libs/http: include path when logging auth failures

tests: try to make restored-datadir comparison tests not flaky v2
This test occasionally fails with a difference in "pg_xact/0000" file between the local and restored datadirs. My hypothesis is that something changed in the database between the last explicit checkpoint and the shutdown. I suspect autovacuum, it could certainly create transactions. To fix, be more precise about the point in time that we compare. Shut down the endpoint first, then read the last LSN (i.e. the shutdown checkpoint's LSN), from the local disk with pg_controldata. And use exactly that LSN in the basebackup. Closes #559
2026-05-04 06:40:37 +00:00 · 2024-03-12 10:55:12 +00:00 · 2024-03-11 23:29:32 +04:00 · 2024-03-11 23:29:32 +04:00 · 2024-03-11 22:36:52 +04:00 · 2024-03-11 18:36:39 +00:00
157 changed files with 5524 additions and 2712 deletions
--- a/.github/workflows/build_and_test.yml
+++ b/.github/workflows/build_and_test.yml
@@ -474,7 +474,7 @@ jobs:
          TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
          CHECK_ONDISK_DATA_COMPATIBILITY: nonempty
          BUILD_TAG: ${{ needs.tag.outputs.build-tag }}
-          PAGESERVER_VIRTUAL_FILE_IO_ENGINE: std-fs
+          PAGESERVER_VIRTUAL_FILE_IO_ENGINE: tokio-epoll-uring
          PAGESERVER_GET_VECTORED_IMPL: vectored

      # Temporary disable this step until we figure out why it's so flaky
@@ -554,7 +554,7 @@ jobs:
          VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
          PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
          TEST_RESULT_CONNSTR: "${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}"
-          PAGESERVER_VIRTUAL_FILE_IO_ENGINE: std-fs
+          PAGESERVER_VIRTUAL_FILE_IO_ENGINE: tokio-epoll-uring
      # XXX: no coverage data handling here, since benchmarks are run on release builds,
      # while coverage is currently collected for the debug ones

--- a/Cargo.lock
+++ b/Cargo.lock
@@ -241,7 +241,7 @@ checksum = "16e62a023e7c117e27523144c5d2459f4397fcc3cab0085af8e2224f643a0193"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -252,7 +252,7 @@ checksum = "b9ccdd8f2a161be9bd5c023df56f1b2a0bd1d83872ae53b71a84a12c9bf6e842"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -626,7 +626,7 @@ dependencies = [
 "once_cell",
 "pin-project-lite",
 "pin-utils",
- "rustls",
+ "rustls 0.21.9",
 "tokio",
 "tracing",
 ]
@@ -907,6 +907,16 @@ version = "1.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8c3c1a368f70d6cf7302d78f8f7093da241fb8e8807c05cc9e51a125895a6d5b"

+[[package]]
+name = "bcder"
+version = "0.7.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c627747a6774aab38beb35990d88309481378558875a41da1a4b2e373c906ef0"
+dependencies = [
+ "bytes",
+ "smallvec",
+]
+
 [[package]]
 name = "bincode"
 version = "1.3.3"
@@ -935,7 +945,7 @@ dependencies = [
 "regex",
 "rustc-hash",
 "shlex",
- "syn 2.0.32",
+ "syn 2.0.52",
 "which",
 ]

@@ -986,9 +996,9 @@ checksum = "14c189c53d098945499cdfa7ecc63567cf3886b3332b312a5b4585d8d3a6a610"

 [[package]]
 name = "bytes"
-version = "1.4.0"
+version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "89b2fd2a0dcf38d7971e2194b6b6eebab45ae01067456a7fd93d5547a61b70be"
+checksum = "a2bd12c1caf447e69cd4528f47f94d203fd2582878ecb9e9465484c4148a8223"
 dependencies = [
 "serde",
 ]
@@ -1149,7 +1159,7 @@ dependencies = [
 "heck",
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -1574,7 +1584,7 @@ dependencies = [
 "proc-macro2",
 "quote",
 "strsim",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -1585,7 +1595,7 @@ checksum = "29a358ff9f12ec09c3e61fef9b5a9902623a695a46a917b07f269bff1445611a"
 dependencies = [
 "darling_core",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -1627,6 +1637,16 @@ dependencies = [
 "zeroize",
 ]

+[[package]]
+name = "der"
+version = "0.7.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fffa369a668c8af7dbf8b5e56c9f744fbd399949ed171606040001947de40b1c"
+dependencies = [
+ "const-oid",
+ "zeroize",
+]
+
 [[package]]
 name = "der-parser"
 version = "8.2.0"
@@ -1681,7 +1701,7 @@ dependencies = [
 "diesel_table_macro_syntax",
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -1701,7 +1721,7 @@ version = "0.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "fc5557efc453706fed5e4fa85006fe9817c224c3f480a34c7e5959fd700921c5"
 dependencies = [
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -1723,7 +1743,7 @@ checksum = "487585f4d0c6655fe74905e2504d8ad6908e4db67f744eb140876906c2f3175d"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -1747,10 +1767,10 @@ version = "0.14.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "413301934810f597c1d19ca71c8710e99a3f1ba28a0d2ebc01551a2daeea3c5c"
 dependencies = [
- "der",
+ "der 0.6.1",
 "elliptic-curve",
 "rfc6979",
- "signature",
+ "signature 1.6.4",
 ]

 [[package]]
@@ -1767,7 +1787,7 @@ checksum = "e7bb888ab5300a19b8e5bceef25ac745ad065f3c9f7efc6de1b91958110891d3"
 dependencies = [
 "base16ct",
 "crypto-bigint 0.4.9",
- "der",
+ "der 0.6.1",
 "digest",
 "ff",
 "generic-array",
@@ -1827,7 +1847,7 @@ dependencies = [
 "darling",
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -2087,7 +2107,7 @@ checksum = "87750cf4b7a4c0625b1529e4c543c2182106e4dedc60a2a6455e00d212c489ac"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -2470,10 +2490,10 @@ dependencies = [
 "http 0.2.9",
 "hyper",
 "log",
- "rustls",
+ "rustls 0.21.9",
 "rustls-native-certs",
 "tokio",
- "tokio-rustls",
+ "tokio-rustls 0.24.0",
 ]

 [[package]]
@@ -2711,7 +2731,7 @@ checksum = "5c7ea04a7c5c055c175f189b6dc6ba036fd62306b58c66c9f6389036c503a3f4"
 dependencies = [
 "base64 0.21.1",
 "js-sys",
- "pem 3.0.3",
+ "pem",
 "ring 0.17.6",
 "serde",
 "serde_json",
@@ -3234,7 +3254,7 @@ checksum = "a948666b637a0f465e8564c73e89d4dde00d72d4d473cc972f390fc3dcee7d9c"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -3716,7 +3736,7 @@ dependencies = [
 "parquet",
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -3754,16 +3774,6 @@ version = "0.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "19b17cddbe7ec3f8bc800887bab5e717348c95ea2ca0b1bf0837fb964dc67099"

-[[package]]
-name = "pem"
-version = "2.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6b13fe415cdf3c8e44518e18a7c95a13431d9bdf6d15367d82b23c377fdd441a"
-dependencies = [
- "base64 0.21.1",
- "serde",
-]
-
 [[package]]
 name = "pem"
 version = "3.0.3"
@@ -3825,7 +3835,7 @@ checksum = "39407670928234ebc5e6e580247dd567ad73a3578460c5990f9503df207e8f07"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -3846,8 +3856,8 @@ version = "0.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9eca2c590a5f85da82668fa685c09ce2888b9430e83299debf1f34b65fd4a4ba"
 dependencies = [
- "der",
- "spki",
+ "der 0.6.1",
+ "spki 0.6.0",
 ]

 [[package]]
@@ -3946,14 +3956,14 @@ dependencies = [
 "futures",
 "once_cell",
 "pq_proto",
- "rustls",
- "rustls-pemfile",
+ "rustls 0.22.2",
+ "rustls-pemfile 2.1.1",
 "serde",
 "thiserror",
 "tokio",
 "tokio-postgres",
 "tokio-postgres-rustls",
- "tokio-rustls",
+ "tokio-rustls 0.25.0",
 "tracing",
 "workspace_hack",
 ]
@@ -4042,7 +4052,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3b69d39aab54d069e7f2fe8cb970493e7834601ca2d8c65fd7bbd183578080d1"
 dependencies = [
 "proc-macro2",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -4053,9 +4063,9 @@ checksum = "dc375e1527247fe1a97d8b7156678dfe7c1af2fc075c9a4db3690ecd2a148068"

 [[package]]
 name = "proc-macro2"
-version = "1.0.66"
+version = "1.0.78"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "18fb31db3f9bddb2ea821cde30a9f70117e3f119938b5ee630b7403aa6e2ead9"
+checksum = "e2422ad645d89c99f8f3e6b88a9fdeca7fabeac836b1002371c4367c8f984aae"
 dependencies = [
 "unicode-ident",
 ]
@@ -4202,8 +4212,8 @@ dependencies = [
 "routerify",
 "rstest",
 "rustc-hash",
- "rustls",
- "rustls-pemfile",
+ "rustls 0.22.2",
+ "rustls-pemfile 2.1.1",
 "scopeguard",
 "serde",
 "serde_json",
@@ -4216,11 +4226,10 @@ dependencies = [
 "thiserror",
 "tikv-jemalloc-ctl",
 "tikv-jemallocator",
- "tls-listener",
 "tokio",
 "tokio-postgres",
 "tokio-postgres-rustls",
- "tokio-rustls",
+ "tokio-rustls 0.25.0",
 "tokio-util",
 "tracing",
 "tracing-opentelemetry",
@@ -4248,9 +4257,9 @@ dependencies = [

 [[package]]
 name = "quote"
-version = "1.0.32"
+version = "1.0.35"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "50f3b39ccfb720540debaa0164757101c08ecb8d326b15358ce76a62c7e85965"
+checksum = "291ec9ab5efd934aaf503a6466c5d5251535d108ee747472c3977cc5acc868ef"
 dependencies = [
 "proc-macro2",
 ]
@@ -4371,12 +4380,12 @@ dependencies = [

 [[package]]
 name = "rcgen"
-version = "0.11.1"
+version = "0.12.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4954fbc00dcd4d8282c987710e50ba513d351400dbdd00e803a05172a90d8976"
+checksum = "48406db8ac1f3cbc7dcdb56ec355343817958a356ff430259bb07baf7607e1e1"
 dependencies = [
- "pem 2.0.1",
- "ring 0.16.20",
+ "pem",
+ "ring 0.17.6",
 "time",
 "yasna",
 ]
@@ -4394,15 +4403,15 @@ dependencies = [
 "itoa",
 "percent-encoding",
 "pin-project-lite",
- "rustls",
+ "rustls 0.21.9",
 "rustls-native-certs",
- "rustls-pemfile",
+ "rustls-pemfile 1.0.2",
 "rustls-webpki 0.101.7",
 "ryu",
 "sha1_smol",
 "socket2 0.4.9",
 "tokio",
- "tokio-rustls",
+ "tokio-rustls 0.24.0",
 "tokio-util",
 "url",
 ]
@@ -4548,14 +4557,14 @@ dependencies = [
 "once_cell",
 "percent-encoding",
 "pin-project-lite",
- "rustls",
- "rustls-pemfile",
+ "rustls 0.21.9",
+ "rustls-pemfile 1.0.2",
 "serde",
 "serde_json",
 "serde_urlencoded",
 "tokio",
 "tokio-native-tls",
- "tokio-rustls",
+ "tokio-rustls 0.24.0",
 "tokio-util",
 "tower-service",
 "url",
@@ -4721,7 +4730,7 @@ dependencies = [
 "regex",
 "relative-path",
 "rustc_version",
- "syn 2.0.32",
+ "syn 2.0.52",
 "unicode-ident",
 ]

@@ -4805,6 +4814,20 @@ dependencies = [
 "sct",
 ]

+[[package]]
+name = "rustls"
+version = "0.22.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e87c9956bd9807afa1f77e0f7594af32566e830e088a5576d27c5b6f30f49d41"
+dependencies = [
+ "log",
+ "ring 0.17.6",
+ "rustls-pki-types",
+ "rustls-webpki 0.102.2",
+ "subtle",
+ "zeroize",
+]
+
 [[package]]
 name = "rustls-native-certs"
 version = "0.6.2"
@@ -4812,7 +4835,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0167bac7a9f490495f3c33013e7722b53cb087ecbe082fb0c6387c96f634ea50"
 dependencies = [
 "openssl-probe",
- "rustls-pemfile",
+ "rustls-pemfile 1.0.2",
 "schannel",
 "security-framework",
 ]
@@ -4826,6 +4849,22 @@ dependencies = [
 "base64 0.21.1",
 ]

+[[package]]
+name = "rustls-pemfile"
+version = "2.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f48172685e6ff52a556baa527774f61fcaa884f59daf3375c62a3f1cd2549dab"
+dependencies = [
+ "base64 0.21.1",
+ "rustls-pki-types",
+]
+
+[[package]]
+name = "rustls-pki-types"
+version = "1.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5ede67b28608b4c60685c7d54122d4400d90f62b40caee7700e700380a390fa8"
+
 [[package]]
 name = "rustls-webpki"
 version = "0.100.2"
@@ -4846,6 +4885,17 @@ dependencies = [
 "untrusted 0.9.0",
 ]

+[[package]]
+name = "rustls-webpki"
+version = "0.102.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "faaa0a62740bedb9b2ef5afa303da42764c012f743917351dc9a237ea1663610"
+dependencies = [
+ "ring 0.17.6",
+ "rustls-pki-types",
+ "untrusted 0.9.0",
+]
+
 [[package]]
 name = "rustversion"
 version = "1.0.12"
@@ -4888,7 +4938,7 @@ dependencies = [
 "serde_with",
 "thiserror",
 "tokio",
- "tokio-rustls",
+ "tokio-rustls 0.25.0",
 "tokio-stream",
 "tracing",
 "tracing-appender",
@@ -5023,7 +5073,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3be24c1842290c45df0a7bf069e0c268a747ad05a192f2fd7dcfdbc1cba40928"
 dependencies = [
 "base16ct",
- "der",
+ "der 0.6.1",
 "generic-array",
 "pkcs8",
 "subtle",
@@ -5067,7 +5117,7 @@ checksum = "2e95efd0cefa32028cdb9766c96de71d96671072f9fb494dc9fb84c0ef93e52b"
 dependencies = [
 "httpdate",
 "reqwest",
- "rustls",
+ "rustls 0.21.9",
 "sentry-backtrace",
 "sentry-contexts",
 "sentry-core",
@@ -5189,7 +5239,7 @@ checksum = "aafe972d60b0b9bee71a91b92fee2d4fb3c9d7e8f6b179aa99f27203d99a4816"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -5270,7 +5320,7 @@ dependencies = [
 "darling",
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -5356,6 +5406,15 @@ dependencies = [
 "rand_core 0.6.4",
 ]

