Bytes, which uses \0
+    // for null bytes.
+    #[allow(clippy::octal_escapes)]
+    pub fn short_input() -> Request {
+        let pg_record = Self::pg_record;
+        Request {
         key: Key {
             field1: 0,
             field2: 1663,
@@ -194,13 +224,14 @@ fn short() -> Request {
         ],
         pg_version: 14,
     }
-}
+    }
 
-/// Medium sized payload, serializes as 26393 bytes.
-// see [`short`]
-#[allow(clippy::octal_escapes)]
-fn medium() -> Request {
-    Request {
+    /// Medium sized payload, serializes as 26393 bytes.
+    // see [`short`]
+    #[allow(clippy::octal_escapes)]
+    pub fn medium_input() -> Request {
+        let pg_record = Self::pg_record;
+        Request {
         key: Key {
             field1: 0,
             field2: 1663,
@@ -442,37 +473,5 @@ fn medium() -> Request {
         ],
         pg_version: 14,
     }
-}
-
-fn pg_record(will_init: bool, bytes: &'static [u8]) -> NeonWalRecord {
-    let rec = Bytes::from_static(bytes);
-    NeonWalRecord::Postgres { will_init, rec }
-}
-
-/// Simple wrapper around `WalRedoManager::request_redo`.
-///
-/// In benchmarks this is cloned around.
-#[derive(Clone)]
-struct Request {
-    key: Key,
-    lsn: Lsn,
-    base_img: Option<(Lsn, Bytes)>,
-    records: Vec<(Lsn, NeonWalRecord)>,
-    pg_version: u32,
-}
-
-impl Request {
-    async fn execute(self, manager: &PostgresRedoManager) -> anyhow::Result<Bytes> {
-        let Request {
-            key,
-            lsn,
-            base_img,
-            records,
-            pg_version,
-        } = self;
-
-        manager
-            .request_redo(key, lsn, base_img, records, pg_version)
-            .await
     }
 }

pageserver/src/bin/pageserver.rs

@@ -600,32 +600,37 @@ fn start_pageserver(
             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();
+            {
+                let tenant_manager = tenant_manager.clone();
+                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,
-                    local_disk_storage,
-                    cancel,
-                    metrics_ctx,
-                )
-                .instrument(info_span!("metrics_collection"))
-                .await?;
-                Ok(())
+                    pageserver::consumption_metrics::collect_metrics(
+                        tenant_manager,
+                        metric_collection_endpoint,
+                        &conf.metric_collection_bucket,
+                        conf.metric_collection_interval,
+                        conf.cached_metric_collection_interval,
+                        conf.synthetic_size_calculation_interval,
+                        conf.id,
+                        local_disk_storage,
+                        cancel,
+                        metrics_ctx,
+                    )
+                    .instrument(info_span!("metrics_collection"))
+                    .await?;
+                    Ok(())
+                }
             },
         );
     }

pageserver/src/config.rs

@@ -95,6 +95,8 @@ pub mod defaults {
 
     pub const DEFAULT_VALIDATE_VECTORED_GET: bool = true;
 
+    pub const DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB: usize = 0;
+
     ///
     /// Default built-in configuration file.
     ///
@@ -156,6 +158,8 @@ pub mod defaults {
 #heatmap_upload_concurrency = {DEFAULT_HEATMAP_UPLOAD_CONCURRENCY}
 #secondary_download_concurrency = {DEFAULT_SECONDARY_DOWNLOAD_CONCURRENCY}
 
+#ephemeral_bytes_per_memory_kb = {DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB}
+
 [remote_storage]
 
 "#
@@ -234,6 +238,7 @@ pub struct PageServerConf {
     // How often to send unchanged cached metrics to the metrics endpoint.
     pub cached_metric_collection_interval: Duration,
     pub metric_collection_endpoint: Option<Url>,
+    pub metric_collection_bucket: Option<RemoteStorageConfig>,
     pub synthetic_size_calculation_interval: Duration,
 
     pub disk_usage_based_eviction: Option<DiskUsageEvictionTaskConfig>,
@@ -278,6 +283,13 @@ pub struct PageServerConf {
     pub max_vectored_read_bytes: MaxVectoredReadBytes,
 
     pub validate_vectored_get: bool,
+
+    /// How many bytes of ephemeral layer content will we allow per kilobyte of RAM.  When this
+    /// is exceeded, we start proactively closing ephemeral layers to limit the total amount
+    /// of ephemeral data.
+    ///
+    /// Setting this to zero disables limits on total ephemeral layer size.
+    pub ephemeral_bytes_per_memory_kb: usize,
 }
 
 /// We do not want to store this in a PageServerConf because the latter may be logged
@@ -373,6 +385,7 @@ struct PageServerConfigBuilder {
     cached_metric_collection_interval: BuilderValue<Duration>,
     metric_collection_endpoint: BuilderValue<Option<Url>>,
     synthetic_size_calculation_interval: BuilderValue<Duration>,
+    metric_collection_bucket: BuilderValue<Option<RemoteStorageConfig>>,
 
     disk_usage_based_eviction: BuilderValue<Option<DiskUsageEvictionTaskConfig>>,
 
@@ -398,6 +411,8 @@ struct PageServerConfigBuilder {
     max_vectored_read_bytes: BuilderValue<MaxVectoredReadBytes>,
 
     validate_vectored_get: BuilderValue<bool>,
+
+    ephemeral_bytes_per_memory_kb: BuilderValue<usize>,
 }
 
 impl PageServerConfigBuilder {
@@ -455,6 +470,8 @@ impl PageServerConfigBuilder {
             .expect("cannot parse default synthetic size calculation interval")),
             metric_collection_endpoint: Set(DEFAULT_METRIC_COLLECTION_ENDPOINT),
 
+            metric_collection_bucket: Set(None),
+
             disk_usage_based_eviction: Set(None),
 
             test_remote_failures: Set(0),
@@ -482,6 +499,7 @@ impl PageServerConfigBuilder {
                 NonZeroUsize::new(DEFAULT_MAX_VECTORED_READ_BYTES).unwrap(),
             )),
             validate_vectored_get: Set(DEFAULT_VALIDATE_VECTORED_GET),
+            ephemeral_bytes_per_memory_kb: Set(DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB),
         }
     }
 }
@@ -586,6 +604,13 @@ impl PageServerConfigBuilder {
         self.metric_collection_endpoint = BuilderValue::Set(metric_collection_endpoint)
     }
 
+    pub fn metric_collection_bucket(
+        &mut self,
+        metric_collection_bucket: Option<RemoteStorageConfig>,
+    ) {
+        self.metric_collection_bucket = BuilderValue::Set(metric_collection_bucket)
+    }
+
     pub fn synthetic_size_calculation_interval(
         &mut self,
         synthetic_size_calculation_interval: Duration,
@@ -654,6 +679,10 @@ impl PageServerConfigBuilder {
         self.validate_vectored_get = BuilderValue::Set(value);
     }
 
+    pub fn get_ephemeral_bytes_per_memory_kb(&mut self, value: usize) {
+        self.ephemeral_bytes_per_memory_kb = BuilderValue::Set(value);
+    }
+
     pub fn build(self) -> anyhow::Result<PageServerConf> {
         let default = Self::default_values();
 
@@ -694,6 +723,7 @@ impl PageServerConfigBuilder {
                 metric_collection_interval,
                 cached_metric_collection_interval,
                 metric_collection_endpoint,
+                metric_collection_bucket,
                 synthetic_size_calculation_interval,
                 disk_usage_based_eviction,
                 test_remote_failures,
@@ -708,6 +738,7 @@ impl PageServerConfigBuilder {
                 get_vectored_impl,
                 max_vectored_read_bytes,
                 validate_vectored_get,
+                ephemeral_bytes_per_memory_kb,
             }
             CUSTOM LOGIC
             {
@@ -942,6 +973,9 @@ impl PageServerConf {
                     let endpoint = parse_toml_string(key, item)?.parse().context("failed to parse metric_collection_endpoint")?;
                     builder.metric_collection_endpoint(Some(endpoint));
                 },
+                "metric_collection_bucket" => {
+                    builder.metric_collection_bucket(RemoteStorageConfig::from_toml(item)?)
+                }
                 "synthetic_size_calculation_interval" =>
                     builder.synthetic_size_calculation_interval(parse_toml_duration(key, item)?),
                 "test_remote_failures" => builder.test_remote_failures(parse_toml_u64(key, item)?),
@@ -995,6 +1029,9 @@ impl PageServerConf {
                 "validate_vectored_get" => {
                     builder.get_validate_vectored_get(parse_toml_bool("validate_vectored_get", item)?)
                 }
+                "ephemeral_bytes_per_memory_kb" => {
+                    builder.get_ephemeral_bytes_per_memory_kb(parse_toml_u64("ephemeral_bytes_per_memory_kb", item)? as usize)
+                }
                 _ => bail!("unrecognized pageserver option '{key}'"),
             }
         }
@@ -1057,6 +1094,7 @@ impl PageServerConf {
             metric_collection_interval: Duration::from_secs(60),
             cached_metric_collection_interval: Duration::from_secs(60 * 60),
             metric_collection_endpoint: defaults::DEFAULT_METRIC_COLLECTION_ENDPOINT,
+            metric_collection_bucket: None,
             synthetic_size_calculation_interval: Duration::from_secs(60),
             disk_usage_based_eviction: None,
             test_remote_failures: 0,
@@ -1075,6 +1113,7 @@ impl PageServerConf {
                     .expect("Invalid default constant"),
             ),
             validate_vectored_get: defaults::DEFAULT_VALIDATE_VECTORED_GET,
+            ephemeral_bytes_per_memory_kb: defaults::DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB,
         }
     }
 }
@@ -1289,6 +1328,7 @@ background_task_maximum_delay = '334 s'
                     defaults::DEFAULT_CACHED_METRIC_COLLECTION_INTERVAL
                 )?,
                 metric_collection_endpoint: defaults::DEFAULT_METRIC_COLLECTION_ENDPOINT,
+                metric_collection_bucket: None,
                 synthetic_size_calculation_interval: humantime::parse_duration(
                     defaults::DEFAULT_SYNTHETIC_SIZE_CALCULATION_INTERVAL
                 )?,
@@ -1311,6 +1351,7 @@ background_task_maximum_delay = '334 s'
                         .expect("Invalid default constant")
                 ),
                 validate_vectored_get: defaults::DEFAULT_VALIDATE_VECTORED_GET,
+                ephemeral_bytes_per_memory_kb: defaults::DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB
             },
             "Correct defaults should be used when no config values are provided"
         );
@@ -1363,6 +1404,7 @@ background_task_maximum_delay = '334 s'
                 metric_collection_interval: Duration::from_secs(222),
                 cached_metric_collection_interval: Duration::from_secs(22200),
                 metric_collection_endpoint: Some(Url::parse("http://localhost:80/metrics")?),
+                metric_collection_bucket: None,
                 synthetic_size_calculation_interval: Duration::from_secs(333),
                 disk_usage_based_eviction: None,
                 test_remote_failures: 0,
@@ -1381,6 +1423,7 @@ background_task_maximum_delay = '334 s'
                         .expect("Invalid default constant")
                 ),
                 validate_vectored_get: defaults::DEFAULT_VALIDATE_VECTORED_GET,
+                ephemeral_bytes_per_memory_kb: defaults::DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB
             },
             "Should be able to parse all basic config values correctly"
         );

pageserver/src/consumption_metrics.rs

@@ -3,10 +3,13 @@
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::task_mgr::{self, TaskKind, BACKGROUND_RUNTIME};
 use crate::tenant::tasks::BackgroundLoopKind;
-use crate::tenant::{mgr, LogicalSizeCalculationCause, PageReconstructError, Tenant};
+use crate::tenant::{
+    mgr::TenantManager, LogicalSizeCalculationCause, PageReconstructError, Tenant,
+};
 use camino::Utf8PathBuf;
 use consumption_metrics::EventType;
 use pageserver_api::models::TenantState;
+use remote_storage::{GenericRemoteStorage, RemoteStorageConfig};
 use reqwest::Url;
 use std::collections::HashMap;
 use std::sync::Arc;
@@ -40,7 +43,9 @@ type Cache = HashMap<MetricsKey, (EventType, u64)>;
 /// Main thread that serves metrics collection
 #[allow(clippy::too_many_arguments)]
 pub async fn collect_metrics(
+    tenant_manager: Arc<TenantManager>,
     metric_collection_endpoint: &Url,
+    metric_collection_bucket: &Option<RemoteStorageConfig>,
     metric_collection_interval: Duration,
     _cached_metric_collection_interval: Duration,
     synthetic_size_calculation_interval: Duration,
@@ -65,15 +70,19 @@ pub async fn collect_metrics(
         None,
         "synthetic size calculation",
         false,
-        async move {
-            calculate_synthetic_size_worker(
-                synthetic_size_calculation_interval,
-                &cancel,
-                &worker_ctx,
-            )
-            .instrument(info_span!("synthetic_size_worker"))
-            .await?;
-            Ok(())
+        {
+            let tenant_manager = tenant_manager.clone();
+            async move {
+                calculate_synthetic_size_worker(
+                    tenant_manager,
+                    synthetic_size_calculation_interval,
+                    &cancel,
+                    &worker_ctx,
+                )
+                .instrument(info_span!("synthetic_size_worker"))
+                .await?;
+                Ok(())
+            }
         },
     );
 
@@ -94,13 +103,27 @@ pub async fn collect_metrics(
         .build()
         .expect("Failed to create http client with timeout");
 
+    let bucket_client = if let Some(bucket_config) = metric_collection_bucket {
+        match GenericRemoteStorage::from_config(bucket_config) {
+            Ok(client) => Some(client),
+            Err(e) => {
+                // Non-fatal error: if we were given an invalid config, we will proceed
+                // with sending metrics over the network, but not to S3.
+                tracing::warn!("Invalid configuration for metric_collection_bucket: {e}");
+                None
+            }
+        }
+    } else {
+        None
+    };
+
     let node_id = node_id.to_string();
 
     loop {
         let started_at = Instant::now();
 
         // these are point in time, with variable "now"
-        let metrics = metrics::collect_all_metrics(&cached_metrics, &ctx).await;
+        let metrics = metrics::collect_all_metrics(&tenant_manager, &cached_metrics, &ctx).await;
 
         let metrics = Arc::new(metrics);
 
@@ -118,10 +141,18 @@ pub async fn collect_metrics(
                     tracing::error!("failed to persist metrics to {path:?}: {e:#}");
                 }
             }
+
+            if let Some(bucket_client) = &bucket_client {
+                let res =
+                    upload::upload_metrics_bucket(bucket_client, &cancel, &node_id, &metrics).await;
+                if let Err(e) = res {
+                    tracing::error!("failed to upload to S3: {e:#}");
+                }
+            }
         };
 
         let upload = async {
-            let res = upload::upload_metrics(
+            let res = upload::upload_metrics_http(
                 &client,
                 metric_collection_endpoint,
                 &cancel,
@@ -132,7 +163,7 @@ pub async fn collect_metrics(
             .await;
             if let Err(e) = res {
                 // serialization error which should never happen
-                tracing::error!("failed to upload due to {e:#}");
+                tracing::error!("failed to upload via HTTP due to {e:#}");
             }
         };
 
@@ -247,6 +278,7 @@ async fn reschedule(
 
 /// Caclculate synthetic size for each active tenant
 async fn calculate_synthetic_size_worker(
+    tenant_manager: Arc<TenantManager>,
     synthetic_size_calculation_interval: Duration,
     cancel: &CancellationToken,
     ctx: &RequestContext,
@@ -259,7 +291,7 @@ async fn calculate_synthetic_size_worker(
     loop {
         let started_at = Instant::now();
 
-        let tenants = match mgr::list_tenants().await {
+        let tenants = match tenant_manager.list_tenants() {
             Ok(tenants) => tenants,
             Err(e) => {
                 warn!("cannot get tenant list: {e:#}");
@@ -278,10 +310,14 @@ async fn calculate_synthetic_size_worker(
                 continue;
             }
 
-            let Ok(tenant) = mgr::get_tenant(tenant_shard_id, true) else {
+            let Ok(tenant) = tenant_manager.get_attached_tenant_shard(tenant_shard_id) else {
                 continue;
             };
 
+            if !tenant.is_active() {
+                continue;
+            }
+
             // there is never any reason to exit calculate_synthetic_size_worker following any
             // return value -- we don't need to care about shutdown because no tenant is found when
             // pageserver is shut down.
@@ -319,9 +355,7 @@ async fn calculate_and_log(tenant: &Tenant, cancel: &CancellationToken, ctx: &Re
     };
 
     // this error can be returned if timeline is shutting down, but it does not
-    // mean the synthetic size worker should terminate. we do not need any checks
-    // in this function because `mgr::get_tenant` will error out after shutdown has
-    // progressed to shutting down tenants.
+    // mean the synthetic size worker should terminate.
     let shutting_down = matches!(
         e.downcast_ref::<PageReconstructError>(),
         Some(PageReconstructError::Cancelled | PageReconstructError::AncestorStopping(_))

pageserver/src/consumption_metrics/metrics.rs

@@ -1,3 +1,4 @@
+use crate::tenant::mgr::TenantManager;
 use crate::{context::RequestContext, tenant::timeline::logical_size::CurrentLogicalSize};
 use chrono::{DateTime, Utc};
 use consumption_metrics::EventType;
@@ -181,6 +182,7 @@ impl MetricsKey {
 }
 
 pub(super) async fn collect_all_metrics(
+    tenant_manager: &Arc<TenantManager>,
     cached_metrics: &Cache,
     ctx: &RequestContext,
 ) -> Vec<RawMetric> {
@@ -188,7 +190,7 @@ pub(super) async fn collect_all_metrics(
 
     let started_at = std::time::Instant::now();
 
-    let tenants = match crate::tenant::mgr::list_tenants().await {
+    let tenants = match tenant_manager.list_tenants() {
         Ok(tenants) => tenants,
         Err(err) => {
             tracing::error!("failed to list tenants: {:?}", err);
@@ -200,7 +202,8 @@ pub(super) async fn collect_all_metrics(
         if state != TenantState::Active || !id.is_zero() {
             None
         } else {
-            crate::tenant::mgr::get_tenant(id, true)
+            tenant_manager
+                .get_attached_tenant_shard(id)
                 .ok()
                 .map(|tenant| (id.tenant_id, tenant))
         }

pageserver/src/consumption_metrics/upload.rs

@@ -1,4 +1,9 @@
+use std::time::SystemTime;
+
+use chrono::{DateTime, Utc};
 use consumption_metrics::{Event, EventChunk, IdempotencyKey, CHUNK_SIZE};
+use remote_storage::{GenericRemoteStorage, RemotePath};
+use tokio::io::AsyncWriteExt;
 use tokio_util::sync::CancellationToken;
 use tracing::Instrument;
 
@@ -13,8 +18,9 @@ struct Ids {
     pub(super) timeline_id: Option<TimelineId>,
 }
 
+/// Serialize and write metrics to an HTTP endpoint
 #[tracing::instrument(skip_all, fields(metrics_total = %metrics.len()))]
-pub(super) async fn upload_metrics(
+pub(super) async fn upload_metrics_http(
     client: &reqwest::Client,
     metric_collection_endpoint: &reqwest::Url,
     cancel: &CancellationToken,
@@ -74,6 +80,60 @@ pub(super) async fn upload_metrics(
     Ok(())
 }
 
+/// Serialize and write metrics to a remote storage object
+#[tracing::instrument(skip_all, fields(metrics_total = %metrics.len()))]
+pub(super) async fn upload_metrics_bucket(
+    client: &GenericRemoteStorage,
+    cancel: &CancellationToken,
+    node_id: &str,
+    metrics: &[RawMetric],
+) -> anyhow::Result<()> {
+    if metrics.is_empty() {
+        // Skip uploads if we have no metrics, so that readers don't have to handle the edge case
+        // of an empty object.
+        return Ok(());
+    }
+
+    // Compose object path
+    let datetime: DateTime<Utc> = SystemTime::now().into();
+    let ts_prefix = datetime.format("year=%Y/month=%m/day=%d/%H:%M:%SZ");
+    let path = RemotePath::from_string(&format!("{ts_prefix}_{node_id}.ndjson.gz"))?;
+
+    // Set up a gzip writer into a buffer
+    let mut compressed_bytes: Vec<u8> = Vec::new();
+    let compressed_writer = std::io::Cursor::new(&mut compressed_bytes);
+    let mut gzip_writer = async_compression::tokio::write::GzipEncoder::new(compressed_writer);
+
+    // Serialize and write into compressed buffer
+    let started_at = std::time::Instant::now();
+    for res in serialize_in_chunks(CHUNK_SIZE, metrics, node_id) {
+        let (_chunk, body) = res?;
+        gzip_writer.write_all(&body).await?;
+    }
+    gzip_writer.flush().await?;
+    gzip_writer.shutdown().await?;
+    let compressed_length = compressed_bytes.len();
+
+    // Write to remote storage
+    client
+        .upload_storage_object(
+            futures::stream::once(futures::future::ready(Ok(compressed_bytes.into()))),
+            compressed_length,
+            &path,
+            cancel,
+        )
+        .await?;
+    let elapsed = started_at.elapsed();
+
+    tracing::info!(
+        compressed_length,
+        elapsed_ms = elapsed.as_millis(),
+        "write metrics bucket at {path}",
+    );
+
+    Ok(())
+}
+
 // The return type is quite ugly, but we gain testability in isolation
 fn serialize_in_chunks<'a, F>(
     chunk_size: usize,

pageserver/src/control_plane_client.rs

@@ -5,7 +5,8 @@ use pageserver_api::{
     controller_api::NodeRegisterRequest,
     shard::TenantShardId,
     upcall_api::{
-        ReAttachRequest, ReAttachResponse, ValidateRequest, ValidateRequestTenant, ValidateResponse,
+        ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
+        ValidateRequestTenant, ValidateResponse,
     },
 };
 use serde::{de::DeserializeOwned, Serialize};
@@ -37,7 +38,9 @@ pub trait ControlPlaneGenerationsApi {
     fn re_attach(
         &self,
         conf: &PageServerConf,
-    ) -> impl Future<Output = Result<HashMap<TenantShardId, Generation>, RetryForeverError>> + Send;
+    ) -> impl Future<
+        Output = Result<HashMap<TenantShardId, ReAttachResponseTenant>, RetryForeverError>,
+    > + Send;
     fn validate(
         &self,
         tenants: Vec<(TenantShardId, Generation)>,
@@ -118,7 +121,7 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
     async fn re_attach(
         &self,
         conf: &PageServerConf,
-    ) -> Result<HashMap<TenantShardId, Generation>, RetryForeverError> {
+    ) -> Result<HashMap<TenantShardId, ReAttachResponseTenant>, RetryForeverError> {
         let re_attach_path = self
             .base_url
             .join("re-attach")
@@ -181,7 +184,7 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
         Ok(response
             .tenants
             .into_iter()
-            .map(|t| (t.id, Generation::new(t.gen)))
+            .map(|rart| (rart.id, rart))
             .collect::<HashMap<_, _>>())
     }
 

pageserver/src/deletion_queue.rs

@@ -724,8 +724,8 @@ impl DeletionQueue {
 mod test {
     use camino::Utf8Path;
     use hex_literal::hex;
-    use pageserver_api::shard::ShardIndex;
-    use std::io::ErrorKind;
+    use pageserver_api::{shard::ShardIndex, upcall_api::ReAttachResponseTenant};
+    use std::{io::ErrorKind, time::Duration};
     use tracing::info;
 
     use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
@@ -834,9 +834,10 @@ mod test {
         async fn re_attach(
             &self,
             _conf: &PageServerConf,
-        ) -> Result<HashMap<TenantShardId, Generation>, RetryForeverError> {
+        ) -> Result<HashMap<TenantShardId, ReAttachResponseTenant>, RetryForeverError> {
             unimplemented!()
         }
+
         async fn validate(
             &self,
             tenants: Vec<(TenantShardId, Generation)>,

pageserver/src/disk_usage_eviction_task.rs

@@ -61,7 +61,6 @@ use crate::{
     metrics::disk_usage_based_eviction::METRICS,
     task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
     tenant::{
-        self,
         mgr::TenantManager,
         remote_timeline_client::LayerFileMetadata,
         secondary::SecondaryTenant,
@@ -814,8 +813,8 @@ async fn collect_eviction_candidates(
     const LOG_DURATION_THRESHOLD: std::time::Duration = std::time::Duration::from_secs(10);
 
     // get a snapshot of the list of tenants
-    let tenants = tenant::mgr::list_tenants()
-        .await
+    let tenants = tenant_manager
+        .list_tenants()
         .context("get list of tenants")?;
 
     // TODO: avoid listing every layer in every tenant: this loop can block the executor,
@@ -827,8 +826,12 @@ async fn collect_eviction_candidates(
         if cancel.is_cancelled() {
             return Ok(EvictionCandidates::Cancelled);
         }
-        let tenant = match tenant::mgr::get_tenant(tenant_id, true) {
-            Ok(tenant) => tenant,
+        let tenant = match tenant_manager.get_attached_tenant_shard(tenant_id) {
+            Ok(tenant) if tenant.is_active() => tenant,
+            Ok(_) => {
+                debug!(tenant_id=%tenant_id.tenant_id, shard_id=%tenant_id.shard_slug(), "Tenant shard is not active");
+                continue;
+            }
             Err(e) => {
                 // this can happen if tenant has lifecycle transition after we fetched it
                 debug!("failed to get tenant: {e:#}");

pageserver/src/http/openapi_spec.yml

@@ -1038,7 +1038,7 @@ paths:
                   format: hex
       responses:
         "201":
-          description: TimelineInfo
+          description: Timeline was created, or already existed with matching parameters
           content:
             application/json:
               schema:
@@ -1068,11 +1068,17 @@ paths:
               schema:
                 $ref: "#/components/schemas/Error"
         "409":
-          description: Timeline already exists, creation skipped
+          description: Timeline already exists, with different parameters.  Creation cannot proceed.
           content:
             application/json:
               schema:
                 $ref: "#/components/schemas/ConflictError"
+        "429":
+          description: A creation request was sent for the same Timeline Id while a creation was already in progress.  Back off and retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
         "500":
           description: Generic operation error
           content:

pageserver/src/http/routes.rs

@@ -36,6 +36,7 @@ use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::auth::JwtAuth;
 use utils::failpoint_support::failpoints_handler;
+use utils::http::endpoint::prometheus_metrics_handler;
 use utils::http::endpoint::request_span;
 use utils::http::json::json_request_or_empty_body;
 use utils::http::request::{get_request_param, must_get_query_param, parse_query_param};
@@ -48,8 +49,8 @@ use crate::task_mgr::TaskKind;
 use crate::tenant::config::{LocationConf, TenantConfOpt};
 use crate::tenant::mgr::GetActiveTenantError;
 use crate::tenant::mgr::{
-    GetTenantError, SetNewTenantConfigError, TenantManager, TenantMapError, TenantMapInsertError,
-    TenantSlotError, TenantSlotUpsertError, TenantStateError,
+    GetTenantError, TenantManager, TenantMapError, TenantMapInsertError, TenantSlotError,
+    TenantSlotUpsertError, TenantStateError,
 };
 use crate::tenant::mgr::{TenantSlot, UpsertLocationError};
 use crate::tenant::remote_timeline_client;
@@ -248,16 +249,11 @@ impl From<GetTenantError> for ApiError {
     fn from(tse: GetTenantError) -> ApiError {
         match tse {
             GetTenantError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid).into()),
-            GetTenantError::Broken(reason) => {
-                ApiError::InternalServerError(anyhow!("tenant is broken: {}", reason))
-            }
             GetTenantError::NotActive(_) => {
                 // Why is this not `ApiError::NotFound`?
                 // Because we must be careful to never return 404 for a tenant if it does
                 // in fact exist locally. If we did, the caller could draw the conclusion
                 // that it can attach the tenant to another PS and we'd be in split-brain.
-                //
-                // (We can produce this variant only in `mgr::get_tenant(..., active=true)` calls).
                 ApiError::ResourceUnavailable("Tenant not yet active".into())
             }
             GetTenantError::MapState(e) => ApiError::ResourceUnavailable(format!("{e}").into()),
@@ -268,6 +264,9 @@ impl From<GetTenantError> for ApiError {
 impl From<GetActiveTenantError> for ApiError {
     fn from(e: GetActiveTenantError) -> ApiError {
         match e {
+            GetActiveTenantError::Broken(reason) => {
+                ApiError::InternalServerError(anyhow!("tenant is broken: {}", reason))
+            }
             GetActiveTenantError::WillNotBecomeActive(_) => ApiError::Conflict(format!("{}", e)),
             GetActiveTenantError::Cancelled => ApiError::ShuttingDown,
             GetActiveTenantError::NotFound(gte) => gte.into(),
@@ -278,19 +277,6 @@ impl From<GetActiveTenantError> for ApiError {
     }
 }
 
