Merge pull request #7357 from neondatabase/rc/proxy/2024-04-11

Proxy release 2024-04-11

.dockerignore

@@ -22,6 +22,7 @@
 !s3_scrubber/
 !safekeeper/
 !storage_broker/
+!storage_controller/
 !trace/
 !vendor/postgres-*/
 !workspace_hack/

.github/actions/neon-branch-create/action.yml

@@ -10,7 +10,7 @@ inputs:
     required: true
   api_host:
     desctiption: 'Neon API host'
-    default: console.stage.neon.tech
+    default: console-stage.neon.build
 outputs:
   dsn:
     description: 'Created Branch DSN (for main database)'

.github/actions/neon-branch-delete/action.yml

@@ -13,7 +13,7 @@ inputs:
     required: true
   api_host:
     desctiption: 'Neon API host'
-    default: console.stage.neon.tech
+    default: console-stage.neon.build
 
 runs:
   using: "composite"

.github/actions/neon-project-create/action.yml

@@ -13,7 +13,7 @@ inputs:
     default: 15
   api_host:
     desctiption: 'Neon API host'
-    default: console.stage.neon.tech
+    default: console-stage.neon.build
   provisioner:
     desctiption: 'k8s-pod or k8s-neonvm'
     default: 'k8s-pod'

.github/actions/neon-project-delete/action.yml

@@ -10,7 +10,7 @@ inputs:
     required: true
   api_host:
     desctiption: 'Neon API host'
-    default: console.stage.neon.tech
+    default: console-stage.neon.build
 
 runs:
   using: "composite"

CODEOWNERS

@@ -1,5 +1,5 @@
 /compute_tools/ @neondatabase/control-plane @neondatabase/compute
-/control_plane/attachment_service @neondatabase/storage
+/storage_controller @neondatabase/storage
 /libs/pageserver_api/ @neondatabase/storage
 /libs/postgres_ffi/ @neondatabase/compute @neondatabase/safekeepers
 /libs/remote_storage/ @neondatabase/storage

Cargo.lock

@@ -271,43 +271,10 @@ dependencies = [
 ]
 
 [[package]]
-name = "attachment_service"
-version = "0.1.0"
-dependencies = [
- "anyhow",
- "aws-config",
- "bytes",
- "camino",
- "clap",
- "control_plane",
- "diesel",
- "diesel_migrations",
- "fail",
- "futures",
- "git-version",
- "hex",
- "humantime",
- "hyper",
- "itertools",
- "lasso",
- "measured",
- "metrics",
- "once_cell",
- "pageserver_api",
- "pageserver_client",
- "postgres_connection",
- "r2d2",
- "reqwest",
- "routerify",
- "serde",
- "serde_json",
- "thiserror",
- "tokio",
- "tokio-util",
- "tracing",
- "utils",
- "workspace_hack",
-]
+name = "atomic-take"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a8ab6b55fe97976e46f91ddbed8d147d966475dc29b2032757ba47e02376fbc3"
 
 [[package]]
 name = "autocfg"
@@ -337,7 +304,7 @@ dependencies = [
  "fastrand 2.0.0",
  "hex",
  "http 0.2.9",
- "hyper",
+ "hyper 0.14.26",
  "ring 0.17.6",
  "time",
  "tokio",
@@ -374,7 +341,7 @@ dependencies = [
  "bytes",
  "fastrand 2.0.0",
  "http 0.2.9",
- "http-body",
+ "http-body 0.4.5",
  "percent-encoding",
  "pin-project-lite",
  "tracing",
@@ -425,7 +392,7 @@ dependencies = [
  "aws-types",
  "bytes",
  "http 0.2.9",
- "http-body",
+ "http-body 0.4.5",
  "once_cell",
  "percent-encoding",
  "regex-lite",
@@ -553,7 +520,7 @@ dependencies = [
  "crc32fast",
  "hex",
  "http 0.2.9",
- "http-body",
+ "http-body 0.4.5",
  "md-5",
  "pin-project-lite",
  "sha1",
@@ -585,7 +552,7 @@ dependencies = [
  "bytes-utils",
  "futures-core",
  "http 0.2.9",
- "http-body",
+ "http-body 0.4.5",
  "once_cell",
  "percent-encoding",
  "pin-project-lite",
@@ -624,10 +591,10 @@ dependencies = [
  "aws-smithy-types",
  "bytes",
  "fastrand 2.0.0",
- "h2",
+ "h2 0.3.26",
  "http 0.2.9",
- "http-body",
- "hyper",
+ "http-body 0.4.5",
+ "hyper 0.14.26",
  "hyper-rustls",
  "once_cell",
  "pin-project-lite",
@@ -665,7 +632,7 @@ dependencies = [
  "bytes-utils",
  "futures-core",
  "http 0.2.9",
- "http-body",
+ "http-body 0.4.5",
  "itoa",
  "num-integer",
  "pin-project-lite",
@@ -714,8 +681,8 @@ dependencies = [
  "bytes",
  "futures-util",
  "http 0.2.9",
- "http-body",
- "hyper",
+ "http-body 0.4.5",
+ "hyper 0.14.26",
  "itoa",
  "matchit",
  "memchr",
@@ -730,7 +697,7 @@ dependencies = [
  "sha1",
  "sync_wrapper",
  "tokio",
- "tokio-tungstenite",
+ "tokio-tungstenite 0.20.0",
  "tower",
  "tower-layer",
  "tower-service",
@@ -746,7 +713,7 @@ dependencies = [
  "bytes",
  "futures-util",
  "http 0.2.9",
- "http-body",
+ "http-body 0.4.5",
  "mime",
  "rustversion",
  "tower-layer",
@@ -1163,7 +1130,7 @@ version = "4.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "191d9573962933b4027f932c600cd252ce27a8ad5979418fe78e43c07996f27b"
 dependencies = [
- "heck",
+ "heck 0.4.1",
  "proc-macro2",
  "quote",
  "syn 2.0.52",
@@ -1235,7 +1202,7 @@ dependencies = [
  "compute_api",
  "flate2",
  "futures",
- "hyper",
+ "hyper 0.14.26",
  "nix 0.27.1",
  "notify",
  "num_cpus",
@@ -1352,7 +1319,7 @@ dependencies = [
  "git-version",
  "hex",
  "humantime",
- "hyper",
+ "hyper 0.14.26",
  "nix 0.27.1",
  "once_cell",
  "pageserver_api",
@@ -1501,12 +1468,9 @@ dependencies = [
 
 [[package]]
 name = "crossbeam-utils"
-version = "0.8.15"
+version = "0.8.19"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3c063cd8cc95f5c377ed0d4b49a4b21f632396ff690e8470c29b3359b346984b"
-dependencies = [
- "cfg-if",
-]
+checksum = "248e3bacc7dc6baa3b21e405ee045c3047101a49145e7e9eca583ab4c2ca5345"
 
 [[package]]
 name = "crossterm"
@@ -1879,23 +1843,12 @@ checksum = "5443807d6dff69373d433ab9ef5378ad8df50ca6298caf15de6e52e24aaf54d5"
 
 [[package]]
 name = "errno"
-version = "0.3.1"
+version = "0.3.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4bcfec3a70f97c962c307b2d2c56e358cf1d00b558d74262b5f929ee8cc7e73a"
+checksum = "a258e46cdc063eb8519c00b9fc845fc47bcfca4130e2f08e88665ceda8474245"
 dependencies = [
- "errno-dragonfly",
- "libc",
- "windows-sys 0.48.0",
-]
-
-[[package]]
-name = "errno-dragonfly"
-version = "0.1.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "aa68f1b12764fab894d2755d2518754e71b4fd80ecfb822714a1206c2aab39bf"
-dependencies = [
- "cc",
  "libc",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -2252,6 +2205,25 @@ dependencies = [
  "tracing",
 ]
 
+[[package]]
+name = "h2"
+version = "0.4.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "816ec7294445779408f36fe57bc5b7fc1cf59664059096c65f905c1c61f58069"
+dependencies = [
+ "bytes",
+ "fnv",
+ "futures-core",
+ "futures-sink",
+ "futures-util",
+ "http 1.1.0",
+ "indexmap 2.0.1",
+ "slab",
+ "tokio",
+ "tokio-util",
+ "tracing",
+]
+
 [[package]]
 name = "half"
 version = "1.8.2"
@@ -2333,6 +2305,12 @@ version = "0.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "95505c38b4572b2d910cecb0281560f54b440a19336cbbcb27bf6ce6adc6f5a8"
 
+[[package]]
+name = "heck"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
+
 [[package]]
 name = "hermit-abi"
 version = "0.3.3"
@@ -2417,6 +2395,29 @@ dependencies = [
  "pin-project-lite",
 ]
 
+[[package]]
+name = "http-body"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1cac85db508abc24a2e48553ba12a996e87244a0395ce011e62b37158745d643"
+dependencies = [
+ "bytes",
+ "http 1.1.0",
+]
+
+[[package]]
+name = "http-body-util"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "41cb79eb393015dadd30fc252023adb0b2400a0caee0fa2a077e6e21a551e840"
+dependencies = [
+ "bytes",
+ "futures-util",
+ "http 1.1.0",
+ "http-body 1.0.0",
+ "pin-project-lite",
+]
+
 [[package]]
 name = "http-types"
 version = "2.12.0"
@@ -2475,9 +2476,9 @@ dependencies = [
  "futures-channel",
  "futures-core",
  "futures-util",
- "h2",
+ "h2 0.3.26",
  "http 0.2.9",
- "http-body",
+ "http-body 0.4.5",
  "httparse",
  "httpdate",
  "itoa",
@@ -2489,6 +2490,26 @@ dependencies = [
  "want",
 ]
 
+[[package]]
+name = "hyper"
+version = "1.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "186548d73ac615b32a73aafe38fb4f56c0d340e110e5a200bcadbaf2e199263a"
+dependencies = [
+ "bytes",
+ "futures-channel",
+ "futures-util",
+ "h2 0.4.4",
+ "http 1.1.0",
+ "http-body 1.0.0",
+ "httparse",
+ "httpdate",
+ "itoa",
+ "pin-project-lite",
+ "smallvec",
+ "tokio",
+]
+
 [[package]]
 name = "hyper-rustls"
 version = "0.24.0"
@@ -2496,7 +2517,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0646026eb1b3eea4cd9ba47912ea5ce9cc07713d105b1a14698f4e6433d348b7"
 dependencies = [
  "http 0.2.9",
- "hyper",
+ "hyper 0.14.26",
  "log",
  "rustls 0.21.9",
  "rustls-native-certs 0.6.2",
@@ -2510,7 +2531,7 @@ version = "0.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bbb958482e8c7be4bc3cf272a766a2b0bf1a6755e7a6ae777f017a31d11b13b1"
 dependencies = [
- "hyper",
+ "hyper 0.14.26",
  "pin-project-lite",
  "tokio",
  "tokio-io-timeout",
@@ -2523,7 +2544,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d6183ddfa99b85da61a140bea0efc93fdf56ceaa041b37d553518030827f9905"
 dependencies = [
  "bytes",
- "hyper",
+ "hyper 0.14.26",
  "native-tls",
  "tokio",
  "tokio-native-tls",
@@ -2531,15 +2552,33 @@ dependencies = [
 
 [[package]]
 name = "hyper-tungstenite"
-version = "0.11.1"
+version = "0.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7cc7dcb1ab67cd336f468a12491765672e61a3b6b148634dbfe2fe8acd3fe7d9"
+checksum = "7a343d17fe7885302ed7252767dc7bb83609a874b6ff581142241ec4b73957ad"
 dependencies = [
- "hyper",
+ "http-body-util",
+ "hyper 1.2.0",
+ "hyper-util",
  "pin-project-lite",
  "tokio",
- "tokio-tungstenite",
- "tungstenite",
+ "tokio-tungstenite 0.21.0",
+ "tungstenite 0.21.0",
+]
+
+[[package]]
+name = "hyper-util"
+version = "0.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ca38ef113da30126bbff9cd1705f9273e15d45498615d138b0c20279ac7a76aa"
+dependencies = [
+ "bytes",
+ "futures-util",
+ "http 1.1.0",
+ "http-body 1.0.0",
+ "hyper 1.2.0",
+ "pin-project-lite",
+ "socket2 0.5.5",
+ "tokio",
 ]
 
 [[package]]
@@ -2833,6 +2872,12 @@ version = "0.3.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ef53942eb7bf7ff43a617b3e2c1c4a5ecf5944a7c1bc12d7ee39bbb15e5c1519"
 
+[[package]]
+name = "linux-raw-sys"
+version = "0.4.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "01cda141df6706de531b6c46c3a33ecca755538219bd484262fa09410c13539c"
+
 [[package]]
 name = "lock_api"
 version = "0.4.10"
@@ -2887,11 +2932,12 @@ checksum = "490cc448043f947bae3cbee9c203358d62dbee0db12107a74be5c30ccfd09771"
 
 [[package]]
 name = "measured"
-version = "0.0.13"
+version = "0.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f246648d027839a34b420e27c7de1165ace96e19ef894985d0a6ff89a7840a9f"
+checksum = "3cbf033874bea03565f2449572c8640ca37ec26300455faf36001f24755da452"
 dependencies = [
  "bytes",
+ "crossbeam-utils",
  "hashbrown 0.14.0",
  "itoa",
  "lasso",
@@ -2904,16 +2950,27 @@ dependencies = [
 
 [[package]]
 name = "measured-derive"
-version = "0.0.13"
+version = "0.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "edaa5cc22d99d5d6d7d99c3b5b5f7e7f8034c22f1b5d62a1adecd2ed005d9b80"
+checksum = "be9e29b682b38f8af2a89f960455054ab1a9f5a06822f6f3500637ad9fa57def"
 dependencies = [
- "heck",
+ "heck 0.5.0",
  "proc-macro2",
  "quote",
  "syn 2.0.52",
 ]
 
+[[package]]
+name = "measured-process"
+version = "0.0.20"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a20849acdd04c5d6a88f565559044546904648a1842a2937cfff0b48b4ca7ef2"
+dependencies = [
+ "libc",
+ "measured",
+ "procfs 0.16.0",
+]
+
 [[package]]
 name = "memchr"
 version = "2.6.4"
@@ -2953,8 +3010,10 @@ version = "0.1.0"
 dependencies = [
  "chrono",
  "libc",
+ "measured",
+ "measured-process",
  "once_cell",
- "procfs",
+ "procfs 0.14.2",
  "prometheus",
  "rand 0.8.5",
  "rand_distr",
@@ -3504,12 +3563,17 @@ dependencies = [
  "camino",
  "clap",
  "git-version",
+ "humantime",
  "pageserver",
+ "pageserver_api",
  "postgres_ffi",
+ "remote_storage",
  "serde",
  "serde_json",
  "svg_fmt",
  "tokio",
+ "tokio-util",
+ "toml_edit",
  "utils",
  "workspace_hack",
 ]
@@ -3545,7 +3609,7 @@ dependencies = [
  "hex-literal",
  "humantime",
  "humantime-serde",
- "hyper",
+ "hyper 0.14.26",
  "itertools",
  "leaky-bucket",
  "md5",
@@ -3564,7 +3628,7 @@ dependencies = [
  "postgres_connection",
  "postgres_ffi",
  "pq_proto",
- "procfs",
+ "procfs 0.14.2",
  "rand 0.8.5",
  "regex",
  "remote_storage",
@@ -3655,7 +3719,6 @@ dependencies = [
  "anyhow",
  "async-compression",
  "async-stream",
- "async-trait",
  "byteorder",
  "bytes",
  "chrono",
@@ -4125,6 +4188,29 @@ dependencies = [
  "rustix 0.36.16",
 ]
 
+[[package]]
+name = "procfs"
+version = "0.16.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "731e0d9356b0c25f16f33b5be79b1c57b562f141ebfcdb0ad8ac2c13a24293b4"
+dependencies = [
+ "bitflags 2.4.1",
+ "hex",
+ "lazy_static",
+ "procfs-core",
+ "rustix 0.38.28",
+]
+
+[[package]]
+name = "procfs-core"
+version = "0.16.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2d3554923a69f4ce04c4a754260c338f505ce22642d3830e049a399fc2059a29"
+dependencies = [
+ "bitflags 2.4.1",
+ "hex",
+]
+
 [[package]]
 name = "prometheus"
 version = "0.13.3"
@@ -4137,7 +4223,7 @@ dependencies = [
  "libc",
  "memchr",
  "parking_lot 0.12.1",
- "procfs",
+ "procfs 0.14.2",
  "thiserror",
 ]
 
@@ -4158,7 +4244,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "119533552c9a7ffacc21e099c24a0ac8bb19c2a2a3f363de84cd9b844feab270"
 dependencies = [
  "bytes",
- "heck",
+ "heck 0.4.1",
  "itertools",
  "lazy_static",
  "log",
@@ -4202,6 +4288,7 @@ dependencies = [
  "anyhow",
  "async-compression",
  "async-trait",
+ "atomic-take",
  "aws-config",
  "aws-sdk-iam",
  "aws-sigv4",
@@ -4225,9 +4312,12 @@ dependencies = [
  "hmac",
  "hostname",
  "http 1.1.0",
+ "http-body-util",
  "humantime",
- "hyper",
+ "hyper 0.14.26",
+ "hyper 1.2.0",
  "hyper-tungstenite",
+ "hyper-util",
  "ipnet",
  "itertools",
  "lasso",
@@ -4560,7 +4650,7 @@ dependencies = [
  "futures-util",
  "http-types",
  "humantime",
- "hyper",
+ "hyper 0.14.26",
  "itertools",
  "metrics",
  "once_cell",
@@ -4590,10 +4680,10 @@ dependencies = [
  "encoding_rs",
  "futures-core",
  "futures-util",
- "h2",
+ "h2 0.3.26",
  "http 0.2.9",
- "http-body",
- "hyper",
+ "http-body 0.4.5",
+ "hyper 0.14.26",
  "hyper-rustls",
  "hyper-tls",
  "ipnet",
@@ -4651,7 +4741,7 @@ dependencies = [
  "futures",
  "getrandom 0.2.11",
  "http 0.2.9",
- "hyper",
+ "hyper 0.14.26",
  "parking_lot 0.11.2",
  "reqwest",
  "reqwest-middleware",
@@ -4738,7 +4828,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "496c1d3718081c45ba9c31fbfc07417900aa96f4070ff90dc29961836b7a9945"
 dependencies = [
  "http 0.2.9",
- "hyper",
+ "hyper 0.14.26",
  "lazy_static",
  "percent-encoding",
  "regex",
@@ -4850,6 +4940,19 @@ dependencies = [
  "windows-sys 0.48.0",
 ]
 
+[[package]]
+name = "rustix"
+version = "0.38.28"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "72e572a5e8ca657d7366229cdde4bd14c4eb5499a9573d4d366fe1b599daa316"
+dependencies = [
+ "bitflags 2.4.1",
+ "errno",
+ "libc",
+ "linux-raw-sys 0.4.13",
+ "windows-sys 0.52.0",
+]
+
 [[package]]
 name = "rustls"
 version = "0.21.9"
@@ -5030,7 +5133,7 @@ dependencies = [
  "git-version",
  "hex",
  "humantime",
- "hyper",
+ "hyper 0.14.26",
  "metrics",
  "once_cell",
  "parking_lot 0.12.1",
@@ -5515,9 +5618,9 @@ dependencies = [
 
 [[package]]
 name = "smallvec"
-version = "1.11.0"
+version = "1.13.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "62bb4feee49fdd9f707ef802e22365a35de4b7b299de4763d44bfea899442ff9"
+checksum = "e6ecd384b10a64542d77071bd64bd7b231f4ed5940fba55e98c3de13824cf3d7"
 
 [[package]]
 name = "smol_str"
@@ -5609,7 +5712,7 @@ dependencies = [
  "futures-util",
  "git-version",
  "humantime",
- "hyper",
+ "hyper 0.14.26",
  "metrics",
  "once_cell",
  "parking_lot 0.12.1",
@@ -5623,6 +5726,45 @@ dependencies = [
  "workspace_hack",
 ]
 
+[[package]]
+name = "storage_controller"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "aws-config",
+ "bytes",
+ "camino",
+ "clap",
+ "control_plane",
+ "diesel",
+ "diesel_migrations",
+ "fail",
+ "futures",
+ "git-version",
+ "hex",
+ "humantime",
+ "hyper 0.14.26",
+ "itertools",
+ "lasso",
+ "measured",
+ "metrics",
+ "once_cell",
+ "pageserver_api",
+ "pageserver_client",
+ "postgres_connection",
+ "r2d2",
+ "reqwest",
+ "routerify",
+ "serde",
+ "serde_json",
+ "thiserror",
+ "tokio",
+ "tokio-util",
+ "tracing",
+ "utils",
+ "workspace_hack",
+]
+
 [[package]]
 name = "storcon_cli"
 version = "0.1.0"
@@ -5630,7 +5772,7 @@ dependencies = [
  "anyhow",
  "clap",
  "comfy-table",
- "hyper",
+ "hyper 0.14.26",
  "pageserver_api",
  "pageserver_client",
  "reqwest",
@@ -5671,7 +5813,7 @@ version = "0.24.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1e385be0d24f186b4ce2f9982191e7101bb737312ad61c1f2f984f34bcf85d59"
 dependencies = [
- "heck",
+ "heck 0.4.1",
  "proc-macro2",
  "quote",
  "rustversion",
@@ -5799,23 +5941,23 @@ dependencies = [
 
 [[package]]
 name = "test-context"
-version = "0.1.4"
+version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "055831a02a4f5aa28fede67f2902014273eb8c21b958ac5ebbd59b71ef30dbc3"
+checksum = "6676ab8513edfd2601a108621103fdb45cac9098305ca25ec93f7023b06b05d9"
 dependencies = [
- "async-trait",
  "futures",
  "test-context-macros",
 ]
 
 [[package]]
 name = "test-context-macros"
-version = "0.1.4"
+version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8901a55b0a7a06ebc4a674dcca925170da8e613fa3b163a1df804ed10afb154d"
+checksum = "78ea17a2dc368aeca6f554343ced1b1e31f76d63683fa8016e5844bd7a5144a1"
 dependencies = [
+ "proc-macro2",
  "quote",
- "syn 1.0.109",
+ "syn 2.0.52",
 ]
 
 [[package]]
@@ -6113,7 +6255,19 @@ dependencies = [
  "futures-util",
  "log",
  "tokio",
- "tungstenite",
+ "tungstenite 0.20.1",
+]
+
+[[package]]
+name = "tokio-tungstenite"
+version = "0.21.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c83b561d025642014097b66e6c1bb422783339e0909e4429cde4749d1990bc38"
+dependencies = [
+ "futures-util",
+ "log",
+ "tokio",
+ "tungstenite 0.21.0",
 ]
 
 [[package]]
@@ -6180,10 +6334,10 @@ dependencies = [
  "bytes",
  "futures-core",
  "futures-util",
- "h2",
+ "h2 0.3.26",
  "http 0.2.9",
- "http-body",
- "hyper",
+ "http-body 0.4.5",
+ "hyper 0.14.26",
  "hyper-timeout",
  "percent-encoding",
  "pin-project",
@@ -6369,7 +6523,7 @@ dependencies = [
 name = "tracing-utils"
 version = "0.1.0"
 dependencies = [
- "hyper",
+ "hyper 0.14.26",
  "opentelemetry",
  "opentelemetry-otlp",
  "opentelemetry-semantic-conventions",
@@ -6406,6 +6560,25 @@ dependencies = [
  "utf-8",
 ]
 
+[[package]]
+name = "tungstenite"
+version = "0.21.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9ef1a641ea34f399a848dea702823bbecfb4c486f911735368f1f137cb8257e1"
+dependencies = [
+ "byteorder",
+ "bytes",
+ "data-encoding",
+ "http 1.1.0",
+ "httparse",
+ "log",
+ "rand 0.8.5",
+ "sha1",
+ "thiserror",
+ "url",
+ "utf-8",
+]
+
 [[package]]
 name = "twox-hash"
 version = "1.6.3"
@@ -6570,7 +6743,8 @@ dependencies = [
  "heapless",
  "hex",
  "hex-literal",
- "hyper",
+ "humantime",
+ "hyper 0.14.26",
  "jsonwebtoken",
  "leaky-bucket",
  "metrics",
@@ -6930,6 +7104,15 @@ dependencies = [
  "windows-targets 0.48.0",
 ]
 
+[[package]]
+name = "windows-sys"
+version = "0.52.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "282be5f36a8ce781fad8c8ae18fa3f9beff57ec1b52cb3de0789201425d9a33d"
+dependencies = [
+ "windows-targets 0.52.4",
+]
+
 [[package]]
 name = "windows-targets"
 version = "0.42.2"
@@ -6960,6 +7143,21 @@ dependencies = [
  "windows_x86_64_msvc 0.48.0",
 ]
 
+[[package]]
+name = "windows-targets"
+version = "0.52.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7dd37b7e5ab9018759f893a1952c9420d060016fc19a472b4bb20d1bdd694d1b"
+dependencies = [
+ "windows_aarch64_gnullvm 0.52.4",
+ "windows_aarch64_msvc 0.52.4",
+ "windows_i686_gnu 0.52.4",
+ "windows_i686_msvc 0.52.4",
+ "windows_x86_64_gnu 0.52.4",
+ "windows_x86_64_gnullvm 0.52.4",
+ "windows_x86_64_msvc 0.52.4",
+]
+
 [[package]]
 name = "windows_aarch64_gnullvm"
 version = "0.42.2"
@@ -6972,6 +7170,12 @@ version = "0.48.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "91ae572e1b79dba883e0d315474df7305d12f569b400fcf90581b06062f7e1bc"
 
+[[package]]
+name = "windows_aarch64_gnullvm"
+version = "0.52.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bcf46cf4c365c6f2d1cc93ce535f2c8b244591df96ceee75d8e83deb70a9cac9"
+
 [[package]]
 name = "windows_aarch64_msvc"
 version = "0.42.2"
@@ -6984,6 +7188,12 @@ version = "0.48.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b2ef27e0d7bdfcfc7b868b317c1d32c641a6fe4629c171b8928c7b08d98d7cf3"
 
+[[package]]
+name = "windows_aarch64_msvc"
+version = "0.52.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "da9f259dd3bcf6990b55bffd094c4f7235817ba4ceebde8e6d11cd0c5633b675"
+
 [[package]]
 name = "windows_i686_gnu"
 version = "0.42.2"
@@ -6996,6 +7206,12 @@ version = "0.48.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "622a1962a7db830d6fd0a69683c80a18fda201879f0f447f065a3b7467daa241"
 
+[[package]]
+name = "windows_i686_gnu"
+version = "0.52.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b474d8268f99e0995f25b9f095bc7434632601028cf86590aea5c8a5cb7801d3"
+
 [[package]]
 name = "windows_i686_msvc"
 version = "0.42.2"
@@ -7008,6 +7224,12 @@ version = "0.48.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4542c6e364ce21bf45d69fdd2a8e455fa38d316158cfd43b3ac1c5b1b19f8e00"
 
+[[package]]
+name = "windows_i686_msvc"
+version = "0.52.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1515e9a29e5bed743cb4415a9ecf5dfca648ce85ee42e15873c3cd8610ff8e02"
+
 [[package]]
 name = "windows_x86_64_gnu"
 version = "0.42.2"
@@ -7020,6 +7242,12 @@ version = "0.48.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ca2b8a661f7628cbd23440e50b05d705db3686f894fc9580820623656af974b1"
 
+[[package]]
+name = "windows_x86_64_gnu"
+version = "0.52.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5eee091590e89cc02ad514ffe3ead9eb6b660aedca2183455434b93546371a03"
+
 [[package]]
 name = "windows_x86_64_gnullvm"
 version = "0.42.2"
@@ -7032,6 +7260,12 @@ version = "0.48.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7896dbc1f41e08872e9d5e8f8baa8fdd2677f29468c4e156210174edc7f7b953"
 