+[[package]]
+name = "signature"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "77549399552de45a898a580c1b41d445bf730df867cc44e6c0233bbc4b8329de"
+dependencies = [
+ "rand_core 0.6.4",
+]
+
 [[package]]
 name = "simple_asn1"
 version = "0.6.2"
@@ -5440,7 +5499,17 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "67cf02bbac7a337dc36e4f5a693db6c21e7863f45070f7064577eb4367a3212b"
 dependencies = [
 "base64ct",
- "der",
+ "der 0.6.1",
+]
+
+[[package]]
+name = "spki"
+version = "0.7.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d91ed6c858b01f942cd56b37a94b3e0a1798290327d1236e4d9cf4eaca44d29d"
+dependencies = [
+ "base64ct",
+ "der 0.7.8",
 ]

 [[package]]
@@ -5526,9 +5595,9 @@ checksum = "81cdd64d312baedb58e21336b31bc043b77e01cc99033ce76ef539f78e965ebc"

 [[package]]
 name = "svg_fmt"
-version = "0.4.1"
+version = "0.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8fb1df15f412ee2e9dfc1c504260fa695c1c3f10fe9f4a6ee2d2184d7d6450e2"
+checksum = "f83ba502a3265efb76efb89b0a2f7782ad6f2675015d4ce37e4b547dda42b499"

 [[package]]
 name = "syn"
@@ -5543,9 +5612,9 @@ dependencies = [

 [[package]]
 name = "syn"
-version = "2.0.32"
+version = "2.0.52"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "239814284fd6f1a4ffe4ca893952cdd93c224b6a1571c9a9eadd670295c0c9e2"
+checksum = "b699d15b36d1f02c3e7c69f8ffef53de37aefae075d8488d4ba1a7788d574a07"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -5660,22 +5729,22 @@ dependencies = [

 [[package]]
 name = "thiserror"
-version = "1.0.47"
+version = "1.0.57"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "97a802ec30afc17eee47b2855fc72e0c4cd62be9b4efe6591edde0ec5bd68d8f"
+checksum = "1e45bcbe8ed29775f228095caf2cd67af7a4ccf756ebff23a306bf3e8b47b24b"
 dependencies = [
 "thiserror-impl",
 ]

 [[package]]
 name = "thiserror-impl"
-version = "1.0.47"
+version = "1.0.57"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6bb623b56e39ab7dcd4b1b98bb6c8f8d907ed255b18de254088016b27a8ee19b"
+checksum = "a953cb265bef375dae3de6663da4d3804eee9682ea80d8e2542529b73c531c81"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -5794,20 +5863,6 @@ version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1f3ccbac311fea05f86f61904b462b55fb3df8837a366dfc601a0161d0532f20"

-[[package]]
-name = "tls-listener"
-version = "0.7.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "81294c017957a1a69794f506723519255879e15a870507faf45dfed288b763dd"
-dependencies = [
- "futures-util",
- "hyper",
- "pin-project-lite",
- "thiserror",
- "tokio",
- "tokio-rustls",
-]
-
 [[package]]
 name = "tokio"
 version = "1.36.0"
@@ -5860,7 +5915,7 @@ checksum = "5b8a1e28f2deaa14e508979454cb3a223b10b938b45af148bc0986de36f1923b"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -5898,16 +5953,17 @@ dependencies = [

 [[package]]
 name = "tokio-postgres-rustls"
-version = "0.10.0"
+version = "0.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dd5831152cb0d3f79ef5523b357319ba154795d64c7078b2daa95a803b54057f"
+checksum = "0ea13f22eda7127c827983bdaf0d7fff9df21c8817bab02815ac277a21143677"
 dependencies = [
 "futures",
- "ring 0.16.20",
- "rustls",
+ "ring 0.17.6",
+ "rustls 0.22.2",
 "tokio",
 "tokio-postgres",
- "tokio-rustls",
+ "tokio-rustls 0.25.0",
+ "x509-certificate",
 ]

 [[package]]
@@ -5916,7 +5972,18 @@ version = "0.24.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e0d409377ff5b1e3ca6437aa86c1eb7d40c134bfec254e44c830defa92669db5"
 dependencies = [
- "rustls",
+ "rustls 0.21.9",
+ "tokio",
+]
+
+[[package]]
+name = "tokio-rustls"
+version = "0.25.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "775e0c0f0adb3a2f22a00c4745d728b479985fc15ee7ca6a2608388c5569860f"
+dependencies = [
+ "rustls 0.22.2",
+ "rustls-pki-types",
 "tokio",
 ]

@@ -6031,9 +6098,9 @@ dependencies = [
 "pin-project",
 "prost",
 "rustls-native-certs",
- "rustls-pemfile",
+ "rustls-pemfile 1.0.2",
 "tokio",
- "tokio-rustls",
+ "tokio-rustls 0.24.0",
 "tokio-stream",
 "tower",
 "tower-layer",
@@ -6129,7 +6196,7 @@ checksum = "0f57e3ca2a01450b1a921183a9c9cbfda207fd822cef4ccb00a65402cbba7a74"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -6345,7 +6412,7 @@ dependencies = [
 "base64 0.21.1",
 "log",
 "once_cell",
- "rustls",
+ "rustls 0.21.9",
 "rustls-webpki 0.100.2",
 "url",
 "webpki-roots 0.23.1",
@@ -6587,7 +6654,7 @@ dependencies = [
 "once_cell",
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 "wasm-bindgen-shared",
 ]

@@ -6621,7 +6688,7 @@ checksum = "e128beba882dd1eb6200e1dc92ae6c5dbaa4311aa7bb211ca035779e5efc39f8"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 "wasm-bindgen-backend",
 "wasm-bindgen-shared",
 ]
@@ -6954,19 +7021,18 @@ dependencies = [
 "regex-automata 0.4.3",
 "regex-syntax 0.8.2",
 "reqwest",
- "ring 0.16.20",
- "rustls",
+ "rustls 0.21.9",
 "scopeguard",
 "serde",
 "serde_json",
 "smallvec",
 "subtle",
 "syn 1.0.109",
- "syn 2.0.32",
+ "syn 2.0.52",
 "time",
 "time-macros",
 "tokio",
- "tokio-rustls",
+ "tokio-rustls 0.24.0",
 "tokio-util",
 "toml_datetime",
 "toml_edit",
@@ -6977,11 +7043,31 @@ dependencies = [
 "tungstenite",
 "url",
 "uuid",
+ "zeroize",
 "zstd",
 "zstd-safe",
 "zstd-sys",
 ]

+[[package]]
+name = "x509-certificate"
+version = "0.23.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "66534846dec7a11d7c50a74b7cdb208b9a581cad890b7866430d438455847c85"
+dependencies = [
+ "bcder",
+ "bytes",
+ "chrono",
+ "der 0.7.8",
+ "hex",
+ "pem",
+ "ring 0.17.6",
+ "signature 2.2.0",
+ "spki 0.7.3",
+ "thiserror",
+ "zeroize",
+]
+
 [[package]]
 name = "x509-parser"
 version = "0.15.0"
@@ -7040,7 +7126,7 @@ checksum = "b3c129550b3e6de3fd0ba67ba5c81818f9805e58b8d7fee80a3a59d2c9fc601a"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.32",
+ "syn 2.0.52",
 ]

 [[package]]
@@ -7048,6 +7134,20 @@ name = "zeroize"
 version = "1.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2a0956f1ba7c7909bfb66c2e9e4124ab6f6482560f6628b5aaeba39207c9aad9"
+dependencies = [
+ "zeroize_derive",
+]
+
+[[package]]
+name = "zeroize_derive"
+version = "1.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ce36e65b0d2999d2aafac989fb249189a141aee1f53c612c1f37d72631959f69"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.52",
+]

 [[package]]
 name = "zstd"
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -129,8 +129,8 @@ reqwest-retry = "0.2.2"
 routerify = "3"
 rpds = "0.13"
 rustc-hash = "1.1.0"
-rustls = "0.21"
-rustls-pemfile = "1"
+rustls = "0.22"
+rustls-pemfile = "2"
 rustls-split = "0.3"
 scopeguard = "1.1"
 sysinfo = "0.29.2"
@@ -156,12 +156,11 @@ test-context = "0.1"
 thiserror = "1.0"
 tikv-jemallocator = "0.5"
 tikv-jemalloc-ctl = "0.5"
-tls-listener = { version = "0.7", features = ["rustls", "hyper-h1"] }
 tokio = { version = "1.17", features = ["macros"] }
 tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
 tokio-io-timeout = "1.2.0"
-tokio-postgres-rustls = "0.10.0"
-tokio-rustls = "0.24"
+tokio-postgres-rustls = "0.11.0"
+tokio-rustls = "0.25"
 tokio-stream = "0.1"
 tokio-tar = "0.3"
 tokio-util = { version = "0.7.10", features = ["io", "rt"] }
@@ -220,7 +219,7 @@ workspace_hack = { version = "0.1", path = "./workspace_hack/" }

 ## Build dependencies
 criterion = "0.5.1"
-rcgen = "0.11"
+rcgen = "0.12"
 rstest = "0.18"
 camino-tempfile = "1.0.2"
 tonic-build = "0.9"
--- a/4
+++ b/4
@@ -53,7 +53,7 @@ RUN set -e \
      --bin pagectl  \
      --bin safekeeper  \
      --bin storage_broker  \
-      --bin attachment_service  \
+      --bin storage_controller  \
      --bin proxy  \
      --bin neon_local \
      --locked --release \
@@ -81,7 +81,7 @@ COPY --from=build --chown=neon:neon /home/nonroot/target/release/pageserver
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/pagectl             /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/safekeeper          /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/storage_broker      /usr/local/bin
-COPY --from=build --chown=neon:neon /home/nonroot/target/release/attachment_service  /usr/local/bin
+COPY --from=build --chown=neon:neon /home/nonroot/target/release/storage_controller  /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/proxy               /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/neon_local          /usr/local/bin

--- a/clippy.toml
+++ b/clippy.toml
@@ -3,3 +3,10 @@ disallowed-methods = [
    # Allow this for now, to deny it later once we stop using Handle::block_on completely
    # "tokio::runtime::Handle::block_on",
 ]
+
+disallowed-macros = [
+    # use std::pin::pin
+    "futures::pin_mut",
+    # cannot disallow this, because clippy finds used from tokio macros
+    #"tokio::pin",
+]
--- a/compute_tools/src/compute.rs
+++ b/compute_tools/src/compute.rs
@@ -17,6 +17,7 @@ use chrono::{DateTime, Utc};
 use futures::future::join_all;
 use futures::stream::FuturesUnordered;
 use futures::StreamExt;
+use nix::unistd::Pid;
 use postgres::error::SqlState;
 use postgres::{Client, NoTls};
 use tracing::{debug, error, info, instrument, warn};
@@ -396,9 +397,9 @@ impl ComputeNode {
    // Gets the basebackup in a retry loop
    #[instrument(skip_all, fields(%lsn))]
    pub fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
-        let mut retry_period_ms = 500;
+        let mut retry_period_ms = 500.0;
        let mut attempts = 0;
-        let max_attempts = 5;
+        let max_attempts = 10;
        loop {
            let result = self.try_get_basebackup(compute_state, lsn);
            match result {
@@ -410,8 +411,8 @@ impl ComputeNode {
                        "Failed to get basebackup: {} (attempt {}/{})",
                        e, attempts, max_attempts
                    );
-                    std::thread::sleep(std::time::Duration::from_millis(retry_period_ms));
-                    retry_period_ms *= 2;
+                    std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
+                    retry_period_ms *= 1.5;
                }
                Err(_) => {
                    return result;
@@ -722,8 +723,12 @@ impl ComputeNode {
        // Stop it when it's ready
        info!("waiting for postgres");
        wait_for_postgres(&mut pg, Path::new(pgdata))?;
-        pg.kill()?;
-        info!("sent kill signal");
+        // SIGQUIT orders postgres to exit immediately. We don't want to SIGKILL
+        // it to avoid orphaned processes prowling around while datadir is
+        // wiped.
+        let pm_pid = Pid::from_raw(pg.id() as i32);
+        kill(pm_pid, Signal::SIGQUIT)?;
+        info!("sent SIGQUIT signal");
        pg.wait()?;
        info!("done prewarming");

@@ -764,6 +769,26 @@ impl ComputeNode {
        Ok((pg, logs_handle))
    }

+    /// Do post configuration of the already started Postgres. This function spawns a background thread to
+    /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
+    /// version. In the future, it may upgrade all 3rd-party extensions.
+    #[instrument(skip_all)]
+    pub fn post_apply_config(&self) -> Result<()> {
+        let connstr = self.connstr.clone();
+        thread::spawn(move || {
+            let func = || {
+                let mut client = Client::connect(connstr.as_str(), NoTls)?;
+                handle_neon_extension_upgrade(&mut client)
+                    .context("handle_neon_extension_upgrade")?;
+                Ok::<_, anyhow::Error>(())
+            };
+            if let Err(err) = func() {
+                error!("error while post_apply_config: {err:#}");
+            }
+        });
+        Ok(())
+    }
+
    /// Do initial configuration of the already started Postgres.
    #[instrument(skip_all)]
    pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
@@ -998,18 +1023,21 @@ impl ComputeNode {
        let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;

        let config_time = Utc::now();
-        if pspec.spec.mode == ComputeMode::Primary && !pspec.spec.skip_pg_catalog_updates {
-            let pgdata_path = Path::new(&self.pgdata);
-            // temporarily reset max_cluster_size in config
-            // to avoid the possibility of hitting the limit, while we are applying config:
-            // creating new extensions, roles, etc...
-            config::compute_ctl_temp_override_create(pgdata_path, "neon.max_cluster_size=-1")?;
-            self.pg_reload_conf()?;
+        if pspec.spec.mode == ComputeMode::Primary {
+            if !pspec.spec.skip_pg_catalog_updates {
+                let pgdata_path = Path::new(&self.pgdata);
+                // temporarily reset max_cluster_size in config
+                // to avoid the possibility of hitting the limit, while we are applying config:
+                // creating new extensions, roles, etc...
+                config::compute_ctl_temp_override_create(pgdata_path, "neon.max_cluster_size=-1")?;
+                self.pg_reload_conf()?;

-            self.apply_config(&compute_state)?;
+                self.apply_config(&compute_state)?;

-            config::compute_ctl_temp_override_remove(pgdata_path)?;
-            self.pg_reload_conf()?;
+                config::compute_ctl_temp_override_remove(pgdata_path)?;
+                self.pg_reload_conf()?;
+            }
+            self.post_apply_config()?;
        }

        let startup_end_time = Utc::now();
--- a/compute_tools/src/spec.rs
+++ b/compute_tools/src/spec.rs
@@ -302,9 +302,9 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
            RoleAction::Create => {
                // This branch only runs when roles are created through the console, so it is
                // safe to add more permissions here. BYPASSRLS and REPLICATION are inherited
-                // from neon_superuser.
+                // from neon_superuser. (NOTE: REPLICATION has been removed from here for now).
                let mut query: String = format!(
-                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
+                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS IN ROLE neon_superuser",
                    name.pg_quote()
                );
                info!("running role create query: '{}'", &query);
@@ -744,7 +744,17 @@ pub fn handle_extension_neon(client: &mut Client) -> Result<()> {
    // - extension was just installed
    // - extension was already installed and is up to date
    let query = "ALTER EXTENSION neon UPDATE";
-    info!("update neon extension schema with query: {}", query);
+    info!("update neon extension version with query: {}", query);
+    client.simple_query(query)?;
+
+    Ok(())
+}
+
+#[instrument(skip_all)]
+pub fn handle_neon_extension_upgrade(client: &mut Client) -> Result<()> {
+    info!("handle neon extension upgrade");
+    let query = "ALTER EXTENSION neon UPDATE";
+    info!("update neon extension version with query: {}", query);
    client.simple_query(query)?;