-impl From<SetNewTenantConfigError> for ApiError {
-    fn from(e: SetNewTenantConfigError) -> ApiError {
-        match e {
-            SetNewTenantConfigError::GetTenant(tid) => {
-                ApiError::NotFound(anyhow!("tenant {}", tid).into())
-            }
-            e @ (SetNewTenantConfigError::Persist(_) | SetNewTenantConfigError::Other(_)) => {
-                ApiError::InternalServerError(anyhow::Error::new(e))
-            }
-        }
-    }
-}
-
 impl From<crate::tenant::DeleteTimelineError> for ApiError {
     fn from(value: crate::tenant::DeleteTimelineError) -> Self {
         use crate::tenant::DeleteTimelineError::*;
@@ -494,7 +480,7 @@ async fn timeline_create_handler(
     async {
         let tenant = state
             .tenant_manager
-            .get_attached_tenant_shard(tenant_shard_id, false)?;
+            .get_attached_tenant_shard(tenant_shard_id)?;
 
         tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
 
@@ -534,10 +520,13 @@ async fn timeline_create_handler(
                     HttpErrorBody::from_msg("Tenant shutting down".to_string()),
                 )
             }
-            Err(
-                e @ tenant::CreateTimelineError::Conflict
-                | e @ tenant::CreateTimelineError::AlreadyCreating,
-            ) => json_response(StatusCode::CONFLICT, HttpErrorBody::from_msg(e.to_string())),
+            Err(e @ tenant::CreateTimelineError::Conflict) => {
+                json_response(StatusCode::CONFLICT, HttpErrorBody::from_msg(e.to_string()))
+            }
+            Err(e @ tenant::CreateTimelineError::AlreadyCreating) => json_response(
+                StatusCode::TOO_MANY_REQUESTS,
+                HttpErrorBody::from_msg(e.to_string()),
+            ),
             Err(tenant::CreateTimelineError::AncestorLsn(err)) => json_response(
                 StatusCode::NOT_ACCEPTABLE,
                 HttpErrorBody::from_msg(format!("{err:#}")),
@@ -580,7 +569,7 @@ async fn timeline_list_handler(
     let response_data = async {
         let tenant = state
             .tenant_manager
-            .get_attached_tenant_shard(tenant_shard_id, false)?;
+            .get_attached_tenant_shard(tenant_shard_id)?;
 
         tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
 
@@ -618,6 +607,7 @@ async fn timeline_preserve_initdb_handler(
     let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+    let state = get_state(&request);
 
     // Part of the process for disaster recovery from safekeeper-stored WAL:
     // If we don't recover into a new timeline but want to keep the timeline ID,
@@ -625,7 +615,9 @@ async fn timeline_preserve_initdb_handler(
     // location where timeline recreation cand find it.
 
     async {
-        let tenant = mgr::get_tenant(tenant_shard_id, false)?;
+        let tenant = state
+            .tenant_manager
+            .get_attached_tenant_shard(tenant_shard_id)?;
 
         let timeline = tenant
             .get_timeline(timeline_id, false)
@@ -667,7 +659,7 @@ async fn timeline_detail_handler(
     let timeline_info = async {
         let tenant = state
             .tenant_manager
-            .get_attached_tenant_shard(tenant_shard_id, false)?;
+            .get_attached_tenant_shard(tenant_shard_id)?;
 
         tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
 
@@ -854,7 +846,7 @@ async fn timeline_delete_handler(
 
     let tenant = state
         .tenant_manager
-        .get_attached_tenant_shard(tenant_shard_id, false)
+        .get_attached_tenant_shard(tenant_shard_id)
         .map_err(|e| {
             match e {
                 // GetTenantError has a built-in conversion to ApiError, but in this context we don't
@@ -972,10 +964,11 @@ async fn tenant_list_handler(
     _cancel: CancellationToken,
 ) -> Result<Response<Body>, ApiError> {
     check_permission(&request, None)?;
+    let state = get_state(&request);
 
-    let response_data = mgr::list_tenants()
-        .instrument(info_span!("tenant_list"))
-        .await
+    let response_data = state
+        .tenant_manager
+        .list_tenants()
         .map_err(|_| {
             ApiError::ResourceUnavailable("Tenant map is initializing or shutting down".into())
         })?
@@ -998,9 +991,12 @@ async fn tenant_status(
 ) -> Result<Response<Body>, ApiError> {
     let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+    let state = get_state(&request);
 
     let tenant_info = async {
-        let tenant = mgr::get_tenant(tenant_shard_id, false)?;
+        let tenant = state
+            .tenant_manager
+            .get_attached_tenant_shard(tenant_shard_id)?;
 
         // Calculate total physical size of all timelines
         let mut current_physical_size = 0;
@@ -1073,9 +1069,7 @@ async fn tenant_size_handler(
     let inputs_only: Option<bool> = parse_query_param(&request, "inputs_only")?;
     let retention_period: Option<u64> = parse_query_param(&request, "retention_period")?;
     let headers = request.headers();
-
-    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
-    let tenant = mgr::get_tenant(tenant_shard_id, true)?;
+    let state = get_state(&request);
 
     if !tenant_shard_id.is_zero() {
         return Err(ApiError::BadRequest(anyhow!(
@@ -1083,6 +1077,12 @@ async fn tenant_size_handler(
         )));
     }
 
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+    let tenant = state
+        .tenant_manager
+        .get_attached_tenant_shard(tenant_shard_id)?;
+    tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+
     // this can be long operation
     let inputs = tenant
         .gather_size_inputs(
@@ -1151,10 +1151,15 @@ async fn tenant_shard_split_handler(
     let state = get_state(&request);
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
 
+    let tenant = state
+        .tenant_manager
+        .get_attached_tenant_shard(tenant_shard_id)?;
+    tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+
     let new_shards = state
         .tenant_manager
         .shard_split(
-            tenant_shard_id,
+            tenant,
             ShardCount::new(req.new_shard_count),
             req.new_stripe_size,
             &ctx,
@@ -1372,8 +1377,11 @@ async fn get_tenant_config_handler(
 ) -> Result<Response<Body>, ApiError> {
     let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+    let state = get_state(&request);
 
-    let tenant = mgr::get_tenant(tenant_shard_id, false)?;
+    let tenant = state
+        .tenant_manager
+        .get_attached_tenant_shard(tenant_shard_id)?;
 
     let response = HashMap::from([
         (
@@ -1401,15 +1409,31 @@ async fn update_tenant_config_handler(
     let tenant_id = request_data.tenant_id;
     check_permission(&request, Some(tenant_id))?;
 
-    let tenant_conf =
+    let new_tenant_conf =
         TenantConfOpt::try_from(&request_data.config).map_err(ApiError::BadRequest)?;
 
     let state = get_state(&request);
-    state
+
+    let tenant_shard_id = TenantShardId::unsharded(tenant_id);
+
+    let tenant = state
         .tenant_manager
-        .set_new_tenant_config(tenant_conf, tenant_id)
-        .instrument(info_span!("tenant_config", %tenant_id))
-        .await?;
+        .get_attached_tenant_shard(tenant_shard_id)?;
+    tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+
+    // This is a legacy API that only operates on attached tenants: the preferred
+    // API to use is the location_config/ endpoint, which lets the caller provide
+    // the full LocationConf.
+    let location_conf = LocationConf::attached_single(
+        new_tenant_conf.clone(),
+        tenant.get_generation(),
+        &ShardParameters::default(),
+    );
+
+    crate::tenant::Tenant::persist_tenant_config(state.conf, &tenant_shard_id, &location_conf)
+        .await
+        .map_err(ApiError::InternalServerError)?;
+    tenant.set_new_tenant_config(new_tenant_conf);
 
     json_response(StatusCode::OK, ())
 }
@@ -1633,10 +1657,12 @@ async fn handle_tenant_break(
 ) -> Result<Response<Body>, ApiError> {
     let tenant_shard_id: TenantShardId = parse_request_param(&r, "tenant_shard_id")?;
 
-    let tenant = crate::tenant::mgr::get_tenant(tenant_shard_id, true)
-        .map_err(|_| ApiError::Conflict(String::from("no active tenant found")))?;
-
-    tenant.set_broken("broken from test".to_owned()).await;
+    let state = get_state(&r);
+    state
+        .tenant_manager
+        .get_attached_tenant_shard(tenant_shard_id)?
+        .set_broken("broken from test".to_owned())
+        .await;
 
     json_response(StatusCode::OK, ())
 }
@@ -1880,7 +1906,7 @@ async fn active_timeline_of_active_tenant(
     tenant_shard_id: TenantShardId,
     timeline_id: TimelineId,
 ) -> Result<Arc<Timeline>, ApiError> {
-    let tenant = tenant_manager.get_attached_tenant_shard(tenant_shard_id, false)?;
+    let tenant = tenant_manager.get_attached_tenant_shard(tenant_shard_id)?;
 
     tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
 
@@ -2266,6 +2292,7 @@ pub fn make_router(
 
     Ok(router
         .data(state)
+        .get("/metrics", |r| request_span(r, prometheus_metrics_handler))
         .get("/v1/status", |r| api_handler(r, status_handler))
         .put("/v1/failpoints", |r| {
             testing_api_handler("manage failpoints", r, failpoints_handler)

pageserver/src/import_datadir.rs

@@ -2,28 +2,20 @@
 //! Import data and WAL from a PostgreSQL data directory and WAL segments into
 //! a neon Timeline.
 //!
-use std::io::SeekFrom;
 use std::path::{Path, PathBuf};
 
 use anyhow::{bail, ensure, Context, Result};
-use async_compression::tokio::bufread::ZstdDecoder;
-use async_compression::{tokio::write::ZstdEncoder, zstd::CParameter, Level};
 use bytes::Bytes;
 use camino::Utf8Path;
 use futures::StreamExt;
-use nix::NixPath;
-use tokio::fs::{File, OpenOptions};
-use tokio::io::{AsyncBufRead, AsyncRead, AsyncReadExt, AsyncSeekExt, AsyncWriteExt};
+use tokio::io::{AsyncRead, AsyncReadExt};
 use tokio_tar::Archive;
-use tokio_tar::Builder;
-use tokio_tar::HeaderMode;
 use tracing::*;
 use walkdir::WalkDir;
 
 use crate::context::RequestContext;
 use crate::metrics::WAL_INGEST;
 use crate::pgdatadir_mapping::*;
-use crate::tenant::remote_timeline_client::INITDB_PATH;
 use crate::tenant::Timeline;
 use crate::walingest::WalIngest;
 use crate::walrecord::DecodedWALRecord;
@@ -633,65 +625,3 @@ async fn read_all_bytes(reader: &mut (impl AsyncRead + Unpin)) -> Result<Bytes>
     reader.read_to_end(&mut buf).await?;
     Ok(Bytes::from(buf))
 }
-
-pub async fn create_tar_zst(pgdata_path: &Utf8Path, tmp_path: &Utf8Path) -> Result<(File, u64)> {
-    let file = OpenOptions::new()
-        .create(true)
-        .truncate(true)
-        .read(true)
-        .write(true)
-        .open(&tmp_path)
-        .await
-        .with_context(|| format!("tempfile creation {tmp_path}"))?;
-
-    let mut paths = Vec::new();
-    for entry in WalkDir::new(pgdata_path) {
-        let entry = entry?;
-        let metadata = entry.metadata().expect("error getting dir entry metadata");
-        // Also allow directories so that we also get empty directories
-        if !(metadata.is_file() || metadata.is_dir()) {
-            continue;
-        }
-        let path = entry.into_path();
-        paths.push(path);
-    }
-    // Do a sort to get a more consistent listing
-    paths.sort_unstable();
-    let zstd = ZstdEncoder::with_quality_and_params(
-        file,
-        Level::Default,
-        &[CParameter::enable_long_distance_matching(true)],
-    );
-    let mut builder = Builder::new(zstd);
-    // Use reproducible header mode
-    builder.mode(HeaderMode::Deterministic);
-    for path in paths {
-        let rel_path = path.strip_prefix(pgdata_path)?;
-        if rel_path.is_empty() {
-            // The top directory should not be compressed,
-            // the tar crate doesn't like that
-            continue;
-        }
-        builder.append_path_with_name(&path, rel_path).await?;
-    }
-    let mut zstd = builder.into_inner().await?;
-    zstd.shutdown().await?;
-    let mut compressed = zstd.into_inner();
-    let compressed_len = compressed.metadata().await?.len();
-    const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
-    if compressed_len > INITDB_TAR_ZST_WARN_LIMIT {
-        warn!("compressed {INITDB_PATH} size of {compressed_len} is above limit {INITDB_TAR_ZST_WARN_LIMIT}.");
-    }
-    compressed.seek(SeekFrom::Start(0)).await?;
-    Ok((compressed, compressed_len))
-}
-
-pub async fn extract_tar_zst(
-    pgdata_path: &Utf8Path,
-    tar_zst: impl AsyncBufRead + Unpin,
-) -> Result<()> {
-    let tar = Box::pin(ZstdDecoder::new(tar_zst));
-    let mut archive = Archive::new(tar);
-    archive.unpack(pgdata_path).await?;
-    Ok(())
-}

pageserver/src/metrics.rs

@@ -435,7 +435,7 @@ pub(crate) static RESIDENT_PHYSICAL_SIZE_GLOBAL: Lazy<UIntGauge> = Lazy::new(||
 static REMOTE_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
     register_uint_gauge_vec!(
         "pageserver_remote_physical_size",
-        "The size of the layer files present in the remote storage that are listed in the the remote index_part.json.",
+        "The size of the layer files present in the remote storage that are listed in the remote index_part.json.",
         // Corollary: If any files are missing from the index part, they won't be included here.
         &["tenant_id", "shard_id", "timeline_id"]
     )
@@ -699,6 +699,14 @@ pub static STARTUP_IS_LOADING: Lazy<UIntGauge> = Lazy::new(|| {
     .expect("Failed to register pageserver_startup_is_loading")
 });
 
+pub(crate) static TIMELINE_EPHEMERAL_BYTES: Lazy<UIntGauge> = Lazy::new(|| {
+    register_uint_gauge!(
+        "pageserver_timeline_ephemeral_bytes",
+        "Total number of bytes in ephemeral layers, summed for all timelines.  Approximate, lazily updated."
+    )
+    .expect("Failed to register metric")
+});
+
 /// Metrics related to the lifecycle of a [`crate::tenant::Tenant`] object: things
 /// like how long it took to load.
 ///

pageserver/src/page_service.rs

@@ -760,6 +760,7 @@ impl PageServerHandler {
         let mut copyin_reader = pin!(StreamReader::new(self.copyin_stream(pgb, &tenant.cancel)));
         timeline
             .import_basebackup_from_tar(
+                tenant.clone(),
                 &mut copyin_reader,
                 base_lsn,
                 self.broker_client.clone(),

pageserver/src/pgdatadir_mapping.rs

@@ -34,6 +34,7 @@ use strum::IntoEnumIterator;
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, trace, warn};
 use utils::bin_ser::DeserializeError;
+use utils::vec_map::{VecMap, VecMapOrdering};
 use utils::{bin_ser::BeSer, lsn::Lsn};
 
 const MAX_AUX_FILE_DELTAS: usize = 1024;
@@ -1546,12 +1547,13 @@ impl<'a> DatadirModification<'a> {
         if !self.pending_updates.is_empty() {
             // The put_batch call below expects expects the inputs to be sorted by Lsn,
             // so we do that first.
-            let lsn_ordered_batch: Vec<(Key, Lsn, Value)> = self
-                .pending_updates
-                .drain()
-                .map(|(key, vals)| vals.into_iter().map(move |(lsn, val)| (key, lsn, val)))
-                .kmerge_by(|lhs, rhs| lhs.1 .0 < rhs.1 .0)
-                .collect();
+            let lsn_ordered_batch: VecMap<Lsn, (Key, Value)> = VecMap::from_iter(
+                self.pending_updates
+                    .drain()
+                    .map(|(key, vals)| vals.into_iter().map(move |(lsn, val)| (lsn, (key, val))))
+                    .kmerge_by(|lhs, rhs| lhs.0 < rhs.0),
+                VecMapOrdering::GreaterOrEqual,
+            );
 
             writer.put_batch(lsn_ordered_batch, ctx).await?;
         }

pageserver/src/tenant.rs

@@ -43,6 +43,8 @@ use utils::sync::gate::Gate;
 use utils::sync::gate::GateGuard;
 use utils::timeout::timeout_cancellable;
 use utils::timeout::TimeoutCancellableError;
+use utils::zstd::create_zst_tarball;
+use utils::zstd::extract_zst_tarball;
 
 use self::config::AttachedLocationConfig;
 use self::config::AttachmentMode;
@@ -200,6 +202,13 @@ pub(super) struct AttachedTenantConf {
 }
 
 impl AttachedTenantConf {
+    fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
+        Self {
+            tenant_conf,
+            location,
+        }
+    }
+
     fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
         match &location_conf.mode {
             LocationMode::Attached(attach_conf) => Ok(Self {
@@ -676,9 +685,20 @@ impl Tenant {
                 }
 
                 // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
+                enum BrokenVerbosity {
+                    Error,
+                    Info
+                }
                 let make_broken =
-                    |t: &Tenant, err: anyhow::Error| {
-                        error!("attach failed, setting tenant state to Broken: {err:?}");
+                    |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
+                        match verbosity {
+                            BrokenVerbosity::Info => {
+                                info!("attach cancelled, setting tenant state to Broken: {err}");
+                            },
+                            BrokenVerbosity::Error => {
+                                error!("attach failed, setting tenant state to Broken: {err:?}");
+                            }
+                        }
                         t.state.send_modify(|state| {
                             // The Stopping case is for when we have passed control on to DeleteTenantFlow:
                             // if it errors, we will call make_broken when tenant is already in Stopping.
@@ -742,7 +762,7 @@ impl Tenant {
                             // Make the tenant broken so that set_stopping will not hang waiting for it to leave
                             // the Attaching state.  This is an over-reaction (nothing really broke, the tenant is
                             // just shutting down), but ensures progress.
-                            make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"));
+                            make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
                             return Ok(());
                         },
                     )
@@ -764,7 +784,7 @@ impl Tenant {
                         match res {
                             Ok(p) => Some(p),
                             Err(e) => {
-                                make_broken(&tenant_clone, anyhow::anyhow!(e));
+                                make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
                                 return Ok(());
                             }
                         }
@@ -788,7 +808,7 @@ impl Tenant {
                     {
                         Ok(should_resume_deletion) => should_resume_deletion,
                         Err(err) => {
-                            make_broken(&tenant_clone, anyhow::anyhow!(err));
+                            make_broken(&tenant_clone, anyhow::anyhow!(err), BrokenVerbosity::Error);
                             return Ok(());
                         }
                     }
@@ -818,7 +838,7 @@ impl Tenant {
                     .await;
 
                     if let Err(e) = deleted {
-                        make_broken(&tenant_clone, anyhow::anyhow!(e));
+                        make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
                     }
 
                     return Ok(());
@@ -839,7 +859,7 @@ impl Tenant {
                         tenant_clone.activate(broker_client, None, &ctx);
                     }
                     Err(e) => {
-                        make_broken(&tenant_clone, anyhow::anyhow!(e));
+                        make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
                     }
                 }
 
@@ -1391,7 +1411,7 @@ impl Tenant {
     /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
     #[allow(clippy::too_many_arguments)]
     pub(crate) async fn create_timeline(
-        &self,
+        self: &Arc<Tenant>,
         new_timeline_id: TimelineId,
         ancestor_timeline_id: Option<TimelineId>,
         mut ancestor_start_lsn: Option<Lsn>,
@@ -1539,7 +1559,7 @@ impl Tenant {
             })?;
         }
 
-        loaded_timeline.activate(broker_client, None, ctx);
+        loaded_timeline.activate(self.clone(), broker_client, None, ctx);
 
         Ok(loaded_timeline)
     }
@@ -1711,7 +1731,12 @@ impl Tenant {
             let mut activated_timelines = 0;
 
             for timeline in timelines_to_activate {
-                timeline.activate(broker_client.clone(), background_jobs_can_start, ctx);
+                timeline.activate(
+                    self.clone(),
+                    broker_client.clone(),
+                    background_jobs_can_start,
+                    ctx,
+                );
                 activated_timelines += 1;
             }
 
@@ -2043,7 +2068,12 @@ impl Tenant {
                 TenantState::Active { .. } => {
                     return Ok(());
                 }
-                TenantState::Broken { .. } | TenantState::Stopping { .. } => {
+                TenantState::Broken { reason, .. } => {
+                    // This is fatal, and reported distinctly from the general case of "will never be active" because
+                    // it's logically a 500 to external API users (broken is always a bug).
+                    return Err(GetActiveTenantError::Broken(reason));
+                }
+                TenantState::Stopping { .. } => {
                     // There's no chance the tenant can transition back into ::Active
                     return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
                 }
@@ -2121,7 +2151,7 @@ impl Tenant {
 
             // Shut down the timeline's remote client: this means that the indices we write
             // for child shards will not be invalidated by the parent shard deleting layers.
-            tl_client.shutdown().await?;
+            tl_client.shutdown().await;
 
             // Download methods can still be used after shutdown, as they don't flow through the remote client's
             // queue.  In principal the RemoteTimelineClient could provide this without downloading it, but this
@@ -3042,8 +3072,13 @@ impl Tenant {
             }
         }
 
-        let (pgdata_zstd, tar_zst_size) =
-            import_datadir::create_tar_zst(pgdata_path, &temp_path).await?;
+        let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
+        const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
+        if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
+            warn!(
+                "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
+            );
+        }
 
         pausable_failpoint!("before-initdb-upload");
 
@@ -3143,7 +3178,7 @@ impl Tenant {
 
             let buf_read =
                 BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
-            import_datadir::extract_tar_zst(&pgdata_path, buf_read)
+            extract_zst_tarball(&pgdata_path, buf_read)
                 .await
                 .context("extract initdb tar")?;
         } else {

pageserver/src/tenant/config.rs

@@ -196,16 +196,17 @@ impl LocationConf {
     /// For use when attaching/re-attaching: update the generation stored in this
     /// structure.  If we were in a secondary state, promote to attached (posession
     /// of a fresh generation implies this).
-    pub(crate) fn attach_in_generation(&mut self, generation: Generation) {
+    pub(crate) fn attach_in_generation(&mut self, mode: AttachmentMode, generation: Generation) {
         match &mut self.mode {
             LocationMode::Attached(attach_conf) => {
                 attach_conf.generation = generation;
+                attach_conf.attach_mode = mode;
             }
             LocationMode::Secondary(_) => {
                 // We are promoted to attached by the control plane's re-attach response
                 self.mode = LocationMode::Attached(AttachedLocationConfig {
                     generation,
-                    attach_mode: AttachmentMode::Single,
+                    attach_mode: mode,
                 })
             }
         }

pageserver/src/tenant/delete.rs

@@ -111,6 +111,7 @@ async fn create_local_delete_mark(
     let _ = std::fs::OpenOptions::new()
         .write(true)
         .create(true)
+        .truncate(true)
         .open(&marker_path)
         .with_context(|| format!("could not create delete marker file {marker_path:?}"))?;
 

pageserver/src/tenant/mgr.rs

@@ -2,13 +2,13 @@
 //! page server.
 
 use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
-use futures::stream::StreamExt;
 use itertools::Itertools;
 use pageserver_api::key::Key;
-use pageserver_api::models::ShardParameters;
+use pageserver_api::models::LocationConfigMode;
 use pageserver_api::shard::{
     ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId,
 };
+use pageserver_api::upcall_api::ReAttachResponseTenant;
 use rand::{distributions::Alphanumeric, Rng};
 use std::borrow::Cow;
 use std::cmp::Ordering;
@@ -16,6 +16,7 @@ use std::collections::{BTreeMap, HashMap};
 use std::ops::Deref;
 use std::sync::Arc;
 use std::time::{Duration, Instant};
+use sysinfo::SystemExt;
 use tokio::fs;
 use utils::timeout::{timeout_cancellable, TimeoutCancellableError};
 
@@ -39,10 +40,10 @@ use crate::metrics::{TENANT, TENANT_MANAGER as METRICS};
 use crate::task_mgr::{self, TaskKind};
 use crate::tenant::config::{
     AttachedLocationConfig, AttachmentMode, LocationConf, LocationMode, SecondaryLocationConfig,
-    TenantConfOpt,
 };
 use crate::tenant::delete::DeleteTenantFlow;
 use crate::tenant::span::debug_assert_current_span_has_tenant_id;
+use crate::tenant::storage_layer::inmemory_layer;
 use crate::tenant::{AttachedTenantConf, SpawnMode, Tenant, TenantState};
 use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TEMP_FILE_SUFFIX};
 