+[[package]]
+name = "windows_x86_64_gnullvm"
+version = "0.52.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "77ca79f2451b49fa9e2af39f0747fe999fcda4f5e241b2898624dca97a1f2177"
+
 [[package]]
 name = "windows_x86_64_msvc"
 version = "0.42.2"
@@ -7044,6 +7278,12 @@ version = "0.48.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1a515f5799fe4961cb532f983ce2b23082366b898e52ffbce459c86f67c8378a"
 
+[[package]]
+name = "windows_x86_64_msvc"
+version = "0.52.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32b752e52a2da0ddfbdbcc6fceadfeede4c939ed16d13e648833a61dfb611ed8"
+
 [[package]]
 name = "winnow"
 version = "0.4.6"
@@ -7092,11 +7332,10 @@ dependencies = [
  "futures-sink",
  "futures-util",
  "getrandom 0.2.11",
- "hashbrown 0.13.2",
  "hashbrown 0.14.0",
  "hex",
  "hmac",
- "hyper",
+ "hyper 0.14.26",
  "indexmap 1.9.3",
  "itertools",
  "libc",
@@ -7134,7 +7373,6 @@ dependencies = [
  "tower",
  "tracing",
  "tracing-core",
- "tungstenite",
  "url",
  "uuid",
  "zeroize",

Cargo.toml

@@ -3,7 +3,6 @@ resolver = "2"
 members = [
     "compute_tools",
     "control_plane",
-    "control_plane/attachment_service",
     "control_plane/storcon_cli",
     "pageserver",
     "pageserver/compaction",
@@ -13,6 +12,7 @@ members = [
     "proxy",
     "safekeeper",
     "storage_broker",
+    "storage_controller",
     "s3_scrubber",
     "workspace_hack",
     "trace",
@@ -44,6 +44,7 @@ license = "Apache-2.0"
 anyhow = { version = "1.0", features = ["backtrace"] }
 arc-swap = "1.6"
 async-compression = { version = "0.4.0", features = ["tokio", "gzip", "zstd"] }
+atomic-take = "1.1.0"
 azure_core = "0.18"
 azure_identity = "0.18"
 azure_storage = "0.18"
@@ -97,7 +98,7 @@ http-types = { version = "2", default-features = false }
 humantime = "2.1"
 humantime-serde = "1.1.1"
 hyper = "0.14"
-hyper-tungstenite = "0.11"
+hyper-tungstenite = "0.13.0"
 inotify = "0.10.2"
 ipnet = "2.9.0"
 itertools = "0.10"
@@ -106,7 +107,8 @@ lasso = "0.7"
 leaky-bucket = "1.0.1"
 libc = "0.2"
 md5 = "0.7.0"
-measured = { version = "0.0.13", features=["default", "lasso"] }
+measured = { version = "0.0.20", features=["lasso"] }
+measured-process = { version = "0.0.20" }
 memoffset = "0.8"
 native-tls = "0.2"
 nix = { version = "0.27", features = ["fs", "process", "socket", "signal", "poll"] }
@@ -159,7 +161,7 @@ svg_fmt = "0.4.1"
 sync_wrapper = "0.1.2"
 tar = "0.4"
 task-local-extensions = "0.1.4"
-test-context = "0.1"
+test-context = "0.3"
 thiserror = "1.0"
 tikv-jemallocator = "0.5"
 tikv-jemalloc-ctl = "0.5"

compute_tools/src/config.rs

@@ -6,8 +6,8 @@ use std::path::Path;
 use anyhow::Result;
 
 use crate::pg_helpers::escape_conf_value;
-use crate::pg_helpers::PgOptionsSerialize;
-use compute_api::spec::{ComputeMode, ComputeSpec};
+use crate::pg_helpers::{GenericOptionExt, PgOptionsSerialize};
+use compute_api::spec::{ComputeMode, ComputeSpec, GenericOption};
 
 /// Check that `line` is inside a text file and put it there if it is not.
 /// Create file if it doesn't exist.
@@ -92,6 +92,27 @@ pub fn write_postgres_conf(
         }
     }
 
+    if cfg!(target_os = "linux") {
+        // Check /proc/sys/vm/overcommit_memory -- if it equals 2 (i.e. linux memory overcommit is
+        // disabled), then the control plane has enabled swap and we should set
+        // dynamic_shared_memory_type = 'mmap'.
+        //
+        // This is (maybe?) temporary - for more, see https://github.com/neondatabase/cloud/issues/12047.
+        let overcommit_memory_contents = std::fs::read_to_string("/proc/sys/vm/overcommit_memory")
+            // ignore any errors - they may be expected to occur under certain situations (e.g. when
+            // not running in Linux).
+            .unwrap_or_else(|_| String::new());
+        if overcommit_memory_contents.trim() == "2" {
+            let opt = GenericOption {
+                name: "dynamic_shared_memory_type".to_owned(),
+                value: Some("mmap".to_owned()),
+                vartype: "enum".to_owned(),
+            };
+
+            write!(file, "{}", opt.to_pg_setting())?;
+        }
+    }
+
     // If there are any extra options in the 'settings' field, append those
     if spec.cluster.settings.is_some() {
         writeln!(file, "# Managed by compute_ctl: begin")?;

compute_tools/src/pg_helpers.rs

@@ -44,7 +44,7 @@ pub fn escape_conf_value(s: &str) -> String {
     format!("'{}'", res)
 }
 
-trait GenericOptionExt {
+pub trait GenericOptionExt {
     fn to_pg_option(&self) -> String;
     fn to_pg_setting(&self) -> String;
 }

control_plane/src/background_process.rs

@@ -86,7 +86,10 @@ where
         .stdout(process_log_file)
         .stderr(same_file_for_stderr)
         .args(args);
-    let filled_cmd = fill_remote_storage_secrets_vars(fill_rust_env_vars(background_command));
+
+    let filled_cmd = fill_env_vars_prefixed_neon(fill_remote_storage_secrets_vars(
+        fill_rust_env_vars(background_command),
+    ));
     filled_cmd.envs(envs);
 
     let pid_file_to_check = match &initial_pid_file {
@@ -268,6 +271,15 @@ fn fill_remote_storage_secrets_vars(mut cmd: &mut Command) -> &mut Command {
     cmd
 }
 
+fn fill_env_vars_prefixed_neon(mut cmd: &mut Command) -> &mut Command {
+    for (var, val) in std::env::vars() {
+        if var.starts_with("NEON_PAGESERVER_") {
+            cmd = cmd.env(var, val);
+        }
+    }
+    cmd
+}
+
 /// Add a `pre_exec` to the cmd that, inbetween fork() and exec(),
 /// 1. Claims a pidfile with a fcntl lock on it and
 /// 2. Sets up the pidfile's file descriptor so that it (and the lock)

control_plane/storcon_cli/src/main.rs

@@ -223,7 +223,7 @@ impl Client {
         }
     }
 
-    /// Simple HTTP request wrapper for calling into attachment service
+    /// Simple HTTP request wrapper for calling into storage controller
     async fn dispatch<RQ, RS>(
         &self,
         method: hyper::Method,

diesel.toml

@@ -2,8 +2,8 @@
 # see https://diesel.rs/guides/configuring-diesel-cli
 
 [print_schema]
-file = "control_plane/attachment_service/src/schema.rs"
+file = "storage_controller/src/schema.rs"
 custom_type_derives = ["diesel::query_builder::QueryId"]
 
 [migrations_directory]
-dir = "control_plane/attachment_service/migrations"
+dir = "storage_controller/migrations"

docs/sourcetree.md

@@ -7,6 +7,11 @@ Below you will find a brief overview of each subdir in the source tree in alphab
 Neon storage broker, providing messaging between safekeepers and pageservers.
 [storage_broker.md](./storage_broker.md)
 
+`storage_controller`:
+
+Neon storage controller, manages a cluster of pageservers and exposes an API that enables
+managing a many-sharded tenant as a single entity.
+
 `/control_plane`:
 
 Local control plane.

libs/metrics/Cargo.toml

@@ -10,11 +10,13 @@ libc.workspace = true
 once_cell.workspace = true
 chrono.workspace = true
 twox-hash.workspace = true
+measured.workspace = true
 
 workspace_hack.workspace = true
 
 [target.'cfg(target_os = "linux")'.dependencies]
 procfs.workspace = true
+measured-process.workspace = true
 
 [dev-dependencies]
 rand = "0.8"

libs/metrics/src/lib.rs

@@ -4,6 +4,17 @@
 //! a default registry.
 #![deny(clippy::undocumented_unsafe_blocks)]
 
+use measured::{
+    label::{LabelGroupVisitor, LabelName, NoLabels},
+    metric::{
+        counter::CounterState,
+        gauge::GaugeState,
+        group::{Encoding, MetricValue},
+        name::{MetricName, MetricNameEncoder},
+        MetricEncoding, MetricFamilyEncoding,
+    },
+    FixedCardinalityLabel, LabelGroup, MetricGroup,
+};
 use once_cell::sync::Lazy;
 use prometheus::core::{
     Atomic, AtomicU64, Collector, GenericCounter, GenericCounterVec, GenericGauge, GenericGaugeVec,
@@ -11,6 +22,7 @@ use prometheus::core::{
 pub use prometheus::opts;
 pub use prometheus::register;
 pub use prometheus::Error;
+use prometheus::Registry;
 pub use prometheus::{core, default_registry, proto};
 pub use prometheus::{exponential_buckets, linear_buckets};
 pub use prometheus::{register_counter_vec, Counter, CounterVec};
@@ -23,7 +35,6 @@ pub use prometheus::{register_int_counter_vec, IntCounterVec};
 pub use prometheus::{register_int_gauge, IntGauge};
 pub use prometheus::{register_int_gauge_vec, IntGaugeVec};
 pub use prometheus::{Encoder, TextEncoder};
-use prometheus::{Registry, Result};
 
 pub mod launch_timestamp;
 mod wrappers;
@@ -59,7 +70,7 @@ static INTERNAL_REGISTRY: Lazy<Registry> = Lazy::new(Registry::new);
 /// Register a collector in the internal registry. MUST be called before the first call to `gather()`.
 /// Otherwise, we can have a deadlock in the `gather()` call, trying to register a new collector
 /// while holding the lock.
-pub fn register_internal(c: Box<dyn Collector>) -> Result<()> {
+pub fn register_internal(c: Box<dyn Collector>) -> prometheus::Result<()> {
     INTERNAL_REGISTRY.register(c)
 }
 
@@ -96,6 +107,127 @@ pub const DISK_WRITE_SECONDS_BUCKETS: &[f64] = &[
     0.000_050, 0.000_100, 0.000_500, 0.001, 0.003, 0.005, 0.01, 0.05, 0.1, 0.3, 0.5,
 ];
 
+pub struct BuildInfo {
+    pub revision: &'static str,
+    pub build_tag: &'static str,
+}
+
+// todo: allow label group without the set
+impl LabelGroup for BuildInfo {
+    fn visit_values(&self, v: &mut impl LabelGroupVisitor) {
+        const REVISION: &LabelName = LabelName::from_str("revision");
+        v.write_value(REVISION, &self.revision);
+        const BUILD_TAG: &LabelName = LabelName::from_str("build_tag");
+        v.write_value(BUILD_TAG, &self.build_tag);
+    }
+}
+
+impl<T: Encoding> MetricFamilyEncoding<T> for BuildInfo
+where
+    GaugeState: MetricEncoding<T>,
+{
+    fn collect_family_into(
+        &self,
+        name: impl measured::metric::name::MetricNameEncoder,
+        enc: &mut T,
+    ) -> Result<(), T::Err> {
+        enc.write_help(&name, "Build/version information")?;
+        GaugeState::write_type(&name, enc)?;
+        GaugeState {
+            count: std::sync::atomic::AtomicI64::new(1),
+        }
+        .collect_into(&(), self, name, enc)
+    }
+}
+
+#[derive(MetricGroup)]
+#[metric(new(build_info: BuildInfo))]
+pub struct NeonMetrics {
+    #[cfg(target_os = "linux")]
+    #[metric(namespace = "process")]
+    #[metric(init = measured_process::ProcessCollector::for_self())]
+    process: measured_process::ProcessCollector,
+
+    #[metric(namespace = "libmetrics")]
+    #[metric(init = LibMetrics::new(build_info))]
+    libmetrics: LibMetrics,
+}
+
+#[derive(MetricGroup)]
+#[metric(new(build_info: BuildInfo))]
+pub struct LibMetrics {
+    #[metric(init = build_info)]
+    build_info: BuildInfo,
+
+    #[metric(flatten)]
+    rusage: Rusage,
+
+    serve_count: CollectionCounter,
+}
+
+fn write_gauge<Enc: Encoding>(
+    x: i64,
+    labels: impl LabelGroup,
+    name: impl MetricNameEncoder,
+    enc: &mut Enc,
+) -> Result<(), Enc::Err> {
+    enc.write_metric_value(name, labels, MetricValue::Int(x))
+}
+
+#[derive(Default)]
+struct Rusage;
+
+#[derive(FixedCardinalityLabel, Clone, Copy)]
+#[label(singleton = "io_operation")]
+enum IoOp {
+    Read,
+    Write,
+}
+
+impl<T: Encoding> MetricGroup<T> for Rusage
+where
+    GaugeState: MetricEncoding<T>,
+{
+    fn collect_group_into(&self, enc: &mut T) -> Result<(), T::Err> {
+        const DISK_IO: &MetricName = MetricName::from_str("disk_io_bytes_total");
+        const MAXRSS: &MetricName = MetricName::from_str("maxrss_kb");
+
+        let ru = get_rusage_stats();
+
+        enc.write_help(
+            DISK_IO,
+            "Bytes written and read from disk, grouped by the operation (read|write)",
+        )?;
+        GaugeState::write_type(DISK_IO, enc)?;
+        write_gauge(ru.ru_inblock * BYTES_IN_BLOCK, IoOp::Read, DISK_IO, enc)?;
+        write_gauge(ru.ru_oublock * BYTES_IN_BLOCK, IoOp::Write, DISK_IO, enc)?;
+
+        enc.write_help(MAXRSS, "Memory usage (Maximum Resident Set Size)")?;
+        GaugeState::write_type(MAXRSS, enc)?;
+        write_gauge(ru.ru_maxrss, IoOp::Read, MAXRSS, enc)?;
+
+        Ok(())
+    }
+}
+
+#[derive(Default)]
+struct CollectionCounter(CounterState);
+
+impl<T: Encoding> MetricFamilyEncoding<T> for CollectionCounter
+where
+    CounterState: MetricEncoding<T>,
+{
+    fn collect_family_into(
+        &self,
+        name: impl measured::metric::name::MetricNameEncoder,
+        enc: &mut T,
+    ) -> Result<(), T::Err> {
+        self.0.inc();
+        enc.write_help(&name, "Number of metric requests made")?;
+        self.0.collect_into(&(), NoLabels, name, enc)
+    }
+}
+
 pub fn set_build_info_metric(revision: &str, build_tag: &str) {
     let metric = register_int_gauge_vec!(
         "libmetrics_build_info",
@@ -105,6 +237,7 @@ pub fn set_build_info_metric(revision: &str, build_tag: &str) {
     .expect("Failed to register build info metric");
     metric.with_label_values(&[revision, build_tag]).set(1);
 }
+const BYTES_IN_BLOCK: i64 = 512;
 
 // Records I/O stats in a "cross-platform" way.
 // Compiles both on macOS and Linux, but current macOS implementation always returns 0 as values for I/O stats.
@@ -117,7 +250,6 @@ pub fn set_build_info_metric(revision: &str, build_tag: &str) {
 fn update_rusage_metrics() {
     let rusage_stats = get_rusage_stats();
 
-    const BYTES_IN_BLOCK: i64 = 512;
     DISK_IO_BYTES
         .with_label_values(&["read"])
         .set(rusage_stats.ru_inblock * BYTES_IN_BLOCK);
@@ -151,6 +283,7 @@ macro_rules! register_int_counter_pair_vec {
         }
     }};
 }
+
 /// Create an [`IntCounterPair`] and registers to default registry.
 #[macro_export(local_inner_macros)]
 macro_rules! register_int_counter_pair {
@@ -188,7 +321,10 @@ impl<P: Atomic> GenericCounterPairVec<P> {
     ///
     /// An error is returned if the number of label values is not the same as the
     /// number of VariableLabels in Desc.
-    pub fn get_metric_with_label_values(&self, vals: &[&str]) -> Result<GenericCounterPair<P>> {
+    pub fn get_metric_with_label_values(
+        &self,
+        vals: &[&str],
+    ) -> prometheus::Result<GenericCounterPair<P>> {
         Ok(GenericCounterPair {
             inc: self.inc.get_metric_with_label_values(vals)?,
             dec: self.dec.get_metric_with_label_values(vals)?,
@@ -201,7 +337,7 @@ impl<P: Atomic> GenericCounterPairVec<P> {
         self.get_metric_with_label_values(vals).unwrap()
     }
 
-    pub fn remove_label_values(&self, res: &mut [Result<()>; 2], vals: &[&str]) {
+    pub fn remove_label_values(&self, res: &mut [prometheus::Result<()>; 2], vals: &[&str]) {
         res[0] = self.inc.remove_label_values(vals);
         res[1] = self.dec.remove_label_values(vals);
     }

libs/pageserver_api/src/controller_api.rs

@@ -2,7 +2,7 @@ use std::str::FromStr;
 
 /// Request/response types for the storage controller
 /// API (`/control/v1` prefix).  Implemented by the server
-/// in [`attachment_service::http`]
+/// in [`storage_controller::http`]
 use serde::{Deserialize, Serialize};
 use utils::id::{NodeId, TenantId};
 

libs/pageserver_api/src/models.rs

@@ -20,6 +20,7 @@ use utils::{
     history_buffer::HistoryBufferWithDropCounter,
     id::{NodeId, TenantId, TimelineId},
     lsn::Lsn,
+    serde_system_time,
 };
 
 use crate::controller_api::PlacementPolicy;
@@ -758,11 +759,7 @@ pub struct WalRedoManagerStatus {
 #[derive(Default, Debug, Serialize, Deserialize, Clone)]
 pub struct SecondaryProgress {
     /// The remote storage LastModified time of the heatmap object we last downloaded.
-    #[serde(
-        serialize_with = "opt_ser_rfc3339_millis",
-        deserialize_with = "opt_deser_rfc3339_millis"
-    )]
-    pub heatmap_mtime: Option<SystemTime>,
+    pub heatmap_mtime: Option<serde_system_time::SystemTime>,
 
     /// The number of layers currently on-disk
     pub layers_downloaded: usize,
@@ -775,29 +772,6 @@ pub struct SecondaryProgress {
     pub bytes_total: u64,
 }
 
-fn opt_ser_rfc3339_millis<S: serde::Serializer>(
-    ts: &Option<SystemTime>,
-    serializer: S,
-) -> Result<S::Ok, S::Error> {
-    match ts {
-        Some(ts) => serializer.collect_str(&humantime::format_rfc3339_millis(*ts)),
-        None => serializer.serialize_none(),
-    }
-}
-
-fn opt_deser_rfc3339_millis<'de, D>(deserializer: D) -> Result<Option<SystemTime>, D::Error>
-where
-    D: serde::de::Deserializer<'de>,
-{
-    let s: Option<String> = serde::de::Deserialize::deserialize(deserializer)?;
-    match s {
-        None => Ok(None),
-        Some(s) => humantime::parse_rfc3339(&s)
-            .map_err(serde::de::Error::custom)
-            .map(Some),
-    }
-}
-
 pub mod virtual_file {
     #[derive(
         Copy,

libs/pageserver_api/src/models/utilization.rs

@@ -1,4 +1,4 @@
-use std::time::SystemTime;
+use utils::serde_system_time::SystemTime;
 
 /// Pageserver current utilization and scoring for how good candidate the pageserver would be for
 /// the next tenant.
@@ -21,28 +21,9 @@ pub struct PageserverUtilization {
     /// When was this snapshot captured, pageserver local time.
     ///
     /// Use millis to give confidence that the value is regenerated often enough.
-    #[serde(
-        serialize_with = "ser_rfc3339_millis",
-        deserialize_with = "deser_rfc3339_millis"
-    )]
     pub captured_at: SystemTime,
 }
 
-fn ser_rfc3339_millis<S: serde::Serializer>(
-    ts: &SystemTime,
-    serializer: S,
-) -> Result<S::Ok, S::Error> {
-    serializer.collect_str(&humantime::format_rfc3339_millis(*ts))
-}
-
-fn deser_rfc3339_millis<'de, D>(deserializer: D) -> Result<SystemTime, D::Error>
-where
-    D: serde::de::Deserializer<'de>,
-{
-    let s: String = serde::de::Deserialize::deserialize(deserializer)?;
-    humantime::parse_rfc3339(&s).map_err(serde::de::Error::custom)
-}
-
 /// openapi knows only `format: int64`, so avoid outputting a non-parseable value by generated clients.
 ///
 /// Instead of newtype, use this because a newtype would get require handling deserializing values
@@ -69,7 +50,9 @@ mod tests {
             disk_usage_bytes: u64::MAX,
             free_space_bytes: 0,
             utilization_score: u64::MAX,
-            captured_at: SystemTime::UNIX_EPOCH + Duration::from_secs(1708509779),
+            captured_at: SystemTime(
+                std::time::SystemTime::UNIX_EPOCH + Duration::from_secs(1708509779),
+            ),
         };
 
         let s = serde_json::to_string(&doc).unwrap();

libs/remote_storage/tests/test_real_azure.rs

@@ -57,7 +57,6 @@ enum MaybeEnabledStorage {
     Disabled,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorage {
     async fn setup() -> Self {
         ensure_logging_ready();
@@ -86,7 +85,6 @@ struct AzureWithTestBlobs {
     remote_blobs: HashSet<RemotePath>,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorageWithTestBlobs {
     async fn setup() -> Self {
         ensure_logging_ready();
@@ -148,7 +146,6 @@ struct AzureWithSimpleTestBlobs {
     remote_blobs: HashSet<RemotePath>,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorageWithSimpleTestBlobs {
     async fn setup() -> Self {
         ensure_logging_ready();

libs/remote_storage/tests/test_real_s3.rs

@@ -219,7 +219,6 @@ enum MaybeEnabledStorage {
     Disabled,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorage {
     async fn setup() -> Self {
         ensure_logging_ready();
@@ -248,7 +247,6 @@ struct S3WithTestBlobs {
     remote_blobs: HashSet<RemotePath>,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorageWithTestBlobs {
     async fn setup() -> Self {
         ensure_logging_ready();
@@ -310,7 +308,6 @@ struct S3WithSimpleTestBlobs {
     remote_blobs: HashSet<RemotePath>,
 }
 
-#[async_trait::async_trait]
 impl AsyncTestContext for MaybeEnabledStorageWithSimpleTestBlobs {
     async fn setup() -> Self {
         ensure_logging_ready();

libs/utils/Cargo.toml

@@ -22,6 +22,7 @@ camino.workspace = true
 chrono.workspace = true
 heapless.workspace = true
 hex = { workspace = true, features = ["serde"] }
+humantime.workspace = true
 hyper = { workspace = true, features = ["full"] }
 fail.workspace = true
 futures = { workspace = true}

libs/utils/src/env.rs

@@ -0,0 +1,21 @@
+//! Wrapper around `std::env::var` for parsing environment variables.
+
+use std::{fmt::Display, str::FromStr};
+
+pub fn var<V, E>(varname: &str) -> Option<V>
+where
+    V: FromStr<Err = E>,
+    E: Display,
+{
+    match std::env::var(varname) {
+        Ok(s) => Some(
+            s.parse()
+                .map_err(|e| format!("failed to parse env var {varname}: {e:#}"))
+                .unwrap(),
+        ),
+        Err(std::env::VarError::NotPresent) => None,
+        Err(std::env::VarError::NotUnicode(_)) => {
+            panic!("env var {varname} is not unicode")
+        }
+    }
+}

libs/utils/src/lib.rs

@@ -63,6 +63,7 @@ pub mod measured_stream;
 
 pub mod serde_percent;
 pub mod serde_regex;
+pub mod serde_system_time;
 
 pub mod pageserver_feedback;
 
@@ -89,6 +90,8 @@ pub mod yielding_loop;
 
 pub mod zstd;
 
+pub mod env;
+
 /// 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/serde_system_time.rs

@@ -0,0 +1,55 @@
+//! A `serde::{Deserialize,Serialize}` type for SystemTime with RFC3339 format and millisecond precision.
+
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, serde::Serialize, serde::Deserialize)]
+#[serde(transparent)]
+pub struct SystemTime(
+    #[serde(
+        deserialize_with = "deser_rfc3339_millis",
+        serialize_with = "ser_rfc3339_millis"
+    )]
+    pub std::time::SystemTime,
+);
+
+fn ser_rfc3339_millis<S: serde::ser::Serializer>(
+    ts: &std::time::SystemTime,
+    serializer: S,
+) -> Result<S::Ok, S::Error> {
+    serializer.collect_str(&humantime::format_rfc3339_millis(*ts))
+}
+
+fn deser_rfc3339_millis<'de, D>(deserializer: D) -> Result<std::time::SystemTime, D::Error>
+where
+    D: serde::de::Deserializer<'de>,
+{
+    let s: String = serde::de::Deserialize::deserialize(deserializer)?;
+    humantime::parse_rfc3339(&s).map_err(serde::de::Error::custom)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Helper function to make a SystemTime have millisecond precision by truncating additional nanoseconds.
+    fn to_millisecond_precision(time: SystemTime) -> SystemTime {
+        match time.0.duration_since(std::time::SystemTime::UNIX_EPOCH) {
+            Ok(duration) => {
+                let total_millis = duration.as_secs() * 1_000 + u64::from(duration.subsec_millis());
+                SystemTime(
+                    std::time::SystemTime::UNIX_EPOCH
+                        + std::time::Duration::from_millis(total_millis),
+                )
+            }
+            Err(_) => time,
+        }
+    }
+
+    #[test]
+    fn test_serialize_deserialize() {
+        let input = SystemTime(std::time::SystemTime::now());
+        let expected_serialized = format!("\"{}\"", humantime::format_rfc3339_millis(input.0));
+        let serialized = serde_json::to_string(&input).unwrap();
+        assert_eq!(expected_serialized, serialized);
+        let deserialized: SystemTime = serde_json::from_str(&expected_serialized).unwrap();
+        assert_eq!(to_millisecond_precision(input), deserialized);
+    }
+}

pageserver/benches/bench_walredo.rs

@@ -27,25 +27,25 @@
 //!
 //! # Reference Numbers
 //!
-//! 2024-03-20 on i3en.3xlarge
+//! 2024-04-04 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]
+//! short/1                 time:   [25.925 µs 26.060 µs 26.209 µs]
+//! short/2                 time:   [31.277 µs 31.483 µs 31.722 µs]
+//! short/4                 time:   [45.496 µs 45.831 µs 46.182 µs]
+//! short/8                 time:   [84.298 µs 84.920 µs 85.566 µs]
+//! short/16                time:   [185.04 µs 186.41 µs 187.88 µs]
+//! short/32                time:   [385.01 µs 386.77 µs 388.70 µs]
+//! short/64                time:   [770.24 µs 773.04 µs 776.04 µs]
+//! short/128               time:   [1.5017 ms 1.5064 ms 1.5113 ms]
+//! medium/1                time:   [106.65 µs 107.20 µs 107.85 µs]
+//! medium/2                time:   [153.28 µs 154.24 µs 155.56 µs]
+//! medium/4                time:   [325.67 µs 327.01 µs 328.71 µs]
+//! medium/8                time:   [646.82 µs 650.17 µs 653.91 µs]
+//! medium/16               time:   [1.2645 ms 1.2701 ms 1.2762 ms]
+//! medium/32               time:   [2.4409 ms 2.4550 ms 2.4692 ms]
+//! medium/64               time:   [4.6814 ms 4.7114 ms 4.7408 ms]
+//! medium/128              time:   [8.7790 ms 8.9037 ms 9.0282 ms]
 //! ```
 
 use bytes::{Buf, Bytes};

pageserver/client/src/mgmt_api.rs

@@ -128,12 +128,12 @@ impl Client {
 
     pub async fn timeline_info(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
         force_await_logical_size: ForceAwaitLogicalSize,
     ) -> Result<pageserver_api::models::TimelineInfo> {
         let uri = format!(
-            "{}/v1/tenant/{tenant_id}/timeline/{timeline_id}",
+            "{}/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}",
             self.mgmt_api_endpoint
         );
 