    Ok(())
@@ -795,6 +805,18 @@ $$;"#,
        "",
        "",
        // Add new migrations below.
+        r#"
+DO $$
+DECLARE
+    role_name TEXT;
+BEGIN
+    FOR role_name IN SELECT rolname FROM pg_roles WHERE rolreplication IS TRUE
+    LOOP
+        RAISE NOTICE 'EXECUTING ALTER ROLE % NOREPLICATION', quote_ident(role_name);
+        EXECUTE 'ALTER ROLE ' || quote_ident(role_name) || ' NOREPLICATION';
+    END LOOP;
+END
+$$;"#,
    ];

    let mut query = "CREATE SCHEMA IF NOT EXISTS neon_migration";
--- a/control_plane/attachment_service/Cargo.toml
+++ b/control_plane/attachment_service/Cargo.toml
@@ -4,6 +4,10 @@ version = "0.1.0"
 edition.workspace = true
 license.workspace = true

+[[bin]]
+name = "storage_controller"
+path = "src/main.rs"
+
 [features]
 default = []
 # Enables test-only APIs and behaviors
--- a/control_plane/attachment_service/src/compute_hook.rs
+++ b/control_plane/attachment_service/src/compute_hook.rs
@@ -3,7 +3,7 @@ use std::{collections::HashMap, time::Duration};
 use control_plane::endpoint::{ComputeControlPlane, EndpointStatus};
 use control_plane::local_env::LocalEnv;
 use hyper::{Method, StatusCode};
-use pageserver_api::shard::{ShardIndex, ShardNumber, TenantShardId};
+use pageserver_api::shard::{ShardCount, ShardNumber, ShardStripeSize, TenantShardId};
 use postgres_connection::parse_host_port;
 use serde::{Deserialize, Serialize};
 use tokio_util::sync::CancellationToken;
@@ -19,8 +19,66 @@ const SLOWDOWN_DELAY: Duration = Duration::from_secs(5);

 pub(crate) const API_CONCURRENCY: usize = 32;

-pub(super) struct ComputeHookTenant {
-    shards: Vec<(ShardIndex, NodeId)>,
+struct ShardedComputeHookTenant {
+    stripe_size: ShardStripeSize,
+    shard_count: ShardCount,
+    shards: Vec<(ShardNumber, NodeId)>,
+}
+
+enum ComputeHookTenant {
+    Unsharded(NodeId),
+    Sharded(ShardedComputeHookTenant),
+}
+
+impl ComputeHookTenant {
+    /// Construct with at least one shard's information
+    fn new(tenant_shard_id: TenantShardId, stripe_size: ShardStripeSize, node_id: NodeId) -> Self {
+        if tenant_shard_id.shard_count.count() > 1 {
+            Self::Sharded(ShardedComputeHookTenant {
+                shards: vec![(tenant_shard_id.shard_number, node_id)],
+                stripe_size,
+                shard_count: tenant_shard_id.shard_count,
+            })
+        } else {
+            Self::Unsharded(node_id)
+        }
+    }
+
+    /// Set one shard's location.  If stripe size or shard count have changed, Self is reset
+    /// and drops existing content.
+    fn update(
+        &mut self,
+        tenant_shard_id: TenantShardId,
+        stripe_size: ShardStripeSize,
+        node_id: NodeId,
+    ) {
+        match self {
+            Self::Unsharded(existing_node_id) if tenant_shard_id.shard_count.count() == 1 => {
+                *existing_node_id = node_id
+            }
+            Self::Sharded(sharded_tenant)
+                if sharded_tenant.stripe_size == stripe_size
+                    && sharded_tenant.shard_count == tenant_shard_id.shard_count =>
+            {
+                if let Some(existing) = sharded_tenant
+                    .shards
+                    .iter()
+                    .position(|s| s.0 == tenant_shard_id.shard_number)
+                {
+                    sharded_tenant.shards.get_mut(existing).unwrap().1 = node_id;
+                } else {
+                    sharded_tenant
+                        .shards
+                        .push((tenant_shard_id.shard_number, node_id));
+                    sharded_tenant.shards.sort_by_key(|s| s.0)
+                }
+            }
+            _ => {
+                // Shard count changed: reset struct.
+                *self = Self::new(tenant_shard_id, stripe_size, node_id);
+            }
+        }
+    }
 }

 #[derive(Serialize, Deserialize, Debug)]
@@ -33,6 +91,7 @@ struct ComputeHookNotifyRequestShard {
 #[derive(Serialize, Deserialize, Debug)]
 struct ComputeHookNotifyRequest {
    tenant_id: TenantId,
+    stripe_size: Option<ShardStripeSize>,
    shards: Vec<ComputeHookNotifyRequestShard>,
 }

@@ -63,42 +122,43 @@ pub(crate) enum NotifyError {
 }

 impl ComputeHookTenant {
-    async fn maybe_reconfigure(&mut self, tenant_id: TenantId) -> Option<ComputeHookNotifyRequest> {
-        // Find the highest shard count and drop any shards that aren't
-        // for that shard count.
-        let shard_count = self.shards.iter().map(|(k, _v)| k.shard_count).max();
-        let Some(shard_count) = shard_count else {
-            // No shards, nothing to do.
-            tracing::info!("ComputeHookTenant::maybe_reconfigure: no shards");
-            return None;
-        };
-
-        self.shards.retain(|(k, _v)| k.shard_count == shard_count);
-        self.shards
-            .sort_by_key(|(shard, _node_id)| shard.shard_number);
-
-        if self.shards.len() == shard_count.count() as usize || shard_count.is_unsharded() {
-            // We have pageservers for all the shards: emit a configuration update
-            return Some(ComputeHookNotifyRequest {
+    fn maybe_reconfigure(&self, tenant_id: TenantId) -> Option<ComputeHookNotifyRequest> {
+        match self {
+            Self::Unsharded(node_id) => Some(ComputeHookNotifyRequest {
                tenant_id,
-                shards: self
-                    .shards
-                    .iter()
-                    .map(|(shard, node_id)| ComputeHookNotifyRequestShard {
-                        shard_number: shard.shard_number,
-                        node_id: *node_id,
-                    })
-                    .collect(),
-            });
-        } else {
-            tracing::info!(
-                "ComputeHookTenant::maybe_reconfigure: not enough shards ({}/{})",
-                self.shards.len(),
-                shard_count.count()
-            );
-        }
+                shards: vec![ComputeHookNotifyRequestShard {
+                    shard_number: ShardNumber(0),
+                    node_id: *node_id,
+                }],
+                stripe_size: None,
+            }),
+            Self::Sharded(sharded_tenant)
+                if sharded_tenant.shards.len() == sharded_tenant.shard_count.count() as usize =>
+            {
+                Some(ComputeHookNotifyRequest {
+                    tenant_id,
+                    shards: sharded_tenant
+                        .shards
+                        .iter()
+                        .map(|(shard_number, node_id)| ComputeHookNotifyRequestShard {
+                            shard_number: *shard_number,
+                            node_id: *node_id,
+                        })
+                        .collect(),
+                    stripe_size: Some(sharded_tenant.stripe_size),
+                })
+            }
+            Self::Sharded(sharded_tenant) => {
+                // Sharded tenant doesn't yet have information for all its shards

-        None
+                tracing::info!(
+                    "ComputeHookTenant::maybe_reconfigure: not enough shards ({}/{})",
+                    sharded_tenant.shards.len(),
+                    sharded_tenant.shard_count.count()
+                );
+                None
+            }
+        }
    }
 }

@@ -139,7 +199,11 @@ impl ComputeHook {
        };
        let cplane =
            ComputeControlPlane::load(env.clone()).expect("Error loading compute control plane");
-        let ComputeHookNotifyRequest { tenant_id, shards } = reconfigure_request;
+        let ComputeHookNotifyRequest {
+            tenant_id,
+            shards,
+            stripe_size,
+        } = reconfigure_request;

        let compute_pageservers = shards
            .into_iter()
@@ -156,7 +220,9 @@ impl ComputeHook {
        for (endpoint_name, endpoint) in &cplane.endpoints {
            if endpoint.tenant_id == tenant_id && endpoint.status() == EndpointStatus::Running {
                tracing::info!("Reconfiguring endpoint {}", endpoint_name,);
-                endpoint.reconfigure(compute_pageservers.clone()).await?;
+                endpoint
+                    .reconfigure(compute_pageservers.clone(), stripe_size)
+                    .await?;
            }
        }

@@ -271,30 +337,26 @@ impl ComputeHook {
        &self,
        tenant_shard_id: TenantShardId,
        node_id: NodeId,
+        stripe_size: ShardStripeSize,
        cancel: &CancellationToken,
    ) -> Result<(), NotifyError> {
        let mut locked = self.state.lock().await;
-        let entry = locked
-            .entry(tenant_shard_id.tenant_id)
-            .or_insert_with(|| ComputeHookTenant { shards: Vec::new() });

-        let shard_index = ShardIndex {
-            shard_count: tenant_shard_id.shard_count,
-            shard_number: tenant_shard_id.shard_number,
+        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
+            }
        };

-        let mut set = false;
-        for (existing_shard, existing_node) in &mut entry.shards {
-            if *existing_shard == shard_index {
-                *existing_node = node_id;
-                set = true;
-            }
-        }
-        if !set {
-            entry.shards.push((shard_index, node_id));
-        }
-
-        let reconfigure_request = entry.maybe_reconfigure(tenant_shard_id.tenant_id).await;
+        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.
@@ -316,3 +378,85 @@ impl ComputeHook {
        }
    }
 }
+
+#[cfg(test)]
+pub(crate) mod tests {
+    use pageserver_api::shard::{ShardCount, ShardNumber};
+    use utils::id::TenantId;
+
+    use super::*;
+
+    #[test]
+    fn tenant_updates() -> anyhow::Result<()> {
+        let tenant_id = TenantId::generate();
+        let mut tenant_state = ComputeHookTenant::new(
+            TenantShardId {
+                tenant_id,
+                shard_count: ShardCount::new(0),
+                shard_number: ShardNumber(0),
+            },
+            ShardStripeSize(12345),
+            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());
+
+        // 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
+        // see all shards)
+        tenant_state.update(
+            TenantShardId {
+                tenant_id,
+                shard_count: ShardCount::new(2),
+                shard_number: ShardNumber(1),
+            },
+            ShardStripeSize(32768),
+            NodeId(1),
+        );
+        assert!(tenant_state.maybe_reconfigure(tenant_id).is_none());
+
+        // Writing the second shard makes it ready to notify
+        tenant_state.update(
+            TenantShardId {
+                tenant_id,
+                shard_count: ShardCount::new(2),
+                shard_number: ShardNumber(0),
+            },
+            ShardStripeSize(32768),
+            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))
+        );
+
+        Ok(())
+    }
+}
--- a/control_plane/attachment_service/src/http.rs
+++ b/control_plane/attachment_service/src/http.rs
@@ -1,6 +1,5 @@
 use crate::reconciler::ReconcileError;
 use crate::service::{Service, STARTUP_RECONCILE_TIMEOUT};
-use crate::PlacementPolicy;
 use hyper::{Body, Request, Response};
 use hyper::{StatusCode, Uri};
 use pageserver_api::models::{
@@ -119,13 +118,9 @@ async fn handle_tenant_create(

    let create_req = json_request::<TenantCreateRequest>(&mut req).await?;

-    // TODO: enable specifying this.  Using Single as a default helps legacy tests to work (they
-    // have no expectation of HA).
-    let placement_policy = PlacementPolicy::Single;
-
    json_response(
        StatusCode::CREATED,
-        service.tenant_create(create_req, placement_policy).await?,
+        service.tenant_create(create_req).await?,
    )
 }

--- a/control_plane/attachment_service/src/lib.rs
+++ b/control_plane/attachment_service/src/lib.rs
@@ -1,4 +1,4 @@
-use serde::{Deserialize, Serialize};
+use serde::Serialize;
 use utils::seqwait::MonotonicCounter;

 mod auth;
@@ -13,23 +13,6 @@ mod schema;
 pub mod service;
 mod tenant_state;

-#[derive(Clone, Serialize, Deserialize, Debug, PartialEq, Eq)]
-enum PlacementPolicy {
-    /// Cheapest way to attach a tenant: just one pageserver, no secondary
-    Single,
-    /// Production-ready way to attach a tenant: one attached pageserver and
-    /// some number of secondaries.
-    Double(usize),
-    /// Create one secondary mode locations. This is useful when onboarding
-    /// a tenant, or for an idle tenant that we might want to bring online quickly.
-    Secondary,
-
-    /// Do not attach to any pageservers.  This is appropriate for tenants that
-    /// have been idle for a long time, where we do not mind some delay in making
-    /// them available in future.
-    Detached,
-}
-
 #[derive(Ord, PartialOrd, Eq, PartialEq, Copy, Clone, Serialize)]
 struct Sequence(u64);

@@ -66,9 +49,3 @@ impl Sequence {
        Sequence(self.0 + 1)
    }
 }
-
-impl Default for PlacementPolicy {
-    fn default() -> Self {
-        PlacementPolicy::Double(1)
-    }
-}
--- a/control_plane/attachment_service/src/node.rs
+++ b/control_plane/attachment_service/src/node.rs
@@ -1,6 +1,16 @@
-use pageserver_api::controller_api::{NodeAvailability, NodeSchedulingPolicy};
+use std::{str::FromStr, time::Duration};
+
+use hyper::StatusCode;
+use pageserver_api::{
+    controller_api::{
+        NodeAvailability, NodeRegisterRequest, NodeSchedulingPolicy, TenantLocateResponseShard,
+    },
+    shard::TenantShardId,
+};
+use pageserver_client::mgmt_api;
 use serde::Serialize;
-use utils::id::NodeId;
+use tokio_util::sync::CancellationToken;
+use utils::{backoff, id::NodeId};

 use crate::persistence::NodePersistence;

@@ -12,16 +22,29 @@ use crate::persistence::NodePersistence;
 /// implementation of serialization on this type is only for debug dumps.
 #[derive(Clone, Serialize)]
 pub(crate) struct Node {
-    pub(crate) id: NodeId,
+    id: NodeId,

-    pub(crate) availability: NodeAvailability,
-    pub(crate) scheduling: NodeSchedulingPolicy,
+    availability: NodeAvailability,
+    scheduling: NodeSchedulingPolicy,