@@ -125,6 +126,46 @@ pub(crate) enum ShardSelector {
     Page(Key),
 }
 
+/// A convenience for use with the re_attach ControlPlaneClient function: rather
+/// than the serializable struct, we build this enum that encapsulates
+/// the invariant that attached tenants always have generations.
+///
+/// This represents the subset of a LocationConfig that we receive during re-attach.
+pub(crate) enum TenantStartupMode {
+    Attached((AttachmentMode, Generation)),
+    Secondary,
+}
+
+impl TenantStartupMode {
+    /// Return the generation & mode that should be used when starting
+    /// this tenant.
+    ///
+    /// If this returns None, the re-attach struct is in an invalid state and
+    /// should be ignored in the response.
+    fn from_reattach_tenant(rart: ReAttachResponseTenant) -> Option<Self> {
+        match (rart.mode, rart.gen) {
+            (LocationConfigMode::Detached, _) => None,
+            (LocationConfigMode::Secondary, _) => Some(Self::Secondary),
+            (LocationConfigMode::AttachedMulti, Some(g)) => {
+                Some(Self::Attached((AttachmentMode::Multi, Generation::new(g))))
+            }
+            (LocationConfigMode::AttachedSingle, Some(g)) => {
+                Some(Self::Attached((AttachmentMode::Single, Generation::new(g))))
+            }
+            (LocationConfigMode::AttachedStale, Some(g)) => {
+                Some(Self::Attached((AttachmentMode::Stale, Generation::new(g))))
+            }
+            _ => {
+                tracing::warn!(
+                    "Received invalid re-attach state for tenant {}: {rart:?}",
+                    rart.id
+                );
+                None
+            }
+        }
+    }
+}
+
 impl TenantsMap {
     /// Convenience function for typical usage, where we want to get a `Tenant` object, for
     /// working with attached tenants.  If the TenantId is in the map but in Secondary state,
@@ -271,7 +312,7 @@ pub struct TenantManager {
 
 fn emergency_generations(
     tenant_confs: &HashMap<TenantShardId, anyhow::Result<LocationConf>>,
-) -> HashMap<TenantShardId, Generation> {
+) -> HashMap<TenantShardId, TenantStartupMode> {
     tenant_confs
         .iter()
         .filter_map(|(tid, lc)| {
@@ -279,12 +320,15 @@ fn emergency_generations(
                 Ok(lc) => lc,
                 Err(_) => return None,
             };
-            let gen = match &lc.mode {
-                LocationMode::Attached(alc) => Some(alc.generation),
-                LocationMode::Secondary(_) => None,
-            };
-
-            gen.map(|g| (*tid, g))
+            Some((
+                *tid,
+                match &lc.mode {
+                    LocationMode::Attached(alc) => {
+                        TenantStartupMode::Attached((alc.attach_mode, alc.generation))
+                    }
+                    LocationMode::Secondary(_) => TenantStartupMode::Secondary,
+                },
+            ))
         })
         .collect()
 }
@@ -294,7 +338,7 @@ async fn init_load_generations(
     tenant_confs: &HashMap<TenantShardId, anyhow::Result<LocationConf>>,
     resources: &TenantSharedResources,
     cancel: &CancellationToken,
-) -> anyhow::Result<Option<HashMap<TenantShardId, Generation>>> {
+) -> anyhow::Result<Option<HashMap<TenantShardId, TenantStartupMode>>> {
     let generations = if conf.control_plane_emergency_mode {
         error!(
             "Emergency mode!  Tenants will be attached unsafely using their last known generation"
@@ -304,7 +348,12 @@ async fn init_load_generations(
         info!("Calling control plane API to re-attach tenants");
         // If we are configured to use the control plane API, then it is the source of truth for what tenants to load.
         match client.re_attach(conf).await {
-            Ok(tenants) => tenants,
+            Ok(tenants) => tenants
+                .into_iter()
+                .flat_map(|(id, rart)| {
+                    TenantStartupMode::from_reattach_tenant(rart).map(|tsm| (id, tsm))
+                })
+                .collect(),
             Err(RetryForeverError::ShuttingDown) => {
                 anyhow::bail!("Shut down while waiting for control plane re-attach response")
             }
@@ -322,9 +371,17 @@ async fn init_load_generations(
     // Must only do this if remote storage is enabled, otherwise deletion queue
     // is not running and channel push will fail.
     if resources.remote_storage.is_some() {
-        resources
-            .deletion_queue_client
-            .recover(generations.clone())?;
+        let attached_tenants = generations
+            .iter()
+            .flat_map(|(id, start_mode)| {
+                match start_mode {
+                    TenantStartupMode::Attached((_mode, generation)) => Some(generation),
+                    TenantStartupMode::Secondary => None,
+                }
+                .map(|gen| (*id, *gen))
+            })
+            .collect();
+        resources.deletion_queue_client.recover(attached_tenants)?;
     }
 
     Ok(Some(generations))
@@ -487,12 +544,23 @@ pub async fn init_tenant_mgr(
 
     let ctx = RequestContext::todo_child(TaskKind::Startup, DownloadBehavior::Warn);
 
+    // Initialize dynamic limits that depend on system resources
+    let system_memory =
+        sysinfo::System::new_with_specifics(sysinfo::RefreshKind::new().with_memory())
+            .total_memory();
+    let max_ephemeral_layer_bytes =
+        conf.ephemeral_bytes_per_memory_kb as u64 * (system_memory / 1024);
+    tracing::info!("Initialized ephemeral layer size limit to {max_ephemeral_layer_bytes}, for {system_memory} bytes of memory");
+    inmemory_layer::GLOBAL_RESOURCES.max_dirty_bytes.store(
+        max_ephemeral_layer_bytes,
+        std::sync::atomic::Ordering::Relaxed,
+    );
+
     // Scan local filesystem for attached tenants
     let tenant_configs = init_load_tenant_configs(conf).await?;
 
-    // Determine which tenants are to be attached
-    let tenant_generations =
-        init_load_generations(conf, &tenant_configs, &resources, &cancel).await?;
+    // Determine which tenants are to be secondary or attached, and in which generation
+    let tenant_modes = init_load_generations(conf, &tenant_configs, &resources, &cancel).await?;
 
     tracing::info!(
         "Attaching {} tenants at startup, warming up {} at a time",
@@ -522,97 +590,102 @@ pub async fn init_tenant_mgr(
             }
         };
 
-        let generation = if let Some(generations) = &tenant_generations {
+        // FIXME: if we were attached, and get demoted to secondary on re-attach, we
+        // don't have a place to get a config.
+        // (https://github.com/neondatabase/neon/issues/5377)
+        const DEFAULT_SECONDARY_CONF: SecondaryLocationConfig =
+            SecondaryLocationConfig { warm: true };
+
+        // Update the location config according to the re-attach response
+        if let Some(tenant_modes) = &tenant_modes {
             // We have a generation map: treat it as the authority for whether
             // this tenant is really attached.
-            if let Some(gen) = generations.get(&tenant_shard_id) {
-                if let LocationMode::Attached(attached) = &location_conf.mode {
-                    if attached.generation > *gen {
+            match tenant_modes.get(&tenant_shard_id) {
+                None => {
+                    info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Detaching tenant, control plane omitted it in re-attach response");
+                    if let Err(e) = safe_remove_tenant_dir_all(&tenant_dir_path).await {
+                        error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
+                            "Failed to remove detached tenant directory '{tenant_dir_path}': {e:?}",
+                        );
+                    }
+
+                    // We deleted local content: move on to next tenant, don't try and spawn this one.
+                    continue;
+                }
+                Some(TenantStartupMode::Secondary) => {
+                    if !matches!(location_conf.mode, LocationMode::Secondary(_)) {
+                        location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
+                    }
+                }
+                Some(TenantStartupMode::Attached((attach_mode, generation))) => {
+                    let old_gen_higher = match &location_conf.mode {
+                        LocationMode::Attached(AttachedLocationConfig {
+                            generation: old_generation,
+                            attach_mode: _attach_mode,
+                        }) => {
+                            if old_generation > generation {
+                                Some(old_generation)
+                            } else {
+                                None
+                            }
+                        }
+                        _ => None,
+                    };
+                    if let Some(old_generation) = old_gen_higher {
                         tracing::error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
-                            "Control plane gave decreasing generation ({gen:?}) in re-attach response for tenant that was attached in generation {:?}, demoting to secondary",
-                            attached.generation
+                            "Control plane gave decreasing generation ({generation:?}) in re-attach response for tenant that was attached in generation {:?}, demoting to secondary",
+                            old_generation
                         );
 
                         // We cannot safely attach this tenant given a bogus generation number, but let's avoid throwing away
                         // local disk content: demote to secondary rather than detaching.
-                        tenants.insert(
-                            tenant_shard_id,
-                            TenantSlot::Secondary(SecondaryTenant::new(
-                                tenant_shard_id,
-                                location_conf.shard,
-                                location_conf.tenant_conf.clone(),
-                                &SecondaryLocationConfig { warm: false },
-                            )),
-                        );
+                        location_conf.mode = LocationMode::Secondary(DEFAULT_SECONDARY_CONF);
+                    } else {
+                        location_conf.attach_in_generation(*attach_mode, *generation);
                     }
                 }
-                *gen
-            } else {
-                match &location_conf.mode {
-                    LocationMode::Secondary(secondary_config) => {
-                        // We do not require the control plane's permission for secondary mode
-                        // tenants, because they do no remote writes and hence require no
-                        // generation number
-                        info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Loaded tenant in secondary mode");
-                        tenants.insert(
-                            tenant_shard_id,
-                            TenantSlot::Secondary(SecondaryTenant::new(
-                                tenant_shard_id,
-                                location_conf.shard,
-                                location_conf.tenant_conf,
-                                secondary_config,
-                            )),
-                        );
-                    }
-                    LocationMode::Attached(_) => {
-                        // TODO: augment re-attach API to enable the control plane to
-                        // instruct us about secondary attachments.  That way, instead of throwing
-                        // away local state, we can gracefully fall back to secondary here, if the control
-                        // plane tells us so.
-                        // (https://github.com/neondatabase/neon/issues/5377)
-                        info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Detaching tenant, control plane omitted it in re-attach response");
-                        if let Err(e) = safe_remove_tenant_dir_all(&tenant_dir_path).await {
-                            error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
-                                "Failed to remove detached tenant directory '{tenant_dir_path}': {e:?}",
-                            );
-                        }
-                    }
-                };
-
-                continue;
             }
         } else {
             // Legacy mode: no generation information, any tenant present
             // on local disk may activate
             info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Starting tenant in legacy mode, no generation",);
-            Generation::none()
         };
 
         // Presence of a generation number implies attachment: attach the tenant
         // if it wasn't already, and apply the generation number.
-        location_conf.attach_in_generation(generation);
         Tenant::persist_tenant_config(conf, &tenant_shard_id, &location_conf).await?;
 
         let shard_identity = location_conf.shard;
-        match tenant_spawn(
-            conf,
-            tenant_shard_id,
-            &tenant_dir_path,
-            resources.clone(),
-            AttachedTenantConf::try_from(location_conf)?,
-            shard_identity,
-            Some(init_order.clone()),
-            &TENANTS,
-            SpawnMode::Lazy,
-            &ctx,
-        ) {
-            Ok(tenant) => {
-                tenants.insert(tenant_shard_id, TenantSlot::Attached(tenant));
+        let slot = match location_conf.mode {
+            LocationMode::Attached(attached_conf) => {
+                match tenant_spawn(
+                    conf,
+                    tenant_shard_id,
+                    &tenant_dir_path,
+                    resources.clone(),
+                    AttachedTenantConf::new(location_conf.tenant_conf, attached_conf),
+                    shard_identity,
+                    Some(init_order.clone()),
+                    &TENANTS,
+                    SpawnMode::Lazy,
+                    &ctx,
+                ) {
+                    Ok(tenant) => TenantSlot::Attached(tenant),
+                    Err(e) => {
+                        error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Failed to start tenant: {e:#}");
+                        continue;
+                    }
+                }
             }
-            Err(e) => {
-                error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Failed to start tenant: {e:#}");
-            }
-        }
+            LocationMode::Secondary(secondary_conf) => TenantSlot::Secondary(SecondaryTenant::new(
+                tenant_shard_id,
+                shard_identity,
+                location_conf.tenant_conf,
+                &secondary_conf,
+            )),
+        };
+
+        tenants.insert(tenant_shard_id, slot);
     }
 
     info!("Processed {} local tenants at startup", tenants.len());
@@ -815,16 +888,6 @@ async fn shutdown_all_tenants0(tenants: &std::sync::RwLock<TenantsMap>) {
     // caller will log how long we took
 }
 
-#[derive(Debug, thiserror::Error)]
-pub(crate) enum SetNewTenantConfigError {
-    #[error(transparent)]
-    GetTenant(#[from] GetTenantError),
-    #[error(transparent)]
-    Persist(anyhow::Error),
-    #[error(transparent)]
-    Other(anyhow::Error),
-}
-
 #[derive(thiserror::Error, Debug)]
 pub(crate) enum UpsertLocationError {
     #[error("Bad config request: {0}")]
@@ -850,32 +913,21 @@ impl TenantManager {
         self.conf
     }
 
-    /// Gets the attached tenant from the in-memory data, erroring if it's absent, in secondary mode, or is not fitting to the query.
-    /// `active_only = true` allows to query only tenants that are ready for operations, erroring on other kinds of tenants.
+    /// Gets the attached tenant from the in-memory data, erroring if it's absent, in secondary mode, or currently
+    /// undergoing a state change (i.e. slot is InProgress).
+    ///
+    /// The return Tenant is not guaranteed to be active: check its status after obtaing it, or
+    /// use [`Tenant::wait_to_become_active`] before using it if you will do I/O on it.
     pub(crate) fn get_attached_tenant_shard(
         &self,
         tenant_shard_id: TenantShardId,
-        active_only: bool,
     ) -> Result<Arc<Tenant>, GetTenantError> {
         let locked = self.tenants.read().unwrap();
 
         let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)?;
 
         match peek_slot {
-            Some(TenantSlot::Attached(tenant)) => match tenant.current_state() {
-                TenantState::Broken {
-                    reason,
-                    backtrace: _,
-                } if active_only => Err(GetTenantError::Broken(reason)),
-                TenantState::Active => Ok(Arc::clone(tenant)),
-                _ => {
-                    if active_only {
-                        Err(GetTenantError::NotActive(tenant_shard_id))
-                    } else {
-                        Ok(Arc::clone(tenant))
-                    }
-                }
-            },
+            Some(TenantSlot::Attached(tenant)) => Ok(Arc::clone(tenant)),
             Some(TenantSlot::InProgress(_)) => Err(GetTenantError::NotActive(tenant_shard_id)),
             None | Some(TenantSlot::Secondary(_)) => {
                 Err(GetTenantError::NotFound(tenant_shard_id.tenant_id))
@@ -1368,7 +1420,8 @@ impl TenantManager {
                     .wait_to_become_active(activation_timeout)
                     .await
                     .map_err(|e| match e {
-                        GetActiveTenantError::WillNotBecomeActive(_) => {
+                        GetActiveTenantError::WillNotBecomeActive(_)
+                        | GetActiveTenantError::Broken(_) => {
                             DeleteTenantError::InvalidState(tenant.current_state())
                         }
                         GetActiveTenantError::Cancelled => DeleteTenantError::Cancelled,
@@ -1395,29 +1448,30 @@ impl TenantManager {
         result
     }
 
-    #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), new_shard_count=%new_shard_count.literal()))]
+    #[instrument(skip_all, fields(tenant_id=%tenant.get_tenant_shard_id().tenant_id, shard_id=%tenant.get_tenant_shard_id().shard_slug(), new_shard_count=%new_shard_count.literal()))]
     pub(crate) async fn shard_split(
         &self,
-        tenant_shard_id: TenantShardId,
+        tenant: Arc<Tenant>,
         new_shard_count: ShardCount,
         new_stripe_size: Option<ShardStripeSize>,
         ctx: &RequestContext,
     ) -> anyhow::Result<Vec<TenantShardId>> {
+        let tenant_shard_id = *tenant.get_tenant_shard_id();
         let r = self
-            .do_shard_split(tenant_shard_id, new_shard_count, new_stripe_size, ctx)
+            .do_shard_split(tenant, new_shard_count, new_stripe_size, ctx)
             .await;
         if r.is_err() {
             // Shard splitting might have left the original shard in a partially shut down state (it
             // stops the shard's remote timeline client).  Reset it to ensure we leave things in
             // a working state.
             if self.get(tenant_shard_id).is_some() {
-                tracing::warn!("Resetting {tenant_shard_id} after shard split failure");
+                tracing::warn!("Resetting after shard split failure");
                 if let Err(e) = self.reset_tenant(tenant_shard_id, false, ctx).await {
                     // Log this error because our return value will still be the original error, not this one.  This is
                     // a severe error: if this happens, we might be leaving behind a tenant that is not fully functional
                     // (e.g. has uploads disabled).  We can't do anything else: if reset fails then shutting the tenant down or
                     // setting it broken probably won't help either.
-                    tracing::error!("Failed to reset {tenant_shard_id}: {e}");
+                    tracing::error!("Failed to reset: {e}");
                 }
             }
         }
@@ -1427,12 +1481,12 @@ impl TenantManager {
 
     pub(crate) async fn do_shard_split(
         &self,
-        tenant_shard_id: TenantShardId,
+        tenant: Arc<Tenant>,
         new_shard_count: ShardCount,
         new_stripe_size: Option<ShardStripeSize>,
         ctx: &RequestContext,
     ) -> anyhow::Result<Vec<TenantShardId>> {
-        let tenant = get_tenant(tenant_shard_id, true)?;
+        let tenant_shard_id = *tenant.get_tenant_shard_id();
 
         // Validate the incoming request
         if new_shard_count.count() <= tenant_shard_id.shard_count.count() {
@@ -1478,7 +1532,6 @@ impl TenantManager {
             // If [`Tenant::split_prepare`] fails, we must reload the tenant, because it might
             // have been left in a partially-shut-down state.
             tracing::warn!("Failed to prepare for split: {e}, reloading Tenant before returning");
-            self.reset_tenant(tenant_shard_id, false, ctx).await?;
             return Err(e);
         }
 
@@ -1662,9 +1715,9 @@ impl TenantManager {
                     .layers
                     .read()
                     .await
-                    .resident_layers()
-                    .collect::<Vec<_>>()
-                    .await;
+                    .likely_resident_layers()
+                    .collect::<Vec<_>>();
+
                 for layer in timeline_layers {
                     let relative_path = layer
                         .local_path()
@@ -1876,38 +1929,23 @@ impl TenantManager {
         removal_result
     }
 
-    pub(crate) async fn set_new_tenant_config(
+    pub(crate) fn list_tenants(
         &self,
-        new_tenant_conf: TenantConfOpt,
-        tenant_id: TenantId,
-    ) -> Result<(), SetNewTenantConfigError> {
-        // Legacy API: does not support sharding
-        let tenant_shard_id = TenantShardId::unsharded(tenant_id);
-
-        info!("configuring tenant {tenant_id}");
-        let tenant = get_tenant(tenant_shard_id, true)?;
-
-        if !tenant.tenant_shard_id().shard_count.is_unsharded() {
-            // Note that we use ShardParameters::default below.
-            return Err(SetNewTenantConfigError::Other(anyhow::anyhow!(
-            "This API may only be used on single-sharded tenants, use the /location_config API for sharded tenants"
-        )));
-        }
-
-        // This is a legacy API that only operates on attached tenants: the preferred
-        // API to use is the location_config/ endpoint, which lets the caller provide
-        // the full LocationConf.
-        let location_conf = LocationConf::attached_single(
-            new_tenant_conf.clone(),
-            tenant.generation,
-            &ShardParameters::default(),
-        );
-
-        Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &location_conf)
-            .await
-            .map_err(SetNewTenantConfigError::Persist)?;
-        tenant.set_new_tenant_config(new_tenant_conf);
-        Ok(())
+    ) -> Result<Vec<(TenantShardId, TenantState, Generation)>, TenantMapListError> {
+        let tenants = TENANTS.read().unwrap();
+        let m = match &*tenants {
+            TenantsMap::Initializing => return Err(TenantMapListError::Initializing),
+            TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m,
+        };
+        Ok(m.iter()
+            .filter_map(|(id, tenant)| match tenant {
+                TenantSlot::Attached(tenant) => {
+                    Some((*id, tenant.current_state(), tenant.generation()))
+                }
+                TenantSlot::Secondary(_) => None,
+                TenantSlot::InProgress(_) => None,
+            })
+            .collect())
     }
 }
 
@@ -1920,51 +1958,12 @@ pub(crate) enum GetTenantError {
 
     #[error("Tenant {0} is not active")]
     NotActive(TenantShardId),
-    /// Broken is logically a subset of NotActive, but a distinct error is useful as
-    /// NotActive is usually a retryable state for API purposes, whereas Broken
-    /// is a stuck error state
-    #[error("Tenant is broken: {0}")]
-    Broken(String),
 
     // Initializing or shutting down: cannot authoritatively say whether we have this tenant
     #[error("Tenant map is not available: {0}")]
     MapState(#[from] TenantMapError),
 }
 
-/// Gets the tenant from the in-memory data, erroring if it's absent or is not fitting to the query.
-/// `active_only = true` allows to query only tenants that are ready for operations, erroring on other kinds of tenants.
-///
-/// This method is cancel-safe.
-pub(crate) fn get_tenant(
-    tenant_shard_id: TenantShardId,
-    active_only: bool,
-) -> Result<Arc<Tenant>, GetTenantError> {
-    let locked = TENANTS.read().unwrap();
-
-    let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)?;
-
-    match peek_slot {
-        Some(TenantSlot::Attached(tenant)) => match tenant.current_state() {
-            TenantState::Broken {
-                reason,
-                backtrace: _,
-            } if active_only => Err(GetTenantError::Broken(reason)),
-            TenantState::Active => Ok(Arc::clone(tenant)),
-            _ => {
-                if active_only {
-                    Err(GetTenantError::NotActive(tenant_shard_id))
-                } else {
-                    Ok(Arc::clone(tenant))
-                }
-            }
-        },
-        Some(TenantSlot::InProgress(_)) => Err(GetTenantError::NotActive(tenant_shard_id)),
-        None | Some(TenantSlot::Secondary(_)) => {
-            Err(GetTenantError::NotFound(tenant_shard_id.tenant_id))
-        }
-    }
-}
-
 #[derive(thiserror::Error, Debug)]
 pub(crate) enum GetActiveTenantError {
     /// We may time out either while TenantSlot is InProgress, or while the Tenant
@@ -1988,6 +1987,12 @@ pub(crate) enum GetActiveTenantError {
     /// Tenant exists, but is in a state that cannot become active (e.g. Stopping, Broken)
     #[error("will not become active.  Current state: {0}")]
     WillNotBecomeActive(TenantState),
+
+    /// Broken is logically a subset of WillNotBecomeActive, but a distinct error is useful as
+    /// WillNotBecomeActive is a permitted error under some circumstances, whereas broken should
+    /// never happen.
+    #[error("Tenant is broken: {0}")]
+    Broken(String),
 }
 
 /// Get a [`Tenant`] in its active state. If the tenant_id is currently in [`TenantSlot::InProgress`]
@@ -2143,7 +2148,7 @@ pub(crate) async fn load_tenant(
 
     let mut location_conf =
         Tenant::load_tenant_config(conf, &tenant_shard_id).map_err(TenantMapInsertError::Other)?;
-    location_conf.attach_in_generation(generation);
+    location_conf.attach_in_generation(AttachmentMode::Single, generation);
 
     Tenant::persist_tenant_config(conf, &tenant_shard_id, &location_conf).await?;
 
@@ -2207,27 +2212,6 @@ pub(crate) enum TenantMapListError {
     Initializing,
 }
 
-///
-/// Get list of tenants, for the mgmt API
-///
-pub(crate) async fn list_tenants(
-) -> Result<Vec<(TenantShardId, TenantState, Generation)>, TenantMapListError> {
-    let tenants = TENANTS.read().unwrap();
-    let m = match &*tenants {
-        TenantsMap::Initializing => return Err(TenantMapListError::Initializing),
-        TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => m,
-    };
-    Ok(m.iter()
-        .filter_map(|(id, tenant)| match tenant {
-            TenantSlot::Attached(tenant) => {
-                Some((*id, tenant.current_state(), tenant.generation()))
-            }
-            TenantSlot::Secondary(_) => None,
-            TenantSlot::InProgress(_) => None,
-        })
-        .collect())
-}
-
 #[derive(Debug, thiserror::Error)]
 pub(crate) enum TenantMapInsertError {
     #[error(transparent)]

pageserver/src/tenant/remote_timeline_client.rs

@@ -217,7 +217,7 @@ use crate::task_mgr::shutdown_token;
 use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::remote_timeline_client::download::download_retry;
 use crate::tenant::storage_layer::AsLayerDesc;
-use crate::tenant::upload_queue::Delete;
+use crate::tenant::upload_queue::{Delete, UploadQueueStoppedDeletable};
 use crate::tenant::TIMELINES_SEGMENT_NAME;
 use crate::{
     config::PageServerConf,
@@ -266,15 +266,6 @@ pub enum MaybeDeletedIndexPart {
     Deleted(IndexPart),
 }
 
-/// Errors that can arise when calling [`RemoteTimelineClient::stop`].
-#[derive(Debug, thiserror::Error)]
-pub enum StopError {
-    /// Returned if the upload queue was never initialized.
-    /// See [`RemoteTimelineClient::init_upload_queue`] and [`RemoteTimelineClient::init_upload_queue_for_empty_remote`].
-    #[error("queue is not initialized")]
-    QueueUninitialized,
-}
-
 #[derive(Debug, thiserror::Error)]
 pub enum PersistIndexPartWithDeletedFlagError {
     #[error("another task is already setting the deleted_flag, started at {0:?}")]
@@ -399,15 +390,10 @@ impl RemoteTimelineClient {
             "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.initialize_with_current_remote_index_part(index_part)?;
+        self.update_remote_physical_size_gauge(Some(index_part));
+        self.stop_impl(&mut upload_queue);
 
         upload_queue
             .stopped_mut()
@@ -421,7 +407,8 @@ impl RemoteTimelineClient {
         match &mut *self.upload_queue.lock().unwrap() {
             UploadQueue::Uninitialized => None,
             UploadQueue::Initialized(q) => q.get_last_remote_consistent_lsn_projected(),
-            UploadQueue::Stopped(q) => q
+            UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => None,
+            UploadQueue::Stopped(UploadQueueStopped::Deletable(q)) => q
                 .upload_queue_for_deletion
                 .get_last_remote_consistent_lsn_projected(),
         }
@@ -431,7 +418,8 @@ impl RemoteTimelineClient {
         match &mut *self.upload_queue.lock().unwrap() {
             UploadQueue::Uninitialized => None,
             UploadQueue::Initialized(q) => Some(q.get_last_remote_consistent_lsn_visible()),
-            UploadQueue::Stopped(q) => Some(
+            UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => None,
+            UploadQueue::Stopped(UploadQueueStopped::Deletable(q)) => Some(
                 q.upload_queue_for_deletion
                     .get_last_remote_consistent_lsn_visible(),
             ),
@@ -898,7 +886,7 @@ impl RemoteTimelineClient {
     /// Wait for all previously scheduled operations to complete, and then stop.
     ///
     /// Not cancellation safe
-    pub(crate) async fn shutdown(self: &Arc<Self>) -> Result<(), StopError> {
+    pub(crate) async fn shutdown(self: &Arc<Self>) {
         // On cancellation the queue is left in ackward state of refusing new operations but
         // proper stop is yet to be called. On cancel the original or some later task must call
         // `stop` or `shutdown`.
@@ -909,8 +897,12 @@ impl RemoteTimelineClient {
         let fut = {
             let mut guard = self.upload_queue.lock().unwrap();
             let upload_queue = match &mut *guard {
-                UploadQueue::Stopped(_) => return Ok(()),
-                UploadQueue::Uninitialized => return Err(StopError::QueueUninitialized),
+                UploadQueue::Stopped(_) => return,
+                UploadQueue::Uninitialized => {
+                    // transition into Stopped state
+                    self.stop_impl(&mut guard);
+                    return;
+                }
                 UploadQueue::Initialized(ref mut init) => init,
             };
 