@@ -151,11 +151,11 @@ impl Client {
 
     pub async fn keyspace(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
     ) -> Result<pageserver_api::models::partitioning::Partitioning> {
         let uri = format!(
-            "{}/v1/tenant/{tenant_id}/timeline/{timeline_id}/keyspace",
+            "{}/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/keyspace",
             self.mgmt_api_endpoint
         );
         self.get(&uri)

pageserver/compaction/Cargo.toml

@@ -11,7 +11,6 @@ default = []
 anyhow.workspace = true
 async-compression.workspace = true
 async-stream.workspace = true
-async-trait.workspace = true
 byteorder.workspace = true
 bytes.workspace = true
 chrono = { workspace = true, features = ["serde"] }

pageserver/compaction/src/helpers.rs

@@ -180,7 +180,7 @@ where
                 match top.deref_mut() {
                     LazyLoadLayer::Unloaded(ref mut l) => {
                         let fut = l.load_keys(this.ctx);
-                        this.load_future.set(Some(fut));
+                        this.load_future.set(Some(Box::pin(fut)));
                         continue;
                     }
                     LazyLoadLayer::Loaded(ref mut entries) => {

pageserver/compaction/src/interface.rs

@@ -3,7 +3,6 @@
 //!
 //! All the heavy lifting is done by the create_image and create_delta
 //! functions that the implementor provides.
-use async_trait::async_trait;
 use futures::Future;
 use pageserver_api::{key::Key, keyspace::key_range_size};
 use std::ops::Range;
@@ -141,18 +140,16 @@ pub trait CompactionLayer<K: CompactionKey + ?Sized> {
 
     fn is_delta(&self) -> bool;
 }
-
-#[async_trait]
 pub trait CompactionDeltaLayer<E: CompactionJobExecutor + ?Sized>: CompactionLayer<E::Key> {
     type DeltaEntry<'a>: CompactionDeltaEntry<'a, E::Key>
     where
         Self: 'a;
 
     /// Return all keys in this delta layer.
-    async fn load_keys<'a>(
+    fn load_keys<'a>(
         &self,
         ctx: &E::RequestContext,
-    ) -> anyhow::Result<Vec<Self::DeltaEntry<'_>>>;
+    ) -> impl Future<Output = anyhow::Result<Vec<Self::DeltaEntry<'_>>>> + Send;
 }
 
 pub trait CompactionImageLayer<E: CompactionJobExecutor + ?Sized>: CompactionLayer<E::Key> {}

pageserver/compaction/src/simulator.rs

@@ -2,7 +2,6 @@ mod draw;
 
 use draw::{LayerTraceEvent, LayerTraceFile, LayerTraceOp};
 
-use async_trait::async_trait;
 use futures::StreamExt;
 use rand::Rng;
 use tracing::info;
@@ -139,7 +138,6 @@ impl interface::CompactionLayer<Key> for Arc<MockDeltaLayer> {
     }
 }
 
-#[async_trait]
 impl interface::CompactionDeltaLayer<MockTimeline> for Arc<MockDeltaLayer> {
     type DeltaEntry<'a> = MockRecord;
 

pageserver/ctl/Cargo.toml

@@ -12,9 +12,14 @@ bytes.workspace = true
 camino.workspace = true
 clap = { workspace = true, features = ["string"] }
 git-version.workspace = true
+humantime.workspace = true
 pageserver = { path = ".." }
+pageserver_api.workspace = true
+remote_storage = { path = "../../libs/remote_storage" }
 postgres_ffi.workspace = true
 tokio.workspace = true
+tokio-util.workspace = true
+toml_edit.workspace = true
 utils.workspace = true
 svg_fmt.workspace = true
 workspace_hack.workspace = true

pageserver/ctl/src/main.rs

@@ -9,6 +9,11 @@ mod index_part;
 mod layer_map_analyzer;
 mod layers;
 
+use std::{
+    str::FromStr,
+    time::{Duration, SystemTime},
+};
+
 use camino::{Utf8Path, Utf8PathBuf};
 use clap::{Parser, Subcommand};
 use index_part::IndexPartCmd;
@@ -20,8 +25,16 @@ use pageserver::{
     tenant::{dump_layerfile_from_path, metadata::TimelineMetadata},
     virtual_file,
 };
+use pageserver_api::shard::TenantShardId;
 use postgres_ffi::ControlFileData;
-use utils::{lsn::Lsn, project_git_version};
+use remote_storage::{RemotePath, RemoteStorageConfig};
+use tokio_util::sync::CancellationToken;
+use utils::{
+    id::TimelineId,
+    logging::{self, LogFormat, TracingErrorLayerEnablement},
+    lsn::Lsn,
+    project_git_version,
+};
 
 project_git_version!(GIT_VERSION);
 
@@ -43,6 +56,7 @@ enum Commands {
     #[command(subcommand)]
     IndexPart(IndexPartCmd),
     PrintLayerFile(PrintLayerFileCmd),
+    TimeTravelRemotePrefix(TimeTravelRemotePrefixCmd),
     DrawTimeline {},
     AnalyzeLayerMap(AnalyzeLayerMapCmd),
     #[command(subcommand)]
@@ -68,6 +82,26 @@ struct PrintLayerFileCmd {
     path: Utf8PathBuf,
 }
 
+/// Roll back the time for the specified prefix using S3 history.
+///
+/// The command is fairly low level and powerful. Validation is only very light,
+/// so it is more powerful, and thus potentially more dangerous.
+#[derive(Parser)]
+struct TimeTravelRemotePrefixCmd {
+    /// A configuration string for the remote_storage configuration.
+    ///
+    /// Example: `remote_storage = { bucket_name = "aws-storage-bucket-name", bucket_region = "us-east-2" }`
+    config_toml_str: String,
+    /// remote prefix to time travel recover. For safety reasons, we require it to contain
+    /// a timeline or tenant ID in the prefix.
+    prefix: String,
+    /// Timestamp to travel to. Given in format like `2024-01-20T10:45:45Z`. Assumes UTC and second accuracy.
+    travel_to: String,
+    /// Timestamp of the start of the operation, must be after any changes we want to roll back and after.
+    /// You can use a few seconds before invoking the command. Same format as `travel_to`.
+    done_if_after: Option<String>,
+}
+
 #[derive(Parser)]
 struct AnalyzeLayerMapCmd {
     /// Pageserver data path
@@ -78,6 +112,14 @@ struct AnalyzeLayerMapCmd {
 
 #[tokio::main]
 async fn main() -> anyhow::Result<()> {
+    logging::init(
+        LogFormat::Plain,
+        TracingErrorLayerEnablement::EnableWithRustLogFilter,
+        logging::Output::Stdout,
+    )?;
+
+    logging::replace_panic_hook_with_tracing_panic_hook().forget();
+
     let cli = CliOpts::parse();
 
     match cli.command {
@@ -105,6 +147,42 @@ async fn main() -> anyhow::Result<()> {
                 print_layerfile(&cmd.path).await?;
             }
         }
+        Commands::TimeTravelRemotePrefix(cmd) => {
+            let timestamp = humantime::parse_rfc3339(&cmd.travel_to)
+                .map_err(|_e| anyhow::anyhow!("Invalid time for travel_to: '{}'", cmd.travel_to))?;
+
+            let done_if_after = if let Some(done_if_after) = &cmd.done_if_after {
+                humantime::parse_rfc3339(done_if_after).map_err(|_e| {
+                    anyhow::anyhow!("Invalid time for done_if_after: '{}'", done_if_after)
+                })?
+            } else {
+                const SAFETY_MARGIN: Duration = Duration::from_secs(3);
+                tokio::time::sleep(SAFETY_MARGIN).await;
+                // Convert to string representation and back to get rid of sub-second values
+                let done_if_after = SystemTime::now();
+                tokio::time::sleep(SAFETY_MARGIN).await;
+                done_if_after
+            };
+
+            let timestamp = strip_subsecond(timestamp);
+            let done_if_after = strip_subsecond(done_if_after);
+
+            let Some(prefix) = validate_prefix(&cmd.prefix) else {
+                println!("specified prefix '{}' failed validation", cmd.prefix);
+                return Ok(());
+            };
+            let toml_document = toml_edit::Document::from_str(&cmd.config_toml_str)?;
+            let toml_item = toml_document
+                .get("remote_storage")
+                .expect("need remote_storage");
+            let config = RemoteStorageConfig::from_toml(toml_item)?.expect("incomplete config");
+            let storage = remote_storage::GenericRemoteStorage::from_config(&config);
+            let cancel = CancellationToken::new();
+            storage
+                .unwrap()
+                .time_travel_recover(Some(&prefix), timestamp, done_if_after, &cancel)
+                .await?;
+        }
     };
     Ok(())
 }
@@ -185,3 +263,89 @@ fn handle_metadata(
 
     Ok(())
 }
+
+/// Ensures that the given S3 prefix is sufficiently constrained.
+/// The command is very risky already and we don't want to expose something
+/// that allows usually unintentional and quite catastrophic time travel of
+/// an entire bucket, which would be a major catastrophy and away
+/// by only one character change (similar to "rm -r /home /username/foobar").
+fn validate_prefix(prefix: &str) -> Option<RemotePath> {
+    if prefix.is_empty() {
+        // Empty prefix means we want to specify the *whole* bucket
+        return None;
+    }
+    let components = prefix.split('/').collect::<Vec<_>>();
+    let (last, components) = {
+        let last = components.last()?;
+        if last.is_empty() {
+            (
+                components.iter().nth_back(1)?,
+                &components[..(components.len() - 1)],
+            )
+        } else {
+            (last, &components[..])
+        }
+    };
+    'valid: {
+        if let Ok(_timeline_id) = TimelineId::from_str(last) {
+            // Ends in either a tenant or timeline ID
+            break 'valid;
+        }
+        if *last == "timelines" {
+            if let Some(before_last) = components.iter().nth_back(1) {
+                if let Ok(_tenant_id) = TenantShardId::from_str(before_last) {
+                    // Has a valid tenant id
+                    break 'valid;
+                }
+            }
+        }
+
+        return None;
+    }
+    RemotePath::from_string(prefix).ok()
+}
+
+fn strip_subsecond(timestamp: SystemTime) -> SystemTime {
+    let ts_str = humantime::format_rfc3339_seconds(timestamp).to_string();
+    humantime::parse_rfc3339(&ts_str).expect("can't parse just created timestamp")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_validate_prefix() {
+        assert_eq!(validate_prefix(""), None);
+        assert_eq!(validate_prefix("/"), None);
+        #[track_caller]
+        fn assert_valid(prefix: &str) {
+            let remote_path = RemotePath::from_string(prefix).unwrap();
+            assert_eq!(validate_prefix(prefix), Some(remote_path));
+        }
+        assert_valid("wal/3aa8fcc61f6d357410b7de754b1d9001/641e5342083b2235ee3deb8066819683/");
+        // Path is not relative but absolute
+        assert_eq!(
+            validate_prefix(
+                "/wal/3aa8fcc61f6d357410b7de754b1d9001/641e5342083b2235ee3deb8066819683/"
+            ),
+            None
+        );
+        assert_valid("wal/3aa8fcc61f6d357410b7de754b1d9001/");
+        // Partial tenant IDs should be invalid, S3 will match all tenants with the specific ID prefix
+        assert_eq!(validate_prefix("wal/3aa8fcc61f6d357410b7d"), None);
+        assert_eq!(validate_prefix("wal"), None);
+        assert_eq!(validate_prefix("/wal/"), None);
+        assert_valid("pageserver/v1/tenants/3aa8fcc61f6d357410b7de754b1d9001");
+        // Partial tenant ID
+        assert_eq!(
+            validate_prefix("pageserver/v1/tenants/3aa8fcc61f6d357410b"),
+            None
+        );
+        assert_valid("pageserver/v1/tenants/3aa8fcc61f6d357410b7de754b1d9001/timelines");
+        assert_valid("pageserver/v1/tenants/3aa8fcc61f6d357410b7de754b1d9001-0004/timelines");
+        assert_valid("pageserver/v1/tenants/3aa8fcc61f6d357410b7de754b1d9001/timelines/");
+        assert_valid("pageserver/v1/tenants/3aa8fcc61f6d357410b7de754b1d9001/timelines/641e5342083b2235ee3deb8066819683");
+        assert_eq!(validate_prefix("pageserver/v1/tenants/"), None);
+    }
+}

pageserver/pagebench/src/cmd/basebackup.rs

@@ -1,4 +1,5 @@
 use anyhow::Context;
+use pageserver_api::shard::TenantShardId;
 use pageserver_client::mgmt_api::ForceAwaitLogicalSize;
 use pageserver_client::page_service::BasebackupRequest;
 
@@ -95,7 +96,7 @@ async fn main_impl(
             let timeline = *timeline;
             let info = mgmt_api_client
                 .timeline_info(
-                    timeline.tenant_id,
+                    TenantShardId::unsharded(timeline.tenant_id),
                     timeline.timeline_id,
                     ForceAwaitLogicalSize::No,
                 )

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -4,6 +4,7 @@ use pageserver_api::key::{is_rel_block_key, key_to_rel_block, Key};
 use pageserver_api::keyspace::KeySpaceAccum;
 use pageserver_api::models::PagestreamGetPageRequest;
 
+use pageserver_api::shard::TenantShardId;
 use tokio_util::sync::CancellationToken;
 use utils::id::TenantTimelineId;
 use utils::lsn::Lsn;
@@ -173,7 +174,10 @@ async fn main_impl(
                 let timeline = *timeline;
                 async move {
                     let partitioning = mgmt_api_client
-                        .keyspace(timeline.tenant_id, timeline.timeline_id)
+                        .keyspace(
+                            TenantShardId::unsharded(timeline.tenant_id),
+                            timeline.timeline_id,
+                        )
                         .await?;
                     let lsn = partitioning.at_lsn;
                     let start = Instant::now();

pageserver/pagebench/src/cmd/trigger_initial_size_calculation.rs

@@ -1,6 +1,7 @@
 use std::sync::Arc;
 
 use humantime::Duration;
+use pageserver_api::shard::TenantShardId;
 use tokio::task::JoinSet;
 use utils::id::TenantTimelineId;
 
@@ -59,7 +60,11 @@ async fn main_impl(args: Args) -> anyhow::Result<()> {
         let mgmt_api_client = Arc::clone(&mgmt_api_client);
         js.spawn(async move {
             let info = mgmt_api_client
-                .timeline_info(tl.tenant_id, tl.timeline_id, ForceAwaitLogicalSize::Yes)
+                .timeline_info(
+                    TenantShardId::unsharded(tl.tenant_id),
+                    tl.timeline_id,
+                    ForceAwaitLogicalSize::Yes,
+                )
                 .await
                 .unwrap();
 
@@ -74,7 +79,11 @@ async fn main_impl(args: Args) -> anyhow::Result<()> {
                 while !info.current_logical_size_is_accurate {
                     ticker.tick().await;
                     info = mgmt_api_client
-                        .timeline_info(tl.tenant_id, tl.timeline_id, ForceAwaitLogicalSize::Yes)
+                        .timeline_info(
+                            TenantShardId::unsharded(tl.tenant_id),
+                            tl.timeline_id,
+                            ForceAwaitLogicalSize::Yes,
+                        )
                         .await
                         .unwrap();
                 }

pageserver/src/bin/pageserver.rs

@@ -18,6 +18,7 @@ use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
 use pageserver::task_mgr::WALRECEIVER_RUNTIME;
 use pageserver::tenant::{secondary, TenantSharedResources};
 use remote_storage::GenericRemoteStorage;
+use tokio::signal::unix::SignalKind;
 use tokio::time::Instant;
 use tracing::*;
 
@@ -671,42 +672,37 @@ fn start_pageserver(
     let mut shutdown_pageserver = Some(shutdown_pageserver.drop_guard());
 
     // All started up! Now just sit and wait for shutdown signal.
-    {
-        use signal_hook::consts::*;
-        let signal_handler = BACKGROUND_RUNTIME.spawn_blocking(move || {
-            let mut signals =
-                signal_hook::iterator::Signals::new([SIGINT, SIGTERM, SIGQUIT]).unwrap();
-            return signals
-                .forever()
-                .next()
-                .expect("forever() never returns None unless explicitly closed");
-        });
-        let signal = BACKGROUND_RUNTIME
-            .block_on(signal_handler)
-            .expect("join error");
-        match signal {
-            SIGQUIT => {
-                info!("Got signal {signal}. Terminating in immediate shutdown mode",);
-                std::process::exit(111);
-            }
-            SIGINT | SIGTERM => {
-                info!("Got signal {signal}. Terminating gracefully in fast shutdown mode",);
 
-                // This cancels the `shutdown_pageserver` cancellation tree.
-                // Right now that tree doesn't reach very far, and `task_mgr` is used instead.
-                // The plan is to change that over time.
-                shutdown_pageserver.take();
-                let bg_remote_storage = remote_storage.clone();
-                let bg_deletion_queue = deletion_queue.clone();
-                BACKGROUND_RUNTIME.block_on(pageserver::shutdown_pageserver(
-                    &tenant_manager,
-                    bg_remote_storage.map(|_| bg_deletion_queue),
-                    0,
-                ));
-                unreachable!()
-            }
-            _ => unreachable!(),
-        }
+    {
+        BACKGROUND_RUNTIME.block_on(async move {
+            let mut sigint = tokio::signal::unix::signal(SignalKind::interrupt()).unwrap();
+            let mut sigterm = tokio::signal::unix::signal(SignalKind::terminate()).unwrap();
+            let mut sigquit = tokio::signal::unix::signal(SignalKind::quit()).unwrap();
+            let signal = tokio::select! {
+                _ = sigquit.recv() => {
+                    info!("Got signal SIGQUIT. Terminating in immediate shutdown mode",);
+                    std::process::exit(111);
+                }
+                _ = sigint.recv() => { "SIGINT" },
+                _ = sigterm.recv() => { "SIGTERM" },
+            };
+
+            info!("Got signal {signal}. Terminating gracefully in fast shutdown mode",);
+
+            // This cancels the `shutdown_pageserver` cancellation tree.
+            // Right now that tree doesn't reach very far, and `task_mgr` is used instead.
+            // The plan is to change that over time.
+            shutdown_pageserver.take();
+            let bg_remote_storage = remote_storage.clone();
+            let bg_deletion_queue = deletion_queue.clone();
+            pageserver::shutdown_pageserver(
+                &tenant_manager,
+                bg_remote_storage.map(|_| bg_deletion_queue),
+                0,
+            )
+            .await;
+            unreachable!()
+        })
     }
 }
 

pageserver/src/control_plane_client.rs

@@ -12,7 +12,7 @@ use pageserver_api::{
 use serde::{de::DeserializeOwned, Serialize};
 use tokio_util::sync::CancellationToken;
 use url::Url;
-use utils::{backoff, generation::Generation, id::NodeId};
+use utils::{backoff, failpoint_support, generation::Generation, id::NodeId};
 
 use crate::{
     config::{NodeMetadata, PageServerConf},
@@ -210,7 +210,10 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
                 .collect(),
         };
 
-        fail::fail_point!("control-plane-client-validate");
+        failpoint_support::sleep_millis_async!("control-plane-client-validate-sleep", &self.cancel);
+        if self.cancel.is_cancelled() {
+            return Err(RetryForeverError::ShuttingDown);
+        }
 
         let response: ValidateResponse = self.retry_http_forever(&re_attach_path, request).await?;
 

pageserver/src/http/openapi_spec.yml

@@ -1629,7 +1629,7 @@ components:
           type: integer
           format: int64
           minimum: 0
-          description: The amount of disk space currently utilized by layer files.
+          description: The amount of disk space currently used.
         free_space_bytes:
           type: integer
           format: int64

pageserver/src/import_datadir.rs

@@ -8,6 +8,7 @@ use anyhow::{bail, ensure, Context, Result};
 use bytes::Bytes;
 use camino::Utf8Path;
 use futures::StreamExt;
+use pageserver_api::key::rel_block_to_key;
 use tokio::io::{AsyncRead, AsyncReadExt};
 use tokio_tar::Archive;
 use tracing::*;
@@ -170,7 +171,10 @@ async fn import_rel(
         let r = reader.read_exact(&mut buf).await;
         match r {
             Ok(_) => {
-                modification.put_rel_page_image(rel, blknum, Bytes::copy_from_slice(&buf))?;
+                let key = rel_block_to_key(rel, blknum);
+                if modification.tline.get_shard_identity().is_key_local(&key) {
+                    modification.put_rel_page_image(rel, blknum, Bytes::copy_from_slice(&buf))?;
+                }
             }
 
             // TODO: UnexpectedEof is expected

pageserver/src/metrics.rs

@@ -2100,6 +2100,7 @@ pub(crate) fn remove_tenant_metrics(tenant_shard_id: &TenantShardId) {
 use futures::Future;
 use pin_project_lite::pin_project;
 use std::collections::HashMap;
+use std::num::NonZeroUsize;
 use std::pin::Pin;
 use std::sync::{Arc, Mutex};
 use std::task::{Context, Poll};
@@ -2669,6 +2670,26 @@ pub(crate) mod disk_usage_based_eviction {
     pub(crate) static METRICS: Lazy<Metrics> = Lazy::new(Metrics::default);
 }
 
+static TOKIO_EXECUTOR_THREAD_COUNT: Lazy<UIntGaugeVec> = Lazy::new(|| {
+    register_uint_gauge_vec!(
+        "pageserver_tokio_executor_thread_configured_count",
+        "Total number of configued tokio executor threads in the process.
+         The `setup` label denotes whether we're running with multiple runtimes or a single runtime.",
+        &["setup"],
+    )
+    .unwrap()
+});
+
+pub(crate) fn set_tokio_runtime_setup(setup: &str, num_threads: NonZeroUsize) {
+    static SERIALIZE: std::sync::Mutex<()> = std::sync::Mutex::new(());
+    let _guard = SERIALIZE.lock().unwrap();
+    TOKIO_EXECUTOR_THREAD_COUNT.reset();
+    TOKIO_EXECUTOR_THREAD_COUNT
+        .get_metric_with_label_values(&[setup])
+        .unwrap()
+        .set(u64::try_from(num_threads.get()).unwrap());
+}
+
 pub fn preinitialize_metrics() {
     // Python tests need these and on some we do alerting.
     //

pageserver/src/task_mgr.rs

@@ -33,13 +33,14 @@
 use std::collections::HashMap;
 use std::fmt;
 use std::future::Future;
+use std::num::NonZeroUsize;
 use std::panic::AssertUnwindSafe;
+use std::str::FromStr;
 use std::sync::atomic::{AtomicU64, Ordering};
 use std::sync::{Arc, Mutex};
 
 use futures::FutureExt;
 use pageserver_api::shard::TenantShardId;
-use tokio::runtime::Runtime;
 use tokio::task::JoinHandle;
 use tokio::task_local;
 use tokio_util::sync::CancellationToken;
@@ -48,8 +49,11 @@ use tracing::{debug, error, info, warn};
 
 use once_cell::sync::Lazy;
 
+use utils::env;
 use utils::id::TimelineId;
 
+use crate::metrics::set_tokio_runtime_setup;
+
 //
 // There are four runtimes:
 //
@@ -98,52 +102,119 @@ use utils::id::TimelineId;
 // other operations, if the upload tasks e.g. get blocked on locks. It shouldn't
 // happen, but still.
 //
-pub static COMPUTE_REQUEST_RUNTIME: Lazy<Runtime> = Lazy::new(|| {
-    tokio::runtime::Builder::new_multi_thread()
-        .thread_name("compute request worker")
-        .enable_all()
-        .build()
-        .expect("Failed to create compute request runtime")
-});
 
-pub static MGMT_REQUEST_RUNTIME: Lazy<Runtime> = Lazy::new(|| {
-    tokio::runtime::Builder::new_multi_thread()
-        .thread_name("mgmt request worker")
-        .enable_all()
-        .build()
-        .expect("Failed to create mgmt request runtime")
-});
-
-pub static WALRECEIVER_RUNTIME: Lazy<Runtime> = Lazy::new(|| {
-    tokio::runtime::Builder::new_multi_thread()
-        .thread_name("walreceiver worker")
-        .enable_all()
-        .build()
-        .expect("Failed to create walreceiver runtime")
-});
-
-pub static BACKGROUND_RUNTIME: Lazy<Runtime> = Lazy::new(|| {
-    tokio::runtime::Builder::new_multi_thread()
-        .thread_name("background op worker")
-        // if you change the number of worker threads please change the constant below
-        .enable_all()
-        .build()
-        .expect("Failed to create background op runtime")
-});
-
-pub(crate) static BACKGROUND_RUNTIME_WORKER_THREADS: Lazy<usize> = Lazy::new(|| {
-    // force init and thus panics
-    let _ = BACKGROUND_RUNTIME.handle();
+pub(crate) static TOKIO_WORKER_THREADS: Lazy<NonZeroUsize> = Lazy::new(|| {
     // replicates tokio-1.28.1::loom::sys::num_cpus which is not available publicly
     // tokio would had already panicked for parsing errors or NotUnicode
     //
     // this will be wrong if any of the runtimes gets their worker threads configured to something
     // else, but that has not been needed in a long time.
-    std::env::var("TOKIO_WORKER_THREADS")
-        .map(|s| s.parse::<usize>().unwrap())
-        .unwrap_or_else(|_e| usize::max(2, num_cpus::get()))
+    NonZeroUsize::new(
+        std::env::var("TOKIO_WORKER_THREADS")
+            .map(|s| s.parse::<usize>().unwrap())
+            .unwrap_or_else(|_e| usize::max(2, num_cpus::get())),
+    )
+    .expect("the max() ensures that this is not zero")
 });
 