-    pub(crate) listen_http_addr: String,
-    pub(crate) listen_http_port: u16,
+    listen_http_addr: String,
+    listen_http_port: u16,

-    pub(crate) listen_pg_addr: String,
-    pub(crate) listen_pg_port: u16,
+    listen_pg_addr: String,
+    listen_pg_port: u16,
+
+    // This cancellation token means "stop any RPCs in flight to this node, and don't start
+    // any more". It is not related to process shutdown.
+    #[serde(skip)]
+    cancel: CancellationToken,
+}
+
+/// When updating [`Node::availability`] we use this type to indicate to the caller
+/// whether/how they changed it.
+pub(crate) enum AvailabilityTransition {
+    ToActive,
+    ToOffline,
+    Unchanged,
 }

 impl Node {
@@ -29,6 +52,71 @@ impl Node {
        format!("http://{}:{}", self.listen_http_addr, self.listen_http_port)
    }

+    pub(crate) fn get_id(&self) -> NodeId {
+        self.id
+    }
+
+    pub(crate) fn set_scheduling(&mut self, scheduling: NodeSchedulingPolicy) {
+        self.scheduling = scheduling
+    }
+
+    /// Does this registration request match `self`?  This is used when deciding whether a registration
+    /// request should be allowed to update an existing record with the same node ID.
+    pub(crate) fn registration_match(&self, register_req: &NodeRegisterRequest) -> bool {
+        self.id == register_req.node_id
+            && self.listen_http_addr == register_req.listen_http_addr
+            && self.listen_http_port == register_req.listen_http_port
+            && self.listen_pg_addr == register_req.listen_pg_addr
+            && self.listen_pg_port == register_req.listen_pg_port
+    }
+
+    /// For a shard located on this node, populate a response object
+    /// with this node's address information.
+    pub(crate) fn shard_location(&self, shard_id: TenantShardId) -> TenantLocateResponseShard {
+        TenantLocateResponseShard {
+            shard_id,
+            node_id: self.id,
+            listen_http_addr: self.listen_http_addr.clone(),
+            listen_http_port: self.listen_http_port,
+            listen_pg_addr: self.listen_pg_addr.clone(),
+            listen_pg_port: self.listen_pg_port,
+        }
+    }
+
+    pub(crate) fn set_availability(
+        &mut self,
+        availability: NodeAvailability,
+    ) -> AvailabilityTransition {
+        use NodeAvailability::*;
+        let transition = match (self.availability, availability) {
+            (Offline, Active) => {
+                // Give the node a new cancellation token, effectively resetting it to un-cancelled.  Any
+                // users of previously-cloned copies of the node will still see the old cancellation
+                // state.  For example, Reconcilers in flight will have to complete and be spawned
+                // again to realize that the node has become available.
+                self.cancel = CancellationToken::new();
+                AvailabilityTransition::ToActive
+            }
+            (Active, Offline) => {
+                // Fire the node's cancellation token to cancel any in-flight API requests to it
+                self.cancel.cancel();
+                AvailabilityTransition::ToOffline
+            }
+            _ => AvailabilityTransition::Unchanged,
+        };
+        self.availability = availability;
+        transition
+    }
+
+    /// Whether we may send API requests to this node.
+    pub(crate) fn is_available(&self) -> bool {
+        // When we clone a node, [`Self::availability`] is a snapshot, but [`Self::cancel`] holds
+        // a reference to the original Node's cancellation status.  Checking both of these results
+        // in a "pessimistic" check where we will consider a Node instance unavailable if it was unavailable
+        // when we cloned it, or if the original Node instance's cancellation token was fired.
+        matches!(self.availability, NodeAvailability::Active) && !self.cancel.is_cancelled()
+    }
+
    /// Is this node elegible to have work scheduled onto it?
    pub(crate) fn may_schedule(&self) -> bool {
        match self.availability {
@@ -44,6 +132,26 @@ impl Node {
        }
    }

+    pub(crate) fn new(
+        id: NodeId,
+        listen_http_addr: String,
+        listen_http_port: u16,
+        listen_pg_addr: String,
+        listen_pg_port: u16,
+    ) -> Self {
+        Self {
+            id,
+            listen_http_addr,
+            listen_http_port,
+            listen_pg_addr,
+            listen_pg_port,
+            scheduling: NodeSchedulingPolicy::Filling,
+            // TODO: we shouldn't really call this Active until we've heartbeated it.
+            availability: NodeAvailability::Active,
+            cancel: CancellationToken::new(),
+        }
+    }
+
    pub(crate) fn to_persistent(&self) -> NodePersistence {
        NodePersistence {
            node_id: self.id.0 as i64,
@@ -54,4 +162,96 @@ impl Node {
            listen_pg_port: self.listen_pg_port as i32,
        }
    }
+
+    pub(crate) fn from_persistent(np: NodePersistence) -> Self {
+        Self {
+            id: NodeId(np.node_id as u64),
+            // At startup we consider a node offline until proven otherwise.
+            availability: NodeAvailability::Offline,
+            scheduling: NodeSchedulingPolicy::from_str(&np.scheduling_policy)
+                .expect("Bad scheduling policy in DB"),
+            listen_http_addr: np.listen_http_addr,
+            listen_http_port: np.listen_http_port as u16,
+            listen_pg_addr: np.listen_pg_addr,
+            listen_pg_port: np.listen_pg_port as u16,
+            cancel: CancellationToken::new(),
+        }
+    }
+
+    /// Wrapper for issuing requests to pageserver management API: takes care of generic
+    /// retry/backoff for retryable HTTP status codes.
+    ///
+    /// This will return None to indicate cancellation.  Cancellation may happen from
+    /// the cancellation token passed in, or from Self's cancellation token (i.e. node
+    /// going offline).
+    pub(crate) async fn with_client_retries<T, O, F>(
+        &self,
+        mut op: O,
+        jwt: &Option<String>,
+        warn_threshold: u32,
+        max_retries: u32,
+        timeout: Duration,
+        cancel: &CancellationToken,
+    ) -> Option<mgmt_api::Result<T>>
+    where
+        O: FnMut(mgmt_api::Client) -> F,
+        F: std::future::Future<Output = mgmt_api::Result<T>>,
+    {
+        fn is_fatal(e: &mgmt_api::Error) -> bool {
+            use mgmt_api::Error::*;
+            match e {
+                ReceiveBody(_) | ReceiveErrorBody(_) => false,
+                ApiError(StatusCode::SERVICE_UNAVAILABLE, _)
+                | ApiError(StatusCode::GATEWAY_TIMEOUT, _)
+                | ApiError(StatusCode::REQUEST_TIMEOUT, _) => false,
+                ApiError(_, _) => true,
+                Cancelled => true,
+            }
+        }
+
+        backoff::retry(
+            || {
+                let http_client = reqwest::ClientBuilder::new()
+                    .timeout(timeout)
+                    .build()
+                    .expect("Failed to construct HTTP client");
+
+                let client =
+                    mgmt_api::Client::from_client(http_client, self.base_url(), jwt.as_deref());
+
+                let node_cancel_fut = self.cancel.cancelled();
+
+                let op_fut = op(client);
+
+                async {
+                    tokio::select! {
+                        r = op_fut=> {r},
+                        _ = node_cancel_fut => {
+                        Err(mgmt_api::Error::Cancelled)
+                    }}
+                }
+            },
+            is_fatal,
+            warn_threshold,
+            max_retries,
+            &format!(
+                "Call to node {} ({}:{}) management API",
+                self.id, self.listen_http_addr, self.listen_http_port
+            ),
+            cancel,
+        )
+        .await
+    }
+}
+
+impl std::fmt::Display for Node {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{} ({})", self.id, self.listen_http_addr)
+    }
+}
+
+impl std::fmt::Debug for Node {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{} ({})", self.id, self.listen_http_addr)
+    }
 }
--- a/control_plane/attachment_service/src/persistence.rs
+++ b/control_plane/attachment_service/src/persistence.rs
@@ -7,11 +7,9 @@ use self::split_state::SplitState;
 use camino::Utf8Path;
 use camino::Utf8PathBuf;
 use diesel::pg::PgConnection;
-use diesel::{
-    Connection, ExpressionMethods, Insertable, QueryDsl, QueryResult, Queryable, RunQueryDsl,
-    Selectable, SelectableHelper,
-};
-use pageserver_api::controller_api::NodeSchedulingPolicy;
+use diesel::prelude::*;
+use diesel::Connection;
+use pageserver_api::controller_api::{NodeSchedulingPolicy, PlacementPolicy};
 use pageserver_api::models::TenantConfig;
 use pageserver_api::shard::{ShardCount, ShardNumber, TenantShardId};
 use serde::{Deserialize, Serialize};
@@ -19,7 +17,6 @@ use utils::generation::Generation;
 use utils::id::{NodeId, TenantId};

 use crate::node::Node;
-use crate::PlacementPolicy;

 /// ## What do we store?
 ///
@@ -210,7 +207,7 @@ impl Persistence {
                tenant.tenant_id = tenant_id.to_string();
                tenant.config = serde_json::to_string(&TenantConfig::default())
                    .map_err(|e| DatabaseError::Logical(format!("Serialization error: {e}")))?;
-                tenant.placement_policy = serde_json::to_string(&PlacementPolicy::default())
+                tenant.placement_policy = serde_json::to_string(&PlacementPolicy::Single)
                    .map_err(|e| DatabaseError::Logical(format!("Serialization error: {e}")))?;
            }
        }
--- a/control_plane/attachment_service/src/reconciler.rs
+++ b/control_plane/attachment_service/src/reconciler.rs
@@ -1,6 +1,5 @@
 use crate::persistence::Persistence;
 use crate::service;
-use pageserver_api::controller_api::NodeAvailability;
 use pageserver_api::models::{
    LocationConfig, LocationConfigMode, LocationConfigSecondary, TenantConfig,
 };
@@ -28,15 +27,16 @@ pub(super) struct Reconciler {
    pub(crate) shard: ShardIdentity,
    pub(crate) generation: Option<Generation>,
    pub(crate) intent: TargetState,
+
+    /// Nodes not referenced by [`Self::intent`], from which we should try
+    /// to detach this tenant shard.
+    pub(crate) detach: Vec<Node>,
+
    pub(crate) config: TenantConfig,
    pub(crate) observed: ObservedState,

    pub(crate) service_config: service::Config,

-    /// A snapshot of the pageservers as they were when we were asked
-    /// to reconcile.
-    pub(crate) pageservers: Arc<HashMap<NodeId, Node>>,
-
    /// A hook to notify the running postgres instances when we change the location
    /// of a tenant.  Use this via [`Self::compute_notify`] to update our failure flag
    /// and guarantee eventual retries.
@@ -67,29 +67,37 @@ pub(super) struct Reconciler {
 /// and the TargetState is just the instruction for a particular Reconciler run.
 #[derive(Debug)]
 pub(crate) struct TargetState {
-    pub(crate) attached: Option<NodeId>,
-    pub(crate) secondary: Vec<NodeId>,
+    pub(crate) attached: Option<Node>,
+    pub(crate) secondary: Vec<Node>,
 }

 impl TargetState {
-    pub(crate) fn from_intent(intent: &IntentState) -> Self {
+    pub(crate) fn from_intent(nodes: &HashMap<NodeId, Node>, intent: &IntentState) -> Self {
        Self {
-            attached: *intent.get_attached(),
-            secondary: intent.get_secondary().clone(),
+            attached: intent.get_attached().map(|n| {
+                nodes
+                    .get(&n)
+                    .expect("Intent attached referenced non-existent node")
+                    .clone()
+            }),
+            secondary: intent
+                .get_secondary()
+                .iter()
+                .map(|n| {
+                    nodes
+                        .get(n)
+                        .expect("Intent secondary referenced non-existent node")
+                        .clone()
+                })
+                .collect(),
        }
    }
-
-    fn all_pageservers(&self) -> Vec<NodeId> {
-        let mut result = self.secondary.clone();
-        if let Some(node_id) = &self.attached {
-            result.push(*node_id);
-        }
-        result
-    }
 }

 #[derive(thiserror::Error, Debug)]
 pub(crate) enum ReconcileError {
+    #[error(transparent)]
+    Remote(#[from] mgmt_api::Error),
    #[error(transparent)]
    Notify(#[from] NotifyError),
    #[error("Cancelled")]
@@ -101,44 +109,83 @@ pub(crate) enum ReconcileError {
 impl Reconciler {
    async fn location_config(
        &mut self,
-        node_id: NodeId,
+        node: &Node,
        config: LocationConfig,
        flush_ms: Option<Duration>,
-    ) -> anyhow::Result<()> {
-        let node = self
-            .pageservers
-            .get(&node_id)
-            .expect("Pageserver may not be removed while referenced");
+        lazy: bool,
+    ) -> Result<(), ReconcileError> {
+        self.observed
+            .locations
+            .insert(node.get_id(), ObservedStateLocation { conf: None });
+
+        // TODO: amend locations that use long-polling: they will hit this timeout.
+        let timeout = Duration::from_secs(25);
+
+        tracing::info!("location_config({node}) calling: {:?}", config);
+        let tenant_shard_id = self.tenant_shard_id;
+        let config_ref = &config;
+        match node
+            .with_client_retries(
+                |client| async move {
+                    let config = config_ref.clone();
+                    client
+                        .location_config(tenant_shard_id, config.clone(), flush_ms, lazy)
+                        .await
+                },
+                &self.service_config.jwt_token,
+                1,
+                3,
+                timeout,
+                &self.cancel,
+            )
+            .await
+        {
+            Some(Ok(_)) => {}
+            Some(Err(e)) => return Err(e.into()),
+            None => return Err(ReconcileError::Cancel),
+        };
+        tracing::info!("location_config({node}) complete: {:?}", config);

        self.observed
            .locations
-            .insert(node.id, ObservedStateLocation { conf: None });
-
-        tracing::info!("location_config({}) calling: {:?}", node_id, config);
-        let client =
-            mgmt_api::Client::new(node.base_url(), self.service_config.jwt_token.as_deref());
-        client
-            .location_config(self.tenant_shard_id, config.clone(), flush_ms)
-            .await?;
-        tracing::info!("location_config({}) complete: {:?}", node_id, config);
-
-        self.observed
-            .locations
-            .insert(node.id, ObservedStateLocation { conf: Some(config) });
+            .insert(node.get_id(), ObservedStateLocation { conf: Some(config) });