@@ -942,7 +934,7 @@ impl RemoteTimelineClient {
             }
         }
 
-        self.stop()
+        self.stop();
     }
 
     /// Set the deleted_at field in the remote index file.
@@ -1324,12 +1316,7 @@ impl RemoteTimelineClient {
             // upload finishes or times out soon enough.
             if cancel.is_cancelled() {
                 info!("upload task cancelled by shutdown request");
-                match self.stop() {
-                    Ok(()) => {}
-                    Err(StopError::QueueUninitialized) => {
-                        unreachable!("we never launch an upload task if the queue is uninitialized, and once it is initialized, we never go back")
-                    }
-                }
+                self.stop();
                 return;
             }
 
@@ -1584,17 +1571,23 @@ impl RemoteTimelineClient {
     /// In-progress operations will still be running after this function returns.
     /// Use `task_mgr::shutdown_tasks(None, Some(self.tenant_id), Some(timeline_id))`
     /// to wait for them to complete, after calling this function.
-    pub(crate) fn stop(&self) -> Result<(), StopError> {
+    pub(crate) fn stop(&self) {
         // Whichever *task* for this RemoteTimelineClient grabs the mutex first will transition the queue
         // into stopped state, thereby dropping all off the queued *ops* which haven't become *tasks* yet.
         // The other *tasks* will come here and observe an already shut down queue and hence simply wrap up their business.
         let mut guard = self.upload_queue.lock().unwrap();
-        match &mut *guard {
-            UploadQueue::Uninitialized => Err(StopError::QueueUninitialized),
+        self.stop_impl(&mut guard);
+    }
+
+    fn stop_impl(&self, guard: &mut std::sync::MutexGuard<UploadQueue>) {
+        match &mut **guard {
+            UploadQueue::Uninitialized => {
+                info!("UploadQueue is in state Uninitialized, nothing to do");
+                **guard = UploadQueue::Stopped(UploadQueueStopped::Uninitialized);
+            }
             UploadQueue::Stopped(_) => {
                 // nothing to do
                 info!("another concurrent task already shut down the queue");
-                Ok(())
             }
             UploadQueue::Initialized(initialized) => {
                 info!("shutting down upload queue");
@@ -1627,11 +1620,13 @@ impl RemoteTimelineClient {
                     };
 
                     let upload_queue = std::mem::replace(
-                        &mut *guard,
-                        UploadQueue::Stopped(UploadQueueStopped {
-                            upload_queue_for_deletion,
-                            deleted_at: SetDeletedFlagProgress::NotRunning,
-                        }),
+                        &mut **guard,
+                        UploadQueue::Stopped(UploadQueueStopped::Deletable(
+                            UploadQueueStoppedDeletable {
+                                upload_queue_for_deletion,
+                                deleted_at: SetDeletedFlagProgress::NotRunning,
+                            },
+                        )),
                     );
                     if let UploadQueue::Initialized(qi) = upload_queue {
                         qi
@@ -1660,10 +1655,6 @@ impl RemoteTimelineClient {
                     // which is exactly what we want to happen.
                     drop(op);
                 }
-
-                // We're done.
-                drop(guard);
-                Ok(())
             }
         }
     }

pageserver/src/tenant/secondary/downloader.rs

@@ -15,6 +15,7 @@ use crate::{
     tenant::{
         config::SecondaryLocationConfig,
         debug_assert_current_span_has_tenant_and_timeline_id,
+        ephemeral_file::is_ephemeral_file,
         remote_timeline_client::{
             index::LayerFileMetadata, is_temp_download_file, FAILED_DOWNLOAD_WARN_THRESHOLD,
             FAILED_REMOTE_OP_RETRIES,
@@ -534,7 +535,11 @@ impl<'a> TenantDownloader<'a> {
             .await
             .maybe_fatal_err(&context_msg)?;
 
-        tracing::debug!("Wrote local heatmap to {}", heatmap_path);
+        tracing::debug!(
+            "Wrote local heatmap to {}, with {} timelines",
+            heatmap_path,
+            heatmap.timelines.len()
+        );
 
         // Clean up any local layers that aren't in the heatmap.  We do this first for all timelines, on the general
         // principle that deletions should be done before writes wherever possible, and so that we can use this
@@ -547,6 +552,10 @@ impl<'a> TenantDownloader<'a> {
         // Download the layers in the heatmap
         for timeline in heatmap.timelines {
             if self.secondary_state.cancel.is_cancelled() {
+                tracing::debug!(
+                    "Cancelled before downloading timeline {}",
+                    timeline.timeline_id
+                );
                 return Ok(());
             }
 
@@ -764,10 +773,13 @@ impl<'a> TenantDownloader<'a> {
             }
         };
 
+        tracing::debug!(timeline_id=%timeline.timeline_id, "Downloading layers, {} in heatmap", timeline.layers.len());
+
         // Download heatmap layers that are not present on local disk, or update their
         // access time if they are already present.
         for layer in timeline.layers {
             if self.secondary_state.cancel.is_cancelled() {
+                tracing::debug!("Cancelled -- dropping out of layer loop");
                 return Ok(());
             }
 
@@ -950,7 +962,10 @@ async fn init_timeline_state(
             // Secondary mode doesn't use local metadata files, but they might have been left behind by an attached tenant.
             warn!(path=?dentry.path(), "found legacy metadata file, these should have been removed in load_tenant_config");
             continue;
-        } else if crate::is_temporary(&file_path) || is_temp_download_file(&file_path) {
+        } else if crate::is_temporary(&file_path)
+            || is_temp_download_file(&file_path)
+            || is_ephemeral_file(file_name)
+        {
             // Temporary files are frequently left behind from restarting during downloads
             tracing::info!("Cleaning up temporary file {file_path}");
             if let Err(e) = tokio::fs::remove_file(&file_path)

pageserver/src/tenant/secondary/heatmap_uploader.rs

@@ -9,6 +9,7 @@ use crate::{
     metrics::SECONDARY_MODE,
     tenant::{
         config::AttachmentMode,
+        mgr::GetTenantError,
         mgr::TenantManager,
         remote_timeline_client::remote_heatmap_path,
         span::debug_assert_current_span_has_tenant_id,
@@ -292,8 +293,11 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
             "Starting heatmap write on command");
         let tenant = self
             .tenant_manager
-            .get_attached_tenant_shard(*tenant_shard_id, true)
+            .get_attached_tenant_shard(*tenant_shard_id)
             .map_err(|e| anyhow::anyhow!(e))?;
+        if !tenant.is_active() {
+            return Err(GetTenantError::NotActive(*tenant_shard_id).into());
+        }
 
         Ok(UploadPending {
             // Ignore our state for last digest: this forces an upload even if nothing has changed

pageserver/src/tenant/secondary/scheduler.rs

@@ -300,6 +300,7 @@ where
 
         let tenant_shard_id = job.get_tenant_shard_id();
         let barrier = if let Some(barrier) = self.get_running(tenant_shard_id) {
+            tracing::info!("Command already running, waiting for it");
             barrier
         } else {
             let running = self.spawn_now(job);

pageserver/src/tenant/storage_layer.rs

@@ -3,7 +3,7 @@
 pub mod delta_layer;
 mod filename;
 pub mod image_layer;
-mod inmemory_layer;
+pub(crate) mod inmemory_layer;
 pub(crate) mod layer;
 mod layer_desc;
 

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -23,8 +23,12 @@ use tracing::*;
 use utils::{bin_ser::BeSer, id::TimelineId, lsn::Lsn, vec_map::VecMap};
 // avoid binding to Write (conflicts with std::io::Write)
 // while being able to use std::fmt::Write's methods
+use crate::metrics::TIMELINE_EPHEMERAL_BYTES;
+use std::cmp::Ordering;
 use std::fmt::Write as _;
 use std::ops::Range;
+use std::sync::atomic::Ordering as AtomicOrdering;
+use std::sync::atomic::{AtomicU64, AtomicUsize};
 use tokio::sync::{RwLock, RwLockWriteGuard};
 
 use super::{
@@ -70,6 +74,8 @@ pub struct InMemoryLayerInner {
     /// Each serialized Value is preceded by a 'u32' length field.
     /// PerSeg::page_versions map stores offsets into this file.
     file: EphemeralFile,
+
+    resource_units: GlobalResourceUnits,
 }
 
 impl std::fmt::Debug for InMemoryLayerInner {
@@ -78,6 +84,121 @@ impl std::fmt::Debug for InMemoryLayerInner {
     }
 }
 
+/// State shared by all in-memory (ephemeral) layers.  Updated infrequently during background ticks in Timeline,
+/// to minimize contention.
+///
+/// This global state is used to implement behaviors that require a global view of the system, e.g.
+/// rolling layers proactively to limit the total amount of dirty data.
+pub(crate) struct GlobalResources {
+    // Limit on how high dirty_bytes may grow before we start freezing layers to reduce it.
+    // Zero means unlimited.
+    pub(crate) max_dirty_bytes: AtomicU64,
+    // How many bytes are in all EphemeralFile objects
+    dirty_bytes: AtomicU64,
+    // How many layers are contributing to dirty_bytes
+    dirty_layers: AtomicUsize,
+}
+
+// Per-timeline RAII struct for its contribution to [`GlobalResources`]
+struct GlobalResourceUnits {
+    // How many dirty bytes have I added to the global dirty_bytes: this guard object is responsible
+    // for decrementing the global counter by this many bytes when dropped.
+    dirty_bytes: u64,
+}
+
+impl GlobalResourceUnits {
+    // Hint for the layer append path to update us when the layer size differs from the last
+    // call to update_size by this much.  If we don't reach this threshold, we'll still get
+    // updated when the Timeline "ticks" in the background.
+    const MAX_SIZE_DRIFT: u64 = 10 * 1024 * 1024;
+
+    fn new() -> Self {
+        GLOBAL_RESOURCES
+            .dirty_layers
+            .fetch_add(1, AtomicOrdering::Relaxed);
+        Self { dirty_bytes: 0 }
+    }
+
+    /// Do not call this frequently: all timelines will write to these same global atomics,
+    /// so this is a relatively expensive operation.  Wait at least a few seconds between calls.
+    ///
+    /// Returns the effective layer size limit that should be applied, if any, to keep
+    /// the total number of dirty bytes below the configured maximum.
+    fn publish_size(&mut self, size: u64) -> Option<u64> {
+        let new_global_dirty_bytes = match size.cmp(&self.dirty_bytes) {
+            Ordering::Equal => GLOBAL_RESOURCES.dirty_bytes.load(AtomicOrdering::Relaxed),
+            Ordering::Greater => {
+                let delta = size - self.dirty_bytes;
+                let old = GLOBAL_RESOURCES
+                    .dirty_bytes
+                    .fetch_add(delta, AtomicOrdering::Relaxed);
+                old + delta
+            }
+            Ordering::Less => {
+                let delta = self.dirty_bytes - size;
+                let old = GLOBAL_RESOURCES
+                    .dirty_bytes
+                    .fetch_sub(delta, AtomicOrdering::Relaxed);
+                old - delta
+            }
+        };
+
+        // This is a sloppy update: concurrent updates to the counter will race, and the exact
+        // value of the metric might not be the exact latest value of GLOBAL_RESOURCES::dirty_bytes.
+        // That's okay: as long as the metric contains some recent value, it doesn't have to always
+        // be literally the last update.
+        TIMELINE_EPHEMERAL_BYTES.set(new_global_dirty_bytes);
+
+        self.dirty_bytes = size;
+
+        let max_dirty_bytes = GLOBAL_RESOURCES
+            .max_dirty_bytes
+            .load(AtomicOrdering::Relaxed);
+        if max_dirty_bytes > 0 && new_global_dirty_bytes > max_dirty_bytes {
+            // Set the layer file limit to the average layer size: this implies that all above-average
+            // sized layers will be elegible for freezing.  They will be frozen in the order they
+            // next enter publish_size.
+            Some(
+                new_global_dirty_bytes
+                    / GLOBAL_RESOURCES.dirty_layers.load(AtomicOrdering::Relaxed) as u64,
+            )
+        } else {
+            None
+        }
+    }
+
+    // Call publish_size if the input size differs from last published size by more than
+    // the drift limit
+    fn maybe_publish_size(&mut self, size: u64) {
+        let publish = match size.cmp(&self.dirty_bytes) {
+            Ordering::Equal => false,
+            Ordering::Greater => size - self.dirty_bytes > Self::MAX_SIZE_DRIFT,
+            Ordering::Less => self.dirty_bytes - size > Self::MAX_SIZE_DRIFT,
+        };
+
+        if publish {
+            self.publish_size(size);
+        }
+    }
+}
+
+impl Drop for GlobalResourceUnits {
+    fn drop(&mut self) {
+        GLOBAL_RESOURCES
+            .dirty_layers
+            .fetch_sub(1, AtomicOrdering::Relaxed);
+
+        // Subtract our contribution to the global total dirty bytes
+        self.publish_size(0);
+    }
+}
+
+pub(crate) static GLOBAL_RESOURCES: GlobalResources = GlobalResources {
+    max_dirty_bytes: AtomicU64::new(0),
+    dirty_bytes: AtomicU64::new(0),
+    dirty_layers: AtomicUsize::new(0),
+};
+
 impl InMemoryLayer {
     pub(crate) fn get_timeline_id(&self) -> TimelineId {
         self.timeline_id
@@ -93,6 +214,10 @@ impl InMemoryLayer {
         }
     }
 
+    pub(crate) fn try_len(&self) -> Option<u64> {
+        self.inner.try_read().map(|i| i.file.len()).ok()
+    }
+
     pub(crate) fn assert_writable(&self) {
         assert!(self.end_lsn.get().is_none());
     }
@@ -328,6 +453,7 @@ impl InMemoryLayer {
             inner: RwLock::new(InMemoryLayerInner {
                 index: HashMap::new(),
                 file,
+                resource_units: GlobalResourceUnits::new(),
             }),
         })
     }
@@ -378,9 +504,18 @@ impl InMemoryLayer {
             warn!("Key {} at {} already exists", key, lsn);
         }
 
+        let size = locked_inner.file.len();
+        locked_inner.resource_units.maybe_publish_size(size);
+
         Ok(())
     }
 
+    pub(crate) async fn tick(&self) -> Option<u64> {
+        let mut inner = self.inner.write().await;
+        let size = inner.file.len();
+        inner.resource_units.publish_size(size)
+    }
+
     pub(crate) async fn put_tombstones(&self, _key_ranges: &[(Range<Key>, Lsn)]) -> Result<()> {
         // TODO: Currently, we just leak the storage for any deleted keys
         Ok(())

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

@@ -0,0 +1,119 @@
+//! failpoints for unit tests, implying `#[cfg(test)]`.
+//!
+//! These are not accessible over http.
+
+use super::*;
+
+impl Layer {
+    /// Enable a failpoint from a unit test.
+    pub(super) fn enable_failpoint(&self, failpoint: Failpoint) {
+        self.0.failpoints.lock().unwrap().push(failpoint);
+    }
+}
+
+impl LayerInner {
+    /// Query if this failpoint is enabled, as in, arrive at a failpoint.
+    ///
+    /// Calls to this method need to be `#[cfg(test)]` guarded.
+    pub(super) async fn failpoint(&self, kind: FailpointKind) -> Result<(), FailpointHit> {
+        let fut = {
+            let mut fps = self.failpoints.lock().unwrap();
+            // find the *last* failpoint for cases in which we need to use multiple for the same
+            // thing (two blocked evictions)
+            let fp = fps.iter_mut().rfind(|x| x.kind() == kind);
+
+            let Some(fp) = fp else {
+                return Ok(());
+            };
+
+            fp.hit()
+        };
+
+        fut.await
+    }
+}
+
+#[derive(Debug, PartialEq, Eq)]
+pub(crate) enum FailpointKind {
+    /// Failpoint acts as an accurate cancelled by drop here; see the only site of use.
+    AfterDeterminingLayerNeedsNoDownload,
+    /// Failpoint for stalling eviction starting
+    WaitBeforeStartingEvicting,
+    /// Failpoint hit in the spawned task
+    WaitBeforeDownloading,
+}
+
+pub(crate) enum Failpoint {
+    AfterDeterminingLayerNeedsNoDownload,
+    WaitBeforeStartingEvicting(
+        Option<utils::completion::Completion>,
+        utils::completion::Barrier,
+    ),
+    WaitBeforeDownloading(
+        Option<utils::completion::Completion>,
+        utils::completion::Barrier,
+    ),
+}
+
+impl Failpoint {
+    fn kind(&self) -> FailpointKind {
+        match self {
+            Failpoint::AfterDeterminingLayerNeedsNoDownload => {
+                FailpointKind::AfterDeterminingLayerNeedsNoDownload
+            }
+            Failpoint::WaitBeforeStartingEvicting(..) => FailpointKind::WaitBeforeStartingEvicting,
+            Failpoint::WaitBeforeDownloading(..) => FailpointKind::WaitBeforeDownloading,
+        }
+    }
+
+    fn hit(&mut self) -> impl std::future::Future<Output = Result<(), FailpointHit>> + 'static {
+        use futures::future::FutureExt;
+
+        // use boxed futures to avoid Either hurdles
+        match self {
+            Failpoint::AfterDeterminingLayerNeedsNoDownload => {
+                let kind = self.kind();
+
+                async move { Err(FailpointHit(kind)) }.boxed()
+            }
+            Failpoint::WaitBeforeStartingEvicting(arrival, b)
+            | Failpoint::WaitBeforeDownloading(arrival, b) => {
+                // first one signals arrival
+                drop(arrival.take());
+
+                let b = b.clone();
+
+                async move {
+                    tracing::trace!("waiting on a failpoint barrier");
+                    b.wait().await;
+                    tracing::trace!("done waiting on a failpoint barrier");
+                    Ok(())
+                }
+                .boxed()
+            }
+        }
+    }
+}
+
+impl std::fmt::Display for FailpointKind {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        std::fmt::Debug::fmt(self, f)
+    }
+}
+
+#[derive(Debug)]
+pub(crate) struct FailpointHit(FailpointKind);
+
+impl std::fmt::Display for FailpointHit {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        std::fmt::Debug::fmt(self, f)
+    }
+}
+
+impl std::error::Error for FailpointHit {}
+
+impl From<FailpointHit> for DownloadError {
+    fn from(value: FailpointHit) -> Self {
+        DownloadError::Failpoint(value.0)
+    }
+}

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

@@ -1,14 +1,13 @@
-use futures::StreamExt;
 use pageserver_api::key::CONTROLFILE_KEY;
 use tokio::task::JoinSet;
-use tracing::Instrument;
 use utils::{
     completion::{self, Completion},
     id::TimelineId,
 };
 
+use super::failpoints::{Failpoint, FailpointKind};
 use super::*;
-use crate::{context::DownloadBehavior, task_mgr::BACKGROUND_RUNTIME};
+use crate::context::DownloadBehavior;
 use crate::{task_mgr::TaskKind, tenant::harness::TenantHarness};
 
 /// Used in tests to advance a future to wanted await point, and not futher.
@@ -21,7 +20,7 @@ const FOREVER: std::time::Duration = std::time::Duration::from_secs(ADVANCE.as_s
 /// Demonstrate the API and resident -> evicted -> resident -> deleted transitions.
 #[tokio::test]
 async fn smoke_test() {
-    let handle = BACKGROUND_RUNTIME.handle();
+    let handle = tokio::runtime::Handle::current();
 
     let h = TenantHarness::create("smoke_test").unwrap();
     let span = h.span();
@@ -38,7 +37,7 @@ async fn smoke_test() {
     let layer = {
         let mut layers = {
             let layers = timeline.layers.read().await;
-            layers.resident_layers().collect::<Vec<_>>().await
+            layers.likely_resident_layers().collect::<Vec<_>>()
         };
 
         assert_eq!(layers.len(), 1);
@@ -88,7 +87,7 @@ async fn smoke_test() {
     //
     // ZERO for timeout does not work reliably, so first take up all spawn_blocking slots to
     // artificially slow it down.
-    let helper = SpawnBlockingPoolHelper::consume_all_spawn_blocking_threads(handle).await;
+    let helper = SpawnBlockingPoolHelper::consume_all_spawn_blocking_threads(&handle).await;
 
     match layer
         .evict_and_wait(std::time::Duration::ZERO)
@@ -99,7 +98,7 @@ async fn smoke_test() {
             // expected, but note that the eviction is "still ongoing"
             helper.release().await;
             // exhaust spawn_blocking pool to ensure it is now complete
-            SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads(handle)
+            SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads(&handle)
                 .await;
         }
         other => unreachable!("{other:?}"),
@@ -108,7 +107,7 @@ async fn smoke_test() {
     // only way to query if a layer is resident is to acquire a ResidentLayer instance.
     // Layer::keep_resident never downloads, but it might initialize if the layer file is found
     // downloaded locally.
-    let none = layer.keep_resident().await.unwrap();
+    let none = layer.keep_resident().await;
     assert!(
         none.is_none(),
         "Expected none, because eviction removed the local file, found: {none:?}"
@@ -167,6 +166,7 @@ async fn smoke_test() {
     rtc.wait_completion().await.unwrap();
 
     assert_eq!(rtc.get_remote_physical_size(), 0);
+    assert_eq!(0, LAYER_IMPL_METRICS.inits_cancelled.get())
 }
 
 /// This test demonstrates a previous hang when a eviction and deletion were requested at the same
@@ -174,7 +174,7 @@ async fn smoke_test() {
 #[tokio::test(start_paused = true)]
 async fn evict_and_wait_on_wanted_deleted() {
     // this is the runtime on which Layer spawns the blocking tasks on
-    let handle = BACKGROUND_RUNTIME.handle();
+    let handle = tokio::runtime::Handle::current();
 
     let h = TenantHarness::create("evict_and_wait_on_wanted_deleted").unwrap();
     utils::logging::replace_panic_hook_with_tracing_panic_hook().forget();
@@ -188,7 +188,7 @@ async fn evict_and_wait_on_wanted_deleted() {
     let layer = {
         let mut layers = {
             let layers = timeline.layers.read().await;
-            layers.resident_layers().collect::<Vec<_>>().await
+            layers.likely_resident_layers().collect::<Vec<_>>()
         };
 
         assert_eq!(layers.len(), 1);
@@ -213,11 +213,11 @@ async fn evict_and_wait_on_wanted_deleted() {
         drop(resident);
 
         // make sure the eviction task gets to run
-        SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads(handle).await;
+        SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads(&handle).await;
 
         let resident = layer.keep_resident().await;
         assert!(
-            matches!(resident, Ok(None)),
+            resident.is_none(),
             "keep_resident should not have re-initialized: {resident:?}"
         );
 
@@ -235,24 +235,332 @@ async fn evict_and_wait_on_wanted_deleted() {
         layers.finish_gc_timeline(&[layer]);
     }
 
-    SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads(handle).await;
+    SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads(&handle).await;
 
     assert_eq!(1, LAYER_IMPL_METRICS.started_deletes.get());
     assert_eq!(1, LAYER_IMPL_METRICS.completed_deletes.get());
     assert_eq!(1, LAYER_IMPL_METRICS.started_evictions.get());
     assert_eq!(1, LAYER_IMPL_METRICS.completed_evictions.get());
+    assert_eq!(0, LAYER_IMPL_METRICS.inits_cancelled.get())
 }
 