+enum TokioRuntimeMode {
+    SingleThreaded,
+    MultiThreaded { num_workers: NonZeroUsize },
+}
+
+impl FromStr for TokioRuntimeMode {
+    type Err = String;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "current_thread" => Ok(TokioRuntimeMode::SingleThreaded),
+            s => match s.strip_prefix("multi_thread:") {
+                Some("default") => Ok(TokioRuntimeMode::MultiThreaded {
+                    num_workers: *TOKIO_WORKER_THREADS,
+                }),
+                Some(suffix) => {
+                    let num_workers = suffix.parse::<NonZeroUsize>().map_err(|e| {
+                        format!(
+                            "invalid number of multi-threaded runtime workers ({suffix:?}): {e}",
+                        )
+                    })?;
+                    Ok(TokioRuntimeMode::MultiThreaded { num_workers })
+                }
+                None => Err(format!("invalid runtime config: {s:?}")),
+            },
+        }
+    }
+}
+
+static ONE_RUNTIME: Lazy<Option<tokio::runtime::Runtime>> = Lazy::new(|| {
+    let thread_name = "pageserver-tokio";
+    let Some(mode) = env::var("NEON_PAGESERVER_USE_ONE_RUNTIME") else {
+        // If the env var is not set, leave this static as None.
+        set_tokio_runtime_setup(
+            "multiple-runtimes",
+            NUM_MULTIPLE_RUNTIMES
+                .checked_mul(*TOKIO_WORKER_THREADS)
+                .unwrap(),
+        );
+        return None;
+    };
+    Some(match mode {
+        TokioRuntimeMode::SingleThreaded => {
+            set_tokio_runtime_setup("one-runtime-single-threaded", NonZeroUsize::new(1).unwrap());
+            tokio::runtime::Builder::new_current_thread()
+                .thread_name(thread_name)
+                .enable_all()
+                .build()
+                .expect("failed to create one single runtime")
+        }
+        TokioRuntimeMode::MultiThreaded { num_workers } => {
+            set_tokio_runtime_setup("one-runtime-multi-threaded", num_workers);
+            tokio::runtime::Builder::new_multi_thread()
+                .thread_name(thread_name)
+                .enable_all()
+                .worker_threads(num_workers.get())
+                .build()
+                .expect("failed to create one multi-threaded runtime")
+        }
+    })
+});
+
+/// Declare a lazy static variable named `$varname` that will resolve
+/// to a tokio runtime handle. If the env var `NEON_PAGESERVER_USE_ONE_RUNTIME`
+/// is set, this will resolve to `ONE_RUNTIME`. Otherwise, the macro invocation
+/// declares a separate runtime and the lazy static variable `$varname`
+/// will resolve to that separate runtime.
+///
+/// The result is is that `$varname.spawn()` will use `ONE_RUNTIME` if
+/// `NEON_PAGESERVER_USE_ONE_RUNTIME` is set, and will use the separate runtime
+/// otherwise.
+macro_rules! pageserver_runtime {
+    ($varname:ident, $name:literal) => {
+        pub static $varname: Lazy<&'static tokio::runtime::Runtime> = Lazy::new(|| {
+            if let Some(runtime) = &*ONE_RUNTIME {
+                return runtime;
+            }
+            static RUNTIME: Lazy<tokio::runtime::Runtime> = Lazy::new(|| {
+                tokio::runtime::Builder::new_multi_thread()
+                    .thread_name($name)
+                    .worker_threads(TOKIO_WORKER_THREADS.get())
+                    .enable_all()
+                    .build()
+                    .expect(std::concat!("Failed to create runtime ", $name))
+            });
+            &*RUNTIME
+        });
+    };
+}
+
+pageserver_runtime!(COMPUTE_REQUEST_RUNTIME, "compute request worker");
+pageserver_runtime!(MGMT_REQUEST_RUNTIME, "mgmt request worker");
+pageserver_runtime!(WALRECEIVER_RUNTIME, "walreceiver worker");
+pageserver_runtime!(BACKGROUND_RUNTIME, "background op worker");
+// Bump this number when adding a new pageserver_runtime!
+// SAFETY: it's obviously correct
+const NUM_MULTIPLE_RUNTIMES: NonZeroUsize = unsafe { NonZeroUsize::new_unchecked(4) };
+
 #[derive(Debug, Clone, Copy)]
 pub struct PageserverTaskId(u64);
 

pageserver/src/tenant/remote_timeline_client.rs

@@ -200,6 +200,7 @@ use utils::backoff::{
 use std::collections::{HashMap, VecDeque};
 use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::{Arc, Mutex};
+use std::time::Duration;
 
 use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath, TimeoutOrCancel};
 use std::ops::DerefMut;
@@ -207,7 +208,7 @@ use tracing::{debug, error, info, instrument, warn};
 use tracing::{info_span, Instrument};
 use utils::lsn::Lsn;
 
-use crate::deletion_queue::DeletionQueueClient;
+use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
 use crate::metrics::{
     MeasureRemoteOp, RemoteOpFileKind, RemoteOpKind, RemoteTimelineClientMetrics,
     RemoteTimelineClientMetricsCallTrackSize, REMOTE_ONDEMAND_DOWNLOADED_BYTES,
@@ -261,6 +262,10 @@ pub(crate) const INITDB_PRESERVED_PATH: &str = "initdb-preserved.tar.zst";
 /// Default buffer size when interfacing with [`tokio::fs::File`].
 pub(crate) const BUFFER_SIZE: usize = 32 * 1024;
 
+/// Doing non-essential flushes of deletion queue is subject to this timeout, after
+/// which we warn and skip.
+const DELETION_QUEUE_FLUSH_TIMEOUT: Duration = Duration::from_secs(10);
+
 pub enum MaybeDeletedIndexPart {
     IndexPart(IndexPart),
     Deleted(IndexPart),
@@ -588,14 +593,14 @@ impl RemoteTimelineClient {
         upload_queue: &mut UploadQueueInitialized,
         metadata: TimelineMetadata,
     ) {
+        let disk_consistent_lsn = upload_queue.latest_metadata.disk_consistent_lsn();
+
         info!(
-            "scheduling metadata upload with {} files ({} changed)",
+            "scheduling metadata upload up to consistent LSN {disk_consistent_lsn} with {} files ({} changed)",
             upload_queue.latest_files.len(),
             upload_queue.latest_files_changes_since_metadata_upload_scheduled,
         );
 
-        let disk_consistent_lsn = upload_queue.latest_metadata.disk_consistent_lsn();
-
         let index_part = IndexPart::new(
             upload_queue.latest_files.clone(),
             disk_consistent_lsn,
@@ -1050,6 +1055,26 @@ impl RemoteTimelineClient {
         Ok(())
     }
 
+    async fn flush_deletion_queue(&self) -> Result<(), DeletionQueueError> {
+        match tokio::time::timeout(
+            DELETION_QUEUE_FLUSH_TIMEOUT,
+            self.deletion_queue_client.flush_immediate(),
+        )
+        .await
+        {
+            Ok(result) => result,
+            Err(_timeout) => {
+                // Flushing remote deletions is not mandatory: we flush here to make the system easier to test, and
+                // to ensure that _usually_ objects are really gone after a DELETE is acked.  However, in case of deletion
+                // queue issues (https://github.com/neondatabase/neon/issues/6440), we don't want to wait indefinitely here.
+                tracing::warn!(
+                    "Timed out waiting for deletion queue flush, acking deletion anyway"
+                );
+                Ok(())
+            }
+        }
+    }
+
     /// Prerequisites: UploadQueue should be in stopped state and deleted_at should be successfuly set.
     /// The function deletes layer files one by one, then lists the prefix to see if we leaked something
     /// deletes leaked files if any and proceeds with deletion of index file at the end.
@@ -1099,7 +1124,7 @@ impl RemoteTimelineClient {
 
         // Execute all pending deletions, so that when we proceed to do a list_prefixes below, we aren't
         // taking the burden of listing all the layers that we already know we should delete.
-        self.deletion_queue_client.flush_immediate().await?;
+        self.flush_deletion_queue().await?;
 
         let cancel = shutdown_token();
 
@@ -1173,7 +1198,7 @@ impl RemoteTimelineClient {
 
         // Timeline deletion is rare and we have probably emitted a reasonably number of objects: wait
         // for a flush to a persistent deletion list so that we may be sure deletion will occur.
-        self.deletion_queue_client.flush_immediate().await?;
+        self.flush_deletion_queue().await?;
 
         fail::fail_point!("timeline-delete-after-index-delete", |_| {
             Err(anyhow::anyhow!(

pageserver/src/tenant/secondary/downloader.rs

@@ -51,7 +51,7 @@ use tokio_util::sync::CancellationToken;
 use tracing::{info_span, instrument, warn, Instrument};
 use utils::{
     backoff, completion::Barrier, crashsafe::path_with_suffix_extension, failpoint_support, fs_ext,
-    id::TimelineId,
+    id::TimelineId, serde_system_time,
 };
 
 use super::{
@@ -591,7 +591,7 @@ impl<'a> TenantDownloader<'a> {
         let mut progress = SecondaryProgress {
             layers_total: heatmap_stats.layers,
             bytes_total: heatmap_stats.bytes,
-            heatmap_mtime: Some(heatmap_mtime),
+            heatmap_mtime: Some(serde_system_time::SystemTime(heatmap_mtime)),
             layers_downloaded: 0,
             bytes_downloaded: 0,
         };

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -19,6 +19,7 @@ use pageserver_api::models::InMemoryLayerInfo;
 use pageserver_api::shard::TenantShardId;
 use std::collections::{BinaryHeap, HashMap, HashSet};
 use std::sync::{Arc, OnceLock};
+use std::time::Instant;
 use tracing::*;
 use utils::{bin_ser::BeSer, id::TimelineId, lsn::Lsn, vec_map::VecMap};
 // avoid binding to Write (conflicts with std::io::Write)
@@ -53,6 +54,8 @@ pub struct InMemoryLayer {
     /// Writes are only allowed when this is `None`.
     end_lsn: OnceLock<Lsn>,
 
+    opened_at: Instant,
+
     /// The above fields never change, except for `end_lsn`, which is only set once.
     /// All other changing parts are in `inner`, and protected by a mutex.
     inner: RwLock<InMemoryLayerInner>,
@@ -460,6 +463,7 @@ impl InMemoryLayer {
             tenant_shard_id,
             start_lsn,
             end_lsn: OnceLock::new(),
+            opened_at: Instant::now(),
             inner: RwLock::new(InMemoryLayerInner {
                 index: HashMap::new(),
                 file,
@@ -520,6 +524,10 @@ impl InMemoryLayer {
         Ok(())
     }
 
+    pub(crate) fn get_opened_at(&self) -> Instant {
+        self.opened_at
+    }
+
     pub(crate) async fn tick(&self) -> Option<u64> {
         let mut inner = self.inner.write().await;
         let size = inner.file.len();

pageserver/src/tenant/tasks.rs

@@ -18,7 +18,7 @@ use utils::{backoff, completion};
 
 static CONCURRENT_BACKGROUND_TASKS: once_cell::sync::Lazy<tokio::sync::Semaphore> =
     once_cell::sync::Lazy::new(|| {
-        let total_threads = *task_mgr::BACKGROUND_RUNTIME_WORKER_THREADS;
+        let total_threads = task_mgr::TOKIO_WORKER_THREADS.get();
         let permits = usize::max(
             1,
             // while a lot of the work is done on spawn_blocking, we still do
@@ -72,6 +72,7 @@ pub(crate) async fn concurrent_background_tasks_rate_limit_permit(
         loop_kind == BackgroundLoopKind::InitialLogicalSizeCalculation
     );
 
+    // TODO: assert that we run on BACKGROUND_RUNTIME; requires tokio_unstable Handle::id();
     match CONCURRENT_BACKGROUND_TASKS.acquire().await {
         Ok(permit) => permit,
         Err(_closed) => unreachable!("we never close the semaphore"),

pageserver/src/tenant/timeline.rs

@@ -282,10 +282,12 @@ pub struct Timeline {
     pub(super) flush_loop_state: Mutex<FlushLoopState>,
 
     /// layer_flush_start_tx can be used to wake up the layer-flushing task.
-    /// The value is a counter, incremented every time a new flush cycle is requested.
-    /// The flush cycle counter is sent back on the layer_flush_done channel when
-    /// the flush finishes. You can use that to wait for the flush to finish.
-    layer_flush_start_tx: tokio::sync::watch::Sender<u64>,
+    /// - The u64 value is a counter, incremented every time a new flush cycle is requested.
+    ///   The flush cycle counter is sent back on the layer_flush_done channel when
+    ///   the flush finishes. You can use that to wait for the flush to finish.
+    /// - The LSN is updated to max() of its current value and the latest disk_consistent_lsn
+    ///   read by whoever sends an update
+    layer_flush_start_tx: tokio::sync::watch::Sender<(u64, Lsn)>,
     /// to be notified when layer flushing has finished, subscribe to the layer_flush_done channel
     layer_flush_done_tx: tokio::sync::watch::Sender<(u64, Result<(), FlushLayerError>)>,
 
@@ -1169,8 +1171,8 @@ impl Timeline {
     /// Flush to disk all data that was written with the put_* functions
     #[instrument(skip(self), fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%self.timeline_id))]
     pub(crate) async fn freeze_and_flush(&self) -> anyhow::Result<()> {
-        self.freeze_inmem_layer(false).await;
-        self.flush_frozen_layers_and_wait().await
+        let to_lsn = self.freeze_inmem_layer(false).await;
+        self.flush_frozen_layers_and_wait(to_lsn).await
     }
 
     /// If there is no writer, and conditions for rolling the latest layer are met, then freeze it.
@@ -1190,7 +1192,39 @@ impl Timeline {
         };
 
         let Some(open_layer) = &layers_guard.layer_map().open_layer else {
-            // No open layer, no work to do.
+            // If there is no open layer, we have no layer freezing to do.  However, we might need to generate
+            // some updates to disk_consistent_lsn and remote_consistent_lsn, in case we ingested some WAL regions
+            // that didn't result in writes to this shard.
+
+            // Must not hold the layers lock while waiting for a flush.
+            drop(layers_guard);
+
+            let last_record_lsn = self.get_last_record_lsn();
+            let disk_consistent_lsn = self.get_disk_consistent_lsn();
+            if last_record_lsn > disk_consistent_lsn {
+                // We have no open layer, but disk_consistent_lsn is behind the last record: this indicates
+                // we are a sharded tenant and have skipped some WAL
+                let last_freeze_ts = *self.last_freeze_ts.read().unwrap();
+                if last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
+                    // This should be somewhat rare, so we log it at INFO level.
+                    //
+                    // We checked for checkpoint timeout so that a shard without any
+                    // data ingested (yet) doesn't write a remote index as soon as it
+                    // sees its LSN advance: we only do this if we've been layer-less
+                    // for some time.
+                    tracing::info!(
+                        "Advancing disk_consistent_lsn past WAL ingest gap {} -> {}",
+                        disk_consistent_lsn,
+                        last_record_lsn
+                    );
+
+                    // The flush loop will update remote consistent LSN as well as disk consistent LSN.
+                    self.flush_frozen_layers_and_wait(last_record_lsn)
+                        .await
+                        .ok();
+                }
+            }
+
             return;
         };
 
@@ -1223,7 +1257,7 @@ impl Timeline {
             checkpoint_distance,
             self.get_last_record_lsn(),
             self.last_freeze_at.load(),
-            *self.last_freeze_ts.read().unwrap(),
+            open_layer.get_opened_at(),
         ) {
             match open_layer.info() {
                 InMemoryLayerInfo::Frozen { lsn_start, lsn_end } => {
@@ -1588,7 +1622,7 @@ impl Timeline {
         checkpoint_distance: u64,
         projected_lsn: Lsn,
         last_freeze_at: Lsn,
-        last_freeze_ts: Instant,
+        opened_at: Instant,
     ) -> bool {
         let distance = projected_lsn.widening_sub(last_freeze_at);
 
@@ -1614,13 +1648,13 @@ impl Timeline {
             );
 
             true
-        } else if distance > 0 && last_freeze_ts.elapsed() >= self.get_checkpoint_timeout() {
+        } else if distance > 0 && opened_at.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()
-            );
+                    "Will roll layer at {} with layer size {} due to time since first write to the layer ({:?})",
+                    projected_lsn,
+                    layer_size,
+                    opened_at.elapsed()
+                );
 
             true
         } else {
@@ -1769,7 +1803,7 @@ impl Timeline {
         let disk_consistent_lsn = metadata.disk_consistent_lsn();
         let (state, _) = watch::channel(state);
 
-        let (layer_flush_start_tx, _) = tokio::sync::watch::channel(0);
+        let (layer_flush_start_tx, _) = tokio::sync::watch::channel((0, disk_consistent_lsn));
         let (layer_flush_done_tx, _) = tokio::sync::watch::channel((0, Ok(())));
 
         let evictions_low_residence_duration_metric_threshold = {
@@ -3174,7 +3208,9 @@ impl Timeline {
         self.last_record_lsn.advance(new_lsn);
     }
 
-    async fn freeze_inmem_layer(&self, write_lock_held: bool) {
+    /// Whether there was a layer to freeze or not, return the value of get_last_record_lsn
+    /// before we attempted the freeze: this guarantees that ingested data is frozen up to this lsn (inclusive).
+    async fn freeze_inmem_layer(&self, write_lock_held: bool) -> Lsn {
         // Freeze the current open in-memory layer. It will be written to disk on next
         // iteration.
 
@@ -3184,7 +3220,9 @@ impl Timeline {
             Some(self.write_lock.lock().await)
         };
 
-        self.freeze_inmem_layer_at(self.get_last_record_lsn()).await;
+        let to_lsn = self.get_last_record_lsn();
+        self.freeze_inmem_layer_at(to_lsn).await;
+        to_lsn
     }
 
     async fn freeze_inmem_layer_at(&self, at: Lsn) {
@@ -3197,7 +3235,7 @@ impl Timeline {
     /// Layer flusher task's main loop.
     async fn flush_loop(
         self: &Arc<Self>,
-        mut layer_flush_start_rx: tokio::sync::watch::Receiver<u64>,
+        mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
         ctx: &RequestContext,
     ) {
         info!("started flush loop");
@@ -3210,7 +3248,11 @@ impl Timeline {
                 _ = layer_flush_start_rx.changed() => {}
             }
             trace!("waking up");
-            let flush_counter = *layer_flush_start_rx.borrow();
+            let (flush_counter, frozen_to_lsn) = *layer_flush_start_rx.borrow();
+
+            // The highest LSN to which we flushed in the loop over frozen layers
+            let mut flushed_to_lsn = Lsn(0);
+
             let result = loop {
                 if self.cancel.is_cancelled() {
                     info!("dropping out of flush loop for timeline shutdown");
@@ -3231,7 +3273,9 @@ impl Timeline {
                     break Ok(());
                 };
                 match self.flush_frozen_layer(layer_to_flush, ctx).await {
-                    Ok(()) => {}
+                    Ok(this_layer_to_lsn) => {
+                        flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
+                    }
                     Err(FlushLayerError::Cancelled) => {
                         info!("dropping out of flush loop for timeline shutdown");
                         return;
@@ -3240,11 +3284,36 @@ impl Timeline {
                         FlushLayerError::Other(_) | FlushLayerError::CreateImageLayersError(_),
                     ) => {
                         error!("could not flush frozen layer: {err:?}");
-                        break err;
+                        break err.map(|_| ());
                     }
                 }
                 timer.stop_and_record();
             };
+
+            // Unsharded tenants should never advance their LSN beyond the end of the
+            // highest layer they write: such gaps between layer data and the frozen LSN
+            // are only legal on sharded tenants.
+            debug_assert!(
+                self.shard_identity.count.count() > 1
+                    || flushed_to_lsn >= frozen_to_lsn
+                    || !flushed_to_lsn.is_valid()
+            );
+
+            if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
+                // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
+                // to us via layer_flush_start_rx, then advance it here.
+                //
+                // This path is only taken for tenants with multiple shards: single sharded tenants should
+                // never encounter a gap in the wal.
+                let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
+                tracing::debug!("Advancing disk_consistent_lsn across layer gap {old_disk_consistent_lsn}->{frozen_to_lsn}");
+                if self.set_disk_consistent_lsn(frozen_to_lsn) {
+                    if let Err(e) = self.schedule_uploads(frozen_to_lsn, vec![]) {
+                        tracing::warn!("Failed to schedule metadata upload after updating disk_consistent_lsn: {e}");
+                    }
+                }
+            }
+
             // Notify any listeners that we're done
             let _ = self
                 .layer_flush_done_tx
@@ -3252,7 +3321,13 @@ impl Timeline {
         }
     }
 
-    async fn flush_frozen_layers_and_wait(&self) -> anyhow::Result<()> {
+    /// Request the flush loop to write out all frozen layers up to `to_lsn` as Delta L0 files to disk.
+    /// The caller is responsible for the freezing, e.g., [`Self::freeze_inmem_layer`].
+    ///
+    /// `last_record_lsn` may be higher than the highest LSN of a frozen layer: if this is the case,
+    /// it means no data will be written between the top of the highest frozen layer and to_lsn,
+    /// e.g. because this tenant shard has ingested up to to_lsn and not written any data locally for that part of the WAL.
+    async fn flush_frozen_layers_and_wait(&self, last_record_lsn: Lsn) -> anyhow::Result<()> {
         let mut rx = self.layer_flush_done_tx.subscribe();
 
         // Increment the flush cycle counter and wake up the flush task.
@@ -3266,9 +3341,10 @@ impl Timeline {
             anyhow::bail!("cannot flush frozen layers when flush_loop is not running, state is {flush_loop_state:?}")
         }
 
-        self.layer_flush_start_tx.send_modify(|counter| {
+        self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
             my_flush_request = *counter + 1;
             *counter = my_flush_request;
+            *lsn = std::cmp::max(last_record_lsn, *lsn);
         });
 
         loop {
@@ -3305,16 +3381,22 @@ impl Timeline {
     }
 
     fn flush_frozen_layers(&self) {
-        self.layer_flush_start_tx.send_modify(|val| *val += 1);
+        self.layer_flush_start_tx.send_modify(|(counter, lsn)| {
+            *counter += 1;
+
+            *lsn = std::cmp::max(*lsn, Lsn(self.last_freeze_at.load().0 - 1));
+        });
     }
 
     /// Flush one frozen in-memory layer to disk, as a new delta layer.
+    ///
+    /// Return value is the last lsn (inclusive) of the layer that was frozen.
     #[instrument(skip_all, fields(layer=%frozen_layer))]
     async fn flush_frozen_layer(
         self: &Arc<Self>,
         frozen_layer: Arc<InMemoryLayer>,
         ctx: &RequestContext,
-    ) -> Result<(), FlushLayerError> {
+    ) -> Result<Lsn, FlushLayerError> {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
         // As a special case, when we have just imported an image into the repository,
@@ -3389,7 +3471,6 @@ impl Timeline {
         }
 
         let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
-        let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
 
         // The new on-disk layers are now in the layer map. We can remove the
         // in-memory layer from the map now. The flushed layer is stored in
@@ -3403,10 +3484,7 @@ impl Timeline {
 
             guard.finish_flush_l0_layer(delta_layer_to_add.as_ref(), &frozen_layer, &self.metrics);
 
-            if disk_consistent_lsn != old_disk_consistent_lsn {
-                assert!(disk_consistent_lsn > old_disk_consistent_lsn);
-                self.disk_consistent_lsn.store(disk_consistent_lsn);
-
+            if self.set_disk_consistent_lsn(disk_consistent_lsn) {
                 // Schedule remote uploads that will reflect our new disk_consistent_lsn
                 self.schedule_uploads(disk_consistent_lsn, layers_to_upload)?;
             }
@@ -3423,7 +3501,22 @@ impl Timeline {
         // This failpoint is used by another test case `test_pageserver_recovery`.
         fail_point!("flush-frozen-exit");
 
-        Ok(())
+        Ok(Lsn(lsn_range.end.0 - 1))
+    }
+
+    /// Return true if the value changed
+    ///
+    /// This function must only be used from the layer flush task, and may not be called concurrently.
+    fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
+        // We do a simple load/store cycle: that's why this function isn't safe for concurrent use.
+        let old_value = self.disk_consistent_lsn.load();
+        if new_value != old_value {
+            assert!(new_value >= old_value);
+            self.disk_consistent_lsn.store(new_value);
+            true
+        } else {
+            false
+        }
     }
 
     /// Update metadata file
@@ -4610,23 +4703,16 @@ struct TimelineWriterState {
     max_lsn: Option<Lsn>,
     // Cached details of the last freeze. Avoids going trough the atomic/lock on every put.
     cached_last_freeze_at: Lsn,
-    cached_last_freeze_ts: Instant,
 }
 
 impl TimelineWriterState {
-    fn new(
-        open_layer: Arc<InMemoryLayer>,
-        current_size: u64,
-        last_freeze_at: Lsn,
-        last_freeze_ts: Instant,
-    ) -> Self {
+    fn new(open_layer: Arc<InMemoryLayer>, current_size: u64, last_freeze_at: Lsn) -> Self {
         Self {
             open_layer,
             current_size,
             prev_lsn: None,
             max_lsn: None,
             cached_last_freeze_at: last_freeze_at,
-            cached_last_freeze_ts: last_freeze_ts,
         }
     }
 }
@@ -4725,12 +4811,10 @@ impl<'a> TimelineWriter<'a> {
         let initial_size = 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(
             layer,
             initial_size,
             last_freeze_at,
-            last_freeze_ts,
         ));
 
         Ok(())
@@ -4777,7 +4861,7 @@ impl<'a> TimelineWriter<'a> {
             self.get_checkpoint_distance(),
             lsn,
             state.cached_last_freeze_at,
-            state.cached_last_freeze_ts,
+            state.open_layer.get_opened_at(),
         ) {
             OpenLayerAction::Roll
         } else {

pageserver/src/tenant/timeline/compaction.rs

@@ -12,7 +12,6 @@ use super::layer_manager::LayerManager;
 use super::{CompactFlags, DurationRecorder, RecordedDuration, Timeline};
 
 use anyhow::{anyhow, Context};
-use async_trait::async_trait;
 use enumset::EnumSet;
 use fail::fail_point;
 use itertools::Itertools;
@@ -1122,7 +1121,6 @@ impl CompactionLayer<Key> for ResidentDeltaLayer {
     }
 }
 
-#[async_trait]
 impl CompactionDeltaLayer<TimelineAdaptor> for ResidentDeltaLayer {
     type DeltaEntry<'a> = DeltaEntry<'a>;
 

pageserver/src/tenant/timeline/layer_manager.rs

@@ -120,9 +120,10 @@ impl LayerManager {
     /// Called from `freeze_inmem_layer`, returns true if successfully frozen.
     pub(crate) async fn try_freeze_in_memory_layer(
         &mut self,
-        Lsn(last_record_lsn): Lsn,
+        lsn: Lsn,
         last_freeze_at: &AtomicLsn,
     ) {
+        let Lsn(last_record_lsn) = lsn;
         let end_lsn = Lsn(last_record_lsn + 1);
 
         if let Some(open_layer) = &self.layer_map.open_layer {
@@ -135,8 +136,11 @@ impl LayerManager {
             self.layer_map.frozen_layers.push_back(open_layer_rc);
             self.layer_map.open_layer = None;
             self.layer_map.next_open_layer_at = Some(end_lsn);
-            last_freeze_at.store(end_lsn);
         }
+
+        // Even if there was no layer to freeze, advance last_freeze_at to last_record_lsn+1: this
+        // accounts for regions in the LSN range where we might have ingested no data due to sharding.
+        last_freeze_at.store(end_lsn);
     }
 
     /// Add image layers to the layer map, called from `create_image_layers`.

pageserver/src/utilization.rs

@@ -15,11 +15,23 @@ pub(crate) fn regenerate(tenants_path: &Path) -> anyhow::Result<PageserverUtiliz
         .map_err(std::io::Error::from)
         .context("statvfs tenants directory")?;
 
-    let blocksz = statvfs.block_size();
+    // https://unix.stackexchange.com/a/703650
+    let blocksz = if statvfs.fragment_size() > 0 {
+        statvfs.fragment_size()
+    } else {
+        statvfs.block_size()
+    };
 
     #[cfg_attr(not(target_os = "macos"), allow(clippy::unnecessary_cast))]
     let free = statvfs.blocks_available() as u64 * blocksz;
-    let used = crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.get();
+
+    #[cfg_attr(not(target_os = "macos"), allow(clippy::unnecessary_cast))]
+    let used = statvfs
+        .blocks()
+        // use blocks_free instead of available here to match df in case someone compares
+        .saturating_sub(statvfs.blocks_free()) as u64
+        * blocksz;
+
     let captured_at = std::time::SystemTime::now();
 
     let doc = PageserverUtilization {
@@ -29,7 +41,7 @@ pub(crate) fn regenerate(tenants_path: &Path) -> anyhow::Result<PageserverUtiliz
         //
         // note that u64::MAX will be output as i64::MAX as u64, but that should not matter
         utilization_score: u64::MAX,
-        captured_at,
+        captured_at: utils::serde_system_time::SystemTime(captured_at),
     };
 