        Ok(())
    }

+    fn get_node(&self, node_id: &NodeId) -> Option<&Node> {
+        if let Some(node) = self.intent.attached.as_ref() {
+            if node.get_id() == *node_id {
+                return Some(node);
+            }
+        }
+
+        if let Some(node) = self
+            .intent
+            .secondary
+            .iter()
+            .find(|n| n.get_id() == *node_id)
+        {
+            return Some(node);
+        }
+
+        if let Some(node) = self.detach.iter().find(|n| n.get_id() == *node_id) {
+            return Some(node);
+        }
+
+        None
+    }
+
    async fn maybe_live_migrate(&mut self) -> Result<(), ReconcileError> {
-        let destination = if let Some(node_id) = self.intent.attached {
-            match self.observed.locations.get(&node_id) {
+        let destination = if let Some(node) = &self.intent.attached {
+            match self.observed.locations.get(&node.get_id()) {
                Some(conf) => {
                    // We will do a live migration only if the intended destination is not
                    // currently in an attached state.
                    match &conf.conf {
                        Some(conf) if conf.mode == LocationConfigMode::Secondary => {
                            // Fall through to do a live migration
-                            node_id
+                            node
                        }
                        None | Some(_) => {
                            // Attached or uncertain: don't do a live migration, proceed
@@ -151,7 +198,7 @@ impl Reconciler {
                None => {
                    // Our destination is not attached: maybe live migrate if some other
                    // node is currently attached.  Fall through.
-                    node_id
+                    node
                }
            }
        } else {
@@ -164,15 +211,13 @@ impl Reconciler {
        for (node_id, state) in &self.observed.locations {
            if let Some(observed_conf) = &state.conf {
                if observed_conf.mode == LocationConfigMode::AttachedSingle {
-                    let node = self
-                        .pageservers
-                        .get(node_id)
-                        .expect("Nodes may not be removed while referenced");
                    // We will only attempt live migration if the origin is not offline: this
                    // avoids trying to do it while reconciling after responding to an HA failover.
-                    if !matches!(node.availability, NodeAvailability::Offline) {
-                        origin = Some(*node_id);
-                        break;
+                    if let Some(node) = self.get_node(node_id) {
+                        if node.is_available() {
+                            origin = Some(node.clone());
+                            break;
+                        }
                    }
                }
            }
@@ -185,7 +230,7 @@ impl Reconciler {

        // We have an origin and a destination: proceed to do the live migration
        tracing::info!("Live migrating {}->{}", origin, destination);
-        self.live_migrate(origin, destination).await?;
+        self.live_migrate(origin, destination.clone()).await?;

        Ok(())
    }
@@ -193,13 +238,8 @@ impl Reconciler {
    async fn get_lsns(
        &self,
        tenant_shard_id: TenantShardId,
-        node_id: &NodeId,
+        node: &Node,
    ) -> anyhow::Result<HashMap<TimelineId, Lsn>> {
-        let node = self
-            .pageservers
-            .get(node_id)
-            .expect("Pageserver may not be removed while referenced");
-
        let client =
            mgmt_api::Client::new(node.base_url(), self.service_config.jwt_token.as_deref());

@@ -210,19 +250,27 @@ impl Reconciler {
            .collect())
    }

-    async fn secondary_download(&self, tenant_shard_id: TenantShardId, node_id: &NodeId) {
-        let node = self
-            .pageservers
-            .get(node_id)
-            .expect("Pageserver may not be removed while referenced");
-
-        let client =
-            mgmt_api::Client::new(node.base_url(), self.service_config.jwt_token.as_deref());
-
-        match client.tenant_secondary_download(tenant_shard_id).await {
-            Ok(()) => {}
-            Err(_) => {
-                tracing::info!("  (skipping, destination wasn't in secondary mode)")
+    async fn secondary_download(
+        &self,
+        tenant_shard_id: TenantShardId,
+        node: &Node,
+    ) -> Result<(), ReconcileError> {
+        match node
+            .with_client_retries(
+                |client| async move { client.tenant_secondary_download(tenant_shard_id).await },
+                &self.service_config.jwt_token,
+                1,
+                1,
+                Duration::from_secs(60),
+                &self.cancel,
+            )
+            .await
+        {
+            None => Err(ReconcileError::Cancel),
+            Some(Ok(_)) => Ok(()),
+            Some(Err(e)) => {
+                tracing::info!("  (skipping destination download: {})", e);
+                Ok(())
            }
        }
    }
@@ -230,17 +278,14 @@ impl Reconciler {
    async fn await_lsn(
        &self,
        tenant_shard_id: TenantShardId,
-        pageserver_id: &NodeId,
+        node: &Node,
        baseline: HashMap<TimelineId, Lsn>,
    ) -> anyhow::Result<()> {
        loop {
-            let latest = match self.get_lsns(tenant_shard_id, pageserver_id).await {
+            let latest = match self.get_lsns(tenant_shard_id, node).await {
                Ok(l) => l,
                Err(e) => {
-                    println!(
-                        "🕑 Can't get LSNs on pageserver {} yet, waiting ({e})",
-                        pageserver_id
-                    );
+                    tracing::info!("🕑 Can't get LSNs on node {node} yet, waiting ({e})",);
                    std::thread::sleep(Duration::from_millis(500));
                    continue;
                }
@@ -250,7 +295,7 @@ impl Reconciler {
            for (timeline_id, baseline_lsn) in &baseline {
                match latest.get(timeline_id) {
                    Some(latest_lsn) => {
-                        println!("🕑 LSN origin {baseline_lsn} vs destination {latest_lsn}");
+                        tracing::info!("🕑 LSN origin {baseline_lsn} vs destination {latest_lsn}");
                        if latest_lsn < baseline_lsn {
                            any_behind = true;
                        }
@@ -265,7 +310,7 @@ impl Reconciler {
            }

            if !any_behind {
-                println!("✅ LSN caught up.  Proceeding...");
+                tracing::info!("✅ LSN caught up.  Proceeding...");
                break;
            } else {
                std::thread::sleep(Duration::from_millis(500));
@@ -277,11 +322,11 @@ impl Reconciler {

    pub async fn live_migrate(
        &mut self,
-        origin_ps_id: NodeId,
-        dest_ps_id: NodeId,
-    ) -> anyhow::Result<()> {
+        origin_ps: Node,
+        dest_ps: Node,
+    ) -> Result<(), ReconcileError> {
        // `maybe_live_migrate` is responsibble for sanity of inputs
-        assert!(origin_ps_id != dest_ps_id);
+        assert!(origin_ps.get_id() != dest_ps.get_id());

        fn build_location_config(
            shard: &ShardIdentity,
@@ -301,10 +346,7 @@ impl Reconciler {
            }
        }

-        tracing::info!(
-            "🔁 Switching origin pageserver {} to stale mode",
-            origin_ps_id
-        );
+        tracing::info!("🔁 Switching origin node {origin_ps} to stale mode",);

        // FIXME: it is incorrect to use self.generation here, we should use the generation
        // from the ObservedState of the origin pageserver (it might be older than self.generation)
@@ -315,21 +357,18 @@ impl Reconciler {
            self.generation,
            None,
        );
-        self.location_config(origin_ps_id, stale_conf, Some(Duration::from_secs(10)))
+        self.location_config(&origin_ps, stale_conf, Some(Duration::from_secs(10)), false)
            .await?;

-        let baseline_lsns = Some(self.get_lsns(self.tenant_shard_id, &origin_ps_id).await?);
+        let baseline_lsns = Some(self.get_lsns(self.tenant_shard_id, &origin_ps).await?);

        // If we are migrating to a destination that has a secondary location, warm it up first
-        if let Some(destination_conf) = self.observed.locations.get(&dest_ps_id) {
+        if let Some(destination_conf) = self.observed.locations.get(&dest_ps.get_id()) {
            if let Some(destination_conf) = &destination_conf.conf {
                if destination_conf.mode == LocationConfigMode::Secondary {
-                    tracing::info!(
-                        "🔁 Downloading latest layers to destination pageserver {}",
-                        dest_ps_id,
-                    );
-                    self.secondary_download(self.tenant_shard_id, &dest_ps_id)
-                        .await;
+                    tracing::info!("🔁 Downloading latest layers to destination node {dest_ps}",);
+                    self.secondary_download(self.tenant_shard_id, &dest_ps)
+                        .await?;
                }
            }
        }
@@ -337,7 +376,7 @@ impl Reconciler {
        // Increment generation before attaching to new pageserver
        self.generation = Some(
            self.persistence
-                .increment_generation(self.tenant_shard_id, dest_ps_id)
+                .increment_generation(self.tenant_shard_id, dest_ps.get_id())
                .await?,
        );

@@ -349,22 +388,23 @@ impl Reconciler {
            None,
        );

-        tracing::info!("🔁 Attaching to pageserver {}", dest_ps_id);
-        self.location_config(dest_ps_id, dest_conf, None).await?;
+        tracing::info!("🔁 Attaching to pageserver {dest_ps}");
+        self.location_config(&dest_ps, dest_conf, None, false)
+            .await?;

        if let Some(baseline) = baseline_lsns {
            tracing::info!("🕑 Waiting for LSN to catch up...");
-            self.await_lsn(self.tenant_shard_id, &dest_ps_id, baseline)
+            self.await_lsn(self.tenant_shard_id, &dest_ps, baseline)
                .await?;
        }

-        tracing::info!("🔁 Notifying compute to use pageserver {}", dest_ps_id);
+        tracing::info!("🔁 Notifying compute to use pageserver {dest_ps}");

        // During a live migration it is unhelpful to proceed if we couldn't notify compute: if we detach
        // the origin without notifying compute, we will render the tenant unavailable.
        while let Err(e) = self.compute_notify().await {
            match e {
-                NotifyError::Fatal(_) => return Err(anyhow::anyhow!(e)),
+                NotifyError::Fatal(_) => return Err(ReconcileError::Notify(e)),
                _ => {
                    tracing::warn!(
                        "Live migration blocked by compute notification error, retrying: {e}"
@@ -382,22 +422,19 @@ impl Reconciler {
            None,
            Some(LocationConfigSecondary { warm: true }),
        );
-        self.location_config(origin_ps_id, origin_secondary_conf.clone(), None)
+        self.location_config(&origin_ps, origin_secondary_conf.clone(), None, false)
            .await?;
        // TODO: we should also be setting the ObservedState on earlier API calls, in case we fail
        // partway through.  In fact, all location conf API calls should be in a wrapper that sets
        // the observed state to None, then runs, then sets it to what we wrote.
        self.observed.locations.insert(
-            origin_ps_id,
+            origin_ps.get_id(),
            ObservedStateLocation {
                conf: Some(origin_secondary_conf),
            },
        );

-        println!(
-            "🔁 Switching to AttachedSingle mode on pageserver {}",
-            dest_ps_id
-        );
+        tracing::info!("🔁 Switching to AttachedSingle mode on node {dest_ps}",);
        let dest_final_conf = build_location_config(
            &self.shard,
            &self.config,
@@ -405,16 +442,61 @@ impl Reconciler {
            self.generation,
            None,
        );
-        self.location_config(dest_ps_id, dest_final_conf.clone(), None)
+        self.location_config(&dest_ps, dest_final_conf.clone(), None, false)
            .await?;
        self.observed.locations.insert(
-            dest_ps_id,
+            dest_ps.get_id(),
            ObservedStateLocation {
                conf: Some(dest_final_conf),
            },
        );

-        println!("✅ Migration complete");
+        tracing::info!("✅ Migration complete");
+
+        Ok(())
+    }
+
+    async fn maybe_refresh_observed(&mut self) -> Result<(), ReconcileError> {
+        // If the attached node has uncertain state, read it from the pageserver before proceeding: this
+        // is important to avoid spurious generation increments.
+        //
+        // We don't need to do this for secondary/detach locations because it's harmless to just PUT their
+        // location conf, whereas for attached locations it can interrupt clients if we spuriously destroy/recreate
+        // the `Timeline` object in the pageserver.
+
+        let Some(attached_node) = self.intent.attached.as_ref() else {
+            // Nothing to do
+            return Ok(());
+        };
+
+        if matches!(
+            self.observed.locations.get(&attached_node.get_id()),
+            Some(ObservedStateLocation { conf: None })
+        ) {
+            let tenant_shard_id = self.tenant_shard_id;
+            let observed_conf = match attached_node
+                .with_client_retries(
+                    |client| async move { client.get_location_config(tenant_shard_id).await },
+                    &self.service_config.jwt_token,
+                    1,
+                    1,
+                    Duration::from_secs(5),
+                    &self.cancel,
+                )
+                .await
+            {
+                Some(Ok(observed)) => observed,
+                Some(Err(e)) => return Err(e.into()),
+                None => return Err(ReconcileError::Cancel),
+            };
+            tracing::info!("Scanned location configuration on {attached_node}: {observed_conf:?}");
+            self.observed.locations.insert(
+                attached_node.get_id(),
+                ObservedStateLocation {
+                    conf: observed_conf,
+                },
+            );
+        }

        Ok(())
    }
@@ -426,14 +508,14 @@ impl Reconciler {
    /// general case reconciliation where we walk through the intent by pageserver
    /// and call out to the pageserver to apply the desired state.
    pub(crate) async fn reconcile(&mut self) -> Result<(), ReconcileError> {
-        // TODO: if any of self.observed is None, call to remote pageservers
-        // to learn correct state.
+        // Prepare: if we have uncertain `observed` state for our would-be attachement location, then refresh it
+        self.maybe_refresh_observed().await?;

        // Special case: live migration
        self.maybe_live_migrate().await?;

        // If the attached pageserver is not attached, do so now.
-        if let Some(node_id) = self.intent.attached {
+        if let Some(node) = self.intent.attached.as_ref() {
            // If we are in an attached policy, then generation must have been set (null generations
            // are only present when a tenant is initially loaded with a secondary policy)
            debug_assert!(self.generation.is_some());
@@ -444,10 +526,10 @@ impl Reconciler {
            };

            let mut wanted_conf = attached_location_conf(generation, &self.shard, &self.config);
-            match self.observed.locations.get(&node_id) {
+            match self.observed.locations.get(&node.get_id()) {
                Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
                    // Nothing to do
-                    tracing::info!(%node_id, "Observed configuration already correct.")
+                    tracing::info!(node_id=%node.get_id(), "Observed configuration already correct.")
                }
                observed => {
                    // In all cases other than a matching observed configuration, we will
@@ -485,13 +567,21 @@ impl Reconciler {
                    if increment_generation {
                        let generation = self
                            .persistence
-                            .increment_generation(self.tenant_shard_id, node_id)
+                            .increment_generation(self.tenant_shard_id, node.get_id())
                            .await?;
                        self.generation = Some(generation);
                        wanted_conf.generation = generation.into();
                    }
-                    tracing::info!(%node_id, "Observed configuration requires update.");
-                    self.location_config(node_id, wanted_conf, None).await?;
+                    tracing::info!(node_id=%node.get_id(), "Observed configuration requires update.");
+
+                    // Because `node` comes from a ref to &self, clone it before calling into a &mut self
+                    // function: this could be avoided by refactoring the state mutated by location_config into
+                    // a separate type to Self.
+                    let node = node.clone();
+
+                    // Use lazy=true, because we may run many of Self concurrently, and do not want to
+                    // overload the pageserver with logical size calculations.
+                    self.location_config(&node, wanted_conf, None, true).await?;
                    self.compute_notify().await?;
                }
            }
@@ -500,33 +590,27 @@ impl Reconciler {
        // Configure secondary locations: if these were previously attached this
        // implicitly downgrades them from attached to secondary.
        let mut changes = Vec::new();
-        for node_id in &self.intent.secondary {
+        for node in &self.intent.secondary {
            let wanted_conf = secondary_location_conf(&self.shard, &self.config);
-            match self.observed.locations.get(node_id) {
+            match self.observed.locations.get(&node.get_id()) {
                Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
                    // Nothing to do
-                    tracing::info!(%node_id, "Observed configuration already correct.")
+                    tracing::info!(node_id=%node.get_id(), "Observed configuration already correct.")
                }
                _ => {
                    // In all cases other than a matching observed configuration, we will
                    // reconcile this location.
-                    tracing::info!(%node_id, "Observed configuration requires update.");
-                    changes.push((*node_id, wanted_conf))
+                    tracing::info!(node_id=%node.get_id(), "Observed configuration requires update.");
+                    changes.push((node.clone(), wanted_conf))
                }
            }
        }