-/// This test shows that ensures we are able to read the layer while the layer eviction has been
-/// started but not completed due to spawn_blocking pool being blocked.
-///
-/// Here `Layer::keep_resident` is used to "simulate" reads, because it cannot download.
-#[tokio::test(start_paused = true)]
-async fn residency_check_while_evict_and_wait_on_clogged_spawn_blocking() {
-    // this is the runtime on which Layer spawns the blocking tasks on
-    let handle = BACKGROUND_RUNTIME.handle();
-    let h = TenantHarness::create("residency_check_while_evict_and_wait_on_clogged_spawn_blocking")
+/// This test ensures we are able to read the layer while the layer eviction has been
+/// started but not completed.
+#[test]
+fn read_wins_pending_eviction() {
+    let rt = tokio::runtime::Builder::new_current_thread()
+        .max_blocking_threads(1)
+        .enable_all()
+        .start_paused(true)
+        .build()
         .unwrap();
+
+    rt.block_on(async move {
+        // this is the runtime on which Layer spawns the blocking tasks on
+        let handle = tokio::runtime::Handle::current();
+        let h = TenantHarness::create("read_wins_pending_eviction").unwrap();
+        let (tenant, ctx) = h.load().await;
+        let span = h.span();
+        let download_span = span.in_scope(|| tracing::info_span!("downloading", timeline_id = 1));
+
+        let timeline = tenant
+            .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
+            .await
+            .unwrap();
+
+        let layer = {
+            let mut layers = {
+                let layers = timeline.layers.read().await;
+                layers.likely_resident_layers().collect::<Vec<_>>()
+            };
+
+            assert_eq!(layers.len(), 1);
+
+            layers.swap_remove(0)
+        };
+
+        // setup done
+
+        let resident = layer.keep_resident().await.unwrap();
+
+        let mut evict_and_wait = std::pin::pin!(layer.evict_and_wait(FOREVER));
+
+        // drive the future to await on the status channel
+        tokio::time::timeout(ADVANCE, &mut evict_and_wait)
+            .await
+            .expect_err("should had been a timeout since we are holding the layer resident");
+        assert_eq!(1, LAYER_IMPL_METRICS.started_evictions.get());
+
+        let (completion, barrier) = utils::completion::channel();
+        let (arrival, arrived_at_barrier) = utils::completion::channel();
+        layer.enable_failpoint(Failpoint::WaitBeforeStartingEvicting(
+            Some(arrival),
+            barrier,
+        ));
+
+        // now the eviction cannot proceed because the threads are consumed while completion exists
+        drop(resident);
+        arrived_at_barrier.wait().await;
+        assert!(!layer.is_likely_resident());
+
+        // because no actual eviction happened, we get to just reinitialize the DownloadedLayer
+        layer
+            .0
+            .get_or_maybe_download(false, None)
+            .instrument(download_span)
+            .await
+            .expect("should had reinitialized without downloading");
+
+        assert!(layer.is_likely_resident());
+
+        // reinitialization notifies of new resident status, which should error out all evict_and_wait
+        let e = tokio::time::timeout(ADVANCE, &mut evict_and_wait)
+            .await
+            .expect("no timeout, because get_or_maybe_download re-initialized")
+            .expect_err("eviction should not have succeeded because re-initialized");
+
+        // works as intended: evictions lose to "downloads"
+        assert!(matches!(e, EvictionError::Downloaded), "{e:?}");
+        assert_eq!(0, LAYER_IMPL_METRICS.completed_evictions.get());
+
+        // this is not wrong: the eviction is technically still "on the way" as it's still queued
+        // because of a failpoint
+        assert_eq!(
+            0,
+            LAYER_IMPL_METRICS
+                .cancelled_evictions
+                .values()
+                .map(|ctr| ctr.get())
+                .sum::<u64>()
+        );
+
+        drop(completion);
+
+        tokio::time::sleep(ADVANCE).await;
+        SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads0(&handle, 1)
+            .await;
+
+        assert_eq!(0, LAYER_IMPL_METRICS.completed_evictions.get());
+
+        // now we finally can observe the original eviction failing
+        // it would had been possible to observe it earlier, but here it is guaranteed to have
+        // happened.
+        assert_eq!(
+            1,
+            LAYER_IMPL_METRICS
+                .cancelled_evictions
+                .values()
+                .map(|ctr| ctr.get())
+                .sum::<u64>()
+        );
+
+        assert_eq!(
+            1,
+            LAYER_IMPL_METRICS.cancelled_evictions[EvictionCancelled::AlreadyReinitialized].get()
+        );
+
+        assert_eq!(0, LAYER_IMPL_METRICS.inits_cancelled.get())
+    });
+}
+
+/// Use failpoint to delay an eviction starting to get a VersionCheckFailed.
+#[test]
+fn multiple_pending_evictions_in_order() {
+    let name = "multiple_pending_evictions_in_order";
+    let in_order = true;
+    multiple_pending_evictions_scenario(name, in_order);
+}
+
+/// Use failpoint to reorder later eviction before first to get a UnexpectedEvictedState.
+#[test]
+fn multiple_pending_evictions_out_of_order() {
+    let name = "multiple_pending_evictions_out_of_order";
+    let in_order = false;
+    multiple_pending_evictions_scenario(name, in_order);
+}
+
+fn multiple_pending_evictions_scenario(name: &'static str, in_order: bool) {
+    let rt = tokio::runtime::Builder::new_current_thread()
+        .max_blocking_threads(1)
+        .enable_all()
+        .start_paused(true)
+        .build()
+        .unwrap();
+
+    rt.block_on(async move {
+        // this is the runtime on which Layer spawns the blocking tasks on
+        let handle = tokio::runtime::Handle::current();
+        let h = TenantHarness::create(name).unwrap();
+        let (tenant, ctx) = h.load().await;
+        let span = h.span();
+        let download_span = span.in_scope(|| tracing::info_span!("downloading", timeline_id = 1));
+
+        let timeline = tenant
+            .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
+            .await
+            .unwrap();
+
+        let layer = {
+            let mut layers = {
+                let layers = timeline.layers.read().await;
+                layers.likely_resident_layers().collect::<Vec<_>>()
+            };
+
+            assert_eq!(layers.len(), 1);
+
+            layers.swap_remove(0)
+        };
+
+        // setup done
+
+        let resident = layer.keep_resident().await.unwrap();
+
+        let mut evict_and_wait = std::pin::pin!(layer.evict_and_wait(FOREVER));
+
+        // drive the future to await on the status channel
+        tokio::time::timeout(ADVANCE, &mut evict_and_wait)
+            .await
+            .expect_err("should had been a timeout since we are holding the layer resident");
+        assert_eq!(1, LAYER_IMPL_METRICS.started_evictions.get());
+
+        let (completion1, barrier) = utils::completion::channel();
+        let mut completion1 = Some(completion1);
+        let (arrival, arrived_at_barrier) = utils::completion::channel();
+        layer.enable_failpoint(Failpoint::WaitBeforeStartingEvicting(
+            Some(arrival),
+            barrier,
+        ));
+
+        // now the eviction cannot proceed because we are simulating arbitrary long delay for the
+        // eviction task start.
+        drop(resident);
+        assert!(!layer.is_likely_resident());
+
+        arrived_at_barrier.wait().await;
+
+        // because no actual eviction happened, we get to just reinitialize the DownloadedLayer
+        layer
+            .0
+            .get_or_maybe_download(false, None)
+            .instrument(download_span)
+            .await
+            .expect("should had reinitialized without downloading");
+
+        assert!(layer.is_likely_resident());
+
+        // reinitialization notifies of new resident status, which should error out all evict_and_wait
+        let e = tokio::time::timeout(ADVANCE, &mut evict_and_wait)
+            .await
+            .expect("no timeout, because get_or_maybe_download re-initialized")
+            .expect_err("eviction should not have succeeded because re-initialized");
+
+        // works as intended: evictions lose to "downloads"
+        assert!(matches!(e, EvictionError::Downloaded), "{e:?}");
+        assert_eq!(0, LAYER_IMPL_METRICS.completed_evictions.get());
+
+        // this is not wrong: the eviction is technically still "on the way" as it's still queued
+        // because of a failpoint
+        assert_eq!(
+            0,
+            LAYER_IMPL_METRICS
+                .cancelled_evictions
+                .values()
+                .map(|ctr| ctr.get())
+                .sum::<u64>()
+        );
+
+        assert_eq!(0, LAYER_IMPL_METRICS.completed_evictions.get());
+
+        // configure another failpoint for the second eviction -- evictions are per initialization,
+        // so now that we've reinitialized the inner, we get to run two of them at the same time.
+        let (completion2, barrier) = utils::completion::channel();
+        let (arrival, arrived_at_barrier) = utils::completion::channel();
+        layer.enable_failpoint(Failpoint::WaitBeforeStartingEvicting(
+            Some(arrival),
+            barrier,
+        ));
+
+        let mut second_eviction = std::pin::pin!(layer.evict_and_wait(FOREVER));
+
+        // advance to the wait on the queue
+        tokio::time::timeout(ADVANCE, &mut second_eviction)
+            .await
+            .expect_err("timeout because failpoint is blocking");
+
+        arrived_at_barrier.wait().await;
+
+        assert_eq!(2, LAYER_IMPL_METRICS.started_evictions.get());
+
+        let mut release_earlier_eviction = |expected_reason| {
+            assert_eq!(
+                0,
+                LAYER_IMPL_METRICS.cancelled_evictions[expected_reason].get(),
+            );
+
+            drop(completion1.take().unwrap());
+
+            let handle = &handle;
+
+            async move {
+                tokio::time::sleep(ADVANCE).await;
+                SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads0(
+                    handle, 1,
+                )
+                .await;
+
+                assert_eq!(
+                    1,
+                    LAYER_IMPL_METRICS.cancelled_evictions[expected_reason].get(),
+                );
+            }
+        };
+
+        if in_order {
+            release_earlier_eviction(EvictionCancelled::VersionCheckFailed).await;
+        }
+
+        // release the later eviction which is for the current version
+        drop(completion2);
+        tokio::time::sleep(ADVANCE).await;
+        SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads0(&handle, 1)
+            .await;
+
+        if !in_order {
+            release_earlier_eviction(EvictionCancelled::UnexpectedEvictedState).await;
+        }
+
+        tokio::time::timeout(ADVANCE, &mut second_eviction)
+            .await
+            .expect("eviction goes through now that spawn_blocking is unclogged")
+            .expect("eviction should succeed, because version matches");
+
+        assert_eq!(1, LAYER_IMPL_METRICS.completed_evictions.get());
+
+        // ensure the cancelled are unchanged
+        assert_eq!(
+            1,
+            LAYER_IMPL_METRICS
+                .cancelled_evictions
+                .values()
+                .map(|ctr| ctr.get())
+                .sum::<u64>()
+        );
+
+        assert_eq!(0, LAYER_IMPL_METRICS.inits_cancelled.get())
+    });
+}
+
+/// The test ensures with a failpoint that a pending eviction is not cancelled by what is currently
+/// a `Layer::keep_resident` call.
+///
+/// This matters because cancelling the eviction would leave us in a state where the file is on
+/// disk but the layer internal state says it has not been initialized. Futhermore, it allows us to
+/// have non-repairing `Layer::is_likely_resident`.
+#[tokio::test(start_paused = true)]
+async fn cancelled_get_or_maybe_download_does_not_cancel_eviction() {
+    let handle = tokio::runtime::Handle::current();
+    let h =
+        TenantHarness::create("cancelled_get_or_maybe_download_does_not_cancel_eviction").unwrap();
     let (tenant, ctx) = h.load().await;
 
     let timeline = tenant
@@ -263,7 +571,7 @@ async fn residency_check_while_evict_and_wait_on_clogged_spawn_blocking() {
     let layer = {
         let mut layers = {
             let layers = timeline.layers.read().await;
-            layers.resident_layers().collect::<Vec<_>>().await
+            layers.likely_resident_layers().collect::<Vec<_>>()
         };
 
         assert_eq!(layers.len(), 1);
@@ -271,90 +579,154 @@ async fn residency_check_while_evict_and_wait_on_clogged_spawn_blocking() {
         layers.swap_remove(0)
     };
 
-    // setup done
+    // this failpoint will simulate the `get_or_maybe_download` becoming cancelled (by returning an
+    // Err) at the right time as in "during" the `LayerInner::needs_download`.
+    layer.enable_failpoint(Failpoint::AfterDeterminingLayerNeedsNoDownload);
 
-    let resident = layer.keep_resident().await.unwrap();
+    let (completion, barrier) = utils::completion::channel();
+    let (arrival, arrived_at_barrier) = utils::completion::channel();
+
+    layer.enable_failpoint(Failpoint::WaitBeforeStartingEvicting(
+        Some(arrival),
+        barrier,
+    ));
+
+    tokio::time::timeout(ADVANCE, layer.evict_and_wait(FOREVER))
+        .await
+        .expect_err("should had advanced to waiting on channel");
+
+    arrived_at_barrier.wait().await;
+
+    // simulate a cancelled read which is cancelled before it gets to re-initialize
+    let e = layer
+        .0
+        .get_or_maybe_download(false, None)
+        .await
+        .unwrap_err();
+    assert!(
+        matches!(
+            e,
+            DownloadError::Failpoint(FailpointKind::AfterDeterminingLayerNeedsNoDownload)
+        ),
+        "{e:?}"
+    );
+
+    assert!(
+        layer.0.needs_download().await.unwrap().is_none(),
+        "file is still on disk"
+    );
+
+    // release the eviction task
+    drop(completion);
+    tokio::time::sleep(ADVANCE).await;
+    SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads(&handle).await;
+
+    // failpoint is still enabled, but it is not hit
+    let e = layer
+        .0
+        .get_or_maybe_download(false, None)
+        .await
+        .unwrap_err();
+    assert!(matches!(e, DownloadError::DownloadRequired), "{e:?}");
+
+    // failpoint is not counted as cancellation either
+    assert_eq!(0, LAYER_IMPL_METRICS.inits_cancelled.get())
+}
+
+#[tokio::test(start_paused = true)]
+async fn evict_and_wait_does_not_wait_for_download() {
+    // let handle = tokio::runtime::Handle::current();
+    let h = TenantHarness::create("evict_and_wait_does_not_wait_for_download").unwrap();
+    let (tenant, ctx) = h.load().await;
+    let span = h.span();
+    let download_span = span.in_scope(|| tracing::info_span!("downloading", timeline_id = 1));
+
+    let timeline = tenant
+        .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
+        .await
+        .unwrap();
+
+    let layer = {
+        let mut layers = {
+            let layers = timeline.layers.read().await;
+            layers.likely_resident_layers().collect::<Vec<_>>()
+        };
+
+        assert_eq!(layers.len(), 1);
+
+        layers.swap_remove(0)
+    };
+
+    // kind of forced setup: start an eviction but do not allow it progress until we are
+    // downloading
+    let (eviction_can_continue, barrier) = utils::completion::channel();
+    let (arrival, eviction_arrived) = utils::completion::channel();
+    layer.enable_failpoint(Failpoint::WaitBeforeStartingEvicting(
+        Some(arrival),
+        barrier,
+    ));
 
     let mut evict_and_wait = std::pin::pin!(layer.evict_and_wait(FOREVER));
 
-    // drive the future to await on the status channel
+    // use this once-awaited other_evict to synchronize with the eviction
+    let other_evict = layer.evict_and_wait(FOREVER);
+
     tokio::time::timeout(ADVANCE, &mut evict_and_wait)
         .await
-        .expect_err("should had been a timeout since we are holding the layer resident");
-    assert_eq!(1, LAYER_IMPL_METRICS.started_evictions.get());
+        .expect_err("should had advanced");
+    eviction_arrived.wait().await;
+    drop(eviction_can_continue);
+    other_evict.await.unwrap();
 
-    // clog up BACKGROUND_RUNTIME spawn_blocking
-    let helper = SpawnBlockingPoolHelper::consume_all_spawn_blocking_threads(handle).await;
+    // now the layer is evicted, and the "evict_and_wait" is waiting on the receiver
+    assert!(!layer.is_likely_resident());
 
-    // now the eviction cannot proceed because the threads are consumed while completion exists
-    drop(resident);
+    // following new evict_and_wait will fail until we've completed the download
+    let e = layer.evict_and_wait(FOREVER).await.unwrap_err();
+    assert!(matches!(e, EvictionError::NotFound), "{e:?}");
 
-    // because no actual eviction happened, we get to just reinitialize the DownloadedLayer
-    layer
-        .keep_resident()
-        .await
-        .expect("keep_resident should had reinitialized without downloading")
-        .expect("ResidentLayer");
+    let (download_can_continue, barrier) = utils::completion::channel();
+    let (arrival, _download_arrived) = utils::completion::channel();
+    layer.enable_failpoint(Failpoint::WaitBeforeDownloading(Some(arrival), barrier));
 
-    // because the keep_resident check alters wanted evicted without sending a message, we will
-    // never get completed
-    let e = tokio::time::timeout(ADVANCE, &mut evict_and_wait)
-        .await
-        .expect("no timeout, because keep_resident re-initialized")
-        .expect_err("eviction should not have succeeded because re-initialized");
+    let mut download = std::pin::pin!(layer
+        .0
+        .get_or_maybe_download(true, None)
+        .instrument(download_span));
 
-    // works as intended: evictions lose to "downloads"
-    assert!(matches!(e, EvictionError::Downloaded), "{e:?}");
-    assert_eq!(0, LAYER_IMPL_METRICS.completed_evictions.get());
-
-    // this is not wrong: the eviction is technically still "on the way" as it's still queued
-    // because spawn_blocking is clogged up
-    assert_eq!(
-        0,
-        LAYER_IMPL_METRICS
-            .cancelled_evictions
-            .values()
-            .map(|ctr| ctr.get())
-            .sum::<u64>()
+    assert!(
+        !layer.is_likely_resident(),
+        "during download layer is evicted"
     );
 
-    let mut second_eviction = std::pin::pin!(layer.evict_and_wait(FOREVER));
-
-    // advance to the wait on the queue
-    tokio::time::timeout(ADVANCE, &mut second_eviction)
+    tokio::time::timeout(ADVANCE, &mut download)
         .await
-        .expect_err("timeout because spawn_blocking is clogged");
+        .expect_err("should had timed out because of failpoint");
 
-    // in this case we don't leak started evictions, but I think there is still a chance of that
-    // happening, because we could have upgrades race multiple evictions while only one of them
-    // happens?
-    assert_eq!(2, LAYER_IMPL_METRICS.started_evictions.get());
+    // now we finally get to continue, and because the latest state is downloading, we deduce that
+    // original eviction succeeded
+    evict_and_wait.await.unwrap();
 
-    helper.release().await;
+    // however a new evict_and_wait will fail
+    let e = layer.evict_and_wait(FOREVER).await.unwrap_err();
+    assert!(matches!(e, EvictionError::NotFound), "{e:?}");
 
-    // the second_eviction gets to run here
-    //
-    // synchronize to be *strictly* after the second_eviction spawn_blocking run
-    SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads(handle).await;
+    assert!(!layer.is_likely_resident());
 
-    tokio::time::timeout(ADVANCE, &mut second_eviction)
-        .await
-        .expect("eviction goes through now that spawn_blocking is unclogged")
-        .expect("eviction should succeed, because version matches");
+    drop(download_can_continue);
+    download.await.expect("download should had succeeded");
+    assert!(layer.is_likely_resident());
 
-    assert_eq!(1, LAYER_IMPL_METRICS.completed_evictions.get());
+    // only now can we evict
+    layer.evict_and_wait(FOREVER).await.unwrap();
+}
 
-    // now we finally can observe the original spawn_blocking failing
-    // it would had been possible to observe it earlier, but here it is guaranteed to have
-    // happened.
-    assert_eq!(
-        1,
-        LAYER_IMPL_METRICS
-            .cancelled_evictions
-            .values()
-            .map(|ctr| ctr.get())
-            .sum::<u64>()
-    );
+#[test]
+fn layer_size() {
+    assert_eq!(std::mem::size_of::<LayerAccessStats>(), 2040);
+    assert_eq!(std::mem::size_of::<PersistentLayerDesc>(), 104);
+    assert_eq!(std::mem::size_of::<LayerInner>(), 2328);
+    // it also has the utf8 path
 }
 
 struct SpawnBlockingPoolHelper {
@@ -371,31 +743,41 @@ impl SpawnBlockingPoolHelper {
     ///
     /// This should be no issue nowdays, because nextest runs each test in it's own process.
     async fn consume_all_spawn_blocking_threads(handle: &tokio::runtime::Handle) -> Self {
-        let (completion, barrier) = completion::channel();
-        let (tx, mut rx) = tokio::sync::mpsc::channel(8);
+        let default_max_blocking_threads = 512;
 
-        let assumed_max_blocking_threads = 512;
+        Self::consume_all_spawn_blocking_threads0(handle, default_max_blocking_threads).await
+    }
+
+    async fn consume_all_spawn_blocking_threads0(
+        handle: &tokio::runtime::Handle,
+        threads: usize,
+    ) -> Self {
+        assert_ne!(threads, 0);
+
+        let (completion, barrier) = completion::channel();
+        let (started, starts_completed) = completion::channel();
 
         let mut blocking_tasks = JoinSet::new();
 
-        for _ in 0..assumed_max_blocking_threads {
+        for _ in 0..threads {
             let barrier = barrier.clone();
-            let tx = tx.clone();
+            let started = started.clone();
             blocking_tasks.spawn_blocking_on(
                 move || {
-                    tx.blocking_send(()).unwrap();
-                    drop(tx);
+                    drop(started);
                     tokio::runtime::Handle::current().block_on(barrier.wait());
                 },
                 handle,
             );
         }
 
+        drop(started);
+
+        starts_completed.wait().await;
+
         drop(barrier);
 
-        for _ in 0..assumed_max_blocking_threads {
-            rx.recv().await.unwrap();
-        }
+        tracing::trace!("consumed all threads");
 
         SpawnBlockingPoolHelper {
             awaited_by_spawn_blocking_tasks: completion,
@@ -415,13 +797,22 @@ impl SpawnBlockingPoolHelper {
         while let Some(res) = blocking_tasks.join_next().await {
             res.expect("none of the tasks should had panicked");
         }
+
+        tracing::trace!("released all threads");
     }
 
     /// In the tests it is used as an easy way of making sure something scheduled on the target
     /// runtimes `spawn_blocking` has completed, because it must've been scheduled and completed
     /// before our tasks have a chance to schedule and complete.
     async fn consume_and_release_all_of_spawn_blocking_threads(handle: &tokio::runtime::Handle) {
-        Self::consume_all_spawn_blocking_threads(handle)
+        Self::consume_and_release_all_of_spawn_blocking_threads0(handle, 512).await
+    }
+
+    async fn consume_and_release_all_of_spawn_blocking_threads0(
+        handle: &tokio::runtime::Handle,
+        threads: usize,
+    ) {
+        Self::consume_all_spawn_blocking_threads0(handle, threads)
             .await
             .release()
             .await
@@ -435,7 +826,7 @@ fn spawn_blocking_pool_helper_actually_works() {
     // because the amount is not configurable for our helper, expect the same amount as
     // BACKGROUND_RUNTIME using the tokio defaults would have.
     let rt = tokio::runtime::Builder::new_current_thread()
-        .max_blocking_threads(512)
+        .max_blocking_threads(1)
         .enable_all()
         .build()
         .unwrap();
@@ -445,7 +836,8 @@ fn spawn_blocking_pool_helper_actually_works() {
     rt.block_on(async move {
         // this will not return until all threads are spun up and actually executing the code
         // waiting on `consumed` to be `SpawnBlockingPoolHelper::release`'d.
-        let consumed = SpawnBlockingPoolHelper::consume_all_spawn_blocking_threads(handle).await;
+        let consumed =
+            SpawnBlockingPoolHelper::consume_all_spawn_blocking_threads0(handle, 1).await;
 
         println!("consumed");
 

pageserver/src/tenant/timeline.rs

@@ -13,14 +13,13 @@ use bytes::Bytes;
 use camino::Utf8Path;
 use enumset::EnumSet;
 use fail::fail_point;
-use futures::stream::StreamExt;
 use once_cell::sync::Lazy;
 use pageserver_api::{
     key::AUX_FILES_KEY,
     keyspace::KeySpaceAccum,
     models::{
         CompactionAlgorithm, DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest,
-        EvictionPolicy, LayerMapInfo, TimelineState,
+        EvictionPolicy, InMemoryLayerInfo, LayerMapInfo, TimelineState,
     },
     reltag::BlockNumber,
     shard::{ShardIdentity, TenantShardId},
@@ -37,6 +36,7 @@ use tracing::*;
 use utils::{
     bin_ser::BeSer,
     sync::gate::{Gate, GateGuard},
+    vec_map::VecMap,
 };
 
 use std::ops::{Deref, Range};
@@ -54,6 +54,7 @@ use std::{
     ops::ControlFlow,
 };
 
+use crate::deletion_queue::DeletionQueueClient;
 use crate::tenant::timeline::logical_size::CurrentLogicalSize;
 use crate::tenant::{
     layer_map::{LayerMap, SearchResult},
@@ -64,7 +65,6 @@ use crate::{
     disk_usage_eviction_task::DiskUsageEvictionInfo,
     pgdatadir_mapping::CollectKeySpaceError,
 };
-use crate::{deletion_queue::DeletionQueueClient, tenant::remote_timeline_client::StopError};
 use crate::{
     disk_usage_eviction_task::finite_f32,
     tenant::storage_layer::{
@@ -1142,6 +1142,79 @@ impl Timeline {
         self.flush_frozen_layers_and_wait().await
     }
 
+    /// If there is no writer, and conditions for rolling the latest layer are met, then freeze it.
+    ///
+    /// This is for use in background housekeeping, to provide guarantees of layers closing eventually
+    /// even if there are no ongoing writes to drive that.
+    async fn maybe_freeze_ephemeral_layer(&self) {
+        let Ok(_write_guard) = self.write_lock.try_lock() else {
+            // If the write lock is held, there is an active wal receiver: rolling open layers
+            // is their responsibility while they hold this lock.
+            return;
+        };
+
+        let Ok(layers_guard) = self.layers.try_read() else {
+            // Don't block if the layer lock is busy
+            return;
+        };
+
+        let Some(open_layer) = &layers_guard.layer_map().open_layer else {
+            // No open layer, no work to do.
+            return;
+        };
+
+        let Some(current_size) = open_layer.try_len() else {
+            // Unexpected: since we hold the write guard, nobody else should be writing to this layer, so
+            // read lock to get size should always succeed.
+            tracing::warn!("Lock conflict while reading size of open layer");
+            return;
+        };
+
+        let current_lsn = self.get_last_record_lsn();
+
+        let checkpoint_distance_override = open_layer.tick().await;
+
+        if let Some(size_override) = checkpoint_distance_override {
+            if current_size > size_override {
+                // This is not harmful, but it only happens in relatively rare cases where
+                // time-based checkpoints are not happening fast enough to keep the amount of
+                // ephemeral data within configured limits.  It's a sign of stress on the system.
+                tracing::info!("Early-rolling open layer at size {current_size} (limit {size_override}) due to dirty data pressure");
+            }
+        }
+
+        let checkpoint_distance =
+            checkpoint_distance_override.unwrap_or(self.get_checkpoint_distance());
+
+        if self.should_roll(
+            current_size,
+            current_size,
+            checkpoint_distance,
+            self.get_last_record_lsn(),
+            self.last_freeze_at.load(),
+            *self.last_freeze_ts.read().unwrap(),
+        ) {
+            match open_layer.info() {
+                InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
+                    // We may reach this point if the layer was already frozen by not yet flushed: flushing
+                    // happens asynchronously in the background.
+                    tracing::debug!(
+                        "Not freezing open layer, it's already frozen ({lsn_start}..{lsn_end})"
+                    );
+                }
+                InMemoryLayerInfo::Open { .. } => {
+                    // Upgrade to a write lock and freeze the layer
+                    drop(layers_guard);
+                    let mut layers_guard = self.layers.write().await;
+                    layers_guard
+                        .try_freeze_in_memory_layer(current_lsn, &self.last_freeze_at)
+                        .await;
+                }
+            }
+            self.flush_frozen_layers();
+        }
+    }
+
     /// Outermost timeline compaction operation; downloads needed layers.
     pub(crate) async fn compact(
         self: &Arc<Self>,
@@ -1164,6 +1237,11 @@ impl Timeline {
             (guard, permit)
         };
 
+        // Prior to compaction, check if an open ephemeral layer should be closed: this provides
+        // background enforcement of checkpoint interval if there is no active WAL receiver, to avoid keeping
+        // an ephemeral layer open forever when idle.
+        self.maybe_freeze_ephemeral_layer().await;
+
         // this wait probably never needs any "long time spent" logging, because we already nag if
         // compaction task goes over it's period (20s) which is quite often in production.
         let (_guard, _permit) = tokio::select! {
@@ -1196,6 +1274,7 @@ impl Timeline {
 
     pub(crate) fn activate(
         self: &Arc<Self>,
+        parent: Arc<crate::tenant::Tenant>,
         broker_client: BrokerClientChannel,
         background_jobs_can_start: Option<&completion::Barrier>,
         ctx: &RequestContext,
@@ -1206,7 +1285,7 @@ impl Timeline {
         }
         self.launch_wal_receiver(ctx, broker_client);
         self.set_state(TimelineState::Active);
-        self.launch_eviction_task(background_jobs_can_start);
+        self.launch_eviction_task(parent, background_jobs_can_start);
     }
 