     // TODO: make utilization_score into a metric

pageserver/src/walredo.rs

@@ -36,11 +36,12 @@ use bytes::{Bytes, BytesMut};
 use pageserver_api::key::key_to_rel_block;
 use pageserver_api::models::WalRedoManagerStatus;
 use pageserver_api::shard::TenantShardId;
-use std::sync::{Arc, RwLock};
+use std::sync::Arc;
 use std::time::Duration;
 use std::time::Instant;
 use tracing::*;
 use utils::lsn::Lsn;
+use utils::sync::heavier_once_cell;
 
 ///
 /// This is the real implementation that uses a Postgres process to
@@ -53,7 +54,19 @@ pub struct PostgresRedoManager {
     tenant_shard_id: TenantShardId,
     conf: &'static PageServerConf,
     last_redo_at: std::sync::Mutex<Option<Instant>>,
-    redo_process: RwLock<Option<Arc<process::WalRedoProcess>>>,
+    /// The current [`process::WalRedoProcess`] that is used by new redo requests.
+    /// We use [`heavier_once_cell`] for coalescing the spawning, but the redo
+    /// requests don't use the [`heavier_once_cell::Guard`] to keep ahold of the
+    /// their process object; we use [`Arc::clone`] for that.
+    /// This is primarily because earlier implementations that didn't  use [`heavier_once_cell`]
+    /// had that behavior; it's probably unnecessary.
+    /// The only merit of it is that if one walredo process encounters an error,
+    /// it can take it out of rotation (= using [`heavier_once_cell::Guard::take_and_deinit`].
+    /// and retry redo, thereby starting the new process, while other redo tasks might
+    /// still be using the old redo process. But, those other tasks will most likely
+    /// encounter an error as well, and errors are an unexpected condition anyway.
+    /// So, probably we could get rid of the `Arc` in the future.
+    redo_process: heavier_once_cell::OnceCell<Arc<process::WalRedoProcess>>,
 }
 
 ///
@@ -101,6 +114,7 @@ impl PostgresRedoManager {
                         self.conf.wal_redo_timeout,
                         pg_version,
                     )
+                    .await
                 };
                 img = Some(result?);
 
@@ -121,6 +135,7 @@ impl PostgresRedoManager {
                 self.conf.wal_redo_timeout,
                 pg_version,
             )
+            .await
         }
     }
 
@@ -134,7 +149,7 @@ impl PostgresRedoManager {
                     chrono::Utc::now().checked_sub_signed(chrono::Duration::from_std(age).ok()?)
                 })
             },
-            pid: self.redo_process.read().unwrap().as_ref().map(|p| p.id()),
+            pid: self.redo_process.get().map(|p| p.id()),
         })
     }
 }
@@ -152,7 +167,7 @@ impl PostgresRedoManager {
             tenant_shard_id,
             conf,
             last_redo_at: std::sync::Mutex::default(),
-            redo_process: RwLock::new(None),
+            redo_process: heavier_once_cell::OnceCell::default(),
         }
     }
 
@@ -164,8 +179,7 @@ impl PostgresRedoManager {
             if let Some(last_redo_at) = *g {
                 if last_redo_at.elapsed() >= idle_timeout {
                     drop(g);
-                    let mut guard = self.redo_process.write().unwrap();
-                    *guard = None;
+                    drop(self.redo_process.get().map(|guard| guard.take_and_deinit()));
                 }
             }
         }
@@ -174,8 +188,11 @@ impl PostgresRedoManager {
     ///
     /// Process one request for WAL redo using wal-redo postgres
     ///
+    /// # Cancel-Safety
+    ///
+    /// Cancellation safe.
     #[allow(clippy::too_many_arguments)]
-    fn apply_batch_postgres(
+    async fn apply_batch_postgres(
         &self,
         key: Key,
         lsn: Lsn,
@@ -191,42 +208,31 @@ impl PostgresRedoManager {
         const MAX_RETRY_ATTEMPTS: u32 = 1;
         let mut n_attempts = 0u32;
         loop {
-            // launch the WAL redo process on first use
-            let proc: Arc<process::WalRedoProcess> = {
-                let proc_guard = self.redo_process.read().unwrap();
-                match &*proc_guard {
-                    None => {
-                        // "upgrade" to write lock to launch the process
-                        drop(proc_guard);
-                        let mut proc_guard = self.redo_process.write().unwrap();
-                        match &*proc_guard {
-                            None => {
-                                let start = Instant::now();
-                                let proc = Arc::new(
-                                    process::WalRedoProcess::launch(
-                                        self.conf,
-                                        self.tenant_shard_id,
-                                        pg_version,
-                                    )
-                                    .context("launch walredo process")?,
-                                );
-                                let duration = start.elapsed();
-                                WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM
-                                    .observe(duration.as_secs_f64());
-                                info!(
-                                    duration_ms = duration.as_millis(),
-                                    pid = proc.id(),
-                                    "launched walredo process"
-                                );
-                                *proc_guard = Some(Arc::clone(&proc));
-                                proc
-                            }
-                            Some(proc) => Arc::clone(proc),
-                        }
+            let proc: Arc<process::WalRedoProcess> =
+                match self.redo_process.get_or_init_detached().await {
+                    Ok(guard) => Arc::clone(&guard),
+                    Err(permit) => {
+                        // don't hold poison_guard, the launch code can bail
+                        let start = Instant::now();
+                        let proc = Arc::new(
+                            process::WalRedoProcess::launch(
+                                self.conf,
+                                self.tenant_shard_id,
+                                pg_version,
+                            )
+                            .context("launch walredo process")?,
+                        );
+                        let duration = start.elapsed();
+                        WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM.observe(duration.as_secs_f64());
+                        info!(
+                            duration_ms = duration.as_millis(),
+                            pid = proc.id(),
+                            "launched walredo process"
+                        );
+                        self.redo_process.set(Arc::clone(&proc), permit);
+                        proc
                     }
-                    Some(proc) => Arc::clone(proc),
-                }
-            };
+                };
 
             let started_at = std::time::Instant::now();
 
@@ -272,34 +278,34 @@ impl PostgresRedoManager {
                     n_attempts,
                     e,
                 );
-                // Avoid concurrent callers hitting the same issue.
-                // We can't prevent it from happening because we want to enable parallelism.
-                {
-                    let mut guard = self.redo_process.write().unwrap();
-                    match &*guard {
-                        Some(current_field_value) => {
-                            if Arc::ptr_eq(current_field_value, &proc) {
-                                // We're the first to observe an error from `proc`, it's our job to take it out of rotation.
-                                *guard = None;
-                            }
-                        }
-                        None => {
-                            // Another thread was faster to observe the error, and already took the process out of rotation.
-                        }
-                    }
-                }
+                // Avoid concurrent callers hitting the same issue by taking `proc` out of the rotation.
+                // Note that there may be other tasks concurrent with us that also hold `proc`.
+                // We have to deal with that here.
+                // Also read the doc comment on field `self.redo_process`.
+                //
                 // NB: there may still be other concurrent threads using `proc`.
                 // The last one will send SIGKILL when the underlying Arc reaches refcount 0.
-                // NB: it's important to drop(proc) after drop(guard). Otherwise we'd keep
-                // holding the lock while waiting for the process to exit.
-                // NB: the drop impl blocks the current threads with a wait() system call for
-                // the child process. We dropped the `guard` above so that other threads aren't
-                // affected. But, it's good that the current thread _does_ block to wait.
-                // If we instead deferred the waiting into the background / to tokio, it could
-                // happen that if walredo always fails immediately, we spawn processes faster
+                //
+                // NB: the drop impl blocks the dropping thread with a wait() system call for
+                // the child process. In some ways the blocking is actually good: if we
+                // deferred the waiting into the background / to tokio if we used `tokio::process`,
+                // it could happen that if walredo always fails immediately, we spawn processes faster
                 // than we can SIGKILL & `wait` for them to exit. By doing it the way we do here,
                 // we limit this risk of run-away to at most $num_runtimes * $num_executor_threads.
                 // This probably needs revisiting at some later point.
+                match self.redo_process.get() {
+                    None => (),
+                    Some(guard) => {
+                        if Arc::ptr_eq(&proc, &*guard) {
+                            // We're the first to observe an error from `proc`, it's our job to take it out of rotation.
+                            guard.take_and_deinit();
+                        } else {
+                            // Another task already spawned another redo process (further up in this method)
+                            // and put it into `redo_process`. Do nothing, our view of the world is behind.
+                        }
+                    }
+                }
+                // The last task that does this `drop()` of `proc` will do a blocking `wait()` syscall.
                 drop(proc);
             } else if n_attempts != 0 {
                 info!(n_attempts, "retried walredo succeeded");

proxy/Cargo.toml

@@ -12,6 +12,7 @@ testing = []
 anyhow.workspace = true
 async-compression.workspace = true
 async-trait.workspace = true
+atomic-take.workspace = true
 aws-config.workspace = true
 aws-sdk-iam.workspace = true
 aws-sigv4.workspace = true
@@ -36,6 +37,9 @@ http.workspace = true
 humantime.workspace = true
 hyper-tungstenite.workspace = true
 hyper.workspace = true
+hyper1 = { package = "hyper", version = "1.2", features = ["server"] }
+hyper-util = { version = "0.1", features = ["server", "http1", "http2", "tokio"] }
+http-body-util = { version = "0.1" }
 ipnet.workspace = true
 itertools.workspace = true
 lasso = { workspace = true, features = ["multi-threaded"] }

proxy/src/auth/backend/link.rs

@@ -102,8 +102,7 @@ pub(super) async fn authenticate(
 
     ctx.set_user(db_info.user.into());
     ctx.set_project(db_info.aux.clone());
-    let cold_start_info = db_info.aux.cold_start_info.clone().unwrap_or_default();
-    info!(?cold_start_info, "woken up a compute node");
+    info!("woken up a compute node");
 
     // Backwards compatibility. pg_sni_proxy uses "--" in domain names
     // while direct connections do not. Once we migrate to pg_sni_proxy

proxy/src/bin/pg_sni_router.rs

@@ -10,6 +10,7 @@ use itertools::Itertools;
 use proxy::config::TlsServerEndPoint;
 use proxy::context::RequestMonitoring;
 use proxy::proxy::run_until_cancelled;
+use proxy::{BranchId, EndpointId, ProjectId};
 use rustls::pki_types::PrivateKeyDer;
 use tokio::net::TcpListener;
 
@@ -269,7 +270,12 @@ async fn handle_client(
 
     let client = tokio::net::TcpStream::connect(destination).await?;
 
-    let metrics_aux: MetricsAuxInfo = Default::default();
+    let metrics_aux: MetricsAuxInfo = MetricsAuxInfo {
+        endpoint_id: (&EndpointId::from("")).into(),
+        project_id: (&ProjectId::from("")).into(),
+        branch_id: (&BranchId::from("")).into(),
+        cold_start_info: proxy::console::messages::ColdStartInfo::Unknown,
+    };
 
     // doesn't yet matter as pg-sni-router doesn't report analytics logs
     ctx.set_success();

proxy/src/cache/project_info.rs

@@ -16,7 +16,7 @@ use crate::{
     config::ProjectInfoCacheOptions,
     console::AuthSecret,
     intern::{EndpointIdInt, ProjectIdInt, RoleNameInt},
-    EndpointId, ProjectId, RoleName,
+    EndpointId, RoleName,
 };
 
 use super::{Cache, Cached};
@@ -214,14 +214,11 @@ impl ProjectInfoCacheImpl {
     }
     pub fn insert_role_secret(
         &self,
-        project_id: &ProjectId,
-        endpoint_id: &EndpointId,
-        role_name: &RoleName,
+        project_id: ProjectIdInt,
+        endpoint_id: EndpointIdInt,
+        role_name: RoleNameInt,
         secret: Option<AuthSecret>,
     ) {
-        let project_id = ProjectIdInt::from(project_id);
-        let endpoint_id = EndpointIdInt::from(endpoint_id);
-        let role_name = RoleNameInt::from(role_name);
         if self.cache.len() >= self.config.size {
             // If there are too many entries, wait until the next gc cycle.
             return;
@@ -234,12 +231,10 @@ impl ProjectInfoCacheImpl {
     }
     pub fn insert_allowed_ips(
         &self,
-        project_id: &ProjectId,
-        endpoint_id: &EndpointId,
+        project_id: ProjectIdInt,
+        endpoint_id: EndpointIdInt,
         allowed_ips: Arc<Vec<IpPattern>>,
     ) {
-        let project_id = ProjectIdInt::from(project_id);
-        let endpoint_id = EndpointIdInt::from(endpoint_id);
         if self.cache.len() >= self.config.size {
             // If there are too many entries, wait until the next gc cycle.
             return;
@@ -358,7 +353,7 @@ impl Cache for ProjectInfoCacheImpl {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::scram::ServerSecret;
+    use crate::{scram::ServerSecret, ProjectId};
 
     #[tokio::test]
     async fn test_project_info_cache_settings() {
@@ -369,8 +364,8 @@ mod tests {
             ttl: Duration::from_secs(1),
             gc_interval: Duration::from_secs(600),
         });
-        let project_id = "project".into();
-        let endpoint_id = "endpoint".into();
+        let project_id: ProjectId = "project".into();
+        let endpoint_id: EndpointId = "endpoint".into();
         let user1: RoleName = "user1".into();
         let user2: RoleName = "user2".into();
         let secret1 = Some(AuthSecret::Scram(ServerSecret::mock([1; 32])));
@@ -379,9 +374,23 @@ mod tests {
             "127.0.0.1".parse().unwrap(),
             "127.0.0.2".parse().unwrap(),
         ]);
-        cache.insert_role_secret(&project_id, &endpoint_id, &user1, secret1.clone());
-        cache.insert_role_secret(&project_id, &endpoint_id, &user2, secret2.clone());
-        cache.insert_allowed_ips(&project_id, &endpoint_id, allowed_ips.clone());
+        cache.insert_role_secret(
+            (&project_id).into(),
+            (&endpoint_id).into(),
+            (&user1).into(),
+            secret1.clone(),
+        );
+        cache.insert_role_secret(
+            (&project_id).into(),
+            (&endpoint_id).into(),
+            (&user2).into(),
+            secret2.clone(),
+        );
+        cache.insert_allowed_ips(
+            (&project_id).into(),
+            (&endpoint_id).into(),
+            allowed_ips.clone(),
+        );
 
         let cached = cache.get_role_secret(&endpoint_id, &user1).unwrap();
         assert!(cached.cached());
@@ -393,7 +402,12 @@ mod tests {
         // Shouldn't add more than 2 roles.
         let user3: RoleName = "user3".into();
         let secret3 = Some(AuthSecret::Scram(ServerSecret::mock([3; 32])));
-        cache.insert_role_secret(&project_id, &endpoint_id, &user3, secret3.clone());
+        cache.insert_role_secret(
+            (&project_id).into(),
+            (&endpoint_id).into(),
+            (&user3).into(),
+            secret3.clone(),
+        );
         assert!(cache.get_role_secret(&endpoint_id, &user3).is_none());
 
         let cached = cache.get_allowed_ips(&endpoint_id).unwrap();
@@ -421,8 +435,8 @@ mod tests {
         cache.clone().disable_ttl();
         tokio::time::advance(Duration::from_secs(2)).await;
 
-        let project_id = "project".into();
-        let endpoint_id = "endpoint".into();
+        let project_id: ProjectId = "project".into();
+        let endpoint_id: EndpointId = "endpoint".into();
         let user1: RoleName = "user1".into();
         let user2: RoleName = "user2".into();
         let secret1 = Some(AuthSecret::Scram(ServerSecret::mock([1; 32])));
@@ -431,9 +445,23 @@ mod tests {
             "127.0.0.1".parse().unwrap(),
             "127.0.0.2".parse().unwrap(),
         ]);
-        cache.insert_role_secret(&project_id, &endpoint_id, &user1, secret1.clone());
-        cache.insert_role_secret(&project_id, &endpoint_id, &user2, secret2.clone());
-        cache.insert_allowed_ips(&project_id, &endpoint_id, allowed_ips.clone());
+        cache.insert_role_secret(
+            (&project_id).into(),
+            (&endpoint_id).into(),
+            (&user1).into(),
+            secret1.clone(),
+        );
+        cache.insert_role_secret(
+            (&project_id).into(),
+            (&endpoint_id).into(),
+            (&user2).into(),
+            secret2.clone(),
+        );
+        cache.insert_allowed_ips(
+            (&project_id).into(),
+            (&endpoint_id).into(),
+            allowed_ips.clone(),
+        );
 
         tokio::time::advance(Duration::from_secs(2)).await;
         // Nothing should be invalidated.
@@ -470,8 +498,8 @@ mod tests {
             gc_interval: Duration::from_secs(600),
         }));
 
-        let project_id = "project".into();
-        let endpoint_id = "endpoint".into();
+        let project_id: ProjectId = "project".into();
+        let endpoint_id: EndpointId = "endpoint".into();
         let user1: RoleName = "user1".into();
         let user2: RoleName = "user2".into();
         let secret1 = Some(AuthSecret::Scram(ServerSecret::mock([1; 32])));
@@ -480,10 +508,20 @@ mod tests {
             "127.0.0.1".parse().unwrap(),
             "127.0.0.2".parse().unwrap(),
         ]);
-        cache.insert_role_secret(&project_id, &endpoint_id, &user1, secret1.clone());
+        cache.insert_role_secret(
+            (&project_id).into(),
+            (&endpoint_id).into(),
+            (&user1).into(),
+            secret1.clone(),
+        );
         cache.clone().disable_ttl();
         tokio::time::advance(Duration::from_millis(100)).await;
-        cache.insert_role_secret(&project_id, &endpoint_id, &user2, secret2.clone());
+        cache.insert_role_secret(
+            (&project_id).into(),
+            (&endpoint_id).into(),
+            (&user2).into(),
+            secret2.clone(),
+        );
 
         // Added before ttl was disabled + ttl should be still cached.
         let cached = cache.get_role_secret(&endpoint_id, &user1).unwrap();
@@ -497,7 +535,11 @@ mod tests {
         assert!(cache.get_role_secret(&endpoint_id, &user2).is_none());
 
         // Added after ttl was disabled + ttl should not be cached.
-        cache.insert_allowed_ips(&project_id, &endpoint_id, allowed_ips.clone());
+        cache.insert_allowed_ips(
+            (&project_id).into(),
+            (&endpoint_id).into(),
+            allowed_ips.clone(),
+        );
         let cached = cache.get_allowed_ips(&endpoint_id).unwrap();
         assert!(!cached.cached());
 

proxy/src/compute.rs

@@ -276,6 +276,7 @@ impl ConnCfg {
         let stream = connection.stream.into_inner();
 
         info!(
+            cold_start_info = ctx.cold_start_info.as_str(),
             "connected to compute node at {host} ({socket_addr}) sslmode={:?}",
             self.0.get_ssl_mode()
         );

proxy/src/console/messages.rs

@@ -3,7 +3,7 @@ use std::fmt;
 
 use crate::auth::IpPattern;
 
-use crate::{BranchId, EndpointId, ProjectId};
+use crate::intern::{BranchIdInt, EndpointIdInt, ProjectIdInt};
 
 /// Generic error response with human-readable description.
 /// Note that we can't always present it to user as is.
@@ -18,7 +18,7 @@ pub struct ConsoleError {
 pub struct GetRoleSecret {
     pub role_secret: Box<str>,
     pub allowed_ips: Option<Vec<IpPattern>>,
-    pub project_id: Option<ProjectId>,
+    pub project_id: Option<ProjectIdInt>,
 }
 
 // Manually implement debug to omit sensitive info.
@@ -93,22 +93,47 @@ impl fmt::Debug for DatabaseInfo {
 
 /// Various labels for prometheus metrics.
 /// Also known as `ProxyMetricsAuxInfo` in the console.
-#[derive(Debug, Deserialize, Clone, Default)]
+#[derive(Debug, Deserialize, Clone)]
 pub struct MetricsAuxInfo {
-    pub endpoint_id: EndpointId,
-    pub project_id: ProjectId,
-    pub branch_id: BranchId,
-    pub cold_start_info: Option<ColdStartInfo>,
+    pub endpoint_id: EndpointIdInt,
+    pub project_id: ProjectIdInt,
+    pub branch_id: BranchIdInt,
+    #[serde(default)]
+    pub cold_start_info: ColdStartInfo,
 }
 
-#[derive(Debug, Default, Serialize, Deserialize, Clone)]
+#[derive(Debug, Default, Serialize, Deserialize, Clone, Copy)]
 #[serde(rename_all = "snake_case")]
 pub enum ColdStartInfo {
     #[default]
-    Unknown = 0,
-    Warm = 1,
-    PoolHit = 2,
-    PoolMiss = 3,
+    Unknown,
+    /// Compute was already running
+    Warm,
+    #[serde(rename = "pool_hit")]
+    /// Compute was not running but there was an available VM
+    VmPoolHit,
+    #[serde(rename = "pool_miss")]
+    /// Compute was not running and there were no VMs available
+    VmPoolMiss,
+
+    // not provided by control plane
+    /// Connection available from HTTP pool
+    HttpPoolHit,
+    /// Cached connection info
+    WarmCached,
+}
+
+impl ColdStartInfo {
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            ColdStartInfo::Unknown => "unknown",
+            ColdStartInfo::Warm => "warm",
+            ColdStartInfo::VmPoolHit => "pool_hit",
+            ColdStartInfo::VmPoolMiss => "pool_miss",
+            ColdStartInfo::HttpPoolHit => "http_pool_hit",
+            ColdStartInfo::WarmCached => "warm_cached",
+        }
+    }
 }
 
 #[cfg(test)]

proxy/src/console/provider.rs

@@ -12,7 +12,8 @@ use crate::{
     compute,
     config::{CacheOptions, ProjectInfoCacheOptions},
     context::RequestMonitoring,
-    scram, EndpointCacheKey, ProjectId,
+    intern::ProjectIdInt,
+    scram, EndpointCacheKey,
 };
 use dashmap::DashMap;
 use std::{sync::Arc, time::Duration};
@@ -271,7 +272,7 @@ pub struct AuthInfo {
     /// List of IP addresses allowed for the autorization.
     pub allowed_ips: Vec<IpPattern>,
     /// Project ID. This is used for cache invalidation.
-    pub project_id: Option<ProjectId>,
+    pub project_id: Option<ProjectIdInt>,
 }
 
 /// Info for establishing a connection to a compute node.

proxy/src/console/provider/mock.rs

@@ -4,10 +4,16 @@ use super::{
     errors::{ApiError, GetAuthInfoError, WakeComputeError},
     AuthInfo, AuthSecret, CachedNodeInfo, NodeInfo,
 };
-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 crate::{
+    console::{
+        messages::MetricsAuxInfo,
+        provider::{CachedAllowedIps, CachedRoleSecret},
+    },
+    BranchId, EndpointId, ProjectId,
+};
 use futures::TryFutureExt;
 use std::{str::FromStr, sync::Arc};
 use thiserror::Error;
@@ -114,7 +120,12 @@ impl Api {
 
         let node = NodeInfo {
             config,
-            aux: Default::default(),
+            aux: MetricsAuxInfo {
+                endpoint_id: (&EndpointId::from("endpoint")).into(),
+                project_id: (&ProjectId::from("project")).into(),
+                branch_id: (&BranchId::from("branch")).into(),
+                cold_start_info: crate::console::messages::ColdStartInfo::Warm,
+            },
             allow_self_signed_compute: false,
         };
 

proxy/src/console/provider/neon.rs

@@ -181,15 +181,16 @@ impl super::Api for Api {
         }
         let auth_info = self.do_get_auth_info(ctx, user_info).await?;
         if let Some(project_id) = auth_info.project_id {
+            let ep_int = ep.into();
             self.caches.project_info.insert_role_secret(
-                &project_id,
-                ep,
-                user,
+                project_id,
+                ep_int,
+                user.into(),
                 auth_info.secret.clone(),
             );
             self.caches.project_info.insert_allowed_ips(
-                &project_id,
-                ep,
+                project_id,
+                ep_int,
                 Arc::new(auth_info.allowed_ips),
             );
             ctx.set_project_id(project_id);
@@ -217,15 +218,16 @@ impl super::Api for Api {
         let allowed_ips = Arc::new(auth_info.allowed_ips);
         let user = &user_info.user;
         if let Some(project_id) = auth_info.project_id {
+            let ep_int = ep.into();
             self.caches.project_info.insert_role_secret(
-                &project_id,
-                ep,
-                user,
+                project_id,
+                ep_int,
+                user.into(),
                 auth_info.secret.clone(),
             );
             self.caches
                 .project_info
-                .insert_allowed_ips(&project_id, ep, allowed_ips.clone());
+                .insert_allowed_ips(project_id, ep_int, allowed_ips.clone());
             ctx.set_project_id(project_id);
         }
         Ok((
@@ -248,8 +250,7 @@ impl super::Api for Api {
         // which means that we might cache it to reduce the load and latency.
         if let Some(cached) = self.caches.node_info.get(&key) {
             info!(key = &*key, "found cached compute node info");
-            info!("cold_start_info=warm");
-            ctx.set_cold_start_info(ColdStartInfo::Warm);
+            ctx.set_project(cached.aux.clone());
             return Ok(cached);
         }
 
@@ -260,17 +261,21 @@ impl super::Api for Api {
         if permit.should_check_cache() {
             if let Some(cached) = self.caches.node_info.get(&key) {
                 info!(key = &*key, "found cached compute node info");
-                info!("cold_start_info=warm");
-                ctx.set_cold_start_info(ColdStartInfo::Warm);
+                ctx.set_project(cached.aux.clone());
                 return Ok(cached);
             }
         }
 
-        let node = self.do_wake_compute(ctx, user_info).await?;
+        let mut node = self.do_wake_compute(ctx, user_info).await?;
         ctx.set_project(node.aux.clone());
-        let cold_start_info = node.aux.cold_start_info.clone().unwrap_or_default();
-        info!(?cold_start_info, "woken up a compute node");
-        let (_, cached) = self.caches.node_info.insert(key.clone(), node);
+        let cold_start_info = node.aux.cold_start_info;
+        info!("woken up a compute node");
+
+        // store the cached node as 'warm'
+        node.aux.cold_start_info = ColdStartInfo::WarmCached;
+        let (_, mut cached) = self.caches.node_info.insert(key.clone(), node);
+        cached.aux.cold_start_info = cold_start_info;
+
         info!(key = &*key, "created a cache entry for compute node info");
 
         Ok(cached)

proxy/src/context.rs

@@ -11,8 +11,9 @@ use uuid::Uuid;
 use crate::{
     console::messages::{ColdStartInfo, MetricsAuxInfo},
     error::ErrorKind,
+    intern::{BranchIdInt, ProjectIdInt},
     metrics::{LatencyTimer, ENDPOINT_ERRORS_BY_KIND, ERROR_BY_KIND},
-    BranchId, DbName, EndpointId, ProjectId, RoleName,
+    DbName, EndpointId, RoleName,
 };
 
 use self::parquet::RequestData;
@@ -34,8 +35,8 @@ pub struct RequestMonitoring {
     pub span: Span,
 
     // filled in as they are discovered
-    project: Option<ProjectId>,
-    branch: Option<BranchId>,
+    project: Option<ProjectIdInt>,
+    branch: Option<BranchIdInt>,
     endpoint_id: Option<EndpointId>,
     dbname: Option<DbName>,
     user: Option<RoleName>,
@@ -43,7 +44,7 @@ pub struct RequestMonitoring {
     error_kind: Option<ErrorKind>,
     pub(crate) auth_method: Option<AuthMethod>,
     success: bool,
-    cold_start_info: Option<ColdStartInfo>,
+    pub(crate) cold_start_info: ColdStartInfo,
 