        // Detach any extraneous pageservers that are no longer referenced
        // by our intent.
-        let all_pageservers = self.intent.all_pageservers();
-        for node_id in self.observed.locations.keys() {
-            if all_pageservers.contains(node_id) {
-                // We are only detaching pageservers that aren't used at all.
-                continue;
-            }
-
+        for node in &self.detach {
            changes.push((
-                *node_id,
+                node.clone(),
                LocationConfig {
                    mode: LocationConfigMode::Detached,
                    generation: None,
@@ -539,11 +623,11 @@ impl Reconciler {
            ));
        }

-        for (node_id, conf) in changes {
+        for (node, conf) in changes {
            if self.cancel.is_cancelled() {
                return Err(ReconcileError::Cancel);
            }
-            self.location_config(node_id, conf, None).await?;
+            self.location_config(&node, conf, None, false).await?;
        }

        Ok(())
@@ -552,16 +636,21 @@ impl Reconciler {
    pub(crate) async fn compute_notify(&mut self) -> Result<(), NotifyError> {
        // Whenever a particular Reconciler emits a notification, it is always notifying for the intended
        // destination.
-        if let Some(node_id) = self.intent.attached {
+        if let Some(node) = &self.intent.attached {
            let result = self
                .compute_hook
-                .notify(self.tenant_shard_id, node_id, &self.cancel)
+                .notify(
+                    self.tenant_shard_id,
+                    node.get_id(),
+                    self.shard.stripe_size,
+                    &self.cancel,
+                )
                .await;
            if let Err(e) = &result {
                // It is up to the caller whether they want to drop out on this error, but they don't have to:
                // in general we should avoid letting unavailability of the cloud control plane stop us from
                // making progress.
-                tracing::warn!("Failed to notify compute of attached pageserver {node_id}: {e}");
+                tracing::warn!("Failed to notify compute of attached pageserver {node}: {e}");
                // Set this flag so that in our ReconcileResult we will set the flag on the shard that it
                // needs to retry at some point.
                self.compute_notify_failure = true;
--- a/control_plane/attachment_service/src/scheduler.rs
+++ b/control_plane/attachment_service/src/scheduler.rs
@@ -43,7 +43,7 @@ impl Scheduler {
        let mut scheduler_nodes = HashMap::new();
        for node in nodes {
            scheduler_nodes.insert(
-                node.id,
+                node.get_id(),
                SchedulerNode {
                    shard_count: 0,
                    may_schedule: node.may_schedule(),
@@ -68,7 +68,7 @@ impl Scheduler {
        let mut expect_nodes: HashMap<NodeId, SchedulerNode> = HashMap::new();
        for node in nodes {
            expect_nodes.insert(
-                node.id,
+                node.get_id(),
                SchedulerNode {
                    shard_count: 0,
                    may_schedule: node.may_schedule(),
@@ -156,7 +156,7 @@ impl Scheduler {

    pub(crate) fn node_upsert(&mut self, node: &Node) {
        use std::collections::hash_map::Entry::*;
-        match self.nodes.entry(node.id) {
+        match self.nodes.entry(node.get_id()) {
            Occupied(mut entry) => {
                entry.get_mut().may_schedule = node.may_schedule();
            }
@@ -255,7 +255,6 @@ impl Scheduler {
 pub(crate) mod test_utils {

    use crate::node::Node;
-    use pageserver_api::controller_api::{NodeAvailability, NodeSchedulingPolicy};
    use std::collections::HashMap;
    use utils::id::NodeId;
    /// Test helper: synthesize the requested number of nodes, all in active state.
@@ -264,18 +263,17 @@ pub(crate) mod test_utils {
    pub(crate) fn make_test_nodes(n: u64) -> HashMap<NodeId, Node> {
        (1..n + 1)
            .map(|i| {
-                (
-                    NodeId(i),
-                    Node {
-                        id: NodeId(i),
-                        availability: NodeAvailability::Active,
-                        scheduling: NodeSchedulingPolicy::Active,
-                        listen_http_addr: format!("httphost-{i}"),
-                        listen_http_port: 80 + i as u16,
-                        listen_pg_addr: format!("pghost-{i}"),
-                        listen_pg_port: 5432 + i as u16,
-                    },
-                )
+                (NodeId(i), {
+                    let node = Node::new(
+                        NodeId(i),
+                        format!("httphost-{i}"),
+                        80 + i as u16,
+                        format!("pghost-{i}"),
+                        5432 + i as u16,
+                    );
+                    assert!(node.is_available());
+                    node
+                })
            })
            .collect()
    }
--- a/control_plane/attachment_service/src/service.rs
+++ b/control_plane/attachment_service/src/service.rs
--- a/control_plane/attachment_service/src/tenant_state.rs
+++ b/control_plane/attachment_service/src/tenant_state.rs
@@ -1,7 +1,11 @@
-use std::{collections::HashMap, sync::Arc, time::Duration};
+use std::{
+    collections::{HashMap, HashSet},
+    sync::Arc,
+    time::Duration,
+};

 use crate::{metrics, persistence::TenantShardPersistence};
-use pageserver_api::controller_api::NodeAvailability;
+use pageserver_api::controller_api::PlacementPolicy;
 use pageserver_api::{
    models::{LocationConfig, LocationConfigMode, TenantConfig},
    shard::{ShardIdentity, TenantShardId},
@@ -25,7 +29,7 @@ use crate::{
        attached_location_conf, secondary_location_conf, ReconcileError, Reconciler, TargetState,
    },
    scheduler::{ScheduleError, Scheduler},
-    service, PlacementPolicy, Sequence,
+    service, Sequence,
 };

 /// Serialization helper
@@ -370,7 +374,7 @@ impl TenantState {
    /// [`ObservedState`], even if it violates my [`PlacementPolicy`].  Call [`Self::schedule`] next,
    /// to get an intent state that complies with placement policy.  The overall goal is to do scheduling
    /// in a way that makes use of any configured locations that already exist in the outside world.
-    pub(crate) fn intent_from_observed(&mut self) {
+    pub(crate) fn intent_from_observed(&mut self, scheduler: &mut Scheduler) {
        // Choose an attached location by filtering observed locations, and then sorting to get the highest
        // generation
        let mut attached_locs = self
@@ -395,7 +399,7 @@ impl TenantState {

        attached_locs.sort_by_key(|i| i.1);
        if let Some((node_id, _gen)) = attached_locs.into_iter().last() {
-            self.intent.attached = Some(*node_id);
+            self.intent.set_attached(scheduler, Some(*node_id));
        }

        // All remaining observed locations generate secondary intents.  This includes None
@@ -406,7 +410,7 @@ impl TenantState {
        // will take care of promoting one of these secondaries to be attached.
        self.observed.locations.keys().for_each(|node_id| {
            if Some(*node_id) != self.intent.attached {
-                self.intent.secondary.push(*node_id);
+                self.intent.push_secondary(scheduler, *node_id);
            }
        });
    }
@@ -564,7 +568,9 @@ impl TenantState {
        }
    }

-    fn dirty(&self) -> bool {
+    fn dirty(&self, nodes: &Arc<HashMap<NodeId, Node>>) -> bool {
+        let mut dirty_nodes = HashSet::new();
+
        if let Some(node_id) = self.intent.attached {
            // Maybe panic: it is a severe bug if we try to attach while generation is null.
            let generation = self
@@ -575,7 +581,7 @@ impl TenantState {
            match self.observed.locations.get(&node_id) {
                Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
                Some(_) | None => {
-                    return true;
+                    dirty_nodes.insert(node_id);
                }
            }
        }
@@ -585,7 +591,7 @@ impl TenantState {
            match self.observed.locations.get(node_id) {
                Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
                Some(_) | None => {
-                    return true;
+                    dirty_nodes.insert(*node_id);
                }
            }
        }
@@ -593,24 +599,25 @@ impl TenantState {
        for node_id in self.observed.locations.keys() {
            if self.intent.attached != Some(*node_id) && !self.intent.secondary.contains(node_id) {
                // We have observed state that isn't part of our intent: need to clean it up.
-                return true;
+                dirty_nodes.insert(*node_id);
            }
        }

-        // Even if there is no pageserver work to be done, if we have a pending notification to computes,
-        // wake up a reconciler to send it.
-        if self.pending_compute_notification {
-            return true;
-        }
+        dirty_nodes.retain(|node_id| {
+            nodes
+                .get(node_id)
+                .map(|n| n.is_available())
+                .unwrap_or(false)
+        });

-        false
+        !dirty_nodes.is_empty()
    }

    #[allow(clippy::too_many_arguments)]
    #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
    pub(crate) fn maybe_reconcile(
        &mut self,
-        result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResult>,
+        result_tx: &tokio::sync::mpsc::UnboundedSender<ReconcileResult>,
        pageservers: &Arc<HashMap<NodeId, Node>>,
        compute_hook: &Arc<ComputeHook>,
        service_config: &service::Config,
@@ -625,15 +632,20 @@ impl TenantState {
            let node = pageservers
                .get(node_id)
                .expect("Nodes may not be removed while referenced");
-            if observed_loc.conf.is_none()
-                && !matches!(node.availability, NodeAvailability::Offline)
-            {
+            if observed_loc.conf.is_none() && node.is_available() {
                dirty_observed = true;
                break;
            }
        }

-        if !self.dirty() && !dirty_observed {
+        let active_nodes_dirty = self.dirty(pageservers);
+
+        // Even if there is no pageserver work to be done, if we have a pending notification to computes,
+        // wake up a reconciler to send it.
+        let do_reconcile =
+            active_nodes_dirty || dirty_observed || self.pending_compute_notification;
+
+        if !do_reconcile {
            tracing::info!("Not dirty, no reconciliation needed.");
            return None;
        }
@@ -663,6 +675,21 @@ impl TenantState {
            }
        }

+        // Build list of nodes from which the reconciler should detach
+        let mut detach = Vec::new();
+        for node_id in self.observed.locations.keys() {
+            if self.intent.get_attached() != &Some(*node_id)
+                && !self.intent.secondary.contains(node_id)
+            {
+                detach.push(
+                    pageservers
+                        .get(node_id)
+                        .expect("Intent references non-existent pageserver")
+                        .clone(),
+                )
+            }
+        }
+
        // Reconcile in flight for a stale sequence?  Our sequence's task will wait for it before
        // doing our sequence's work.
        let old_handle = self.reconciler.take();
@@ -677,14 +704,15 @@ impl TenantState {
        self.sequence = self.sequence.next();

        let reconciler_cancel = cancel.child_token();
+        let reconciler_intent = TargetState::from_intent(pageservers, &self.intent);
        let mut reconciler = Reconciler {
            tenant_shard_id: self.tenant_shard_id,
            shard: self.shard,
            generation: self.generation,
-            intent: TargetState::from_intent(&self.intent),
+            intent: reconciler_intent,
+            detach,
            config: self.config.clone(),
            observed: self.observed.clone(),
-            pageservers: pageservers.clone(),
            compute_hook: compute_hook.clone(),
            service_config: service_config.clone(),
            _gate_guard: gate_guard,
@@ -701,6 +729,7 @@ impl TenantState {
                                                        tenant_id=%reconciler.tenant_shard_id.tenant_id,
                                                        shard_id=%reconciler.tenant_shard_id.shard_slug());
        metrics::RECONCILER.spawned.inc();
+        let result_tx = result_tx.clone();
        let join_handle = tokio::task::spawn(
            async move {
                // Wait for any previous reconcile task to complete before we start
@@ -819,7 +848,10 @@ impl TenantState {

 #[cfg(test)]
 pub(crate) mod tests {
-    use pageserver_api::shard::{ShardCount, ShardNumber};
+    use pageserver_api::{
+        controller_api::NodeAvailability,
+        shard::{ShardCount, ShardNumber},
+    };
    use utils::id::TenantId;

    use crate::scheduler::test_utils::make_test_nodes;
@@ -878,7 +910,10 @@ pub(crate) mod tests {
        assert_eq!(tenant_state.intent.secondary.len(), 2);

        // Update the scheduler state to indicate the node is offline
-        nodes.get_mut(&attached_node_id).unwrap().availability = NodeAvailability::Offline;
+        nodes
+            .get_mut(&attached_node_id)
+            .unwrap()
+            .set_availability(NodeAvailability::Offline);
        scheduler.node_upsert(nodes.get(&attached_node_id).unwrap());

        // Scheduling the node should promote the still-available secondary node to attached
@@ -897,4 +932,54 @@ pub(crate) mod tests {

        Ok(())
    }
+
+    #[test]
+    fn intent_from_observed() -> anyhow::Result<()> {
+        let nodes = make_test_nodes(3);
+        let mut scheduler = Scheduler::new(nodes.values());
+
+        let mut tenant_state = make_test_tenant_shard(PlacementPolicy::Double(1));
+
+        tenant_state.observed.locations.insert(
+            NodeId(3),
+            ObservedStateLocation {
+                conf: Some(LocationConfig {
+                    mode: LocationConfigMode::AttachedMulti,
+                    generation: Some(2),
+                    secondary_conf: None,
+                    shard_number: tenant_state.shard.number.0,
+                    shard_count: tenant_state.shard.count.literal(),
+                    shard_stripe_size: tenant_state.shard.stripe_size.0,
+                    tenant_conf: TenantConfig::default(),
+                }),
+            },
+        );
+
+        tenant_state.observed.locations.insert(
+            NodeId(2),
+            ObservedStateLocation {
+                conf: Some(LocationConfig {
+                    mode: LocationConfigMode::AttachedStale,
+                    generation: Some(1),
+                    secondary_conf: None,
+                    shard_number: tenant_state.shard.number.0,
+                    shard_count: tenant_state.shard.count.literal(),
+                    shard_stripe_size: tenant_state.shard.stripe_size.0,
+                    tenant_conf: TenantConfig::default(),
+                }),
+            },
+        );
+
+        tenant_state.intent_from_observed(&mut scheduler);
+
+        // The highest generationed attached location gets used as attached
+        assert_eq!(tenant_state.intent.attached, Some(NodeId(3)));
+        // Other locations get used as secondary
+        assert_eq!(tenant_state.intent.secondary, vec![NodeId(2)]);
+
+        scheduler.consistency_check(nodes.values(), [&tenant_state].into_iter())?;
+
+        tenant_state.intent.clear(&mut scheduler);
+        Ok(())
+    }
 }
--- a/control_plane/src/attachment_service.rs
+++ b/control_plane/src/attachment_service.rs
@@ -34,7 +34,7 @@ pub struct AttachmentService {
    client: reqwest::Client,
 }