     /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
@@ -1241,11 +1320,7 @@ impl Timeline {
                     // what is problematic is the shutting down of RemoteTimelineClient, because
                     // obviously it does not make sense to stop while we wait for it, but what
                     // about corner cases like s3 suddenly hanging up?
-                    if let Err(e) = client.shutdown().await {
-                        // Non-fatal.  Shutdown is infallible.  Failures to flush just mean that
-                        // we have some extra WAL replay to do next time the timeline starts.
-                        warn!("failed to flush to remote storage: {e:#}");
-                    }
+                    client.shutdown().await;
                 }
             }
             Err(e) => {
@@ -1282,12 +1357,7 @@ impl Timeline {
         // Shut down remote timeline client: this gracefully moves its metadata into its Stopping state in
         // case our caller wants to use that for a deletion
         if let Some(remote_client) = self.remote_client.as_ref() {
-            match remote_client.stop() {
-                Ok(()) => {}
-                Err(StopError::QueueUninitialized) => {
-                    // Shutting down during initialization is legal
-                }
-            }
+            remote_client.stop();
         }
 
         tracing::debug!("Waiting for tasks...");
@@ -1443,6 +1513,53 @@ impl Timeline {
             Err(EvictionError::Timeout) => Ok(Some(false)),
         }
     }
+
+    fn should_roll(
+        &self,
+        layer_size: u64,
+        projected_layer_size: u64,
+        checkpoint_distance: u64,
+        projected_lsn: Lsn,
+        last_freeze_at: Lsn,
+        last_freeze_ts: Instant,
+    ) -> bool {
+        let distance = projected_lsn.widening_sub(last_freeze_at);
+
+        // Rolling the open layer can be triggered by:
+        // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
+        //    the safekeepers need to store.  For sharded tenants, we multiply by shard count to
+        //    account for how writes are distributed across shards: we expect each node to consume
+        //    1/count of the LSN on average.
+        // 2. The size of the currently open layer.
+        // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
+        //    up and suspend activity.
+        if distance >= checkpoint_distance as i128 * self.shard_identity.count.count() as i128 {
+            info!(
+                "Will roll layer at {} with layer size {} due to LSN distance ({})",
+                projected_lsn, layer_size, distance
+            );
+
+            true
+        } else if projected_layer_size >= checkpoint_distance {
+            info!(
+                "Will roll layer at {} with layer size {} due to layer size ({})",
+                projected_lsn, layer_size, projected_layer_size
+            );
+
+            true
+        } else if distance > 0 && last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
+            info!(
+                "Will roll layer at {} with layer size {} due to time since last flush ({:?})",
+                projected_lsn,
+                layer_size,
+                last_freeze_ts.elapsed()
+            );
+
+            true
+        } else {
+            false
+        }
+    }
 }
 
 /// Number of times we will compute partition within a checkpoint distance.
@@ -2442,7 +2559,7 @@ impl Timeline {
 
         let guard = self.layers.read().await;
 
-        let resident = guard.resident_layers().map(|layer| {
+        let resident = guard.likely_resident_layers().map(|layer| {
             let last_activity_ts = layer.access_stats().latest_activity_or_now();
 
             HeatMapLayer::new(
@@ -2452,7 +2569,7 @@ impl Timeline {
             )
         });
 
-        let layers = resident.collect().await;
+        let layers = resident.collect();
 
         Some(HeatMapTimeline::new(self.timeline_id, layers))
     }
@@ -2596,6 +2713,10 @@ impl Timeline {
                     // Get all the data needed to reconstruct the page version from this layer.
                     // But if we have an older cached page image, no need to go past that.
                     let lsn_floor = max(cached_lsn + 1, start_lsn);
+
+                    let open_layer = open_layer.clone();
+                    drop(guard);
+
                     result = match open_layer
                         .get_value_reconstruct_data(
                             key,
@@ -2613,10 +2734,7 @@ impl Timeline {
                     traversal_path.push((
                         result,
                         cont_lsn,
-                        Box::new({
-                            let open_layer = Arc::clone(open_layer);
-                            move || open_layer.traversal_id()
-                        }),
+                        Box::new(move || open_layer.traversal_id()),
                     ));
                     continue 'outer;
                 }
@@ -2626,6 +2744,10 @@ impl Timeline {
                 if cont_lsn > start_lsn {
                     //info!("CHECKING for {} at {} on frozen layer {}", key, cont_lsn, frozen_layer.filename().display());
                     let lsn_floor = max(cached_lsn + 1, start_lsn);
+
+                    let frozen_layer = frozen_layer.clone();
+                    drop(guard);
+
                     result = match frozen_layer
                         .get_value_reconstruct_data(
                             key,
@@ -2643,10 +2765,7 @@ impl Timeline {
                     traversal_path.push((
                         result,
                         cont_lsn,
-                        Box::new({
-                            let frozen_layer = Arc::clone(frozen_layer);
-                            move || frozen_layer.traversal_id()
-                        }),
+                        Box::new(move || frozen_layer.traversal_id()),
                     ));
                     continue 'outer;
                 }
@@ -2654,6 +2773,8 @@ impl Timeline {
 
             if let Some(SearchResult { lsn_floor, layer }) = layers.search(key, cont_lsn) {
                 let layer = guard.get_from_desc(&layer);
+                drop(guard);
+
                 // Get all the data needed to reconstruct the page version from this layer.
                 // But if we have an older cached page image, no need to go past that.
                 let lsn_floor = max(cached_lsn + 1, lsn_floor);
@@ -4302,7 +4423,7 @@ impl Timeline {
         let mut max_layer_size: Option<u64> = None;
 
         let resident_layers = guard
-            .resident_layers()
+            .likely_resident_layers()
             .map(|layer| {
                 let file_size = layer.layer_desc().file_size;
                 max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
@@ -4315,8 +4436,7 @@ impl Timeline {
                     relative_last_activity: finite_f32::FiniteF32::ZERO,
                 }
             })
-            .collect()
-            .await;
+            .collect();
 
         DiskUsageEvictionInfo {
             max_layer_size,
@@ -4461,49 +4581,6 @@ impl<'a> TimelineWriter<'a> {
         res
     }
 
-    /// "Tick" the timeline writer: it will roll the open layer if required
-    /// and do nothing else.
-    pub(crate) async fn tick(&mut self) -> anyhow::Result<()> {
-        self.open_layer_if_present().await?;
-
-        let last_record_lsn = self.get_last_record_lsn();
-        let action = self.get_open_layer_action(last_record_lsn, 0);
-        if action == OpenLayerAction::Roll {
-            self.roll_layer(last_record_lsn).await?;
-        }
-
-        Ok(())
-    }
-
-    /// Populate the timeline writer state only if an in-memory layer
-    /// is already open.
-    async fn open_layer_if_present(&mut self) -> anyhow::Result<()> {
-        assert!(self.write_guard.is_none());
-
-        let open_layer = {
-            let guard = self.layers.read().await;
-            let layers = guard.layer_map();
-            match layers.open_layer {
-                Some(ref open_layer) => open_layer.clone(),
-                None => {
-                    return Ok(());
-                }
-            }
-        };
-
-        let initial_size = open_layer.size().await?;
-        let last_freeze_at = self.last_freeze_at.load();
-        let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
-        self.write_guard.replace(TimelineWriterState::new(
-            open_layer,
-            initial_size,
-            last_freeze_at,
-            last_freeze_ts,
-        ));
-
-        Ok(())
-    }
-
     async fn handle_open_layer_action(
         &mut self,
         at: Lsn,
@@ -4575,59 +4652,29 @@ impl<'a> TimelineWriter<'a> {
             return OpenLayerAction::None;
         }
 
-        let distance = lsn.widening_sub(state.cached_last_freeze_at);
-        let proposed_open_layer_size = state.current_size + new_value_size;
-
-        // Rolling the open layer can be triggered by:
-        // 1. The distance from the last LSN we rolled at. This bounds the amount of WAL that
-        //    the safekeepers need to store.  For sharded tenants, we multiply by shard count to
-        //    account for how writes are distributed across shards: we expect each node to consume
-        //    1/count of the LSN on average.
-        // 2. The size of the currently open layer.
-        // 3. The time since the last roll. It helps safekeepers to regard pageserver as caught
-        //    up and suspend activity.
-        if distance
-            >= self.get_checkpoint_distance() as i128 * self.shard_identity.count.count() as i128
-        {
-            info!(
-                "Will roll layer at {} with layer size {} due to LSN distance ({})",
-                lsn, state.current_size, distance
-            );
-
-            OpenLayerAction::Roll
-        } else if proposed_open_layer_size >= self.get_checkpoint_distance() {
-            info!(
-                "Will roll layer at {} with layer size {} due to layer size ({})",
-                lsn, state.current_size, proposed_open_layer_size
-            );
-
-            OpenLayerAction::Roll
-        } else if distance > 0
-            && state.cached_last_freeze_ts.elapsed() >= self.get_checkpoint_timeout()
-        {
-            info!(
-                "Will roll layer at {} with layer size {} due to time since last flush ({:?})",
-                lsn,
-                state.current_size,
-                state.cached_last_freeze_ts.elapsed()
-            );
-
+        if self.tl.should_roll(
+            state.current_size,
+            state.current_size + new_value_size,
+            self.get_checkpoint_distance(),
+            lsn,
+            state.cached_last_freeze_at,
+            state.cached_last_freeze_ts,
+        ) {
             OpenLayerAction::Roll
         } else {
             OpenLayerAction::None
         }
     }
 
-    /// Put a batch keys at the specified Lsns.
+    /// Put a batch of keys at the specified Lsns.
     ///
-    /// The batch should be sorted by Lsn such that it's safe
-    /// to roll the open layer mid batch.
+    /// The batch is sorted by Lsn (enforced by usage of [`utils::vec_map::VecMap`].
     pub(crate) async fn put_batch(
         &mut self,
-        batch: Vec<(Key, Lsn, Value)>,
+        batch: VecMap<Lsn, (Key, Value)>,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
-        for (key, lsn, val) in batch {
+        for (lsn, (key, val)) in batch {
             self.put(key, lsn, &val, ctx).await?
         }
 
@@ -4713,7 +4760,6 @@ mod tests {
             .keep_resident()
             .await
             .expect("no download => no downloading errors")
-            .expect("should had been resident")
             .drop_eviction_guard();
 
         let forever = std::time::Duration::from_secs(120);
@@ -4724,7 +4770,7 @@ mod tests {
         let (first, second) = tokio::join!(first, second);
 
         let res = layer.keep_resident().await;
-        assert!(matches!(res, Ok(None)), "{res:?}");
+        assert!(res.is_none(), "{res:?}");
 
         match (first, second) {
             (Ok(()), Ok(())) => {

pageserver/src/tenant/timeline/delete.rs

@@ -16,9 +16,7 @@ use crate::{
     tenant::{
         debug_assert_current_span_has_tenant_and_timeline_id,
         metadata::TimelineMetadata,
-        remote_timeline_client::{
-            self, PersistIndexPartWithDeletedFlagError, RemoteTimelineClient,
-        },
+        remote_timeline_client::{PersistIndexPartWithDeletedFlagError, RemoteTimelineClient},
         CreateTimelineCause, DeleteTimelineError, Tenant,
     },
 };
@@ -50,19 +48,7 @@ async fn stop_tasks(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
 
     // Prevent new uploads from starting.
     if let Some(remote_client) = timeline.remote_client.as_ref() {
-        let res = remote_client.stop();
-        match res {
-            Ok(()) => {}
-            Err(e) => match e {
-                remote_timeline_client::StopError::QueueUninitialized => {
-                    // This case shouldn't happen currently because the
-                    // load and attach code bails out if _any_ of the timeline fails to fetch its IndexPart.
-                    // That is, before we declare the Tenant as Active.
-                    // But we only allow calls to delete_timeline on Active tenants.
-                    return Err(DeleteTimelineError::Other(anyhow::anyhow!("upload queue is uninitialized, likely the timeline was in Broken state prior to this call because it failed to fetch IndexPart during load or attach, check the logs")));
-                }
-            },
-        }
+        remote_client.stop();
     }
 
     // Stop & wait for the remaining timeline tasks, including upload tasks.

pageserver/src/tenant/timeline/eviction_task.rs

@@ -51,6 +51,7 @@ pub struct EvictionTaskTenantState {
 impl Timeline {
     pub(super) fn launch_eviction_task(
         self: &Arc<Self>,
+        parent: Arc<Tenant>,
         background_tasks_can_start: Option<&completion::Barrier>,
     ) {
         let self_clone = Arc::clone(self);
@@ -72,14 +73,14 @@ impl Timeline {
                     _ = completion::Barrier::maybe_wait(background_tasks_can_start) => {}
                 };
 
-                self_clone.eviction_task(cancel).await;
+                self_clone.eviction_task(parent, cancel).await;
                 Ok(())
             },
         );
     }
 
     #[instrument(skip_all, fields(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id))]
-    async fn eviction_task(self: Arc<Self>, cancel: CancellationToken) {
+    async fn eviction_task(self: Arc<Self>, tenant: Arc<Tenant>, cancel: CancellationToken) {
         use crate::tenant::tasks::random_init_delay;
 
         // acquire the gate guard only once within a useful span
@@ -103,7 +104,7 @@ impl Timeline {
         loop {
             let policy = self.get_eviction_policy();
             let cf = self
-                .eviction_iteration(&policy, &cancel, &guard, &ctx)
+                .eviction_iteration(&tenant, &policy, &cancel, &guard, &ctx)
                 .await;
 
             match cf {
@@ -123,6 +124,7 @@ impl Timeline {
     #[instrument(skip_all, fields(policy_kind = policy.discriminant_str()))]
     async fn eviction_iteration(
         self: &Arc<Self>,
+        tenant: &Tenant,
         policy: &EvictionPolicy,
         cancel: &CancellationToken,
         gate: &GateGuard,
@@ -137,7 +139,7 @@ impl Timeline {
             }
             EvictionPolicy::LayerAccessThreshold(p) => {
                 match self
-                    .eviction_iteration_threshold(p, cancel, gate, ctx)
+                    .eviction_iteration_threshold(tenant, p, cancel, gate, ctx)
                     .await
                 {
                     ControlFlow::Break(()) => return ControlFlow::Break(()),
@@ -146,7 +148,11 @@ impl Timeline {
                 (p.period, p.threshold)
             }
             EvictionPolicy::OnlyImitiate(p) => {
-                if self.imitiate_only(p, cancel, gate, ctx).await.is_break() {
+                if self
+                    .imitiate_only(tenant, p, cancel, gate, ctx)
+                    .await
+                    .is_break()
+                {
                     return ControlFlow::Break(());
                 }
                 (p.period, p.threshold)
@@ -175,6 +181,7 @@ impl Timeline {
 
     async fn eviction_iteration_threshold(
         self: &Arc<Self>,
+        tenant: &Tenant,
         p: &EvictionPolicyLayerAccessThreshold,
         cancel: &CancellationToken,
         gate: &GateGuard,
@@ -193,7 +200,10 @@ impl Timeline {
             _ = self.cancel.cancelled() => return ControlFlow::Break(()),
         };
 
-        match self.imitate_layer_accesses(p, cancel, gate, ctx).await {
+        match self
+            .imitate_layer_accesses(tenant, p, cancel, gate, ctx)
+            .await
+        {
             ControlFlow::Break(()) => return ControlFlow::Break(()),
             ControlFlow::Continue(()) => (),
         }
@@ -225,24 +235,18 @@ impl Timeline {
         {
             let guard = self.layers.read().await;
             let layers = guard.layer_map();
-            for hist_layer in layers.iter_historic_layers() {
-                let hist_layer = guard.get_from_desc(&hist_layer);
+            for layer in layers.iter_historic_layers() {
+                let layer = guard.get_from_desc(&layer);
 
                 // guard against eviction while we inspect it; it might be that eviction_task and
                 // disk_usage_eviction_task both select the same layers to be evicted, and
                 // seemingly free up double the space. both succeeding is of no consequence.
-                let guard = match hist_layer.keep_resident().await {
-                    Ok(Some(l)) => l,
-                    Ok(None) => continue,
-                    Err(e) => {
-                        // these should not happen, but we cannot make them statically impossible right
-                        // now.
-                        tracing::warn!(layer=%hist_layer, "failed to keep the layer resident: {e:#}");
-                        continue;
-                    }
-                };
 
-                let last_activity_ts = hist_layer.access_stats().latest_activity_or_now();
+                if !layer.is_likely_resident() {
+                    continue;
+                }
+
+                let last_activity_ts = layer.access_stats().latest_activity_or_now();
 
                 let no_activity_for = match now.duration_since(last_activity_ts) {
                     Ok(d) => d,
@@ -265,9 +269,8 @@ impl Timeline {
                         continue;
                     }
                 };
-                let layer = guard.drop_eviction_guard();
+
                 if no_activity_for > p.threshold {
-                    // this could cause a lot of allocations in some cases
                     js.spawn(async move {
                         layer
                             .evict_and_wait(std::time::Duration::from_secs(5))
@@ -322,6 +325,7 @@ impl Timeline {
     /// disk usage based eviction task.
     async fn imitiate_only(
         self: &Arc<Self>,
+        tenant: &Tenant,
         p: &EvictionPolicyLayerAccessThreshold,
         cancel: &CancellationToken,
         gate: &GateGuard,
@@ -338,7 +342,8 @@ impl Timeline {
             _ = self.cancel.cancelled() => return ControlFlow::Break(()),
         };
 
-        self.imitate_layer_accesses(p, cancel, gate, ctx).await
+        self.imitate_layer_accesses(tenant, p, cancel, gate, ctx)
+            .await
     }
 
     /// If we evict layers but keep cached values derived from those layers, then
@@ -368,6 +373,7 @@ impl Timeline {
     #[instrument(skip_all)]
     async fn imitate_layer_accesses(
         &self,
+        tenant: &Tenant,
         p: &EvictionPolicyLayerAccessThreshold,
         cancel: &CancellationToken,
         gate: &GateGuard,
@@ -403,17 +409,11 @@ impl Timeline {
         // Make one of the tenant's timelines draw the short straw and run the calculation.
         // The others wait until the calculation is done so that they take into account the
         // imitated accesses that the winner made.
-        let tenant = match crate::tenant::mgr::get_tenant(self.tenant_shard_id, true) {
-            Ok(t) => t,
-            Err(_) => {
-                return ControlFlow::Break(());
-            }
-        };
         let mut state = tenant.eviction_task_tenant_state.lock().await;
         match state.last_layer_access_imitation {
             Some(ts) if ts.elapsed() < inter_imitate_period => { /* no need to run */ }
             _ => {
-                self.imitate_synthetic_size_calculation_worker(&tenant, cancel, ctx)
+                self.imitate_synthetic_size_calculation_worker(tenant, cancel, ctx)
                     .await;
                 state.last_layer_access_imitation = Some(tokio::time::Instant::now());
             }
@@ -487,7 +487,7 @@ impl Timeline {
     #[instrument(skip_all)]
     async fn imitate_synthetic_size_calculation_worker(
         &self,
-        tenant: &Arc<Tenant>,
+        tenant: &Tenant,
         cancel: &CancellationToken,
         ctx: &RequestContext,
     ) {

pageserver/src/tenant/timeline/layer_manager.rs

@@ -1,5 +1,4 @@
 use anyhow::{bail, ensure, Context, Result};
-use futures::StreamExt;
 use pageserver_api::shard::TenantShardId;
 use std::{collections::HashMap, sync::Arc};
 use tracing::trace;
@@ -241,29 +240,16 @@ impl LayerManager {
         layer.delete_on_drop();
     }
 
-    pub(crate) fn resident_layers(&self) -> impl futures::stream::Stream<Item = Layer> + '_ {
+    pub(crate) fn likely_resident_layers(&self) -> impl Iterator<Item = Layer> + '_ {
         // for small layer maps, we most likely have all resident, but for larger more are likely
         // to be evicted assuming lots of layers correlated with longer lifespan.
 
-        let layers = self
-            .layer_map()
-            .iter_historic_layers()
-            .map(|desc| self.get_from_desc(&desc));
-
-        let layers = futures::stream::iter(layers);
-
-        layers.filter_map(|layer| async move {
-            // TODO(#6028): this query does not really need to see the ResidentLayer
-            match layer.keep_resident().await {
-                Ok(Some(layer)) => Some(layer.drop_eviction_guard()),
-                Ok(None) => None,
-                Err(e) => {
-                    // these should not happen, but we cannot make them statically impossible right
-                    // now.
-                    tracing::warn!(%layer, "failed to keep the layer resident: {e:#}");
-                    None
-                }
-            }
+        self.layer_map().iter_historic_layers().filter_map(|desc| {
+            self.layer_fmgr
+                .0
+                .get(&desc.key())
+                .filter(|l| l.is_likely_resident())
+                .cloned()
         })
     }
 

pageserver/src/tenant/timeline/uninit.rs

@@ -86,6 +86,7 @@ impl<'t> UninitializedTimeline<'t> {
     /// Prepares timeline data by loading it from the basebackup archive.
     pub(crate) async fn import_basebackup_from_tar(
         self,
+        tenant: Arc<Tenant>,
         copyin_read: &mut (impl tokio::io::AsyncRead + Send + Sync + Unpin),
         base_lsn: Lsn,
         broker_client: storage_broker::BrokerClientChannel,
@@ -114,7 +115,7 @@ impl<'t> UninitializedTimeline<'t> {
 
         // All the data has been imported. Insert the Timeline into the tenant's timelines map
         let tl = self.finish_creation()?;
-        tl.activate(broker_client, None, ctx);
+        tl.activate(tenant, broker_client, None, ctx);
         Ok(tl)
     }
 

pageserver/src/tenant/timeline/walreceiver.rs

@@ -33,11 +33,9 @@ use crate::tenant::timeline::walreceiver::connection_manager::{
 use pageserver_api::shard::TenantShardId;
 use std::future::Future;
 use std::num::NonZeroU64;
-use std::ops::ControlFlow;
 use std::sync::Arc;
 use std::time::Duration;
 use storage_broker::BrokerClientChannel;
-use tokio::select;
 use tokio::sync::watch;
 use tokio_util::sync::CancellationToken;
 use tracing::*;
@@ -91,31 +89,27 @@ impl WalReceiver {
             async move {
                 debug_assert_current_span_has_tenant_and_timeline_id();
                 debug!("WAL receiver manager started, connecting to broker");
+                let cancel = task_mgr::shutdown_token();
                 let mut connection_manager_state = ConnectionManagerState::new(
                     timeline,
                     conf,
                 );
-                loop {
-                    select! {
-                        _ = task_mgr::shutdown_watcher() => {
-                            trace!("WAL receiver shutdown requested, shutting down");
+                while !cancel.is_cancelled() {
+                    let loop_step_result = connection_manager_loop_step(
+                        &mut broker_client,
+                        &mut connection_manager_state,
+                        &walreceiver_ctx,
+                        &cancel,
+                        &loop_status,
+                    ).await;
+                    match loop_step_result {
+                        Ok(()) => continue,
+                        Err(_cancelled) => {
+                            trace!("Connection manager loop ended, shutting down");
                             break;
-                        },
-                        loop_step_result = connection_manager_loop_step(
-                            &mut broker_client,
-                            &mut connection_manager_state,
-                            &walreceiver_ctx,
-                            &loop_status,
-                        ) => match loop_step_result {
-                            ControlFlow::Continue(()) => continue,
-                            ControlFlow::Break(()) => {
-                                trace!("Connection manager loop ended, shutting down");
-                                break;
-                            }
-                        },
+                        }
                     }
                 }
-
                 connection_manager_state.shutdown().await;
                 *loop_status.write().unwrap() = None;
                 Ok(())
@@ -197,6 +191,9 @@ impl<E: Clone> TaskHandle<E> {
         }
     }
 
+    /// # Cancel-Safety
+    ///
+    /// Cancellation-safe.
     async fn next_task_event(&mut self) -> TaskEvent<E> {
         match self.events_receiver.changed().await {
             Ok(()) => TaskEvent::Update((self.events_receiver.borrow()).clone()),

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

@@ -17,7 +17,7 @@ use crate::metrics::{
     WALRECEIVER_ACTIVE_MANAGERS, WALRECEIVER_BROKER_UPDATES, WALRECEIVER_CANDIDATES_ADDED,
     WALRECEIVER_CANDIDATES_REMOVED, WALRECEIVER_SWITCHES,
 };
-use crate::task_mgr::{shutdown_token, TaskKind};
+use crate::task_mgr::TaskKind;
 use crate::tenant::{debug_assert_current_span_has_tenant_and_timeline_id, Timeline};
 use anyhow::Context;
 use chrono::{NaiveDateTime, Utc};
@@ -27,7 +27,7 @@ use storage_broker::proto::SafekeeperTimelineInfo;
 use storage_broker::proto::SubscribeSafekeeperInfoRequest;
 use storage_broker::proto::TenantTimelineId as ProtoTenantTimelineId;
 use storage_broker::{BrokerClientChannel, Code, Streaming};
-use tokio::select;
+use tokio_util::sync::CancellationToken;
 use tracing::*;
 
 use postgres_connection::PgConnectionConfig;
@@ -45,27 +45,33 @@ use super::{
     TaskEvent, TaskHandle,
 };
 
+pub(crate) struct Cancelled;
+
 /// Attempts to subscribe for timeline updates, pushed by safekeepers into the broker.
 /// Based on the updates, desides whether to start, keep or stop a WAL receiver task.
 /// If storage broker subscription is cancelled, exits.
+///
+/// # Cancel-Safety
+///
+/// Not cancellation-safe. Use `cancel` token to request cancellation.
 pub(super) async fn connection_manager_loop_step(
     broker_client: &mut BrokerClientChannel,
     connection_manager_state: &mut ConnectionManagerState,
     ctx: &RequestContext,
+    cancel: &CancellationToken,
     manager_status: &std::sync::RwLock<Option<ConnectionManagerStatus>>,
-) -> ControlFlow<(), ()> {
-    match connection_manager_state
-        .timeline
-        .wait_to_become_active(ctx)
-        .await
-    {
+) -> Result<(), Cancelled> {
+    match tokio::select! {
+        _ = cancel.cancelled() => { return Err(Cancelled); },
+        st = connection_manager_state.timeline.wait_to_become_active(ctx) => { st }
+    } {
         Ok(()) => {}
         Err(new_state) => {
             debug!(
                 ?new_state,
                 "state changed, stopping wal connection manager loop"
             );
-            return ControlFlow::Break(());
+            return Err(Cancelled);
         }
     }
 