     // extra
     // This sender is here to keep the request monitoring channel open while requests are taking place.
@@ -92,7 +93,7 @@ impl RequestMonitoring {
             error_kind: None,
             auth_method: None,
             success: false,
-            cold_start_info: None,
+            cold_start_info: ColdStartInfo::Unknown,
 
             sender: LOG_CHAN.get().and_then(|tx| tx.upgrade()),
             latency_timer: LatencyTimer::new(protocol),
@@ -113,26 +114,31 @@ impl RequestMonitoring {
     }
 
     pub fn set_cold_start_info(&mut self, info: ColdStartInfo) {
-        self.cold_start_info = Some(info);
+        self.cold_start_info = info;
+        self.latency_timer.cold_start_info(info);
     }
 
     pub fn set_project(&mut self, x: MetricsAuxInfo) {
-        self.set_endpoint_id(x.endpoint_id);
+        if self.endpoint_id.is_none() {
+            self.set_endpoint_id(x.endpoint_id.as_str().into())
+        }
         self.branch = Some(x.branch_id);
         self.project = Some(x.project_id);
-        self.cold_start_info = x.cold_start_info;
+        self.set_cold_start_info(x.cold_start_info);
     }
 
-    pub fn set_project_id(&mut self, project_id: ProjectId) {
+    pub fn set_project_id(&mut self, project_id: ProjectIdInt) {
         self.project = Some(project_id);
     }
 
     pub fn set_endpoint_id(&mut self, endpoint_id: EndpointId) {
-        self.span.record("ep", display(&endpoint_id));
-        crate::metrics::CONNECTING_ENDPOINTS
-            .with_label_values(&[self.protocol])
-            .measure(&endpoint_id);
-        self.endpoint_id = Some(endpoint_id);
+        if self.endpoint_id.is_none() {
+            self.span.record("ep", display(&endpoint_id));
+            crate::metrics::CONNECTING_ENDPOINTS
+                .with_label_values(&[self.protocol])
+                .measure(&endpoint_id);
+            self.endpoint_id = Some(endpoint_id);
+        }
     }
 
     pub fn set_application(&mut self, app: Option<SmolStr>) {

proxy/src/context/parquet.rs

@@ -87,7 +87,7 @@ pub struct RequestData {
     /// Or if we make it to proxy_pass
     success: bool,
     /// Indicates if the cplane started the new compute node for this request.
-    cold_start_info: Option<&'static str>,
+    cold_start_info: &'static str,
     /// Tracks time from session start (HTTP request/libpq TCP handshake)
     /// Through to success/failure
     duration_us: u64,
@@ -115,12 +115,7 @@ impl From<&RequestMonitoring> for RequestData {
             region: value.region,
             error: value.error_kind.as_ref().map(|e| e.to_metric_label()),
             success: value.success,
-            cold_start_info: value.cold_start_info.as_ref().map(|x| match x {
-                crate::console::messages::ColdStartInfo::Unknown => "unknown",
-                crate::console::messages::ColdStartInfo::Warm => "warm",
-                crate::console::messages::ColdStartInfo::PoolHit => "pool_hit",
-                crate::console::messages::ColdStartInfo::PoolMiss => "pool_miss",
-            }),
+            cold_start_info: value.cold_start_info.as_str(),
             duration_us: SystemTime::from(value.first_packet)
                 .elapsed()
                 .unwrap_or_default()
@@ -454,7 +449,7 @@ mod tests {
             region: "us-east-1",
             error: None,
             success: rng.gen(),
-            cold_start_info: Some("no"),
+            cold_start_info: "no",
             duration_us: rng.gen_range(0..30_000_000),
         }
     }
@@ -524,15 +519,15 @@ mod tests {
         assert_eq!(
             file_stats,
             [
-                (1314406, 3, 6000),
-                (1314399, 3, 6000),
-                (1314459, 3, 6000),
-                (1314416, 3, 6000),
-                (1314546, 3, 6000),
-                (1314388, 3, 6000),
-                (1314180, 3, 6000),
-                (1314416, 3, 6000),
-                (438359, 1, 2000)
+                (1314385, 3, 6000),
+                (1314378, 3, 6000),
+                (1314438, 3, 6000),
+                (1314395, 3, 6000),
+                (1314525, 3, 6000),
+                (1314367, 3, 6000),
+                (1314159, 3, 6000),
+                (1314395, 3, 6000),
+                (438352, 1, 2000)
             ]
         );
 
@@ -562,11 +557,11 @@ mod tests {
         assert_eq!(
             file_stats,
             [
-                (1220668, 5, 10000),
-                (1226818, 5, 10000),
-                (1228612, 5, 10000),
-                (1227974, 5, 10000),
-                (1219252, 5, 10000)
+                (1220633, 5, 10000),
+                (1226783, 5, 10000),
+                (1228577, 5, 10000),
+                (1227939, 5, 10000),
+                (1219217, 5, 10000)
             ]
         );
 
@@ -598,11 +593,11 @@ mod tests {
         assert_eq!(
             file_stats,
             [
-                (1206315, 5, 10000),
-                (1206046, 5, 10000),
-                (1206339, 5, 10000),
-                (1206327, 5, 10000),
-                (1206582, 5, 10000)
+                (1206280, 5, 10000),
+                (1206011, 5, 10000),
+                (1206304, 5, 10000),
+                (1206292, 5, 10000),
+                (1206547, 5, 10000)
             ]
         );
 
@@ -627,15 +622,15 @@ mod tests {
         assert_eq!(
             file_stats,
             [
-                (1314406, 3, 6000),
-                (1314399, 3, 6000),
-                (1314459, 3, 6000),
-                (1314416, 3, 6000),
-                (1314546, 3, 6000),
-                (1314388, 3, 6000),
-                (1314180, 3, 6000),
-                (1314416, 3, 6000),
-                (438359, 1, 2000)
+                (1314385, 3, 6000),
+                (1314378, 3, 6000),
+                (1314438, 3, 6000),
+                (1314395, 3, 6000),
+                (1314525, 3, 6000),
+                (1314367, 3, 6000),
+                (1314159, 3, 6000),
+                (1314395, 3, 6000),
+                (438352, 1, 2000)
             ]
         );
 
@@ -672,7 +667,7 @@ mod tests {
         // files are smaller than the size threshold, but they took too long to fill so were flushed early
         assert_eq!(
             file_stats,
-            [(658837, 2, 3001), (658551, 2, 3000), (658347, 2, 2999)]
+            [(658823, 2, 3001), (658537, 2, 3000), (658333, 2, 2999)]
         );
 
         tmpdir.close().unwrap();

proxy/src/metrics.rs

@@ -12,6 +12,8 @@ use metrics::{
 use once_cell::sync::Lazy;
 use tokio::time::{self, Instant};
 
+use crate::console::messages::ColdStartInfo;
+
 pub static NUM_DB_CONNECTIONS_GAUGE: Lazy<IntCounterPairVec> = Lazy::new(|| {
     register_int_counter_pair_vec!(
         "proxy_opened_db_connections_total",
@@ -50,8 +52,8 @@ pub static COMPUTE_CONNECTION_LATENCY: Lazy<HistogramVec> = Lazy::new(|| {
         "proxy_compute_connection_latency_seconds",
         "Time it took for proxy to establish a connection to the compute endpoint",
         // http/ws/tcp, true/false, true/false, success/failure, client/client_and_cplane
-        // 3 * 2 * 2 * 2 * 2 = 48 counters
-        &["protocol", "cache_miss", "pool_miss", "outcome", "excluded"],
+        // 3 * 6 * 2 * 2 = 72 counters
+        &["protocol", "cold_start_info", "outcome", "excluded"],
         // largest bucket = 2^16 * 0.5ms = 32s
         exponential_buckets(0.0005, 2.0, 16).unwrap(),
     )
@@ -183,6 +185,20 @@ struct Accumulated {
     compute: time::Duration,
 }
 
+enum Outcome {
+    Success,
+    Failed,
+}
+
+impl Outcome {
+    fn as_str(&self) -> &'static str {
+        match self {
+            Outcome::Success => "success",
+            Outcome::Failed => "failed",
+        }
+    }
+}
+
 pub struct LatencyTimer {
     // time since the stopwatch was started
     start: time::Instant,
@@ -192,9 +208,8 @@ pub struct LatencyTimer {
     accumulated: Accumulated,
     // label data
     protocol: &'static str,
-    cache_miss: bool,
-    pool_miss: bool,
-    outcome: &'static str,
+    cold_start_info: ColdStartInfo,
+    outcome: Outcome,
 }
 
 pub struct LatencyTimerPause<'a> {
@@ -210,11 +225,9 @@ impl LatencyTimer {
             stop: None,
             accumulated: Accumulated::default(),
             protocol,
-            cache_miss: false,
-            // by default we don't do pooling
-            pool_miss: true,
+            cold_start_info: ColdStartInfo::Unknown,
             // assume failed unless otherwise specified
-            outcome: "failed",
+            outcome: Outcome::Failed,
         }
     }
 
@@ -226,12 +239,8 @@ impl LatencyTimer {
         }
     }
 
-    pub fn cache_miss(&mut self) {
-        self.cache_miss = true;
-    }
-
-    pub fn pool_hit(&mut self) {
-        self.pool_miss = false;
+    pub fn cold_start_info(&mut self, cold_start_info: ColdStartInfo) {
+        self.cold_start_info = cold_start_info;
     }
 
     pub fn success(&mut self) {
@@ -239,7 +248,7 @@ impl LatencyTimer {
         self.stop = Some(time::Instant::now());
 
         // success
-        self.outcome = "success";
+        self.outcome = Outcome::Success;
     }
 }
 
@@ -264,9 +273,8 @@ impl Drop for LatencyTimer {
         COMPUTE_CONNECTION_LATENCY
             .with_label_values(&[
                 self.protocol,
-                bool_to_str(self.cache_miss),
-                bool_to_str(self.pool_miss),
-                self.outcome,
+                self.cold_start_info.as_str(),
+                self.outcome.as_str(),
                 "client",
             ])
             .observe((duration.saturating_sub(self.accumulated.client)).as_secs_f64());
@@ -275,9 +283,8 @@ impl Drop for LatencyTimer {
         COMPUTE_CONNECTION_LATENCY
             .with_label_values(&[
                 self.protocol,
-                bool_to_str(self.cache_miss),
-                bool_to_str(self.pool_miss),
-                self.outcome,
+                self.cold_start_info.as_str(),
+                self.outcome.as_str(),
                 "client_and_cplane",
             ])
             .observe((duration.saturating_sub(accumulated_total)).as_secs_f64());

proxy/src/protocol2.rs

@@ -5,19 +5,13 @@ use std::{
     io,
     net::SocketAddr,
     pin::{pin, Pin},
-    sync::Mutex,
     task::{ready, Context, Poll},
 };
 
 use bytes::{Buf, BytesMut};
-use hyper::server::accept::Accept;
-use hyper::server::conn::{AddrIncoming, AddrStream};
-use metrics::IntCounterPairGuard;
+use hyper::server::conn::AddrIncoming;
 use pin_project_lite::pin_project;
 use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, ReadBuf};
-use uuid::Uuid;
-
-use crate::metrics::NUM_CLIENT_CONNECTION_GAUGE;
 
 pub struct ProxyProtocolAccept {
     pub incoming: AddrIncoming,
@@ -331,103 +325,6 @@ impl<T: AsyncRead> AsyncRead for WithClientIp<T> {
     }
 }
 
-impl Accept for ProxyProtocolAccept {
-    type Conn = WithConnectionGuard<WithClientIp<AddrStream>>;
-
-    type Error = io::Error;
-
-    fn poll_accept(
-        mut self: Pin<&mut Self>,
-        cx: &mut Context<'_>,
-    ) -> Poll<Option<Result<Self::Conn, Self::Error>>> {
-        let conn = ready!(Pin::new(&mut self.incoming).poll_accept(cx)?);
-
-        let conn_id = uuid::Uuid::new_v4();
-        let span = tracing::info_span!("http_conn", ?conn_id);
-        {
-            let _enter = span.enter();
-            tracing::info!("accepted new TCP connection");
-        }
-
-        let Some(conn) = conn else {
-            return Poll::Ready(None);
-        };
-
-        Poll::Ready(Some(Ok(WithConnectionGuard {
-            inner: WithClientIp::new(conn),
-            connection_id: Uuid::new_v4(),
-            gauge: Mutex::new(Some(
-                NUM_CLIENT_CONNECTION_GAUGE
-                    .with_label_values(&[self.protocol])
-                    .guard(),
-            )),
-            span,
-        })))
-    }
-}
-
-pin_project! {
-    pub struct WithConnectionGuard<T> {
-        #[pin]
-        pub inner: T,
-        pub connection_id: Uuid,
-        pub gauge: Mutex<Option<IntCounterPairGuard>>,
-        pub span: tracing::Span,
-    }
-
-    impl<S> PinnedDrop for WithConnectionGuard<S> {
-        fn drop(this: Pin<&mut Self>) {
-            let _enter = this.span.enter();
-            tracing::info!("HTTP connection closed")
-        }
-    }
-}
-
-impl<T: AsyncWrite> AsyncWrite for WithConnectionGuard<T> {
-    #[inline]
-    fn poll_write(
-        self: Pin<&mut Self>,
-        cx: &mut Context<'_>,
-        buf: &[u8],
-    ) -> Poll<Result<usize, io::Error>> {
-        self.project().inner.poll_write(cx, buf)
-    }
-
-    #[inline]
-    fn poll_flush(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Result<(), io::Error>> {
-        self.project().inner.poll_flush(cx)
-    }
-
-    #[inline]
-    fn poll_shutdown(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Result<(), io::Error>> {
-        self.project().inner.poll_shutdown(cx)
-    }
-
-    #[inline]
-    fn poll_write_vectored(
-        self: Pin<&mut Self>,
-        cx: &mut Context<'_>,
-        bufs: &[io::IoSlice<'_>],
-    ) -> Poll<Result<usize, io::Error>> {
-        self.project().inner.poll_write_vectored(cx, bufs)
-    }
-
-    #[inline]
-    fn is_write_vectored(&self) -> bool {
-        self.inner.is_write_vectored()
-    }
-}
-
-impl<T: AsyncRead> AsyncRead for WithConnectionGuard<T> {
-    fn poll_read(
-        self: Pin<&mut Self>,
-        cx: &mut Context<'_>,
-        buf: &mut ReadBuf<'_>,
-    ) -> Poll<io::Result<()>> {
-        self.project().inner.poll_read(cx, buf)
-    }
-}
-
 #[cfg(test)]
 mod tests {
     use std::pin::pin;

proxy/src/proxy/connect_compute.rs

@@ -87,7 +87,6 @@ impl ConnectMechanism for TcpMechanism<'_> {
 }
 
 /// Try to connect to the compute node, retrying if necessary.
-/// This function might update `node_info`, so we take it by `&mut`.
 #[tracing::instrument(skip_all)]
 pub async fn connect_to_compute<M: ConnectMechanism, B: ComputeConnectBackend>(
     ctx: &mut RequestMonitoring,
@@ -132,7 +131,6 @@ where
     } else {
         // if we failed to connect, it's likely that the compute node was suspended, wake a new compute node
         info!("compute node's state has likely changed; requesting a wake-up");
-        ctx.latency_timer.cache_miss();
         let old_node_info = invalidate_cache(node_info);
         let mut node_info = wake_compute(&mut num_retries, ctx, user_info).await?;
         node_info.reuse_settings(old_node_info);

proxy/src/proxy/passthrough.rs

@@ -19,8 +19,8 @@ pub async fn proxy_pass(
     aux: MetricsAuxInfo,
 ) -> anyhow::Result<()> {
     let usage = USAGE_METRICS.register(Ids {
-        endpoint_id: aux.endpoint_id.clone(),
-        branch_id: aux.branch_id.clone(),
+        endpoint_id: aux.endpoint_id,
+        branch_id: aux.branch_id,
     });
 
     let m_sent = NUM_BYTES_PROXIED_COUNTER.with_label_values(&["tx"]);

proxy/src/proxy/tests.rs

@@ -12,11 +12,12 @@ use crate::auth::backend::{
 };
 use crate::config::CertResolver;
 use crate::console::caches::NodeInfoCache;
+use crate::console::messages::MetricsAuxInfo;
 use crate::console::provider::{CachedAllowedIps, CachedRoleSecret, ConsoleBackend};
 use crate::console::{self, CachedNodeInfo, NodeInfo};
 use crate::error::ErrorKind;
 use crate::proxy::retry::{retry_after, NUM_RETRIES_CONNECT};
-use crate::{http, sasl, scram};
+use crate::{http, sasl, scram, BranchId, EndpointId, ProjectId};
 use anyhow::{bail, Context};
 use async_trait::async_trait;
 use rstest::rstest;
@@ -512,7 +513,12 @@ impl TestBackend for TestConnectMechanism {
 fn helper_create_cached_node_info(cache: &'static NodeInfoCache) -> CachedNodeInfo {
     let node = NodeInfo {
         config: compute::ConnCfg::new(),
-        aux: Default::default(),
+        aux: MetricsAuxInfo {
+            endpoint_id: (&EndpointId::from("endpoint")).into(),
+            project_id: (&ProjectId::from("project")).into(),
+            branch_id: (&BranchId::from("branch")).into(),
+            cold_start_info: crate::console::messages::ColdStartInfo::Warm,
+        },
         allow_self_signed_compute: false,
     };
     let (_, node) = cache.insert("key".into(), node);

proxy/src/serverless.rs

@@ -4,42 +4,48 @@
 
 mod backend;
 mod conn_pool;
+mod http_util;
 mod json;
 mod sql_over_http;
-pub mod tls_listener;
 mod websocket;
 
+use atomic_take::AtomicTake;
+use bytes::Bytes;
 pub use conn_pool::GlobalConnPoolOptions;
 
-use anyhow::bail;
-use hyper::StatusCode;
-use metrics::IntCounterPairGuard;
+use anyhow::Context;
+use futures::future::{select, Either};
+use futures::TryFutureExt;
+use http::{Method, Response, StatusCode};
+use http_body_util::Full;
+use hyper1::body::Incoming;
+use hyper_util::rt::TokioExecutor;
+use hyper_util::server::conn::auto::Builder;
 use rand::rngs::StdRng;
 use rand::SeedableRng;
 pub use reqwest_middleware::{ClientWithMiddleware, Error};
 pub use reqwest_retry::{policies::ExponentialBackoff, RetryTransientMiddleware};
+use tokio::time::timeout;
+use tokio_rustls::TlsAcceptor;
 use tokio_util::task::TaskTracker;
-use tracing::instrument::Instrumented;
 
 use crate::cancellation::CancellationHandlerMain;
 use crate::config::ProxyConfig;
 use crate::context::RequestMonitoring;
-use crate::protocol2::{ProxyProtocolAccept, WithClientIp, WithConnectionGuard};
+use crate::metrics::{NUM_CLIENT_CONNECTION_GAUGE, TLS_HANDSHAKE_FAILURES};
+use crate::protocol2::WithClientIp;
+use crate::proxy::run_until_cancelled;
 use crate::rate_limiter::EndpointRateLimiter;
 use crate::serverless::backend::PoolingBackend;
-use hyper::{
-    server::conn::{AddrIncoming, AddrStream},
-    Body, Method, Request, Response,
-};
+use crate::serverless::http_util::{api_error_into_response, json_response};
 
-use std::net::IpAddr;
+use std::net::{IpAddr, SocketAddr};
+use std::pin::pin;
 use std::sync::Arc;
-use std::task::Poll;
-use tls_listener::TlsListener;
-use tokio::net::TcpListener;
-use tokio_util::sync::{CancellationToken, DropGuard};
+use tokio::net::{TcpListener, TcpStream};
+use tokio_util::sync::CancellationToken;
 use tracing::{error, info, warn, Instrument};
-use utils::http::{error::ApiError, json::json_response};
+use utils::http::error::ApiError;
 
 pub const SERVERLESS_DRIVER_SNI: &str = "api";
 
@@ -91,161 +97,174 @@ pub async fn task_main(
     tls_server_config.alpn_protocols = vec![b"h2".to_vec(), b"http/1.1".to_vec()];
     let tls_acceptor: tokio_rustls::TlsAcceptor = Arc::new(tls_server_config).into();
 
-    let mut addr_incoming = AddrIncoming::from_listener(ws_listener)?;
-    let _ = addr_incoming.set_nodelay(true);
-    let addr_incoming = ProxyProtocolAccept {
-        incoming: addr_incoming,
-        protocol: "http",
-    };
+    let connections = tokio_util::task::task_tracker::TaskTracker::new();
+    connections.close(); // allows `connections.wait to complete`
 
-    let ws_connections = tokio_util::task::task_tracker::TaskTracker::new();
-    ws_connections.close(); // allows `ws_connections.wait to complete`
+    let server = Builder::new(hyper_util::rt::TokioExecutor::new());
 
-    let tls_listener = TlsListener::new(tls_acceptor, addr_incoming, config.handshake_timeout);
+    while let Some(res) = run_until_cancelled(ws_listener.accept(), &cancellation_token).await {
+        let (conn, peer_addr) = res.context("could not accept TCP stream")?;
+        if let Err(e) = conn.set_nodelay(true) {
+            tracing::error!("could not set nodelay: {e}");
+            continue;
+        }
+        let conn_id = uuid::Uuid::new_v4();
+        let http_conn_span = tracing::info_span!("http_conn", ?conn_id);
 
-    let make_svc = hyper::service::make_service_fn(
-        |stream: &tokio_rustls::server::TlsStream<
-            WithConnectionGuard<WithClientIp<AddrStream>>,
-        >| {
-            let (conn, _) = stream.get_ref();
+        connections.spawn(
+            connection_handler(
+                config,
+                backend.clone(),
+                connections.clone(),
+                cancellation_handler.clone(),
+                endpoint_rate_limiter.clone(),
+                cancellation_token.clone(),
+                server.clone(),
+                tls_acceptor.clone(),
+                conn,
+                peer_addr,
+            )
+            .instrument(http_conn_span),
+        );
+    }
 
-            // this is jank. should dissapear with hyper 1.0 migration.
-            let gauge = conn
-                .gauge
-                .lock()
-                .expect("lock should not be poisoned")
-                .take()
-                .expect("gauge should be set on connection start");
-
-            // Cancel all current inflight HTTP requests if the HTTP connection is closed.
-            let http_cancellation_token = CancellationToken::new();
-            let cancel_connection = http_cancellation_token.clone().drop_guard();
-
-            let span = conn.span.clone();
-            let client_addr = conn.inner.client_addr();
-            let remote_addr = conn.inner.inner.remote_addr();
-            let backend = backend.clone();
-            let ws_connections = ws_connections.clone();
-            let endpoint_rate_limiter = endpoint_rate_limiter.clone();
-            let cancellation_handler = cancellation_handler.clone();
-            async move {
-                let peer_addr = match client_addr {
-                    Some(addr) => addr,
-                    None if config.require_client_ip => bail!("missing required client ip"),
-                    None => remote_addr,
-                };
-                Ok(MetricService::new(
-                    hyper::service::service_fn(move |req: Request<Body>| {
-                        let backend = backend.clone();
-                        let ws_connections2 = ws_connections.clone();
-                        let endpoint_rate_limiter = endpoint_rate_limiter.clone();
-                        let cancellation_handler = cancellation_handler.clone();
-                        let http_cancellation_token = http_cancellation_token.child_token();
-
-                        // `request_handler` is not cancel safe. It expects to be cancelled only at specific times.
-                        // By spawning the future, we ensure it never gets cancelled until it decides to.
-                        ws_connections.spawn(
-                            async move {
-                                // Cancel the current inflight HTTP request if the requets stream is closed.
-                                // This is slightly different to `_cancel_connection` in that
-                                // h2 can cancel individual requests with a `RST_STREAM`.
-                                let _cancel_session = http_cancellation_token.clone().drop_guard();
-
-                                let res = request_handler(
-                                    req,
-                                    config,
-                                    backend,
-                                    ws_connections2,
-                                    cancellation_handler,
-                                    peer_addr.ip(),
-                                    endpoint_rate_limiter,
-                                    http_cancellation_token,
-                                )
-                                .await
-                                .map_or_else(|e| e.into_response(), |r| r);
-
-                                _cancel_session.disarm();
-
-                                res
-                            }
-                            .in_current_span(),
-                        )
-                    }),
-                    gauge,
-                    cancel_connection,
-                    span,
-                ))
-            }
-        },
-    );
-
-    hyper::Server::builder(tls_listener)
-        .serve(make_svc)
-        .with_graceful_shutdown(cancellation_token.cancelled())
-        .await?;
-
-    // await websocket connections
-    ws_connections.wait().await;
+    connections.wait().await;
 
     Ok(())
 }
 
-struct MetricService<S> {
-    inner: S,
-    _gauge: IntCounterPairGuard,
-    _cancel: DropGuard,
-    span: tracing::Span,
-}
+/// Handles the TCP lifecycle.
+///
+/// 1. Parses PROXY protocol V2
+/// 2. Handles TLS handshake
+/// 3. Handles HTTP connection
+///     1. With graceful shutdowns
+///     2. With graceful request cancellation with connection failure
+///     3. With websocket upgrade support.
+#[allow(clippy::too_many_arguments)]
+async fn connection_handler(
+    config: &'static ProxyConfig,
+    backend: Arc<PoolingBackend>,
+    connections: TaskTracker,
+    cancellation_handler: Arc<CancellationHandlerMain>,
+    endpoint_rate_limiter: Arc<EndpointRateLimiter>,
+    cancellation_token: CancellationToken,
+    server: Builder<TokioExecutor>,
+    tls_acceptor: TlsAcceptor,
+    conn: TcpStream,
+    peer_addr: SocketAddr,
+) {
+    let session_id = uuid::Uuid::new_v4();
 
-impl<S> MetricService<S> {
-    fn new(
-        inner: S,
-        _gauge: IntCounterPairGuard,
-        _cancel: DropGuard,
-        span: tracing::Span,
-    ) -> MetricService<S> {
-        MetricService {
-            inner,
-            _gauge,
-            _cancel,
-            span,
+    let _gauge = NUM_CLIENT_CONNECTION_GAUGE
+        .with_label_values(&["http"])
+        .guard();
+
+    // handle PROXY protocol
+    let mut conn = WithClientIp::new(conn);
+    let peer = match conn.wait_for_addr().await {
+        Ok(peer) => peer,
+        Err(e) => {
+            tracing::error!(?session_id, %peer_addr, "failed to accept TCP connection: invalid PROXY protocol V2 header: {e:#}");
+            return;
         }
-    }
-}
+    };
 
-impl<S, ReqBody> hyper::service::Service<Request<ReqBody>> for MetricService<S>
-where
-    S: hyper::service::Service<Request<ReqBody>>,
-{
-    type Response = S::Response;
-    type Error = S::Error;
-    type Future = Instrumented<S::Future>;
+    let peer_addr = peer.unwrap_or(peer_addr).ip();
+    info!(?session_id, %peer_addr, "accepted new TCP connection");
 
-    fn poll_ready(&mut self, cx: &mut std::task::Context<'_>) -> Poll<Result<(), Self::Error>> {
-        self.inner.poll_ready(cx)
-    }
+    // try upgrade to TLS, but with a timeout.
+    let conn = match timeout(config.handshake_timeout, tls_acceptor.accept(conn)).await {
+        Ok(Ok(conn)) => {
+            info!(?session_id, %peer_addr, "accepted new TLS connection");
+            conn
+        }
+        // The handshake failed
+        Ok(Err(e)) => {
+            TLS_HANDSHAKE_FAILURES.inc();
+            warn!(?session_id, %peer_addr, "failed to accept TLS connection: {e:?}");
+            return;
+        }
+        // The handshake timed out
+        Err(e) => {
+            TLS_HANDSHAKE_FAILURES.inc();
+            warn!(?session_id, %peer_addr, "failed to accept TLS connection: {e:?}");
+            return;
+        }
+    };
 