-const COMMAND: &str = "attachment_service";
+const COMMAND: &str = "storage_controller";

 const ATTACHMENT_SERVICE_POSTGRES_VERSION: u32 = 16;

--- a/control_plane/src/bin/neon_local.rs
+++ b/control_plane/src/bin/neon_local.rs
@@ -15,7 +15,7 @@ use control_plane::pageserver::{PageServerNode, PAGESERVER_REMOTE_STORAGE_DIR};
 use control_plane::safekeeper::SafekeeperNode;
 use control_plane::{broker, local_env};
 use pageserver_api::controller_api::{
-    NodeAvailability, NodeConfigureRequest, NodeSchedulingPolicy,
+    NodeAvailability, NodeConfigureRequest, NodeSchedulingPolicy, PlacementPolicy,
 };
 use pageserver_api::models::{
    ShardParameters, TenantCreateRequest, TimelineCreateRequest, TimelineInfo,
@@ -435,6 +435,11 @@ async fn handle_tenant(
            let shard_stripe_size: Option<u32> =
                create_match.get_one::<u32>("shard-stripe-size").cloned();

+            let placement_policy = match create_match.get_one::<String>("placement-policy") {
+                Some(s) if !s.is_empty() => serde_json::from_str::<PlacementPolicy>(s)?,
+                _ => PlacementPolicy::Single,
+            };
+
            let tenant_conf = PageServerNode::parse_config(tenant_conf)?;

            // If tenant ID was not specified, generate one
@@ -456,6 +461,7 @@ async fn handle_tenant(
                            .map(ShardStripeSize)
                            .unwrap_or(ShardParameters::DEFAULT_STRIPE_SIZE),
                    },
+                    placement_policy: Some(placement_policy),
                    config: tenant_conf,
                })
                .await?;
@@ -1024,7 +1030,7 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                        })
                        .collect::<Vec<_>>()
                };
-            endpoint.reconfigure(pageservers).await?;
+            endpoint.reconfigure(pageservers, None).await?;
        }
        "stop" => {
            let endpoint_id = sub_args
@@ -1562,6 +1568,7 @@ fn cli() -> Command {
                    .help("Use this tenant in future CLI commands where tenant_id is needed, but not specified"))
                .arg(Arg::new("shard-count").value_parser(value_parser!(u8)).long("shard-count").action(ArgAction::Set).help("Number of shards in the new tenant (default 1)"))
                .arg(Arg::new("shard-stripe-size").value_parser(value_parser!(u32)).long("shard-stripe-size").action(ArgAction::Set).help("Sharding stripe size in pages"))
+                .arg(Arg::new("placement-policy").value_parser(value_parser!(String)).long("placement-policy").action(ArgAction::Set).help("Placement policy shards in this tenant"))
                )
            .subcommand(Command::new("set-default").arg(tenant_id_arg.clone().required(true))
                .about("Set a particular tenant as default in future CLI commands where tenant_id is needed, but not specified"))
--- a/control_plane/src/endpoint.rs
+++ b/control_plane/src/endpoint.rs
@@ -52,6 +52,7 @@ use compute_api::spec::RemoteExtSpec;
 use compute_api::spec::Role;
 use nix::sys::signal::kill;
 use nix::sys::signal::Signal;
+use pageserver_api::shard::ShardStripeSize;
 use serde::{Deserialize, Serialize};
 use url::Host;
 use utils::id::{NodeId, TenantId, TimelineId};
@@ -655,7 +656,7 @@ impl Endpoint {
        // Wait for it to start
        let mut attempt = 0;
        const ATTEMPT_INTERVAL: Duration = Duration::from_millis(100);
-        const MAX_ATTEMPTS: u32 = 10 * 30; // Wait up to 30 s
+        const MAX_ATTEMPTS: u32 = 10 * 90; // Wait up to 1.5 min
        loop {
            attempt += 1;
            match self.get_status().await {
@@ -735,7 +736,11 @@ impl Endpoint {
        }
    }

-    pub async fn reconfigure(&self, mut pageservers: Vec<(Host, u16)>) -> Result<()> {
+    pub async fn reconfigure(
+        &self,
+        mut pageservers: Vec<(Host, u16)>,
+        stripe_size: Option<ShardStripeSize>,
+    ) -> Result<()> {
        let mut spec: ComputeSpec = {
            let spec_path = self.endpoint_path().join("spec.json");
            let file = std::fs::File::open(spec_path)?;
@@ -765,6 +770,9 @@ impl Endpoint {
        let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
        assert!(!pageserver_connstr.is_empty());
        spec.pageserver_connstring = Some(pageserver_connstr);
+        if stripe_size.is_some() {
+            spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
+        }

        let client = reqwest::Client::new();
        let response = client
--- a/control_plane/src/local_env.rs
+++ b/control_plane/src/local_env.rs
@@ -232,7 +232,7 @@ impl LocalEnv {
        // run from the same location as neon_local.  This means that for compatibility
        // tests that run old pageserver/safekeeper, they still run latest attachment service.
        let neon_local_bin_dir = env::current_exe().unwrap().parent().unwrap().to_owned();
-        neon_local_bin_dir.join("attachment_service")
+        neon_local_bin_dir.join("storage_controller")
    }

    pub fn safekeeper_bin(&self) -> PathBuf {
--- a/control_plane/src/pageserver.rs
+++ b/control_plane/src/pageserver.rs
@@ -429,6 +429,8 @@ impl PageServerNode {
            generation,
            config,
            shard_parameters: ShardParameters::default(),
+            // Placement policy is not meaningful for creations not done via storage controller
+            placement_policy: None,
        };
        if !settings.is_empty() {
            bail!("Unrecognized tenant settings: {settings:?}")
@@ -537,10 +539,11 @@ impl PageServerNode {
        tenant_shard_id: TenantShardId,
        config: LocationConfig,
        flush_ms: Option<Duration>,
+        lazy: bool,
    ) -> anyhow::Result<()> {
        Ok(self
            .http_client
-            .location_config(tenant_shard_id, config, flush_ms)
+            .location_config(tenant_shard_id, config, flush_ms, lazy)
            .await?)
    }

@@ -605,7 +608,7 @@ impl PageServerNode {
                eprintln!("connection error: {}", e);
            }
        });
-        tokio::pin!(client);
+        let client = std::pin::pin!(client);

        // Init base reader
        let (start_lsn, base_tarfile_path) = base;
--- a/libs/metrics/src/lib.rs
+++ b/libs/metrics/src/lib.rs
@@ -29,7 +29,6 @@ pub mod launch_timestamp;
 mod wrappers;
 pub use wrappers::{CountedReader, CountedWriter};
 mod hll;
-pub mod metric_vec_duration;
 pub use hll::{HyperLogLog, HyperLogLogVec};
 #[cfg(target_os = "linux")]
 pub mod more_process_metrics;
--- a/libs/metrics/src/metric_vec_duration.rs
+++ b/libs/metrics/src/metric_vec_duration.rs
@@ -1,23 +0,0 @@
-//! Helpers for observing duration on `HistogramVec` / `CounterVec` / `GaugeVec` / `MetricVec<T>`.
-
-use std::{future::Future, time::Instant};
-
-pub trait DurationResultObserver {
-    fn observe_result<T, E>(&self, res: &Result<T, E>, duration: std::time::Duration);
-}
-
-pub async fn observe_async_block_duration_by_result<
-    T,
-    E,
-    F: Future<Output = Result<T, E>>,
-    O: DurationResultObserver,
->(
-    observer: &O,
-    block: F,
-) -> Result<T, E> {
-    let start = Instant::now();
-    let result = block.await;
-    let duration = start.elapsed();
-    observer.observe_result(&result, duration);
-    result
-}
--- a/libs/pageserver_api/src/controller_api.rs
+++ b/libs/pageserver_api/src/controller_api.rs
@@ -125,5 +125,45 @@ impl From<NodeSchedulingPolicy> for String {
    }
 }

+/// Controls how tenant shards are mapped to locations on pageservers, e.g. whether
+/// to create secondary locations.
+#[derive(Clone, Serialize, Deserialize, Debug, PartialEq, Eq)]
+pub enum PlacementPolicy {
+    /// Cheapest way to attach a tenant: just one pageserver, no secondary
+    Single,
+    /// Production-ready way to attach a tenant: one attached pageserver and
+    /// some number of secondaries.
+    Double(usize),
+    /// Create one secondary mode locations. This is useful when onboarding
+    /// a tenant, or for an idle tenant that we might want to bring online quickly.
+    Secondary,
+
+    /// Do not attach to any pageservers.  This is appropriate for tenants that
+    /// have been idle for a long time, where we do not mind some delay in making
+    /// them available in future.
+    Detached,
+}
+
 #[derive(Serialize, Deserialize, Debug)]
 pub struct TenantShardMigrateResponse {}
+
+#[cfg(test)]
+mod test {
+    use super::*;
+    use serde_json;
+
+    /// Check stability of PlacementPolicy's serialization
+    #[test]
+    fn placement_policy_encoding() -> anyhow::Result<()> {
+        let v = PlacementPolicy::Double(1);
+        let encoded = serde_json::to_string(&v)?;
+        assert_eq!(encoded, "{\"Double\":1}");
+        assert_eq!(serde_json::from_str::<PlacementPolicy>(&encoded)?, v);
+
+        let v = PlacementPolicy::Single;
+        let encoded = serde_json::to_string(&v)?;
+        assert_eq!(encoded, "\"Single\"");
+        assert_eq!(serde_json::from_str::<PlacementPolicy>(&encoded)?, v);
+        Ok(())
+    }
+}
--- a/libs/pageserver_api/src/models.rs
+++ b/libs/pageserver_api/src/models.rs
@@ -21,6 +21,7 @@ use utils::{
    lsn::Lsn,
 };

+use crate::controller_api::PlacementPolicy;
 use crate::{
    reltag::RelTag,
    shard::{ShardCount, ShardStripeSize, TenantShardId},
@@ -242,6 +243,11 @@ pub struct TenantCreateRequest {
    #[serde(skip_serializing_if = "ShardParameters::is_unsharded")]
    pub shard_parameters: ShardParameters,

+    // This parameter is only meaningful in requests sent to the storage controller
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub placement_policy: Option<PlacementPolicy>,
+
    #[serde(flatten)]
    pub config: TenantConfig, // as we have a flattened field, we should reject all unknown fields in it
 }
@@ -435,6 +441,8 @@ pub struct TenantShardLocation {
 #[serde(deny_unknown_fields)]
 pub struct TenantLocationConfigResponse {
    pub shards: Vec<TenantShardLocation>,
+    // If the shards' ShardCount count is >1, stripe_size will be set.
+    pub stripe_size: Option<ShardStripeSize>,
 }

 #[derive(Serialize, Deserialize, Debug)]
--- a/libs/postgres_backend/src/lib.rs
+++ b/libs/postgres_backend/src/lib.rs
@@ -6,7 +6,6 @@
 #![deny(clippy::undocumented_unsafe_blocks)]
 use anyhow::Context;
 use bytes::Bytes;
-use futures::pin_mut;
 use serde::{Deserialize, Serialize};
 use std::io::ErrorKind;
 use std::net::SocketAddr;
@@ -378,8 +377,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
        &mut self,
        cx: &mut std::task::Context<'_>,
    ) -> Poll<Result<(), std::io::Error>> {
-        let flush_fut = self.flush();
-        pin_mut!(flush_fut);
+        let flush_fut = std::pin::pin!(self.flush());
        flush_fut.poll(cx)
    }

--- a/libs/postgres_backend/tests/simple_select.rs
+++ b/libs/postgres_backend/tests/simple_select.rs
@@ -72,14 +72,19 @@ async fn simple_select() {
    }
 }

-static KEY: Lazy<rustls::PrivateKey> = Lazy::new(|| {
+static KEY: Lazy<rustls::pki_types::PrivateKeyDer<'static>> = Lazy::new(|| {
    let mut cursor = Cursor::new(include_bytes!("key.pem"));
-    rustls::PrivateKey(rustls_pemfile::rsa_private_keys(&mut cursor).unwrap()[0].clone())
+    let key = rustls_pemfile::rsa_private_keys(&mut cursor)
+        .next()
+        .unwrap()
+        .unwrap();
+    rustls::pki_types::PrivateKeyDer::Pkcs1(key)
 });

-static CERT: Lazy<rustls::Certificate> = Lazy::new(|| {
+static CERT: Lazy<rustls::pki_types::CertificateDer<'static>> = Lazy::new(|| {
    let mut cursor = Cursor::new(include_bytes!("cert.pem"));
-    rustls::Certificate(rustls_pemfile::certs(&mut cursor).unwrap()[0].clone())
+    let cert = rustls_pemfile::certs(&mut cursor).next().unwrap().unwrap();
+    cert
 });

 // test that basic select with ssl works
@@ -88,9 +93,8 @@ async fn simple_select_ssl() {
    let (client_sock, server_sock) = make_tcp_pair().await;

    let server_cfg = rustls::ServerConfig::builder()
-        .with_safe_defaults()
        .with_no_client_auth()
-        .with_single_cert(vec![CERT.clone()], KEY.clone())
+        .with_single_cert(vec![CERT.clone()], KEY.clone_key())
        .unwrap();
    let tls_config = Some(Arc::new(server_cfg));
    let pgbackend =
@@ -102,10 +106,9 @@ async fn simple_select_ssl() {
    });

    let client_cfg = rustls::ClientConfig::builder()
-        .with_safe_defaults()
        .with_root_certificates({
            let mut store = rustls::RootCertStore::empty();
-            store.add(&CERT).unwrap();
+            store.add(CERT.clone()).unwrap();
            store
        })
        .with_no_client_auth();
--- a/libs/remote_storage/tests/test_real_s3.rs
+++ b/libs/remote_storage/tests/test_real_s3.rs
@@ -17,6 +17,7 @@ use remote_storage::{
 };
 use test_context::test_context;
 use test_context::AsyncTestContext;
+use tokio::io::AsyncBufReadExt;
 use tokio_util::sync::CancellationToken;
 use tracing::info;

@@ -484,32 +485,33 @@ async fn download_is_cancelled(ctx: &mut MaybeEnabledStorage) {
    ))
    .unwrap();

-    let len = upload_large_enough_file(&ctx.client, &path, &cancel).await;
+    let file_len = upload_large_enough_file(&ctx.client, &path, &cancel).await;

    {
-        let mut stream = ctx
+        let stream = ctx
            .client
            .download(&path, &cancel)
            .await
            .expect("download succeeds")
            .download_stream;

-        let first = stream
-            .next()
-            .await
-            .expect("should have the first blob")
-            .expect("should have succeeded");
+        let mut reader = std::pin::pin!(tokio_util::io::StreamReader::new(stream));