@@ -86,7 +92,7 @@ pub(super) async fn connection_manager_loop_step(
     // Subscribe to the broker updates. Stream shares underlying TCP connection
     // with other streams on this client (other connection managers). When
     // object goes out of scope, stream finishes in drop() automatically.
-    let mut broker_subscription = subscribe_for_timeline_updates(broker_client, id).await;
+    let mut broker_subscription = subscribe_for_timeline_updates(broker_client, id, cancel).await?;
     debug!("Subscribed for broker timeline updates");
 
     loop {
@@ -94,6 +100,7 @@ pub(super) async fn connection_manager_loop_step(
 
         // These things are happening concurrently:
         //
+        // - cancellation request
         //  - keep receiving WAL on the current connection
         //      - if the shared state says we need to change connection, disconnect and return
         //      - this runs in a separate task and we receive updates via a watch channel
@@ -101,7 +108,11 @@ pub(super) async fn connection_manager_loop_step(
         //  - receive updates from broker
         //      - this might change the current desired connection
         //  - timeline state changes to something that does not allow walreceiver to run concurrently
-        select! {
+
+        // NB: make sure each of the select expressions are cancellation-safe
+        // (no need for arms to be cancellation-safe).
+        tokio::select! {
+            _ = cancel.cancelled() => { return Err(Cancelled); }
             Some(wal_connection_update) = async {
                 match connection_manager_state.wal_connection.as_mut() {
                     Some(wal_connection) => Some(wal_connection.connection_task.next_task_event().await),
@@ -133,7 +144,7 @@ pub(super) async fn connection_manager_loop_step(
             },
 
             // Got a new update from the broker
-            broker_update = broker_subscription.message() => {
+            broker_update = broker_subscription.message() /* TODO: review cancellation-safety */ => {
                 match broker_update {
                     Ok(Some(broker_update)) => connection_manager_state.register_timeline_update(broker_update),
                     Err(status) => {
@@ -147,16 +158,17 @@ pub(super) async fn connection_manager_loop_step(
                                 warn!("broker subscription failed: {status}");
                             }
                         }
-                        return ControlFlow::Continue(());
+                        return Ok(());
                     }
                     Ok(None) => {
                         error!("broker subscription stream ended"); // can't happen
-                        return ControlFlow::Continue(());
+                        return Ok(());
                     }
                 }
             },
 
             new_event = async {
+                // Reminder: this match arm needs to be cancellation-safe.
                 loop {
                     if connection_manager_state.timeline.current_state() == TimelineState::Loading {
                         warn!("wal connection manager should only be launched after timeline has become active");
@@ -182,11 +194,11 @@ pub(super) async fn connection_manager_loop_step(
                 }
             } => match new_event {
                 ControlFlow::Continue(()) => {
-                    return ControlFlow::Continue(());
+                    return Ok(());
                 }
                 ControlFlow::Break(()) => {
                     debug!("Timeline is no longer active, stopping wal connection manager loop");
-                    return ControlFlow::Break(());
+                    return Err(Cancelled);
                 }
             },
 
@@ -218,16 +230,15 @@ pub(super) async fn connection_manager_loop_step(
 async fn subscribe_for_timeline_updates(
     broker_client: &mut BrokerClientChannel,
     id: TenantTimelineId,
-) -> Streaming<SafekeeperTimelineInfo> {
+    cancel: &CancellationToken,
+) -> Result<Streaming<SafekeeperTimelineInfo>, Cancelled> {
     let mut attempt = 0;
-    let cancel = shutdown_token();
-
     loop {
         exponential_backoff(
             attempt,
             DEFAULT_BASE_BACKOFF_SECONDS,
             DEFAULT_MAX_BACKOFF_SECONDS,
-            &cancel,
+            cancel,
         )
         .await;
         attempt += 1;
@@ -241,9 +252,14 @@ async fn subscribe_for_timeline_updates(
             subscription_key: Some(key),
         };
 
-        match broker_client.subscribe_safekeeper_info(request).await {
+        match {
+            tokio::select! {
+                r = broker_client.subscribe_safekeeper_info(request) => { r }
+                _ = cancel.cancelled() => { return Err(Cancelled); }
+            }
+        } {
             Ok(resp) => {
-                return resp.into_inner();
+                return Ok(resp.into_inner());
             }
             Err(e) => {
                 // Safekeeper nodes can stop pushing timeline updates to the broker, when no new writes happen and
@@ -486,6 +502,10 @@ impl ConnectionManagerState {
 
     /// Drops the current connection (if any) and updates retry timeout for the next
     /// connection attempt to the same safekeeper.
+    ///
+    /// # Cancel-Safety
+    ///
+    /// Not cancellation-safe.
     async fn drop_old_connection(&mut self, needs_shutdown: bool) {
         let wal_connection = match self.wal_connection.take() {
             Some(wal_connection) => wal_connection,
@@ -493,7 +513,14 @@ impl ConnectionManagerState {
         };
 
         if needs_shutdown {
-            wal_connection.connection_task.shutdown().await;
+            wal_connection
+                .connection_task
+                .shutdown()
+                // This here is why this function isn't cancellation-safe.
+                // If we got cancelled here, then self.wal_connection is already None and we lose track of the task.
+                // Even if our caller diligently calls Self::shutdown(), it will find a self.wal_connection=None
+                // and thus be ineffective.
+                .await;
         }
 
         let retry = self
@@ -838,6 +865,9 @@ impl ConnectionManagerState {
         }
     }
 
+    /// # Cancel-Safety
+    ///
+    /// Not cancellation-safe.
     pub(super) async fn shutdown(mut self) {
         if let Some(wal_connection) = self.wal_connection.take() {
             wal_connection.connection_task.shutdown().await;

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

@@ -389,17 +389,6 @@ pub(super) async fn handle_walreceiver_connection(
             }
         }
 
-        {
-            // This is a hack. It piggybacks on the keepalive messages sent by the
-            // safekeeper in order to enforce `checkpoint_timeout` on the currently
-            // open layer. This hack doesn't provide a bound on the total size of
-            // in-memory layers on a pageserver. See https://github.com/neondatabase/neon/issues/6916.
-            let mut writer = timeline.writer().await;
-            if let Err(err) = writer.tick().await {
-                warn!("Timeline writer tick failed: {err}");
-            }
-        }
-
         if let Some(last_lsn) = status_update {
             let timeline_remote_consistent_lsn = timeline
                 .get_remote_consistent_lsn_visible()

pageserver/src/tenant/upload_queue.rs

@@ -121,11 +121,16 @@ pub(super) enum SetDeletedFlagProgress {
     Successful(NaiveDateTime),
 }
 
-pub(super) struct UploadQueueStopped {
+pub(super) struct UploadQueueStoppedDeletable {
     pub(super) upload_queue_for_deletion: UploadQueueInitialized,
     pub(super) deleted_at: SetDeletedFlagProgress,
 }
 
+pub(super) enum UploadQueueStopped {
+    Deletable(UploadQueueStoppedDeletable),
+    Uninitialized,
+}
+
 #[derive(thiserror::Error, Debug)]
 pub(crate) enum NotInitialized {
     #[error("queue is in state Uninitialized")]
@@ -249,12 +254,15 @@ impl UploadQueue {
         }
     }
 
-    pub(crate) fn stopped_mut(&mut self) -> anyhow::Result<&mut UploadQueueStopped> {
+    pub(crate) fn stopped_mut(&mut self) -> anyhow::Result<&mut UploadQueueStoppedDeletable> {
         match self {
             UploadQueue::Initialized(_) | UploadQueue::Uninitialized => {
                 anyhow::bail!("queue is in state {}", self.as_str())
             }
-            UploadQueue::Stopped(stopped) => Ok(stopped),
+            UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => {
+                anyhow::bail!("queue is in state Stopped(Uninitialized)")
+            }
+            UploadQueue::Stopped(UploadQueueStopped::Deletable(deletable)) => Ok(deletable),
         }
     }
 }

pageserver/src/virtual_file.rs

@@ -782,7 +782,7 @@ where
         }
     }
     // NB: don't use `buf.is_empty()` here; it is from the
-    // `impl Deref for Slice { Target = [u8] }`; the the &[u8]
+    // `impl Deref for Slice { Target = [u8] }`; the &[u8]
     // returned by it only covers the initialized portion of `buf`.
     // Whereas we're interested in ensuring that we filled the entire
     // buffer that the user passed in.

pgxn/neon/neon.c

@@ -312,7 +312,7 @@ pg_cluster_size(PG_FUNCTION_ARGS)
 {
 	int64		size;
 
-	size = GetZenithCurrentClusterSize();
+	size = GetNeonCurrentClusterSize();
 
 	if (size == 0)
 		PG_RETURN_NULL();

pgxn/neon/neon.h

@@ -26,6 +26,8 @@ extern void pg_init_libpagestore(void);
 extern void pg_init_walproposer(void);
 
 extern uint64 BackpressureThrottlingTime(void);
+extern void SetNeonCurrentClusterSize(uint64 size);
+extern uint64 GetNeonCurrentClusterSize(void);
 extern void replication_feedback_get_lsns(XLogRecPtr *writeLsn, XLogRecPtr *flushLsn, XLogRecPtr *applyLsn);
 
 extern void PGDLLEXPORT WalProposerSync(int argc, char *argv[]);

pgxn/neon/pagestore_smgr.c

@@ -1831,7 +1831,7 @@ neon_extend(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno,
 		reln->smgr_relpersistence == RELPERSISTENCE_PERMANENT &&
 		!IsAutoVacuumWorkerProcess())
 	{
-		uint64		current_size = GetZenithCurrentClusterSize();
+		uint64		current_size = GetNeonCurrentClusterSize();
 
 		if (current_size >= ((uint64) max_cluster_size) * 1024 * 1024)
 			ereport(ERROR,
@@ -1912,7 +1912,7 @@ neon_zeroextend(SMgrRelation reln, ForkNumber forkNum, BlockNumber blocknum,
 		reln->smgr_relpersistence == RELPERSISTENCE_PERMANENT &&
 		!IsAutoVacuumWorkerProcess())
 	{
-		uint64		current_size = GetZenithCurrentClusterSize();
+		uint64		current_size = GetNeonCurrentClusterSize();
 
 		if (current_size >= ((uint64) max_cluster_size) * 1024 * 1024)
 			ereport(ERROR,

pgxn/neon/walproposer.h

@@ -287,6 +287,7 @@ typedef struct WalproposerShmemState
 	slock_t		mutex;
 	term_t		mineLastElectedTerm;
 	pg_atomic_uint64 backpressureThrottlingTime;
+	pg_atomic_uint64 currentClusterSize;
 
 	/* last feedback from each shard */
 	PageserverFeedback shard_ps_feedback[MAX_SHARDS];

pgxn/neon/walproposer_pg.c

@@ -282,6 +282,7 @@ WalproposerShmemInit(void)
 		memset(walprop_shared, 0, WalproposerShmemSize());
 		SpinLockInit(&walprop_shared->mutex);
 		pg_atomic_init_u64(&walprop_shared->backpressureThrottlingTime, 0);
+		pg_atomic_init_u64(&walprop_shared->currentClusterSize, 0);
 	}
 	LWLockRelease(AddinShmemInitLock);
 
@@ -549,9 +550,10 @@ walprop_pg_init_standalone_sync_safekeepers(void)
 static void
 walprop_sigusr2(SIGNAL_ARGS)
 {
+	int			save_errno = errno;
 	got_SIGUSR2 = true;
-
 	SetLatch(MyLatch);
+	errno = save_errno;
 }
 
 static void
@@ -1971,7 +1973,7 @@ walprop_pg_process_safekeeper_feedback(WalProposer *wp, Safekeeper *sk)
 
 		/* Only one main shard sends non-zero currentClusterSize */
 		if (sk->appendResponse.ps_feedback.currentClusterSize > 0)
-			SetZenithCurrentClusterSize(sk->appendResponse.ps_feedback.currentClusterSize);
+			SetNeonCurrentClusterSize(sk->appendResponse.ps_feedback.currentClusterSize);
 
 		if (min_feedback.disk_consistent_lsn != standby_apply_lsn)
 		{
@@ -2093,6 +2095,18 @@ GetLogRepRestartLSN(WalProposer *wp)
 	return lrRestartLsn;
 }
 
+void SetNeonCurrentClusterSize(uint64 size)
+{
+	pg_atomic_write_u64(&walprop_shared->currentClusterSize, size);
+}
+
+uint64 GetNeonCurrentClusterSize(void)
+{
+	return pg_atomic_read_u64(&walprop_shared->currentClusterSize);
+}
+uint64 GetNeonCurrentClusterSize(void);
+
+
 static const walproposer_api walprop_pg = {
 	.get_shmem_state = walprop_pg_get_shmem_state,
 	.start_streaming = walprop_pg_start_streaming,

proxy/Cargo.toml

@@ -11,6 +11,10 @@ testing = []
 [dependencies]
 anyhow.workspace = true
 async-trait.workspace = true
+aws-config.workspace = true
+aws-sdk-iam.workspace = true
+aws-sigv4.workspace = true
+aws-types.workspace = true
 base64.workspace = true
 bstr.workspace = true
 bytes = { workspace = true, features = ["serde"] }
@@ -27,6 +31,7 @@ hashlink.workspace = true
 hex.workspace = true
 hmac.workspace = true
 hostname.workspace = true
+http.workspace = true
 humantime.workspace = true
 hyper-tungstenite.workspace = true
 hyper.workspace = true
@@ -59,10 +64,11 @@ rustls.workspace = true
 scopeguard.workspace = true
 serde.workspace = true
 serde_json.workspace = true
-sha2.workspace = true
+sha2 = { workspace = true, features = ["asm"] }
 smol_str.workspace = true
 smallvec.workspace = true
 socket2.workspace = true
+subtle.workspace = true
 sync_wrapper.workspace = true
 task-local-extensions.workspace = true
 thiserror.workspace = true
@@ -91,6 +97,7 @@ workspace_hack.workspace = true
 
 [dev-dependencies]
 camino-tempfile.workspace = true
+fallible-iterator.workspace = true
 rcgen.workspace = true
 rstest.workspace = true
 tokio-postgres-rustls.workspace = true

proxy/src/auth/backend.rs

@@ -12,6 +12,8 @@ use crate::console::errors::GetAuthInfoError;
 use crate::console::provider::{CachedRoleSecret, ConsoleBackend};
 use crate::console::{AuthSecret, NodeInfo};
 use crate::context::RequestMonitoring;
+use crate::intern::EndpointIdInt;
+use crate::metrics::{AUTH_RATE_LIMIT_HITS, ENDPOINTS_AUTH_RATE_LIMITED};
 use crate::proxy::connect_compute::ComputeConnectBackend;
 use crate::proxy::NeonOptions;
 use crate::stream::Stream;
@@ -28,7 +30,7 @@ use crate::{
 use crate::{scram, EndpointCacheKey, EndpointId, RoleName};
 use std::sync::Arc;
 use tokio::io::{AsyncRead, AsyncWrite};
-use tracing::info;
+use tracing::{info, warn};
 
 /// Alternative to [`std::borrow::Cow`] but doesn't need `T: ToOwned` as we don't need that functionality
 pub enum MaybeOwned<'a, T> {
@@ -174,6 +176,52 @@ impl TryFrom<ComputeUserInfoMaybeEndpoint> for ComputeUserInfo {
     }
 }
 
+impl AuthenticationConfig {
+    pub fn check_rate_limit(
+        &self,
+
+        ctx: &mut RequestMonitoring,
+        secret: AuthSecret,
+        endpoint: &EndpointId,
+        is_cleartext: bool,
+    ) -> auth::Result<AuthSecret> {
+        // we have validated the endpoint exists, so let's intern it.
+        let endpoint_int = EndpointIdInt::from(endpoint);
+
+        // only count the full hash count if password hack or websocket flow.
+        // in other words, if proxy needs to run the hashing
+        let password_weight = if is_cleartext {
+            match &secret {
+                #[cfg(any(test, feature = "testing"))]
+                AuthSecret::Md5(_) => 1,
+                AuthSecret::Scram(s) => s.iterations + 1,
+            }
+        } else {
+            // validating scram takes just 1 hmac_sha_256 operation.
+            1
+        };
+
+        let limit_not_exceeded = self
+            .rate_limiter
+            .check((endpoint_int, ctx.peer_addr), password_weight);
+
+        if !limit_not_exceeded {
+            warn!(
+                enabled = self.rate_limiter_enabled,
+                "rate limiting authentication"
+            );
+            AUTH_RATE_LIMIT_HITS.inc();
+            ENDPOINTS_AUTH_RATE_LIMITED.measure(endpoint);
+
+            if self.rate_limiter_enabled {
+                return Err(auth::AuthError::too_many_connections());
+            }
+        }
+
+        Ok(secret)
+    }
+}
+
 /// True to its name, this function encapsulates our current auth trade-offs.
 /// Here, we choose the appropriate auth flow based on circumstances.
 ///
@@ -214,14 +262,24 @@ async fn auth_quirks(
         Some(secret) => secret,
         None => api.get_role_secret(ctx, &info).await?,
     };
+    let (cached_entry, secret) = cached_secret.take_value();
+
+    let secret = match secret {
+        Some(secret) => config.check_rate_limit(
+            ctx,
+            secret,
+            &info.endpoint,
+            unauthenticated_password.is_some() || allow_cleartext,
+        )?,
+        None => {
+            // If we don't have an authentication secret, we mock one to
+            // prevent malicious probing (possible due to missing protocol steps).
+            // This mocked secret will never lead to successful authentication.
+            info!("authentication info not found, mocking it");
+            AuthSecret::Scram(scram::ServerSecret::mock(rand::random()))
+        }
+    };
 
-    let secret = cached_secret.value.clone().unwrap_or_else(|| {
-        // If we don't have an authentication secret, we mock one to
-        // prevent malicious probing (possible due to missing protocol steps).
-        // This mocked secret will never lead to successful authentication.
-        info!("authentication info not found, mocking it");
-        AuthSecret::Scram(scram::ServerSecret::mock(&info.user, rand::random()))
-    });
     match authenticate_with_secret(
         ctx,
         secret,
@@ -237,7 +295,7 @@ async fn auth_quirks(
         Err(e) => {
             if e.is_auth_failed() {
                 // The password could have been changed, so we invalidate the cache.
-                cached_secret.invalidate();
+                cached_entry.invalidate();
             }
             Err(e)
         }
@@ -408,3 +466,232 @@ impl ComputeConnectBackend for BackendType<'_, ComputeCredentials, &()> {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use bytes::BytesMut;
+    use fallible_iterator::FallibleIterator;
+    use once_cell::sync::Lazy;
+    use postgres_protocol::{
+        authentication::sasl::{ChannelBinding, ScramSha256},
+        message::{backend::Message as PgMessage, frontend},
+    };
+    use provider::AuthSecret;
+    use tokio::io::{AsyncRead, AsyncReadExt, AsyncWriteExt};
+
+    use crate::{
+        auth::{ComputeUserInfoMaybeEndpoint, IpPattern},
+        config::AuthenticationConfig,
+        console::{
+            self,
+            provider::{self, CachedAllowedIps, CachedRoleSecret},
+            CachedNodeInfo,
+        },
+        context::RequestMonitoring,
+        proxy::NeonOptions,
+        rate_limiter::{AuthRateLimiter, RateBucketInfo},
+        scram::ServerSecret,
+        stream::{PqStream, Stream},
+    };
+
+    use super::auth_quirks;
+
+    struct Auth {
+        ips: Vec<IpPattern>,
+        secret: AuthSecret,
+    }
+
+    impl console::Api for Auth {
+        async fn get_role_secret(
+            &self,
+            _ctx: &mut RequestMonitoring,
+            _user_info: &super::ComputeUserInfo,
+        ) -> Result<CachedRoleSecret, console::errors::GetAuthInfoError> {
+            Ok(CachedRoleSecret::new_uncached(Some(self.secret.clone())))
+        }
+
+        async fn get_allowed_ips_and_secret(
+            &self,
+            _ctx: &mut RequestMonitoring,
+            _user_info: &super::ComputeUserInfo,
+        ) -> Result<(CachedAllowedIps, Option<CachedRoleSecret>), console::errors::GetAuthInfoError>
+        {
+            Ok((
+                CachedAllowedIps::new_uncached(Arc::new(self.ips.clone())),
+                Some(CachedRoleSecret::new_uncached(Some(self.secret.clone()))),
+            ))
+        }
+
+        async fn wake_compute(
+            &self,
+            _ctx: &mut RequestMonitoring,
+            _user_info: &super::ComputeUserInfo,
+        ) -> Result<CachedNodeInfo, console::errors::WakeComputeError> {
+            unimplemented!()
+        }
+    }
+
+    static CONFIG: Lazy<AuthenticationConfig> = Lazy::new(|| AuthenticationConfig {
+        scram_protocol_timeout: std::time::Duration::from_secs(5),
+        rate_limiter_enabled: true,
+        rate_limiter: AuthRateLimiter::new(&RateBucketInfo::DEFAULT_AUTH_SET),
+    });
+
+    async fn read_message(r: &mut (impl AsyncRead + Unpin), b: &mut BytesMut) -> PgMessage {
+        loop {
+            r.read_buf(&mut *b).await.unwrap();
+            if let Some(m) = PgMessage::parse(&mut *b).unwrap() {
+                break m;
+            }
+        }
+    }
+
+    #[tokio::test]
+    async fn auth_quirks_scram() {
+        let (mut client, server) = tokio::io::duplex(1024);
+        let mut stream = PqStream::new(Stream::from_raw(server));
+
+        let mut ctx = RequestMonitoring::test();
+        let api = Auth {
+            ips: vec![],
+            secret: AuthSecret::Scram(ServerSecret::build("my-secret-password").await.unwrap()),
+        };
+
+        let user_info = ComputeUserInfoMaybeEndpoint {
+            user: "conrad".into(),
+            endpoint_id: Some("endpoint".into()),
+            options: NeonOptions::default(),
+        };
+
+        let handle = tokio::spawn(async move {
+            let mut scram = ScramSha256::new(b"my-secret-password", ChannelBinding::unsupported());
+
+            let mut read = BytesMut::new();
+
+            // server should offer scram
+            match read_message(&mut client, &mut read).await {
+                PgMessage::AuthenticationSasl(a) => {
+                    let options: Vec<&str> = a.mechanisms().collect().unwrap();
+                    assert_eq!(options, ["SCRAM-SHA-256"]);
+                }
+                _ => panic!("wrong message"),
+            }
+
+            // client sends client-first-message
+            let mut write = BytesMut::new();
+            frontend::sasl_initial_response("SCRAM-SHA-256", scram.message(), &mut write).unwrap();
+            client.write_all(&write).await.unwrap();
+
+            // server response with server-first-message
+            match read_message(&mut client, &mut read).await {
+                PgMessage::AuthenticationSaslContinue(a) => {
+                    scram.update(a.data()).await.unwrap();
+                }
+                _ => panic!("wrong message"),
+            }
+
+            // client response with client-final-message
+            write.clear();
+            frontend::sasl_response(scram.message(), &mut write).unwrap();
+            client.write_all(&write).await.unwrap();
+
+            // server response with server-final-message
+            match read_message(&mut client, &mut read).await {
+                PgMessage::AuthenticationSaslFinal(a) => {
+                    scram.finish(a.data()).unwrap();
+                }
+                _ => panic!("wrong message"),
+            }
+        });
+
+        let _creds = auth_quirks(&mut ctx, &api, user_info, &mut stream, false, &CONFIG)
+            .await
+            .unwrap();
+
+        handle.await.unwrap();
+    }
+
+    #[tokio::test]
+    async fn auth_quirks_cleartext() {
+        let (mut client, server) = tokio::io::duplex(1024);
+        let mut stream = PqStream::new(Stream::from_raw(server));
+
+        let mut ctx = RequestMonitoring::test();
+        let api = Auth {
+            ips: vec![],
+            secret: AuthSecret::Scram(ServerSecret::build("my-secret-password").await.unwrap()),
+        };
+
+        let user_info = ComputeUserInfoMaybeEndpoint {
+            user: "conrad".into(),
+            endpoint_id: Some("endpoint".into()),
+            options: NeonOptions::default(),
+        };
+
+        let handle = tokio::spawn(async move {
+            let mut read = BytesMut::new();
+            let mut write = BytesMut::new();
+
+            // server should offer cleartext
+            match read_message(&mut client, &mut read).await {
+                PgMessage::AuthenticationCleartextPassword => {}
+                _ => panic!("wrong message"),
+            }
+
+            // client responds with password
+            write.clear();
+            frontend::password_message(b"my-secret-password", &mut write).unwrap();
+            client.write_all(&write).await.unwrap();
+        });
+
+        let _creds = auth_quirks(&mut ctx, &api, user_info, &mut stream, true, &CONFIG)
+            .await
+            .unwrap();
+
+        handle.await.unwrap();
+    }
+
+    #[tokio::test]
+    async fn auth_quirks_password_hack() {
+        let (mut client, server) = tokio::io::duplex(1024);
+        let mut stream = PqStream::new(Stream::from_raw(server));
+
+        let mut ctx = RequestMonitoring::test();
+        let api = Auth {
+            ips: vec![],
+            secret: AuthSecret::Scram(ServerSecret::build("my-secret-password").await.unwrap()),
+        };
+
+        let user_info = ComputeUserInfoMaybeEndpoint {
+            user: "conrad".into(),
+            endpoint_id: None,
+            options: NeonOptions::default(),
+        };
+
+        let handle = tokio::spawn(async move {
+            let mut read = BytesMut::new();
+
+            // server should offer cleartext
+            match read_message(&mut client, &mut read).await {
+                PgMessage::AuthenticationCleartextPassword => {}
+                _ => panic!("wrong message"),
+            }
+
+            // client responds with password
+            let mut write = BytesMut::new();
+            frontend::password_message(b"endpoint=my-endpoint;my-secret-password", &mut write)
+                .unwrap();
+            client.write_all(&write).await.unwrap();
+        });
+
+        let creds = auth_quirks(&mut ctx, &api, user_info, &mut stream, true, &CONFIG)
+            .await
+            .unwrap();
+
+        assert_eq!(creds.info.endpoint, "my-endpoint");
+
+        handle.await.unwrap();
+    }
+}

proxy/src/auth/flow.rs

@@ -194,14 +194,7 @@ pub(crate) async fn validate_password_and_exchange(
         }
         // perform scram authentication as both client and server to validate the keys
         AuthSecret::Scram(scram_secret) => {
-            use postgres_protocol::authentication::sasl::{ChannelBinding, ScramSha256};
-            let sasl_client = ScramSha256::new(password, ChannelBinding::unsupported());
-            let outcome = crate::scram::exchange(
-                &scram_secret,
-                sasl_client,
-                crate::config::TlsServerEndPoint::Undefined,
-            )
-            .await?;
+            let outcome = crate::scram::exchange(&scram_secret, password).await?;
 
             let client_key = match outcome {
                 sasl::Outcome::Success(client_key) => client_key,

proxy/src/bin/proxy.rs

@@ -1,3 +1,10 @@
+use aws_config::environment::EnvironmentVariableCredentialsProvider;
+use aws_config::imds::credentials::ImdsCredentialsProvider;
+use aws_config::meta::credentials::CredentialsProviderChain;
+use aws_config::meta::region::RegionProviderChain;
+use aws_config::profile::ProfileFileCredentialsProvider;
+use aws_config::provider_config::ProviderConfig;
+use aws_config::web_identity_token::WebIdentityTokenCredentialsProvider;
 use futures::future::Either;
 use proxy::auth;
 use proxy::auth::backend::MaybeOwned;
@@ -10,11 +17,15 @@ use proxy::config::ProjectInfoCacheOptions;
 use proxy::console;
 use proxy::context::parquet::ParquetUploadArgs;
 use proxy::http;
+use proxy::metrics::NUM_CANCELLATION_REQUESTS_SOURCE_FROM_CLIENT;
+use proxy::rate_limiter::AuthRateLimiter;
 use proxy::rate_limiter::EndpointRateLimiter;
 use proxy::rate_limiter::RateBucketInfo;
 use proxy::rate_limiter::RateLimiterConfig;
+use proxy::redis::cancellation_publisher::RedisPublisherClient;
+use proxy::redis::connection_with_credentials_provider::ConnectionWithCredentialsProvider;
+use proxy::redis::elasticache;
 use proxy::redis::notifications;
-use proxy::redis::publisher::RedisPublisherClient;
 use proxy::serverless::GlobalConnPoolOptions;
 use proxy::usage_metrics;
 