-    fn call(&mut self, req: Request<ReqBody>) -> Self::Future {
-        self.span
-            .in_scope(|| self.inner.call(req))
-            .instrument(self.span.clone())
+    let session_id = AtomicTake::new(session_id);
+
+    // Cancel all current inflight HTTP requests if the HTTP connection is closed.
+    let http_cancellation_token = CancellationToken::new();
+    let _cancel_connection = http_cancellation_token.clone().drop_guard();
+
+    let conn = server.serve_connection_with_upgrades(
+        hyper_util::rt::TokioIo::new(conn),
+        hyper1::service::service_fn(move |req: hyper1::Request<Incoming>| {
+            // First HTTP request shares the same session ID
+            let session_id = session_id.take().unwrap_or_else(uuid::Uuid::new_v4);
+
+            // Cancel the current inflight HTTP request if the requets stream is closed.
+            // This is slightly different to `_cancel_connection` in that
+            // h2 can cancel individual requests with a `RST_STREAM`.
+            let http_request_token = http_cancellation_token.child_token();
+            let cancel_request = http_request_token.clone().drop_guard();
+
+            // `request_handler` is not cancel safe. It expects to be cancelled only at specific times.
+            // By spawning the future, we ensure it never gets cancelled until it decides to.
+            let handler = connections.spawn(
+                request_handler(
+                    req,
+                    config,
+                    backend.clone(),
+                    connections.clone(),
+                    cancellation_handler.clone(),
+                    session_id,
+                    peer_addr,
+                    endpoint_rate_limiter.clone(),
+                    http_request_token,
+                )
+                .in_current_span()
+                .map_ok_or_else(api_error_into_response, |r| r),
+            );
+
+            async move {
+                let res = handler.await;
+                cancel_request.disarm();
+                res
+            }
+        }),
+    );
+
+    // On cancellation, trigger the HTTP connection handler to shut down.
+    let res = match select(pin!(cancellation_token.cancelled()), pin!(conn)).await {
+        Either::Left((_cancelled, mut conn)) => {
+            conn.as_mut().graceful_shutdown();
+            conn.await
+        }
+        Either::Right((res, _)) => res,
+    };
+
+    match res {
+        Ok(()) => tracing::info!(%peer_addr, "HTTP connection closed"),
+        Err(e) => tracing::warn!(%peer_addr, "HTTP connection error {e}"),
     }
 }
 
 #[allow(clippy::too_many_arguments)]
 async fn request_handler(
-    mut request: Request<Body>,
+    mut request: hyper1::Request<Incoming>,
     config: &'static ProxyConfig,
     backend: Arc<PoolingBackend>,
     ws_connections: TaskTracker,
     cancellation_handler: Arc<CancellationHandlerMain>,
+    session_id: uuid::Uuid,
     peer_addr: IpAddr,
     endpoint_rate_limiter: Arc<EndpointRateLimiter>,
     // used to cancel in-flight HTTP requests. not used to cancel websockets
     http_cancellation_token: CancellationToken,
-) -> Result<Response<Body>, ApiError> {
-    let session_id = uuid::Uuid::new_v4();
-
+) -> Result<Response<Full<Bytes>>, ApiError> {
     let host = request
         .headers()
         .get("host")
@@ -282,14 +301,14 @@ async fn request_handler(
 
         // Return the response so the spawned future can continue.
         Ok(response)
-    } else if request.uri().path() == "/sql" && request.method() == Method::POST {
+    } else if request.uri().path() == "/sql" && *request.method() == Method::POST {
         let ctx = RequestMonitoring::new(session_id, peer_addr, "http", &config.region);
         let span = ctx.span.clone();
 
         sql_over_http::handle(config, ctx, request, backend, http_cancellation_token)
             .instrument(span)
             .await
-    } else if request.uri().path() == "/sql" && request.method() == Method::OPTIONS {
+    } else if request.uri().path() == "/sql" && *request.method() == Method::OPTIONS {
         Response::builder()
             .header("Allow", "OPTIONS, POST")
             .header("Access-Control-Allow-Origin", "*")
@@ -299,7 +318,7 @@ async fn request_handler(
             )
             .header("Access-Control-Max-Age", "86400" /* 24 hours */)
             .status(StatusCode::OK) // 204 is also valid, but see: https://developer.mozilla.org/en-US/docs/Web/HTTP/Methods/OPTIONS#status_code
-            .body(Body::empty())
+            .body(Full::new(Bytes::new()))
             .map_err(|e| ApiError::InternalServerError(e.into()))
     } else {
         json_response(StatusCode::BAD_REQUEST, "query is not supported")

proxy/src/serverless/backend.rs

@@ -9,7 +9,6 @@ use crate::{
     config::ProxyConfig,
     console::{
         errors::{GetAuthInfoError, WakeComputeError},
-        messages::ColdStartInfo,
         CachedNodeInfo,
     },
     context::RequestMonitoring,
@@ -57,7 +56,10 @@ impl PoolingBackend {
         let auth_outcome =
             crate::auth::validate_password_and_exchange(&conn_info.password, secret).await?;
         let res = match auth_outcome {
-            crate::sasl::Outcome::Success(key) => Ok(key),
+            crate::sasl::Outcome::Success(key) => {
+                info!("user successfully authenticated");
+                Ok(key)
+            }
             crate::sasl::Outcome::Failure(reason) => {
                 info!("auth backend failed with an error: {reason}");
                 Err(AuthError::auth_failed(&*conn_info.user_info.user))
@@ -89,8 +91,6 @@ impl PoolingBackend {
         };
 
         if let Some(client) = maybe_client {
-            info!("cold_start_info=warm");
-            ctx.set_cold_start_info(ColdStartInfo::Warm);
             return Ok(client);
         }
         let conn_id = uuid::Uuid::new_v4();

proxy/src/serverless/conn_pool.rs

@@ -17,7 +17,7 @@ use tokio::time::Instant;
 use tokio_postgres::tls::NoTlsStream;
 use tokio_postgres::{AsyncMessage, ReadyForQueryStatus, Socket};
 
-use crate::console::messages::MetricsAuxInfo;
+use crate::console::messages::{ColdStartInfo, MetricsAuxInfo};
 use crate::metrics::{ENDPOINT_POOLS, GC_LATENCY, NUM_OPEN_CLIENTS_IN_HTTP_POOL};
 use crate::usage_metrics::{Ids, MetricCounter, USAGE_METRICS};
 use crate::{
@@ -383,9 +383,12 @@ impl<C: ClientInnerExt> GlobalConnPool<C> {
                     "pid",
                     &tracing::field::display(client.inner.get_process_id()),
                 );
-                info!("pool: reusing connection '{conn_info}'");
+                info!(
+                    cold_start_info = ColdStartInfo::HttpPoolHit.as_str(),
+                    "pool: reusing connection '{conn_info}'"
+                );
                 client.session.send(ctx.session_id)?;
-                ctx.latency_timer.pool_hit();
+                ctx.set_cold_start_info(ColdStartInfo::HttpPoolHit);
                 ctx.latency_timer.success();
                 return Ok(Some(Client::new(client, conn_info.clone(), endpoint_pool)));
             }
@@ -454,8 +457,9 @@ pub fn poll_client<C: ClientInnerExt>(
     let (tx, mut rx) = tokio::sync::watch::channel(session_id);
 
     let span = info_span!(parent: None, "connection", %conn_id);
+    let cold_start_info = ctx.cold_start_info;
     span.in_scope(|| {
-        info!(%conn_info, %session_id, "new connection");
+        info!(cold_start_info = cold_start_info.as_str(), %conn_info, %session_id, "new connection");
     });
     let pool = match conn_info.endpoint_cache_key() {
         Some(endpoint) => Arc::downgrade(&global_pool.get_or_create_endpoint_pool(&endpoint)),
@@ -565,8 +569,8 @@ impl<C: ClientInnerExt> Client<C> {
     pub fn metrics(&self) -> Arc<MetricCounter> {
         let aux = &self.inner.as_ref().unwrap().aux;
         USAGE_METRICS.register(Ids {
-            endpoint_id: aux.endpoint_id.clone(),
-            branch_id: aux.branch_id.clone(),
+            endpoint_id: aux.endpoint_id,
+            branch_id: aux.branch_id,
         })
     }
 }
@@ -666,6 +670,8 @@ impl<C: ClientInnerExt> Drop for Client<C> {
 mod tests {
     use std::{mem, sync::atomic::AtomicBool};
 
+    use crate::{BranchId, EndpointId, ProjectId};
+
     use super::*;
 
     struct MockClient(Arc<AtomicBool>);
@@ -691,7 +697,12 @@ mod tests {
         ClientInner {
             inner: client,
             session: tokio::sync::watch::Sender::new(uuid::Uuid::new_v4()),
-            aux: Default::default(),
+            aux: MetricsAuxInfo {
+                endpoint_id: (&EndpointId::from("endpoint")).into(),
+                project_id: (&ProjectId::from("project")).into(),
+                branch_id: (&BranchId::from("branch")).into(),
+                cold_start_info: crate::console::messages::ColdStartInfo::Warm,
+            },
             conn_id: uuid::Uuid::new_v4(),
         }
     }

proxy/src/serverless/http_util.rs

@@ -0,0 +1,92 @@
+//! Things stolen from `libs/utils/src/http` to add hyper 1.0 compatibility
+//! Will merge back in at some point in the future.
+
+use bytes::Bytes;
+
+use anyhow::Context;
+use http::{Response, StatusCode};
+use http_body_util::Full;
+
+use serde::Serialize;
+use utils::http::error::ApiError;
+
+/// Like [`ApiError::into_response`]
+pub fn api_error_into_response(this: ApiError) -> Response<Full<Bytes>> {
+    match this {
+        ApiError::BadRequest(err) => HttpErrorBody::response_from_msg_and_status(
+            format!("{err:#?}"), // use debug printing so that we give the cause
+            StatusCode::BAD_REQUEST,
+        ),
+        ApiError::Forbidden(_) => {
+            HttpErrorBody::response_from_msg_and_status(this.to_string(), StatusCode::FORBIDDEN)
+        }
+        ApiError::Unauthorized(_) => {
+            HttpErrorBody::response_from_msg_and_status(this.to_string(), StatusCode::UNAUTHORIZED)
+        }
+        ApiError::NotFound(_) => {
+            HttpErrorBody::response_from_msg_and_status(this.to_string(), StatusCode::NOT_FOUND)
+        }
+        ApiError::Conflict(_) => {
+            HttpErrorBody::response_from_msg_and_status(this.to_string(), StatusCode::CONFLICT)
+        }
+        ApiError::PreconditionFailed(_) => HttpErrorBody::response_from_msg_and_status(
+            this.to_string(),
+            StatusCode::PRECONDITION_FAILED,
+        ),
+        ApiError::ShuttingDown => HttpErrorBody::response_from_msg_and_status(
+            "Shutting down".to_string(),
+            StatusCode::SERVICE_UNAVAILABLE,
+        ),
+        ApiError::ResourceUnavailable(err) => HttpErrorBody::response_from_msg_and_status(
+            err.to_string(),
+            StatusCode::SERVICE_UNAVAILABLE,
+        ),
+        ApiError::Timeout(err) => HttpErrorBody::response_from_msg_and_status(
+            err.to_string(),
+            StatusCode::REQUEST_TIMEOUT,
+        ),
+        ApiError::InternalServerError(err) => HttpErrorBody::response_from_msg_and_status(
+            err.to_string(),
+            StatusCode::INTERNAL_SERVER_ERROR,
+        ),
+    }
+}
+
+/// Same as [`utils::http::error::HttpErrorBody`]
+#[derive(Serialize)]
+struct HttpErrorBody {
+    pub msg: String,
+}
+
+impl HttpErrorBody {
+    /// Same as [`utils::http::error::HttpErrorBody::response_from_msg_and_status`]
+    fn response_from_msg_and_status(msg: String, status: StatusCode) -> Response<Full<Bytes>> {
+        HttpErrorBody { msg }.to_response(status)
+    }
+
+    /// Same as [`utils::http::error::HttpErrorBody::to_response`]
+    fn to_response(&self, status: StatusCode) -> Response<Full<Bytes>> {
+        Response::builder()
+            .status(status)
+            .header(http::header::CONTENT_TYPE, "application/json")
+            // we do not have nested maps with non string keys so serialization shouldn't fail
+            .body(Full::new(Bytes::from(serde_json::to_string(self).unwrap())))
+            .unwrap()
+    }
+}
+
+/// Same as [`utils::http::json::json_response`]
+pub fn json_response<T: Serialize>(
+    status: StatusCode,
+    data: T,
+) -> Result<Response<Full<Bytes>>, ApiError> {
+    let json = serde_json::to_string(&data)
+        .context("Failed to serialize JSON response")
+        .map_err(ApiError::InternalServerError)?;
+    let response = Response::builder()
+        .status(status)
+        .header(http::header::CONTENT_TYPE, "application/json")
+        .body(Full::new(Bytes::from(json)))
+        .map_err(|e| ApiError::InternalServerError(e.into()))?;
+    Ok(response)
+}

proxy/src/serverless/sql_over_http.rs

@@ -1,18 +1,22 @@
 use std::pin::pin;
 use std::sync::Arc;
 
+use bytes::Bytes;
 use futures::future::select;
 use futures::future::try_join;
 use futures::future::Either;
 use futures::StreamExt;
 use futures::TryFutureExt;
-use hyper::body::HttpBody;
-use hyper::header;
-use hyper::http::HeaderName;
-use hyper::http::HeaderValue;
-use hyper::Response;
-use hyper::StatusCode;
-use hyper::{Body, HeaderMap, Request};
+use http_body_util::BodyExt;
+use http_body_util::Full;
+use hyper1::body::Body;
+use hyper1::body::Incoming;
+use hyper1::header;
+use hyper1::http::HeaderName;
+use hyper1::http::HeaderValue;
+use hyper1::Response;
+use hyper1::StatusCode;
+use hyper1::{HeaderMap, Request};
 use serde_json::json;
 use serde_json::Value;
 use tokio::time;
@@ -29,7 +33,6 @@ use tracing::error;
 use tracing::info;
 use url::Url;
 use utils::http::error::ApiError;
-use utils::http::json::json_response;
 
 use crate::auth::backend::ComputeUserInfo;
 use crate::auth::endpoint_sni;
@@ -52,6 +55,7 @@ use crate::RoleName;
 use super::backend::PoolingBackend;
 use super::conn_pool::Client;
 use super::conn_pool::ConnInfo;
+use super::http_util::json_response;
 use super::json::json_to_pg_text;
 use super::json::pg_text_row_to_json;
 use super::json::JsonConversionError;
@@ -218,10 +222,10 @@ fn get_conn_info(
 pub async fn handle(
     config: &'static ProxyConfig,
     mut ctx: RequestMonitoring,
-    request: Request<Body>,
+    request: Request<Incoming>,
     backend: Arc<PoolingBackend>,
     cancel: CancellationToken,
-) -> Result<Response<Body>, ApiError> {
+) -> Result<Response<Full<Bytes>>, ApiError> {
     let result = handle_inner(cancel, config, &mut ctx, request, backend).await;
 
     let mut response = match result {
@@ -332,10 +336,9 @@ pub async fn handle(
         }
     };
 
-    response.headers_mut().insert(
-        "Access-Control-Allow-Origin",
-        hyper::http::HeaderValue::from_static("*"),
-    );
+    response
+        .headers_mut()
+        .insert("Access-Control-Allow-Origin", HeaderValue::from_static("*"));
     Ok(response)
 }
 
@@ -396,7 +399,7 @@ impl UserFacingError for SqlOverHttpError {
 #[derive(Debug, thiserror::Error)]
 pub enum ReadPayloadError {
     #[error("could not read the HTTP request body: {0}")]
-    Read(#[from] hyper::Error),
+    Read(#[from] hyper1::Error),
     #[error("could not parse the HTTP request body: {0}")]
     Parse(#[from] serde_json::Error),
 }
@@ -437,7 +440,7 @@ struct HttpHeaders {
 }
 
 impl HttpHeaders {
-    fn try_parse(headers: &hyper::http::HeaderMap) -> Result<Self, SqlOverHttpError> {
+    fn try_parse(headers: &hyper1::http::HeaderMap) -> Result<Self, SqlOverHttpError> {
         // Determine the output options. Default behaviour is 'false'. Anything that is not
         // strictly 'true' assumed to be false.
         let raw_output = headers.get(&RAW_TEXT_OUTPUT) == Some(&HEADER_VALUE_TRUE);
@@ -488,9 +491,9 @@ async fn handle_inner(
     cancel: CancellationToken,
     config: &'static ProxyConfig,
     ctx: &mut RequestMonitoring,
-    request: Request<Body>,
+    request: Request<Incoming>,
     backend: Arc<PoolingBackend>,
-) -> Result<Response<Body>, SqlOverHttpError> {
+) -> Result<Response<Full<Bytes>>, SqlOverHttpError> {
     let _request_gauge = NUM_CONNECTION_REQUESTS_GAUGE
         .with_label_values(&[ctx.protocol])
         .guard();
@@ -528,7 +531,7 @@ async fn handle_inner(
     }
 
     let fetch_and_process_request = async {
-        let body = hyper::body::to_bytes(request.into_body()).await?;
+        let body = request.into_body().collect().await?.to_bytes();
         info!(length = body.len(), "request payload read");
         let payload: Payload = serde_json::from_slice(&body)?;
         Ok::<Payload, ReadPayloadError>(payload) // Adjust error type accordingly
@@ -596,7 +599,7 @@ async fn handle_inner(
     let body = serde_json::to_string(&result).expect("json serialization should not fail");
     let len = body.len();
     let response = response
-        .body(Body::from(body))
+        .body(Full::new(Bytes::from(body)))
         // only fails if invalid status code or invalid header/values are given.
         // these are not user configurable so it cannot fail dynamically
         .expect("building response payload should not fail");
@@ -639,6 +642,7 @@ impl QueryData {
             }
             // The query was cancelled.
             Either::Right((_cancelled, query)) => {
+                tracing::info!("cancelling query");
                 if let Err(err) = cancel_token.cancel_query(NoTls).await {
                     tracing::error!(?err, "could not cancel query");
                 }

proxy/src/serverless/tls_listener.rs

@@ -1,123 +0,0 @@
-use std::{
-    convert::Infallible,
-    pin::Pin,
-    task::{Context, Poll},
-    time::Duration,
-};
-
-use hyper::server::{accept::Accept, conn::AddrStream};
-use pin_project_lite::pin_project;
-use tokio::{
-    io::{AsyncRead, AsyncWrite},
-    task::JoinSet,
-    time::timeout,
-};
-use tokio_rustls::{server::TlsStream, TlsAcceptor};
-use tracing::{info, warn, Instrument};
-
-use crate::{
-    metrics::TLS_HANDSHAKE_FAILURES,
-    protocol2::{WithClientIp, WithConnectionGuard},
-};
-
-pin_project! {
-    /// Wraps a `Stream` of connections (such as a TCP listener) so that each connection is itself
-    /// encrypted using TLS.
-    pub(crate) struct TlsListener<A: Accept> {
-        #[pin]
-        listener: A,
-        tls: TlsAcceptor,
-        waiting: JoinSet<Option<TlsStream<A::Conn>>>,
-        timeout: Duration,
-    }
-}
-
-impl<A: Accept> TlsListener<A> {
-    /// Create a `TlsListener` with default options.
-    pub(crate) fn new(tls: TlsAcceptor, listener: A, timeout: Duration) -> Self {
-        TlsListener {
-            listener,
-            tls,
-            waiting: JoinSet::new(),
-            timeout,
-        }
-    }
-}
-
-impl<A> Accept for TlsListener<A>
-where
-    A: Accept<Conn = WithConnectionGuard<WithClientIp<AddrStream>>>,
-    A::Error: std::error::Error,
-    A::Conn: AsyncRead + AsyncWrite + Unpin + Send + 'static,
-{
-    type Conn = TlsStream<A::Conn>;
-
-    type Error = Infallible;
-
-    fn poll_accept(
-        self: Pin<&mut Self>,
-        cx: &mut Context<'_>,
-    ) -> Poll<Option<Result<Self::Conn, Self::Error>>> {
-        let mut this = self.project();
-
-        loop {
-            match this.listener.as_mut().poll_accept(cx) {
-                Poll::Pending => break,
-                Poll::Ready(Some(Ok(mut conn))) => {
-                    let t = *this.timeout;
-                    let tls = this.tls.clone();
-                    let span = conn.span.clone();
-                    this.waiting.spawn(async move {
-                        let peer_addr = match conn.inner.wait_for_addr().await {
-                            Ok(Some(addr)) => addr,
-                            Err(e) => {
-                                tracing::error!("failed to accept TCP connection: invalid PROXY protocol V2 header: {e:#}");
-                                return None;
-                            }
-                            Ok(None) => conn.inner.inner.remote_addr()
-                        };
-
-                        let accept = tls.accept(conn);
-                        match timeout(t, accept).await {
-                            Ok(Ok(conn)) => {
-                                info!(%peer_addr, "accepted new TLS connection");
-                                Some(conn)
-                            },
-                            // The handshake failed, try getting another connection from the queue
-                            Ok(Err(e)) => {
-                                TLS_HANDSHAKE_FAILURES.inc();
-                                warn!(%peer_addr, "failed to accept TLS connection: {e:?}");
-                                None
-                            }
-                            // The handshake timed out, try getting another connection from the queue
-                            Err(_) => {
-                                TLS_HANDSHAKE_FAILURES.inc();
-                                warn!(%peer_addr, "failed to accept TLS connection: timeout");
-                                None
-                            }
-                        }
-                    }.instrument(span));
-                }
-                Poll::Ready(Some(Err(e))) => {
-                    tracing::error!("error accepting TCP connection: {e}");
-                    continue;
-                }
-                Poll::Ready(None) => return Poll::Ready(None),
-            }
-        }
-
-        loop {
-            return match this.waiting.poll_join_next(cx) {
-                Poll::Ready(Some(Ok(Some(conn)))) => Poll::Ready(Some(Ok(conn))),
-                // The handshake failed to complete, try getting another connection from the queue
-                Poll::Ready(Some(Ok(None))) => continue,
-                // The handshake panicked or was cancelled. ignore and get another connection
-                Poll::Ready(Some(Err(e))) => {
-                    tracing::warn!("handshake aborted: {e}");
-                    continue;
-                }
-                _ => Poll::Pending,
-            };
-        }
-    }
-}

proxy/src/usage_metrics.rs

@@ -3,7 +3,8 @@
 use crate::{
     config::{MetricBackupCollectionConfig, MetricCollectionConfig},
     context::parquet::{FAILED_UPLOAD_MAX_RETRIES, FAILED_UPLOAD_WARN_THRESHOLD},
-    http, BranchId, EndpointId,
+    http,
+    intern::{BranchIdInt, EndpointIdInt},
 };
 use anyhow::Context;
 use async_compression::tokio::write::GzipEncoder;
@@ -43,8 +44,8 @@ const DEFAULT_HTTP_REPORTING_TIMEOUT: Duration = Duration::from_secs(60);
 /// because we enrich the event with project_id in the control-plane endpoint.
 #[derive(Eq, Hash, PartialEq, Serialize, Deserialize, Debug, Clone)]
 pub struct Ids {
-    pub endpoint_id: EndpointId,
-    pub branch_id: BranchId,
+    pub endpoint_id: EndpointIdInt,
+    pub branch_id: BranchIdInt,
 }
 
 pub trait MetricCounterRecorder {
@@ -494,7 +495,7 @@ mod tests {
     use url::Url;
 
     use super::*;
-    use crate::{http, rate_limiter::RateLimiterConfig};
+    use crate::{http, rate_limiter::RateLimiterConfig, BranchId, EndpointId};
 
     #[tokio::test]
     async fn metrics() {
@@ -536,8 +537,8 @@ mod tests {
         // register a new counter
 
         let counter = metrics.register(Ids {
-            endpoint_id: "e1".into(),
-            branch_id: "b1".into(),
+            endpoint_id: (&EndpointId::from("e1")).into(),
+            branch_id: (&BranchId::from("b1")).into(),
         });
 
         // the counter should be observed despite 0 egress

safekeeper/src/wal_backup_partial.rs

@@ -337,6 +337,17 @@ pub async fn main_task(tli: Arc<Timeline>, conf: SafeKeeperConf) {
             }
         }
 
+        // if we don't have any data and zero LSNs, wait for something
+        while flush_lsn_rx.borrow().lsn == Lsn(0) {
+            tokio::select! {
+                _ = cancellation_rx.changed() => {
+                    info!("timeline canceled");
+                    return;
+                }
+                _ = flush_lsn_rx.changed() => {}
+            }
+        }
+
         // fixing the segno and waiting some time to prevent reuploading the same segment too often
         let pending_segno = backup.segno(flush_lsn_rx.borrow().lsn);
         let timeout = tokio::time::sleep(await_duration);

scripts/flaky_tests.py

@@ -15,8 +15,7 @@ FLAKY_TESTS_QUERY = """
         DISTINCT parent_suite, suite, name
     FROM results
     WHERE
-        started_at > CURRENT_DATE - INTERVAL '10' day
-        AND started_at > '2024-03-11 14:50:11.845+00' -- we switched the default PAGESERVER_VIRTUAL_FILE_IO_ENGINE to `tokio-epoll-uring` from `std-fs` on this date, we want to ignore the flaky tests for `std-fs`
+        started_at > CURRENT_DATE - INTERVAL '%s' day
         AND (
             (status IN ('failed', 'broken') AND reference = 'refs/heads/main')
             OR flaky

scripts/sk_cleanup_tenants/script.py

@@ -22,7 +22,7 @@ parser.add_argument("--safekeeper-host", required=True, type=str)
 args = parser.parse_args()
 
 access_key = os.getenv("CONSOLE_API_TOKEN")
-endpoint: str = "https://console.stage.neon.tech/api"
+endpoint: str = "https://console-stage.neon.build/api"
 
 trash_dir: Path = args.trash_dir
 dry_run: bool = args.dry_run

scripts/sk_collect_dumps/readme.md

@@ -3,7 +3,7 @@
 3. Issue admin token (add/remove .stage from url for staging/prod and setting proper API key):
 ```
 # staging:
-AUTH_TOKEN=$(curl https://console.stage.neon.tech/regions/console/api/v1/admin/issue_token -H "Accept: application/json" -H "Content-Type: application/json" -H "Authorization: Bearer $NEON_STAGING_KEY" -X POST -d '{"ttl_seconds": 43200, "scope": "safekeeperdata"}' 2>/dev/null | jq --raw-output '.jwt')
+AUTH_TOKEN=$(curl https://console-stage.neon.build/regions/console/api/v1/admin/issue_token -H "Accept: application/json" -H "Content-Type: application/json" -H "Authorization: Bearer $NEON_STAGING_KEY" -X POST -d '{"ttl_seconds": 43200, "scope": "safekeeperdata"}' 2>/dev/null | jq --raw-output '.jwt')
 # prod:
 AUTH_TOKEN=$(curl https://console.neon.tech/regions/console/api/v1/admin/issue_token -H "Accept: application/json" -H "Content-Type: application/json" -H "Authorization: Bearer $NEON_PROD_KEY" -X POST -d '{"ttl_seconds": 43200, "scope": "safekeeperdata"}' 2>/dev/null | jq --raw-output '.jwt')
 # check

storage_controller/Cargo.toml

@@ -1,5 +1,5 @@
 [package]
-name = "attachment_service"
+name = "storage_controller"
 version = "0.1.0"
 edition.workspace = true
 license.workspace = true
@@ -45,8 +45,8 @@ diesel = { version = "2.1.4", features = ["serde_json", "postgres", "r2d2"] }
 diesel_migrations = { version = "2.1.0" }
 r2d2 = { version = "0.8.10" }
 
-utils = { path = "../../libs/utils/" }
-metrics = { path = "../../libs/metrics/" }
-control_plane = { path = ".." }
-workspace_hack = { version = "0.1", path = "../../workspace_hack" }
+utils = { path = "../libs/utils/" }
+metrics = { path = "../libs/metrics/" }
+control_plane = { path = "../control_plane" }
+workspace_hack = { version = "0.1", path = "../workspace_hack" }
 

storage_controller/src/compute_hook.rs

@@ -1,3 +1,4 @@
+use std::sync::Arc;
 use std::{collections::HashMap, time::Duration};
 
 use control_plane::endpoint::{ComputeControlPlane, EndpointStatus};
@@ -18,14 +19,26 @@ const SLOWDOWN_DELAY: Duration = Duration::from_secs(5);
 
 pub(crate) const API_CONCURRENCY: usize = 32;
 
+struct UnshardedComputeHookTenant {
+    // Which node is this tenant attached to
+    node_id: NodeId,
+
+    // Must hold this lock to send a notification.
+    send_lock: Arc<tokio::sync::Mutex<Option<ComputeHookNotifyRequest>>>,
+}
 struct ShardedComputeHookTenant {
     stripe_size: ShardStripeSize,
     shard_count: ShardCount,
     shards: Vec<(ShardNumber, NodeId)>,
+
+    // Must hold this lock to send a notification.  The contents represent
+    // the last successfully sent notification, and are used to coalesce multiple
+    // updates by only sending when there is a chance since our last successful send.
+    send_lock: Arc<tokio::sync::Mutex<Option<ComputeHookNotifyRequest>>>,
 }
 
 enum ComputeHookTenant {
-    Unsharded(NodeId),
+    Unsharded(UnshardedComputeHookTenant),
     Sharded(ShardedComputeHookTenant),
 }
 
@@ -37,9 +50,20 @@ impl ComputeHookTenant {
                 shards: vec![(tenant_shard_id.shard_number, node_id)],
                 stripe_size,
                 shard_count: tenant_shard_id.shard_count,
+                send_lock: Arc::default(),
             })
         } else {
-            Self::Unsharded(node_id)
+            Self::Unsharded(UnshardedComputeHookTenant {
+                node_id,
+                send_lock: Arc::default(),
+            })
+        }
+    }
+
+    fn get_send_lock(&self) -> &Arc<tokio::sync::Mutex<Option<ComputeHookNotifyRequest>>> {
+        match self {
+            Self::Unsharded(unsharded_tenant) => &unsharded_tenant.send_lock,
+            Self::Sharded(sharded_tenant) => &sharded_tenant.send_lock,
         }
     }
 
@@ -52,8 +76,8 @@ impl ComputeHookTenant {
         node_id: NodeId,
     ) {
         match self {
-            Self::Unsharded(existing_node_id) if tenant_shard_id.shard_count.count() == 1 => {
-                *existing_node_id = node_id
+            Self::Unsharded(unsharded_tenant) if tenant_shard_id.shard_count.count() == 1 => {
+                unsharded_tenant.node_id = node_id
             }
             Self::Sharded(sharded_tenant)
                 if sharded_tenant.stripe_size == stripe_size
@@ -80,14 +104,14 @@ impl ComputeHookTenant {
     }
 }
 
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
 struct ComputeHookNotifyRequestShard {
     node_id: NodeId,
     shard_number: ShardNumber,
 }
 
 /// Request body that we send to the control plane to notify it of where a tenant is attached
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
 struct ComputeHookNotifyRequest {
     tenant_id: TenantId,
     stripe_size: Option<ShardStripeSize>,
@@ -120,14 +144,44 @@ pub(crate) enum NotifyError {
     Fatal(StatusCode),
 }
 
+enum MaybeSendResult {
+    // Please send this request while holding the lock, and if you succeed then write
+    // the request into the lock.
+    Transmit(
+        (
+            ComputeHookNotifyRequest,
+            tokio::sync::OwnedMutexGuard<Option<ComputeHookNotifyRequest>>,
+        ),
+    ),
+    // Something requires sending, but you must wait for a current sender then call again
+    AwaitLock(Arc<tokio::sync::Mutex<Option<ComputeHookNotifyRequest>>>),
+    // Nothing requires sending
+    Noop,
+}
+
 impl ComputeHookTenant {
-    fn maybe_reconfigure(&self, tenant_id: TenantId) -> Option<ComputeHookNotifyRequest> {
-        match self {
-            Self::Unsharded(node_id) => Some(ComputeHookNotifyRequest {
+    fn maybe_send(
+        &self,
+        tenant_id: TenantId,
+        lock: Option<tokio::sync::OwnedMutexGuard<Option<ComputeHookNotifyRequest>>>,
+    ) -> MaybeSendResult {
+        let locked = match lock {
+            Some(already_locked) => already_locked,
+            None => {
+                // Lock order: this _must_ be only a try_lock, because we are called inside of the [`ComputeHook::state`] lock.
+                let Ok(locked) = self.get_send_lock().clone().try_lock_owned() else {
+                    return MaybeSendResult::AwaitLock(self.get_send_lock().clone());
+                };
+                locked
+            }
+        };
+
+        let request = match self {
+            Self::Unsharded(unsharded_tenant) => Some(ComputeHookNotifyRequest {
                 tenant_id,
                 shards: vec![ComputeHookNotifyRequestShard {
                     shard_number: ShardNumber(0),
-                    node_id: *node_id,
+                    node_id: unsharded_tenant.node_id,
                 }],
                 stripe_size: None,
             }),
@@ -151,12 +205,25 @@ impl ComputeHookTenant {
                 // Sharded tenant doesn't yet have information for all its shards
 
                 tracing::info!(
-                    "ComputeHookTenant::maybe_reconfigure: not enough shards ({}/{})",
+                    "ComputeHookTenant::maybe_send: not enough shards ({}/{})",
                     sharded_tenant.shards.len(),
                     sharded_tenant.shard_count.count()
                 );
                 None
             }
+        };
+
+        match request {
+            None => {
+                // Not yet ready to emit a notification
+                tracing::info!("Tenant isn't yet ready to emit a notification");
+                MaybeSendResult::Noop
+            }
+            Some(request) if Some(&request) == locked.as_ref() => {
+                // No change from the last value successfully sent
+                MaybeSendResult::Noop
+            }
+            Some(request) => MaybeSendResult::Transmit((request, locked)),
         }
     }
 }
@@ -166,8 +233,15 @@ impl ComputeHookTenant {
 /// the compute connection string.
 pub(super) struct ComputeHook {
     config: Config,
-    state: tokio::sync::Mutex<HashMap<TenantId, ComputeHookTenant>>,
+    state: std::sync::Mutex<HashMap<TenantId, ComputeHookTenant>>,
     authorization_header: Option<String>,
+
+    // Concurrency limiter, so that we do not overload the cloud control plane when updating
+    // large numbers of tenants (e.g. when failing over after a node failure)
+    api_concurrency: tokio::sync::Semaphore,
+
+    // This lock is only used in testing enviroments, to serialize calls into neon_lock
+    neon_local_lock: tokio::sync::Mutex<()>,
 }
 
 impl ComputeHook {
@@ -181,14 +255,20 @@ impl ComputeHook {
             state: Default::default(),
             config,
             authorization_header,
+            neon_local_lock: Default::default(),
+            api_concurrency: tokio::sync::Semaphore::new(API_CONCURRENCY),
         }
     }
 
     /// For test environments: use neon_local's LocalEnv to update compute
     async fn do_notify_local(
         &self,
-        reconfigure_request: ComputeHookNotifyRequest,
+        reconfigure_request: &ComputeHookNotifyRequest,
     ) -> anyhow::Result<()> {
+        // neon_local updates are not safe to call concurrently, use a lock to serialize
+        // all calls to this function
+        let _locked = self.neon_local_lock.lock().await;
+
         let env = match LocalEnv::load_config() {
             Ok(e) => e,
             Err(e) => {
@@ -205,7 +285,7 @@ impl ComputeHook {
         } = reconfigure_request;
 
         let compute_pageservers = shards
-            .into_iter()
+            .iter()
             .map(|shard| {
                 let ps_conf = env
                     .get_pageserver_conf(shard.node_id)
@@ -217,10 +297,10 @@ impl ComputeHook {
             .collect::<Vec<_>>();
 
         for (endpoint_name, endpoint) in &cplane.endpoints {
-            if endpoint.tenant_id == tenant_id && endpoint.status() == EndpointStatus::Running {
+            if endpoint.tenant_id == *tenant_id && endpoint.status() == EndpointStatus::Running {
                 tracing::info!("Reconfiguring endpoint {}", endpoint_name,);
                 endpoint
-                    .reconfigure(compute_pageservers.clone(), stripe_size)
+                    .reconfigure(compute_pageservers.clone(), *stripe_size)
                     .await?;
             }
         }
@@ -298,12 +378,23 @@ impl ComputeHook {
     async fn do_notify(
         &self,
         url: &String,
-        reconfigure_request: ComputeHookNotifyRequest,
+        reconfigure_request: &ComputeHookNotifyRequest,
         cancel: &CancellationToken,
     ) -> Result<(), NotifyError> {
         let client = reqwest::Client::new();
+
+        // We hold these semaphore units across all retries, rather than only across each
+        // HTTP request: this is to preserve fairness and avoid a situation where a retry might
+        // time out waiting for a semaphore.
+        let _units = self
+            .api_concurrency
+            .acquire()
+            .await
+            // Interpret closed semaphore as shutdown
+            .map_err(|_| NotifyError::ShuttingDown)?;
+
         backoff::retry(
-            || self.do_notify_iteration(&client, url, &reconfigure_request, cancel),
+            || self.do_notify_iteration(&client, url, reconfigure_request, cancel),
             |e| {
                 matches!(
                     e,
@@ -343,42 +434,70 @@ impl ComputeHook {
         stripe_size: ShardStripeSize,
         cancel: &CancellationToken,
     ) -> Result<(), NotifyError> {
-        let mut locked = self.state.lock().await;
+        let maybe_send_result = {
+            let mut state_locked = self.state.lock().unwrap();
 
-        use std::collections::hash_map::Entry;
-        let tenant = match locked.entry(tenant_shard_id.tenant_id) {
-            Entry::Vacant(e) => e.insert(ComputeHookTenant::new(
-                tenant_shard_id,
-                stripe_size,
-                node_id,
-            )),
-            Entry::Occupied(e) => {
-                let tenant = e.into_mut();
-                tenant.update(tenant_shard_id, stripe_size, node_id);
-                tenant
+            use std::collections::hash_map::Entry;
+            let tenant = match state_locked.entry(tenant_shard_id.tenant_id) {
+                Entry::Vacant(e) => e.insert(ComputeHookTenant::new(
+                    tenant_shard_id,
+                    stripe_size,
+                    node_id,
+                )),
+                Entry::Occupied(e) => {
+                    let tenant = e.into_mut();
+                    tenant.update(tenant_shard_id, stripe_size, node_id);
+                    tenant
+                }
+            };
+            tenant.maybe_send(tenant_shard_id.tenant_id, None)
+        };
+
+        // Process result: we may get an update to send, or we may have to wait for a lock
+        // before trying again.
+        let (request, mut send_lock_guard) = match maybe_send_result {
+            MaybeSendResult::Noop => {
+                return Ok(());
             }
+            MaybeSendResult::AwaitLock(send_lock) => {
+                let send_locked = send_lock.lock_owned().await;
+
+                // Lock order: maybe_send is called within the `[Self::state]` lock, and takes the send lock, but here
+                // we have acquired the send lock and take `[Self::state]` lock.  This is safe because maybe_send only uses
+                // try_lock.
+                let state_locked = self.state.lock().unwrap();
+                let Some(tenant) = state_locked.get(&tenant_shard_id.tenant_id) else {
+                    return Ok(());
+                };
+                match tenant.maybe_send(tenant_shard_id.tenant_id, Some(send_locked)) {
+                    MaybeSendResult::AwaitLock(_) => {
+                        unreachable!("We supplied lock guard")
+                    }
+                    MaybeSendResult::Noop => {
+                        return Ok(());
+                    }
+                    MaybeSendResult::Transmit((request, lock)) => (request, lock),
+                }
+            }
+            MaybeSendResult::Transmit((request, lock)) => (request, lock),
         };
 
-        let reconfigure_request = tenant.maybe_reconfigure(tenant_shard_id.tenant_id);
-        let Some(reconfigure_request) = reconfigure_request else {
-            // The tenant doesn't yet have pageservers for all its shards: we won't notify anything
-            // until it does.
-            tracing::info!("Tenant isn't yet ready to emit a notification");
-            return Ok(());
-        };
-
-        if let Some(notify_url) = &self.config.compute_hook_url {
-            self.do_notify(notify_url, reconfigure_request, cancel)
-                .await
+        let result = if let Some(notify_url) = &self.config.compute_hook_url {
+            self.do_notify(notify_url, &request, cancel).await
         } else {
-            self.do_notify_local(reconfigure_request)
-                .await
-                .map_err(|e| {
-                    // This path is for testing only, so munge the error into our prod-style error type.
-                    tracing::error!("Local notification hook failed: {e}");
-                    NotifyError::Fatal(StatusCode::INTERNAL_SERVER_ERROR)
-                })
+            self.do_notify_local(&request).await.map_err(|e| {
+                // This path is for testing only, so munge the error into our prod-style error type.
+                tracing::error!("Local notification hook failed: {e}");
+                NotifyError::Fatal(StatusCode::INTERNAL_SERVER_ERROR)
+            })
+        };
+
+        if result.is_ok() {
+            // Before dropping the send lock, stash the request we just sent so that
+            // subsequent callers can avoid redundantly re-sending the same thing.
+            *send_lock_guard = Some(request);
         }
+        result
     }
 }
 
@@ -402,21 +521,22 @@ pub(crate) mod tests {
             NodeId(1),
         );
 
-        // An unsharded tenant is always ready to emit a notification
-        assert!(tenant_state.maybe_reconfigure(tenant_id).is_some());
-        assert_eq!(
-            tenant_state
-                .maybe_reconfigure(tenant_id)
-                .unwrap()
-                .shards
-                .len(),
-            1
-        );
-        assert!(tenant_state
-            .maybe_reconfigure(tenant_id)
-            .unwrap()
-            .stripe_size
-            .is_none());
+        // An unsharded tenant is always ready to emit a notification, but won't
+        // send the same one twice
+        let send_result = tenant_state.maybe_send(tenant_id, None);
+        let MaybeSendResult::Transmit((request, mut guard)) = send_result else {
+            anyhow::bail!("Wrong send result");
+        };
+        assert_eq!(request.shards.len(), 1);
+        assert!(request.stripe_size.is_none());
+
+        // Simulate successful send
+        *guard = Some(request);
+        drop(guard);
+
+        // Try asking again: this should be a no-op
+        let send_result = tenant_state.maybe_send(tenant_id, None);
+        assert!(matches!(send_result, MaybeSendResult::Noop));
 
         // Writing the first shard of a multi-sharded situation (i.e. in a split)
         // resets the tenant state and puts it in an non-notifying state (need to
@@ -430,7 +550,10 @@ pub(crate) mod tests {
             ShardStripeSize(32768),
             NodeId(1),
         );
-        assert!(tenant_state.maybe_reconfigure(tenant_id).is_none());
+        assert!(matches!(
+            tenant_state.maybe_send(tenant_id, None),
+            MaybeSendResult::Noop
+        ));
 
         // Writing the second shard makes it ready to notify
         tenant_state.update(
@@ -443,22 +566,16 @@ pub(crate) mod tests {
             NodeId(1),
         );
 
-        assert!(tenant_state.maybe_reconfigure(tenant_id).is_some());
-        assert_eq!(
-            tenant_state
-                .maybe_reconfigure(tenant_id)
-                .unwrap()
-                .shards
-                .len(),
-            2
-        );
-        assert_eq!(
-            tenant_state
-                .maybe_reconfigure(tenant_id)
-                .unwrap()
-                .stripe_size,
-            Some(ShardStripeSize(32768))
-        );
+        let send_result = tenant_state.maybe_send(tenant_id, None);
+        let MaybeSendResult::Transmit((request, mut guard)) = send_result else {
+            anyhow::bail!("Wrong send result");
+        };
+        assert_eq!(request.shards.len(), 2);
+        assert_eq!(request.stripe_size, Some(ShardStripeSize(32768)));
+
+        // Simulate successful send
+        *guard = Some(request);
+        drop(guard);
 
         Ok(())
     }

storage_controller/src/http.rs

@@ -8,6 +8,7 @@ use futures::Future;
 use hyper::header::CONTENT_TYPE;
 use hyper::{Body, Request, Response};
 use hyper::{StatusCode, Uri};
+use metrics::{BuildInfo, NeonMetrics};
 use pageserver_api::models::{
     TenantConfigRequest, TenantCreateRequest, TenantLocationConfigRequest, TenantShardSplitRequest,
     TenantTimeTravelRequest, TimelineCreateRequest,
@@ -44,15 +45,19 @@ use control_plane::storage_controller::{AttachHookRequest, InspectRequest};
 use routerify::Middleware;
 
 /// State available to HTTP request handlers
-#[derive(Clone)]
 pub struct HttpState {
     service: Arc<crate::service::Service>,
     auth: Option<Arc<SwappableJwtAuth>>,
+    neon_metrics: NeonMetrics,
     allowlist_routes: Vec<Uri>,
 }
 
 impl HttpState {
-    pub fn new(service: Arc<crate::service::Service>, auth: Option<Arc<SwappableJwtAuth>>) -> Self {
+    pub fn new(
+        service: Arc<crate::service::Service>,
+        auth: Option<Arc<SwappableJwtAuth>>,
+        build_info: BuildInfo,
+    ) -> Self {
         let allowlist_routes = ["/status", "/ready", "/metrics"]
             .iter()
             .map(|v| v.parse().unwrap())
@@ -60,6 +65,7 @@ impl HttpState {
         Self {
             service,
             auth,
+            neon_metrics: NeonMetrics::new(build_info),
             allowlist_routes,
         }
     }
@@ -602,9 +608,17 @@ where
     .await
 }
 
+/// Check if the required scope is held in the request's token, or if the request has
+/// a token with 'admin' scope then always permit it.
 fn check_permissions(request: &Request<Body>, required_scope: Scope) -> Result<(), ApiError> {
     check_permission_with(request, |claims| {
-        crate::auth::check_permission(claims, required_scope)
+        match crate::auth::check_permission(claims, required_scope) {
+            Err(e) => match crate::auth::check_permission(claims, Scope::Admin) {
+                Ok(()) => Ok(()),
+                Err(_) => Err(e),
+            },
+            Ok(()) => Ok(()),
+        }
     })
 }
 
@@ -664,10 +678,11 @@ fn epilogue_metrics_middleware<B: hyper::body::HttpBody + Send + Sync + 'static>
     })
 }
 
-pub async fn measured_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+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 state = get_state(&req);
+    let payload = crate::metrics::METRICS_REGISTRY.encode(&state.neon_metrics);
     let response = Response::builder()
         .status(200)
         .header(CONTENT_TYPE, TEXT_FORMAT)
@@ -696,6 +711,7 @@ where
 pub fn make_router(
     service: Arc<Service>,
     auth: Option<Arc<SwappableJwtAuth>>,
+    build_info: BuildInfo,
 ) -> RouterBuilder<hyper::Body, ApiError> {
     let mut router = endpoint::make_router()
         .middleware(prologue_metrics_middleware())
@@ -712,7 +728,7 @@ pub fn make_router(
     }
 
     router
-        .data(Arc::new(HttpState::new(service, auth)))
+        .data(Arc::new(HttpState::new(service, auth, build_info)))
         .get("/metrics", |r| {
             named_request_span(r, measured_metrics_handler, RequestName("metrics"))
         })

storage_controller/src/lib.rs

@@ -14,7 +14,7 @@ mod reconciler;
 mod scheduler;
 mod schema;
 pub mod service;
-mod tenant_state;
+mod tenant_shard;
 
 #[derive(Ord, PartialOrd, Eq, PartialEq, Copy, Clone, Serialize)]
 struct Sequence(u64);

storage_controller/src/main.rs

@@ -1,18 +1,20 @@
 use anyhow::{anyhow, Context};
-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 camino::Utf8PathBuf;
 use clap::Parser;
 use diesel::Connection;
 use metrics::launch_timestamp::LaunchTimestamp;
+use metrics::BuildInfo;
 use std::sync::Arc;
+use storage_controller::http::make_router;
+use storage_controller::metrics::preinitialize_metrics;
+use storage_controller::persistence::Persistence;
+use storage_controller::service::{Config, Service, MAX_UNAVAILABLE_INTERVAL_DEFAULT};
 use tokio::signal::unix::SignalKind;
 use tokio_util::sync::CancellationToken;
 use utils::auth::{JwtAuth, SwappableJwtAuth};
 use utils::logging::{self, LogFormat};
 
+use utils::sentry_init::init_sentry;
 use utils::{project_build_tag, project_git_version, tcp_listener};
 
 project_git_version!(GIT_VERSION);
@@ -50,7 +52,7 @@ struct Cli {
     #[arg(short, long)]
     path: Option<Utf8PathBuf>,
 
-    /// URL to connect to postgres, like postgresql://localhost:1234/attachment_service
+    /// URL to connect to postgres, like postgresql://localhost:1234/storage_controller
     #[arg(long)]
     database_url: Option<String>,
 
@@ -158,6 +160,8 @@ fn main() -> anyhow::Result<()> {
         std::process::exit(1);
     }));
 
+    let _sentry_guard = init_sentry(Some(GIT_VERSION.into()), &[]);
+
     tokio::runtime::Builder::new_current_thread()
         // We use spawn_blocking for database operations, so require approximately
         // as many blocking threads as we will open database connections.
@@ -189,6 +193,11 @@ async fn async_main() -> anyhow::Result<()> {
         args.listen
     );
 
+    let build_info = BuildInfo {
+        revision: GIT_VERSION,
+        build_tag: BUILD_TAG,
+    };
+
     let strict_mode = if args.dev {
         StrictMode::Dev
     } else {
@@ -250,7 +259,7 @@ async fn async_main() -> anyhow::Result<()> {
     let auth = secrets
         .public_key
         .map(|jwt_auth| Arc::new(SwappableJwtAuth::new(jwt_auth)));
-    let router = make_router(service.clone(), auth)
+    let router = make_router(service.clone(), auth, build_info)
         .build()
         .map_err(|err| anyhow!(err))?;
     let router_service = utils::http::RouterService::new(router).unwrap();

storage_controller/src/metrics.rs

@@ -8,10 +8,8 @@
 //! 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 measured::{label::LabelValue, metric::histogram, FixedCardinalityLabel, MetricGroup};
+use metrics::NeonMetrics;
 use once_cell::sync::Lazy;
 use std::sync::Mutex;
 
@@ -26,13 +24,15 @@ pub fn preinitialize_metrics() {
 
 pub(crate) struct StorageControllerMetrics {
     pub(crate) metrics_group: StorageControllerMetricGroup,
-    encoder: Mutex<measured::text::TextEncoder>,
+    encoder: Mutex<measured::text::BufferedTextEncoder>,
 }
 
 #[derive(measured::MetricGroup)]
+#[metric(new())]
 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>,
@@ -43,7 +43,9 @@ pub(crate) struct StorageControllerMetricGroup {
     /// 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
+    #[metric(metadata = histogram::Thresholds::exponential_buckets(0.1, 2.0))]
     pub(crate) storage_controller_http_request_latency:
         measured::HistogramVec<HttpRequestLatencyLabelGroupSet, 5>,
 
@@ -55,6 +57,7 @@ pub(crate) struct StorageControllerMetricGroup {
     /// Latency of HTTP requests to the pageserver, broken down by pageserver
     /// node id, request name and method. This include both successful and unsuccessful
     /// requests.
+    #[metric(metadata = histogram::Thresholds::exponential_buckets(0.1, 2.0))]
     pub(crate) storage_controller_pageserver_request_latency:
         measured::HistogramVec<PageserverRequestLabelGroupSet, 5>,
 
@@ -66,6 +69,7 @@ pub(crate) struct StorageControllerMetricGroup {
     /// 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.
+    #[metric(metadata = histogram::Thresholds::exponential_buckets(0.1, 2.0))]
     pub(crate) storage_controller_passthrough_request_latency:
         measured::HistogramVec<PageserverRequestLabelGroupSet, 5>,
 
@@ -74,76 +78,34 @@ pub(crate) struct StorageControllerMetricGroup {
         measured::CounterVec<DatabaseQueryErrorLabelGroupSet>,
 
     /// Latency of database queries, broken down by operation.
+    #[metric(metadata = histogram::Thresholds::exponential_buckets(0.1, 2.0))]
     pub(crate) storage_controller_database_query_latency:
         measured::HistogramVec<DatabaseQueryLatencyLabelGroupSet, 5>,
 }
 
 impl StorageControllerMetrics {
-    pub(crate) fn encode(&self) -> Bytes {
+    pub(crate) fn encode(&self, neon_metrics: &NeonMetrics) -> Bytes {
         let mut encoder = self.encoder.lock().unwrap();
-        self.metrics_group.collect_into(&mut *encoder);
+        neon_metrics
+            .collect_group_into(&mut *encoder)
+            .unwrap_or_else(|infallible| match infallible {});
+        self.metrics_group
+            .collect_group_into(&mut *encoder)
+            .unwrap_or_else(|infallible| match infallible {});
         encoder.finish()
     }
 }
 
 impl Default for StorageControllerMetrics {
     fn default() -> Self {
-        Self {
-            metrics_group: StorageControllerMetricGroup::new(),
-            encoder: Mutex::new(measured::text::TextEncoder::new()),
-        }
-    }
-}
+        let mut metrics_group = StorageControllerMetricGroup::new();
+        metrics_group
+            .storage_controller_reconcile_complete
+            .init_all_dense();
 
-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_schedule_optimization: measured::Counter::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),
-            ),
+            metrics_group,
+            encoder: Mutex::new(measured::text::BufferedTextEncoder::new()),
         }
     }
 }
@@ -157,7 +119,7 @@ pub(crate) struct ReconcileCompleteLabelGroup {
 #[derive(measured::LabelGroup)]
 #[label(set = HttpRequestStatusLabelGroupSet)]
 pub(crate) struct HttpRequestStatusLabelGroup<'a> {
-    #[label(dynamic_with = lasso::ThreadedRodeo)]
+    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
     pub(crate) path: &'a str,
     pub(crate) method: Method,
     pub(crate) status: StatusCode,
@@ -166,40 +128,21 @@ pub(crate) struct HttpRequestStatusLabelGroup<'a> {
 #[derive(measured::LabelGroup)]
 #[label(set = HttpRequestLatencyLabelGroupSet)]
 pub(crate) struct HttpRequestLatencyLabelGroup<'a> {
-    #[label(dynamic_with = lasso::ThreadedRodeo)]
+    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
     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)]
+    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
     pub(crate) pageserver_id: &'a str,
-    #[label(dynamic_with = lasso::ThreadedRodeo)]
+    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
     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 {
@@ -213,7 +156,7 @@ pub(crate) struct DatabaseQueryLatencyLabelGroup {
     pub(crate) operation: DatabaseOperation,
 }
 
-#[derive(FixedCardinalityLabel)]
+#[derive(FixedCardinalityLabel, Clone, Copy)]
 pub(crate) enum ReconcileOutcome {
     #[label(rename = "ok")]
     Success,
@@ -221,7 +164,7 @@ pub(crate) enum ReconcileOutcome {
     Cancel,
 }
 
-#[derive(FixedCardinalityLabel, Clone)]
+#[derive(FixedCardinalityLabel, Copy, Clone)]
 pub(crate) enum Method {
     Get,
     Put,
@@ -246,11 +189,12 @@ impl From<hyper::Method> for Method {
     }
 }
 
+#[derive(Clone, Copy)]
 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)
+        v.write_int(self.0.as_u16() as i64)
     }
 }
 
@@ -268,7 +212,7 @@ impl FixedCardinalityLabel for StatusCode {
     }
 }
 
-#[derive(FixedCardinalityLabel)]
+#[derive(FixedCardinalityLabel, Clone, Copy)]
 pub(crate) enum DatabaseErrorLabel {
     Query,
     Connection,