@@ -131,10 +142,16 @@ struct ProxyCliArgs {
     ///
     /// Provided in the form '<Requests Per Second>@<Bucket Duration Size>'.
     /// Can be given multiple times for different bucket sizes.
-    #[clap(long, default_values_t = RateBucketInfo::DEFAULT_SET)]
+    #[clap(long, default_values_t = RateBucketInfo::DEFAULT_ENDPOINT_SET)]
     endpoint_rps_limit: Vec<RateBucketInfo>,
+    /// Whether the auth rate limiter actually takes effect (for testing)
+    #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
+    auth_rate_limit_enabled: bool,
+    /// Authentication rate limiter max number of hashes per second.
+    #[clap(long, default_values_t = RateBucketInfo::DEFAULT_AUTH_SET)]
+    auth_rate_limit: Vec<RateBucketInfo>,
     /// Redis rate limiter max number of requests per second.
-    #[clap(long, default_values_t = RateBucketInfo::DEFAULT_SET)]
+    #[clap(long, default_values_t = RateBucketInfo::DEFAULT_ENDPOINT_SET)]
     redis_rps_limit: Vec<RateBucketInfo>,
     /// Initial limit for dynamic rate limiter. Makes sense only if `rate_limit_algorithm` is *not* `None`.
     #[clap(long, default_value_t = 100)]
@@ -150,9 +167,24 @@ struct ProxyCliArgs {
     /// disable ip check for http requests. If it is too time consuming, it could be turned off.
     #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
     disable_ip_check_for_http: bool,
-    /// redis url for notifications.
+    /// redis url for notifications (if empty, redis_host:port will be used for both notifications and streaming connections)
     #[clap(long)]
     redis_notifications: Option<String>,
+    /// redis host for streaming connections (might be different from the notifications host)
+    #[clap(long)]
+    redis_host: Option<String>,
+    /// redis port for streaming connections (might be different from the notifications host)
+    #[clap(long)]
+    redis_port: Option<u16>,
+    /// redis cluster name, used in aws elasticache
+    #[clap(long)]
+    redis_cluster_name: Option<String>,
+    /// redis user_id, used in aws elasticache
+    #[clap(long)]
+    redis_user_id: Option<String>,
+    /// aws region to retrieve credentials
+    #[clap(long, default_value_t = String::new())]
+    aws_region: String,
     /// cache for `project_info` (use `size=0` to disable)
     #[clap(long, default_value = config::ProjectInfoCacheOptions::CACHE_DEFAULT_OPTIONS)]
     project_info_cache: String,
@@ -216,6 +248,61 @@ async fn main() -> anyhow::Result<()> {
     let config = build_config(&args)?;
 
     info!("Authentication backend: {}", config.auth_backend);
+    info!("Using region: {}", config.aws_region);
+
+    let region_provider = RegionProviderChain::default_provider().or_else(&*config.aws_region); // Replace with your Redis region if needed
+    let provider_conf =
+        ProviderConfig::without_region().with_region(region_provider.region().await);
+    let aws_credentials_provider = {
+        // uses "AWS_ACCESS_KEY_ID", "AWS_SECRET_ACCESS_KEY"
+        CredentialsProviderChain::first_try("env", EnvironmentVariableCredentialsProvider::new())
+            // uses "AWS_PROFILE" / `aws sso login --profile <profile>`
+            .or_else(
+                "profile-sso",
+                ProfileFileCredentialsProvider::builder()
+                    .configure(&provider_conf)
+                    .build(),
+            )
+            // uses "AWS_WEB_IDENTITY_TOKEN_FILE", "AWS_ROLE_ARN", "AWS_ROLE_SESSION_NAME"
+            // needed to access remote extensions bucket
+            .or_else(
+                "token",
+                WebIdentityTokenCredentialsProvider::builder()
+                    .configure(&provider_conf)
+                    .build(),
+            )
+            // uses imds v2
+            .or_else("imds", ImdsCredentialsProvider::builder().build())
+    };
+    let elasticache_credentials_provider = Arc::new(elasticache::CredentialsProvider::new(
+        elasticache::AWSIRSAConfig::new(
+            config.aws_region.clone(),
+            args.redis_cluster_name,
+            args.redis_user_id,
+        ),
+        aws_credentials_provider,
+    ));
+    let redis_notifications_client =
+        match (args.redis_notifications, (args.redis_host, args.redis_port)) {
+            (Some(url), _) => {
+                info!("Starting redis notifications listener ({url})");
+                Some(ConnectionWithCredentialsProvider::new_with_static_credentials(url))
+            }
+            (None, (Some(host), Some(port))) => Some(
+                ConnectionWithCredentialsProvider::new_with_credentials_provider(
+                    host,
+                    port,
+                    elasticache_credentials_provider.clone(),
+                ),
+            ),
+            (None, (None, None)) => {
+                warn!("Redis is disabled");
+                None
+            }
+            _ => {
+                bail!("redis-host and redis-port must be specified together");
+            }
+        };
 
     // Check that we can bind to address before further initialization
     let http_address: SocketAddr = args.http.parse()?;
@@ -233,17 +320,22 @@ async fn main() -> anyhow::Result<()> {
 
     let endpoint_rate_limiter = Arc::new(EndpointRateLimiter::new(&config.endpoint_rps_limit));
     let cancel_map = CancelMap::default();
-    let redis_publisher = match &args.redis_notifications {
-        Some(url) => Some(Arc::new(Mutex::new(RedisPublisherClient::new(
-            url,
+
+    // let redis_notifications_client = redis_notifications_client.map(|x| Box::leak(Box::new(x)));
+    let redis_publisher = match &redis_notifications_client {
+        Some(redis_publisher) => Some(Arc::new(Mutex::new(RedisPublisherClient::new(
+            redis_publisher.clone(),
             args.region.clone(),
             &config.redis_rps_limit,
         )?))),
         None => None,
     };
-    let cancellation_handler = Arc::new(CancellationHandler::new(
+    let cancellation_handler = Arc::new(CancellationHandler::<
+        Option<Arc<tokio::sync::Mutex<RedisPublisherClient>>>,
+    >::new(
         cancel_map.clone(),
         redis_publisher,
+        NUM_CANCELLATION_REQUESTS_SOURCE_FROM_CLIENT,
     ));
 
     // client facing tasks. these will exit on error or on cancellation
@@ -290,17 +382,16 @@ async fn main() -> anyhow::Result<()> {
 
     if let auth::BackendType::Console(api, _) = &config.auth_backend {
         if let proxy::console::provider::ConsoleBackend::Console(api) = &**api {
-            let cache = api.caches.project_info.clone();
-            if let Some(url) = args.redis_notifications {
-                info!("Starting redis notifications listener ({url})");
+            if let Some(redis_notifications_client) = redis_notifications_client {
+                let cache = api.caches.project_info.clone();
                 maintenance_tasks.spawn(notifications::task_main(
-                    url.to_owned(),
+                    redis_notifications_client.clone(),
                     cache.clone(),
                     cancel_map.clone(),
                     args.region.clone(),
                 ));
+                maintenance_tasks.spawn(async move { cache.clone().gc_worker().await });
             }
-            maintenance_tasks.spawn(async move { cache.clone().gc_worker().await });
         }
     }
 
@@ -426,6 +517,8 @@ fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
     };
     let authentication_config = AuthenticationConfig {
         scram_protocol_timeout: args.scram_protocol_timeout,
+        rate_limiter_enabled: args.auth_rate_limit_enabled,
+        rate_limiter: AuthRateLimiter::new(args.auth_rate_limit.clone()),
     };
 
     let mut endpoint_rps_limit = args.endpoint_rps_limit.clone();
@@ -445,8 +538,8 @@ fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
         endpoint_rps_limit,
         redis_rps_limit,
         handshake_timeout: args.handshake_timeout,
-        // TODO: add this argument
         region: args.region.clone(),
+        aws_region: args.aws_region.clone(),
     }));
 
     Ok(config)

proxy/src/cache/common.rs

@@ -43,6 +43,16 @@ impl<C: Cache, V> Cached<C, V> {
         Self { token: None, value }
     }
 
+    pub fn take_value(self) -> (Cached<C, ()>, V) {
+        (
+            Cached {
+                token: self.token,
+                value: (),
+            },
+            self.value,
+        )
+    }
+
     /// Drop this entry from a cache if it's still there.
     pub fn invalidate(self) -> V {
         if let Some((cache, info)) = &self.token {

proxy/src/cache/project_info.rs

@@ -373,10 +373,7 @@ mod tests {
         let endpoint_id = "endpoint".into();
         let user1: RoleName = "user1".into();
         let user2: RoleName = "user2".into();
-        let secret1 = Some(AuthSecret::Scram(ServerSecret::mock(
-            user1.as_str(),
-            [1; 32],
-        )));
+        let secret1 = Some(AuthSecret::Scram(ServerSecret::mock([1; 32])));
         let secret2 = None;
         let allowed_ips = Arc::new(vec![
             "127.0.0.1".parse().unwrap(),
@@ -395,10 +392,7 @@ mod tests {
 
         // Shouldn't add more than 2 roles.
         let user3: RoleName = "user3".into();
-        let secret3 = Some(AuthSecret::Scram(ServerSecret::mock(
-            user3.as_str(),
-            [3; 32],
-        )));
+        let secret3 = Some(AuthSecret::Scram(ServerSecret::mock([3; 32])));
         cache.insert_role_secret(&project_id, &endpoint_id, &user3, secret3.clone());
         assert!(cache.get_role_secret(&endpoint_id, &user3).is_none());
 
@@ -431,14 +425,8 @@ mod tests {
         let endpoint_id = "endpoint".into();
         let user1: RoleName = "user1".into();
         let user2: RoleName = "user2".into();
-        let secret1 = Some(AuthSecret::Scram(ServerSecret::mock(
-            user1.as_str(),
-            [1; 32],
-        )));
-        let secret2 = Some(AuthSecret::Scram(ServerSecret::mock(
-            user2.as_str(),
-            [2; 32],
-        )));
+        let secret1 = Some(AuthSecret::Scram(ServerSecret::mock([1; 32])));
+        let secret2 = Some(AuthSecret::Scram(ServerSecret::mock([2; 32])));
         let allowed_ips = Arc::new(vec![
             "127.0.0.1".parse().unwrap(),
             "127.0.0.2".parse().unwrap(),
@@ -486,14 +474,8 @@ mod tests {
         let endpoint_id = "endpoint".into();
         let user1: RoleName = "user1".into();
         let user2: RoleName = "user2".into();
-        let secret1 = Some(AuthSecret::Scram(ServerSecret::mock(
-            user1.as_str(),
-            [1; 32],
-        )));
-        let secret2 = Some(AuthSecret::Scram(ServerSecret::mock(
-            user2.as_str(),
-            [2; 32],
-        )));
+        let secret1 = Some(AuthSecret::Scram(ServerSecret::mock([1; 32])));
+        let secret2 = Some(AuthSecret::Scram(ServerSecret::mock([2; 32])));
         let allowed_ips = Arc::new(vec![
             "127.0.0.1".parse().unwrap(),
             "127.0.0.2".parse().unwrap(),

proxy/src/cancellation.rs

@@ -1,4 +1,3 @@
-use async_trait::async_trait;
 use dashmap::DashMap;
 use pq_proto::CancelKeyData;
 use std::{net::SocketAddr, sync::Arc};
@@ -10,18 +9,26 @@ use tracing::info;
 use uuid::Uuid;
 
 use crate::{
-    error::ReportableError, metrics::NUM_CANCELLATION_REQUESTS,
-    redis::publisher::RedisPublisherClient,
+    error::ReportableError,
+    metrics::NUM_CANCELLATION_REQUESTS,
+    redis::cancellation_publisher::{
+        CancellationPublisher, CancellationPublisherMut, RedisPublisherClient,
+    },
 };
 
 pub type CancelMap = Arc<DashMap<CancelKeyData, Option<CancelClosure>>>;
+pub type CancellationHandlerMain = CancellationHandler<Option<Arc<Mutex<RedisPublisherClient>>>>;
+pub type CancellationHandlerMainInternal = Option<Arc<Mutex<RedisPublisherClient>>>;
 
 /// Enables serving `CancelRequest`s.
 ///
-/// If there is a `RedisPublisherClient` available, it will be used to publish the cancellation key to other proxy instances.
-pub struct CancellationHandler {
+/// If `CancellationPublisher` is available, cancel request will be used to publish the cancellation key to other proxy instances.
+pub struct CancellationHandler<P> {
     map: CancelMap,
-    redis_client: Option<Arc<Mutex<RedisPublisherClient>>>,
+    client: P,
+    /// This field used for the monitoring purposes.
+    /// Represents the source of the cancellation request.
+    from: &'static str,
 }
 
 #[derive(Debug, Error)]
@@ -44,49 +51,9 @@ impl ReportableError for CancelError {
     }
 }
 
-impl CancellationHandler {
-    pub fn new(map: CancelMap, redis_client: Option<Arc<Mutex<RedisPublisherClient>>>) -> Self {
-        Self { map, redis_client }
-    }
-    /// Cancel a running query for the corresponding connection.
-    pub async fn cancel_session(
-        &self,
-        key: CancelKeyData,
-        session_id: Uuid,
-    ) -> Result<(), CancelError> {
-        let from = "from_client";
-        // NB: we should immediately release the lock after cloning the token.
-        let Some(cancel_closure) = self.map.get(&key).and_then(|x| x.clone()) else {
-            tracing::warn!("query cancellation key not found: {key}");
-            if let Some(redis_client) = &self.redis_client {
-                NUM_CANCELLATION_REQUESTS
-                    .with_label_values(&[from, "not_found"])
-                    .inc();
-                info!("publishing cancellation key to Redis");
-                match redis_client.lock().await.try_publish(key, session_id).await {
-                    Ok(()) => {
-                        info!("cancellation key successfuly published to Redis");
-                    }
-                    Err(e) => {
-                        tracing::error!("failed to publish a message: {e}");
-                        return Err(CancelError::IO(std::io::Error::new(
-                            std::io::ErrorKind::Other,
-                            e.to_string(),
-                        )));
-                    }
-                }
-            }
-            return Ok(());
-        };
-        NUM_CANCELLATION_REQUESTS
-            .with_label_values(&[from, "found"])
-            .inc();
-        info!("cancelling query per user's request using key {key}");
-        cancel_closure.try_cancel_query().await
-    }
-
+impl<P: CancellationPublisher> CancellationHandler<P> {
     /// Run async action within an ephemeral session identified by [`CancelKeyData`].
-    pub fn get_session(self: Arc<Self>) -> Session {
+    pub fn get_session(self: Arc<Self>) -> Session<P> {
         // HACK: We'd rather get the real backend_pid but tokio_postgres doesn't
         // expose it and we don't want to do another roundtrip to query
         // for it. The client will be able to notice that this is not the
@@ -112,9 +79,39 @@ impl CancellationHandler {
             cancellation_handler: self,
         }
     }
+    /// Try to cancel a running query for the corresponding connection.
+    /// If the cancellation key is not found, it will be published to Redis.
+    pub async fn cancel_session(
+        &self,
+        key: CancelKeyData,
+        session_id: Uuid,
+    ) -> Result<(), CancelError> {
+        // NB: we should immediately release the lock after cloning the token.
+        let Some(cancel_closure) = self.map.get(&key).and_then(|x| x.clone()) else {
+            tracing::warn!("query cancellation key not found: {key}");
+            NUM_CANCELLATION_REQUESTS
+                .with_label_values(&[self.from, "not_found"])
+                .inc();
+            match self.client.try_publish(key, session_id).await {
+                Ok(()) => {} // do nothing
+                Err(e) => {
+                    return Err(CancelError::IO(std::io::Error::new(
+                        std::io::ErrorKind::Other,
+                        e.to_string(),
+                    )));
+                }
+            }
+            return Ok(());
+        };
+        NUM_CANCELLATION_REQUESTS
+            .with_label_values(&[self.from, "found"])
+            .inc();
+        info!("cancelling query per user's request using key {key}");
+        cancel_closure.try_cancel_query().await
+    }
 
     #[cfg(test)]
-    fn contains(&self, session: &Session) -> bool {
+    fn contains(&self, session: &Session<P>) -> bool {
         self.map.contains_key(&session.key)
     }
 
@@ -124,31 +121,19 @@ impl CancellationHandler {
     }
 }
 
-#[async_trait]
-pub trait NotificationsCancellationHandler {
-    async fn cancel_session_no_publish(&self, key: CancelKeyData) -> Result<(), CancelError>;
+impl CancellationHandler<()> {
+    pub fn new(map: CancelMap, from: &'static str) -> Self {
+        Self {
+            map,
+            client: (),
+            from,
+        }
+    }
 }
 
-#[async_trait]
-impl NotificationsCancellationHandler for CancellationHandler {
-    async fn cancel_session_no_publish(&self, key: CancelKeyData) -> Result<(), CancelError> {
-        let from = "from_redis";
-        let cancel_closure = self.map.get(&key).and_then(|x| x.clone());
-        match cancel_closure {
-            Some(cancel_closure) => {
-                NUM_CANCELLATION_REQUESTS
-                    .with_label_values(&[from, "found"])
-                    .inc();
-                cancel_closure.try_cancel_query().await
-            }
-            None => {
-                NUM_CANCELLATION_REQUESTS
-                    .with_label_values(&[from, "not_found"])
-                    .inc();
-                tracing::warn!("query cancellation key not found: {key}");
-                Ok(())
-            }
-        }
+impl<P: CancellationPublisherMut> CancellationHandler<Option<Arc<Mutex<P>>>> {
+    pub fn new(map: CancelMap, client: Option<Arc<Mutex<P>>>, from: &'static str) -> Self {
+        Self { map, client, from }
     }
 }
 
@@ -178,14 +163,14 @@ impl CancelClosure {
 }
 
 /// Helper for registering query cancellation tokens.
-pub struct Session {
+pub struct Session<P> {
     /// The user-facing key identifying this session.
     key: CancelKeyData,
     /// The [`CancelMap`] this session belongs to.
-    cancellation_handler: Arc<CancellationHandler>,
+    cancellation_handler: Arc<CancellationHandler<P>>,
 }
 
-impl Session {
+impl<P> Session<P> {
     /// Store the cancel token for the given session.
     /// This enables query cancellation in `crate::proxy::prepare_client_connection`.
     pub fn enable_query_cancellation(&self, cancel_closure: CancelClosure) -> CancelKeyData {
@@ -198,7 +183,7 @@ impl Session {
     }
 }
 
-impl Drop for Session {
+impl<P> Drop for Session<P> {
     fn drop(&mut self) {
         self.cancellation_handler.map.remove(&self.key);
         info!("dropped query cancellation key {}", &self.key);
@@ -207,14 +192,16 @@ impl Drop for Session {
 
 #[cfg(test)]
 mod tests {
+    use crate::metrics::NUM_CANCELLATION_REQUESTS_SOURCE_FROM_REDIS;
+
     use super::*;
 
     #[tokio::test]
     async fn check_session_drop() -> anyhow::Result<()> {
-        let cancellation_handler = Arc::new(CancellationHandler {
-            map: CancelMap::default(),
-            redis_client: None,
-        });
+        let cancellation_handler = Arc::new(CancellationHandler::<()>::new(
+            CancelMap::default(),
+            NUM_CANCELLATION_REQUESTS_SOURCE_FROM_REDIS,
+        ));
 
         let session = cancellation_handler.clone().get_session();
         assert!(cancellation_handler.contains(&session));
@@ -224,4 +211,19 @@ mod tests {
 
         Ok(())
     }
+
+    #[tokio::test]
+    async fn cancel_session_noop_regression() {
+        let handler = CancellationHandler::<()>::new(Default::default(), "local");
+        handler
+            .cancel_session(
+                CancelKeyData {
+                    backend_pid: 0,
+                    cancel_key: 0,
+                },
+                Uuid::new_v4(),
+            )
+            .await
+            .unwrap();
+    }
 }

proxy/src/compute.rs

@@ -82,14 +82,13 @@ pub type ScramKeys = tokio_postgres::config::ScramKeys<32>;
 /// A config for establishing a connection to compute node.
 /// Eventually, `tokio_postgres` will be replaced with something better.
 /// Newtype allows us to implement methods on top of it.
-#[derive(Clone)]
-#[repr(transparent)]
+#[derive(Clone, Default)]
 pub struct ConnCfg(Box<tokio_postgres::Config>);
 
 /// Creation and initialization routines.
 impl ConnCfg {
     pub fn new() -> Self {
-        Self(Default::default())
+        Self::default()
     }
 
     /// Reuse password or auth keys from the other config.
@@ -165,12 +164,6 @@ impl std::ops::DerefMut for ConnCfg {
     }
 }
 
-impl Default for ConnCfg {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
 impl ConnCfg {
     /// Establish a raw TCP connection to the compute node.
     async fn connect_raw(&self, timeout: Duration) -> io::Result<(SocketAddr, TcpStream, &str)> {

proxy/src/config.rs

@@ -1,4 +1,8 @@
-use crate::{auth, rate_limiter::RateBucketInfo, serverless::GlobalConnPoolOptions};
+use crate::{
+    auth,
+    rate_limiter::{AuthRateLimiter, RateBucketInfo},
+    serverless::GlobalConnPoolOptions,
+};
 use anyhow::{bail, ensure, Context, Ok};
 use itertools::Itertools;
 use rustls::{
@@ -28,6 +32,7 @@ pub struct ProxyConfig {
     pub redis_rps_limit: Vec<RateBucketInfo>,
     pub region: String,
     pub handshake_timeout: Duration,
+    pub aws_region: String,
 }
 
 #[derive(Debug)]
@@ -49,6 +54,8 @@ pub struct HttpConfig {
 
 pub struct AuthenticationConfig {
     pub scram_protocol_timeout: tokio::time::Duration,
+    pub rate_limiter_enabled: bool,
+    pub rate_limiter: AuthRateLimiter,
 }
 
 impl TlsConfig {

proxy/src/console.rs

@@ -6,7 +6,7 @@ pub mod messages;
 
 /// Wrappers for console APIs and their mocks.
 pub mod provider;
-pub use provider::{errors, Api, AuthSecret, CachedNodeInfo, NodeInfo};
+pub(crate) use provider::{errors, Api, AuthSecret, CachedNodeInfo, NodeInfo};
 
 /// Various cache-related types.
 pub mod caches {

proxy/src/console/provider.rs

@@ -14,7 +14,6 @@ use crate::{
     context::RequestMonitoring,
     scram, EndpointCacheKey, ProjectId,
 };
-use async_trait::async_trait;
 use dashmap::DashMap;
 use std::{sync::Arc, time::Duration};
 use tokio::sync::{OwnedSemaphorePermit, Semaphore};
@@ -326,8 +325,7 @@ pub type CachedAllowedIps = Cached<&'static ProjectInfoCacheImpl, Arc<Vec<IpPatt
 
 /// This will allocate per each call, but the http requests alone
 /// already require a few allocations, so it should be fine.
-#[async_trait]
-pub trait Api {
+pub(crate) trait Api {
     /// Get the client's auth secret for authentication.
     /// Returns option because user not found situation is special.
     /// We still have to mock the scram to avoid leaking information that user doesn't exist.
@@ -363,7 +361,6 @@ pub enum ConsoleBackend {
     Test(Box<dyn crate::auth::backend::TestBackend>),
 }
 
-#[async_trait]
 impl Api for ConsoleBackend {
     async fn get_role_secret(
         &self,

proxy/src/console/provider/mock.rs

@@ -8,7 +8,6 @@ use crate::console::provider::{CachedAllowedIps, CachedRoleSecret};
 use crate::context::RequestMonitoring;
 use crate::{auth::backend::ComputeUserInfo, compute, error::io_error, scram, url::ApiUrl};
 use crate::{auth::IpPattern, cache::Cached};
-use async_trait::async_trait;
 use futures::TryFutureExt;
 use std::{str::FromStr, sync::Arc};
 use thiserror::Error;
@@ -144,7 +143,6 @@ async fn get_execute_postgres_query(
     Ok(Some(entry))
 }
 
-#[async_trait]
 impl super::Api for Api {
     #[tracing::instrument(skip_all)]
     async fn get_role_secret(

proxy/src/console/provider/neon.rs

@@ -14,7 +14,6 @@ use crate::{
     context::RequestMonitoring,
     metrics::{ALLOWED_IPS_BY_CACHE_OUTCOME, ALLOWED_IPS_NUMBER},
 };
-use async_trait::async_trait;
 use futures::TryFutureExt;
 use std::sync::Arc;
 use tokio::time::Instant;
@@ -56,7 +55,7 @@ impl Api {
         ctx: &mut RequestMonitoring,
         user_info: &ComputeUserInfo,
     ) -> Result<AuthInfo, GetAuthInfoError> {
-        let request_id = uuid::Uuid::new_v4().to_string();
+        let request_id = ctx.session_id.to_string();
         let application_name = ctx.console_application_name();
         async {
             let request = self
@@ -76,6 +75,7 @@ impl Api {
             let start = Instant::now();
             let pause = ctx.latency_timer.pause(crate::metrics::Waiting::Cplane);
             let response = self.endpoint.execute(request).await?;
+            info!("received http response {response:?}");
             drop(pause);
             info!(duration = ?start.elapsed(), "received http response");
             let body = match parse_body::<GetRoleSecret>(response).await {
@@ -113,7 +113,7 @@ impl Api {
         ctx: &mut RequestMonitoring,
         user_info: &ComputeUserInfo,
     ) -> Result<NodeInfo, WakeComputeError> {
-        let request_id = uuid::Uuid::new_v4().to_string();
+        let request_id = ctx.session_id.to_string();
         let application_name = ctx.console_application_name();
         async {
             let mut request_builder = self
@@ -138,6 +138,7 @@ impl Api {
             let start = Instant::now();
             let pause = ctx.latency_timer.pause(crate::metrics::Waiting::Cplane);
             let response = self.endpoint.execute(request).await?;
+            info!("received http response {response:?}");
             drop(pause);
             info!(duration = ?start.elapsed(), "received http response");
             let body = parse_body::<WakeCompute>(response).await?;
@@ -168,7 +169,6 @@ impl Api {
     }
 }
 
-#[async_trait]
 impl super::Api for Api {
     #[tracing::instrument(skip_all)]
     async fn get_role_secret(