-        tracing::info!(len = first.len(), "downloaded first chunk");
+        let first = reader.fill_buf().await.expect("should have the first blob");
+
+        let len = first.len();
+        tracing::info!(len, "downloaded first chunk");

        assert!(
-            first.len() < len,
+            first.len() < file_len,
            "uploaded file is too small, we downloaded all on first chunk"
        );

+        reader.consume(len);
+
        cancel.cancel();

-        let next = stream.next().await.expect("stream should have more");
+        let next = reader.fill_buf().await;

        let e = next.expect_err("expected an error, but got a chunk?");

@@ -520,6 +522,10 @@ async fn download_is_cancelled(ctx: &mut MaybeEnabledStorage) {
                .is_some_and(|e| matches!(e, DownloadError::Cancelled)),
            "{inner:?}"
        );
+
+        let e = DownloadError::from(e);
+
+        assert!(matches!(e, DownloadError::Cancelled), "{e:?}");
    }

    let cancel = CancellationToken::new();
--- a/libs/utils/src/http/endpoint.rs
+++ b/libs/utils/src/http/endpoint.rs
@@ -84,6 +84,9 @@ where
            info!("Handling request");
        }

+        // Take a copy of the path for error logging
+        let path = request.uri().path().to_string();
+
        // No special handling for panics here. There's a `tracing_panic_hook` from another
        // module to do that globally.
        let res = handler(request).await;
@@ -110,7 +113,7 @@ where
                }
                Ok(response)
            }
-            Err(err) => Ok(api_error_handler(err)),
+            Err(err) => Ok(api_error_handler(err, Some(&path))),
        }
    }
    .instrument(request_span)
--- a/libs/utils/src/http/error.rs
+++ b/libs/utils/src/http/error.rs
@@ -108,7 +108,7 @@ impl HttpErrorBody {

 pub async fn route_error_handler(err: routerify::RouteError) -> Response<Body> {
    match err.downcast::<ApiError>() {
-        Ok(api_error) => api_error_handler(*api_error),
+        Ok(api_error) => api_error_handler(*api_error, None),
        Err(other_error) => {
            // We expect all the request handlers to return an ApiError, so this should
            // not be reached. But just in case.
@@ -121,12 +121,16 @@ pub async fn route_error_handler(err: routerify::RouteError) -> Response<Body> {
    }
 }

-pub fn api_error_handler(api_error: ApiError) -> Response<Body> {
+pub fn api_error_handler(api_error: ApiError, path: Option<&str>) -> Response<Body> {
    // Print a stack trace for Internal Server errors

    match api_error {
        ApiError::Forbidden(_) | ApiError::Unauthorized(_) => {
-            warn!("Error processing HTTP request: {api_error:#}")
+            warn!(
+                "Error processing HTTP request: {api_error:#} {}{}",
+                path.as_ref().map(|_| "at").unwrap_or(""),
+                path.unwrap_or("")
+            )
        }
        ApiError::ResourceUnavailable(_) => info!("Error processing HTTP request: {api_error:#}"),
        ApiError::NotFound(_) => info!("Error processing HTTP request: {api_error:#}"),
--- a/libs/walproposer/src/api_bindings.rs
+++ b/libs/walproposer/src/api_bindings.rs
@@ -324,11 +324,11 @@ extern "C" fn finish_sync_safekeepers(wp: *mut WalProposer, lsn: XLogRecPtr) {
    }
 }

-extern "C" fn process_safekeeper_feedback(wp: *mut WalProposer, commit_lsn: XLogRecPtr) {
+extern "C" fn process_safekeeper_feedback(wp: *mut WalProposer) {
    unsafe {
        let callback_data = (*(*wp).config).callback_data;
        let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).process_safekeeper_feedback(&mut (*wp), commit_lsn)
+        (*api).process_safekeeper_feedback(&mut (*wp))
    }
 }

--- a/libs/walproposer/src/walproposer.rs
+++ b/libs/walproposer/src/walproposer.rs
@@ -142,7 +142,7 @@ pub trait ApiImpl {
        todo!()
    }

-    fn process_safekeeper_feedback(&self, _wp: &mut WalProposer, _commit_lsn: u64) {
+    fn process_safekeeper_feedback(&mut self, _wp: &mut WalProposer) {
        todo!()
    }

--- a/pageserver/client/src/mgmt_api.rs
+++ b/pageserver/client/src/mgmt_api.rs
@@ -7,7 +7,7 @@ use utils::{

 pub mod util;

-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct Client {
    mgmt_api_endpoint: String,
    authorization_header: Option<String>,
@@ -24,6 +24,9 @@ pub enum Error {

    #[error("pageserver API: {1}")]
    ApiError(StatusCode, String),
+
+    #[error("Cancelled")]
+    Cancelled,
 }

 pub type Result<T> = std::result::Result<T, Error>;
@@ -251,21 +254,30 @@ impl Client {
        tenant_shard_id: TenantShardId,
        config: LocationConfig,
        flush_ms: Option<std::time::Duration>,
+        lazy: bool,
    ) -> Result<()> {
        let req_body = TenantLocationConfigRequest {
            tenant_id: tenant_shard_id,
            config,
        };
-        let path = format!(
+
+        let mut path = reqwest::Url::parse(&format!(
            "{}/v1/tenant/{}/location_config",
            self.mgmt_api_endpoint, tenant_shard_id
-        );
-        let path = if let Some(flush_ms) = flush_ms {
-            format!("{}?flush_ms={}", path, flush_ms.as_millis())
-        } else {
-            path
-        };
-        self.request(Method::PUT, &path, &req_body).await?;
+        ))
+        // Should always work: mgmt_api_endpoint is configuration, not user input.
+        .expect("Cannot build URL");
+
+        if lazy {
+            path.query_pairs_mut().append_pair("lazy", "true");
+        }
+
+        if let Some(flush_ms) = flush_ms {
+            path.query_pairs_mut()
+                .append_pair("flush_ms", &format!("{}", flush_ms.as_millis()));
+        }
+
+        self.request(Method::PUT, path, &req_body).await?;
        Ok(())
    }

@@ -278,6 +290,21 @@ impl Client {
            .map_err(Error::ReceiveBody)
    }

+    pub async fn get_location_config(
+        &self,
+        tenant_shard_id: TenantShardId,
+    ) -> Result<Option<LocationConfig>> {
+        let path = format!(
+            "{}/v1/location_config/{tenant_shard_id}",
+            self.mgmt_api_endpoint
+        );
+        self.request(Method::GET, &path, ())
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
    pub async fn timeline_create(
        &self,
        tenant_shard_id: TenantShardId,
--- a/pageserver/compaction/src/compact_tiered.rs
+++ b/pageserver/compaction/src/compact_tiered.rs
@@ -63,7 +63,7 @@ pub async fn compact_tiered<E: CompactionJobExecutor>(
        );

        // Identify the range of LSNs that belong to this level. We assume that
-        // each file in this level span an LSN range up to 1.75x target file
+        // each file in this level spans an LSN range up to 1.75x target file
        // size. That should give us enough slop that if we created a slightly
        // oversized L0 layer, e.g. because flushing the in-memory layer was
        // delayed for some reason, we don't consider the oversized layer to
@@ -248,7 +248,6 @@ enum CompactionStrategy {
    CreateImage,
 }

-#[allow(dead_code)] // Todo
 struct CompactionJob<E: CompactionJobExecutor> {
    key_range: Range<E::Key>,
    lsn_range: Range<Lsn>,
@@ -345,7 +344,7 @@ where
    ///
    /// TODO: Currently, this is called exactly once for the level, and we
    /// decide whether to create new image layers to cover the whole level, or
-    /// write a new set of delta. In the future, this should try to partition
+    /// write a new set of deltas. In the future, this should try to partition
    /// the key space, and make the decision separately for each partition.
    async fn divide_job(&mut self, job_id: JobId, ctx: &E::RequestContext) -> anyhow::Result<()> {
        let job = &self.jobs[job_id.0];
@@ -709,18 +708,6 @@ where
    }
 }

-// Sliding window through keyspace and values
-//
-// This is used to decide what layer to write next, from the beginning of the window.
-//
-// Candidates:
-//
-// 1. Create an image layer, snapping to previous images
-// 2. Create a delta layer, snapping to previous images
-// 3. Create an image layer, snapping to
-//
-//
-
 // Take previous partitioning, based on the image layers below.
 //
 // Candidate is at the front:
@@ -739,6 +726,10 @@ struct WindowElement<K> {
    last_key: K,  // inclusive
    accum_size: u64,
 }
+
+// Sliding window through keyspace and values
+//
+// This is used to decide what layer to write next, from the beginning of the window.
 struct Window<K> {
    elems: VecDeque<WindowElement<K>>,

--- a/pageserver/compaction/src/identify_levels.rs
+++ b/pageserver/compaction/src/identify_levels.rs
@@ -1,5 +1,5 @@
-//! An LSM tree consists of multiple levels, each exponential larger than the
-//! previous level. And each level consists of be multiple "tiers". With tiered
+//! An LSM tree consists of multiple levels, each exponentially larger than the
+//! previous level. And each level consists of multiple "tiers". With tiered
 //! compaction, a level is compacted when it has accumulated more than N tiers,
 //! forming one tier on the next level.
 //!
@@ -170,13 +170,6 @@ where
    })
 }

-// helper struct used in depth()
-struct Event<K> {
-    key: K,
-    layer_idx: usize,
-    start: bool,
-}
-
 impl<L> Level<L> {
    /// Count the number of deltas stacked on each other.
    pub fn depth<K>(&self) -> u64
@@ -184,6 +177,11 @@ impl<L> Level<L> {
        K: CompactionKey,
        L: CompactionLayer<K>,
    {
+        struct Event<K> {
+            key: K,
+            layer_idx: usize,
+            start: bool,
+        }
        let mut events: Vec<Event<K>> = Vec::new();
        for (idx, l) in self.layers.iter().enumerate() {
            events.push(Event {
@@ -202,7 +200,7 @@ impl<L> Level<L> {
        // Sweep the key space left to right. Stop at each distinct key, and
        // count the number of deltas on top of the highest image at that key.
        //
-        // This is a little enefficient, as we walk through the active_set on
+        // This is a little inefficient, as we walk through the active_set on
        // every key. We could increment/decrement a counter on each step
        // instead, but that'd require a bit more complex bookkeeping.
        let mut active_set: BTreeSet<(Lsn, bool, usize)> = BTreeSet::new();
@@ -236,6 +234,7 @@ impl<L> Level<L> {
                }
            }
        }
+        debug_assert_eq!(active_set, BTreeSet::new());
        max_depth
    }
 }
--- a/pageserver/compaction/src/interface.rs
+++ b/pageserver/compaction/src/interface.rs
@@ -4,12 +4,12 @@
 //! 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;
 use utils::lsn::Lsn;

 /// Public interface. This is the main thing that the implementor needs to provide
-#[async_trait]
 pub trait CompactionJobExecutor {
    // Type system.
    //
@@ -17,8 +17,7 @@ pub trait CompactionJobExecutor {
    // compaction doesn't distinguish whether they are stored locally or
    // remotely.
    //
-    // The keyspace is defined by CompactionKey trait.
-    //
+    // The keyspace is defined by the CompactionKey trait.
    type Key: CompactionKey;

    type Layer: CompactionLayer<Self::Key> + Clone;
@@ -35,27 +34,27 @@ pub trait CompactionJobExecutor {
    // ----

    /// Return all layers that overlap the given bounding box.
-    async fn get_layers(
+    fn get_layers(
        &mut self,
        key_range: &Range<Self::Key>,
        lsn_range: &Range<Lsn>,
        ctx: &Self::RequestContext,
-    ) -> anyhow::Result<Vec<Self::Layer>>;
+    ) -> impl Future<Output = anyhow::Result<Vec<Self::Layer>>> + Send;

-    async fn get_keyspace(
+    fn get_keyspace(
        &mut self,
        key_range: &Range<Self::Key>,
        lsn: Lsn,
        ctx: &Self::RequestContext,
-    ) -> anyhow::Result<CompactionKeySpace<Self::Key>>;
+    ) -> impl Future<Output = anyhow::Result<CompactionKeySpace<Self::Key>>> + Send;

    /// NB: This is a pretty expensive operation. In the real pageserver
    /// implementation, it downloads the layer, and keeps it resident
    /// until the DeltaLayer is dropped.
-    async fn downcast_delta_layer(
+    fn downcast_delta_layer(
        &self,
        layer: &Self::Layer,
-    ) -> anyhow::Result<Option<Self::DeltaLayer>>;
+    ) -> impl Future<Output = anyhow::Result<Option<Self::DeltaLayer>>> + Send;

    // ----
    // Functions to execute the plan
@@ -63,33 +62,33 @@ pub trait CompactionJobExecutor {

    /// Create a new image layer, materializing all the values in the key range,
    /// at given 'lsn'.
-    async fn create_image(
+    fn create_image(
        &mut self,
        lsn: Lsn,
        key_range: &Range<Self::Key>,
        ctx: &Self::RequestContext,
-    ) -> anyhow::Result<()>;
+    ) -> impl Future<Output = anyhow::Result<()>> + Send;

    /// Create a new delta layer, containing all the values from 'input_layers'
    /// in the given key and LSN range.
-    async fn create_delta(
+    fn create_delta(
        &mut self,
        lsn_range: &Range<Lsn>,
        key_range: &Range<Self::Key>,
        input_layers: &[Self::DeltaLayer],
        ctx: &Self::RequestContext,
-    ) -> anyhow::Result<()>;
+    ) -> impl Future<Output = anyhow::Result<()>> + Send;

    /// Delete a layer. The compaction implementation will call this only after
    /// all the create_image() or create_delta() calls that deletion of this
    /// layer depends on have finished. But if the implementor has extra lazy
-    /// background tasks, like uploading the index json file to remote storage,
+    /// background tasks, like uploading the index json file to remote storage.
    /// it is the implementation's responsibility to track those.
-    async fn delete_layer(
+    fn delete_layer(
        &mut self,
        layer: &Self::Layer,
        ctx: &Self::RequestContext,
-    ) -> anyhow::Result<()>;
+    ) -> impl Future<Output = anyhow::Result<()>> + Send;
 }

 pub trait CompactionKey: std::cmp::Ord + Clone + Copy + std::fmt::Display {
--- a/pageserver/compaction/src/simulator.rs
+++ b/pageserver/compaction/src/simulator.rs
@@ -429,7 +429,6 @@ impl From<&Arc<MockImageLayer>> for MockLayer {
    }
 }

-#[async_trait]
 impl interface::CompactionJobExecutor for MockTimeline {
    type Key = Key;
    type Layer = MockLayer;
--- a/pageserver/src/config.rs
+++ b/pageserver/src/config.rs
@@ -83,6 +83,10 @@ pub mod defaults {

    pub const DEFAULT_INGEST_BATCH_SIZE: u64 = 100;

+    #[cfg(target_os = "linux")]
+    pub const DEFAULT_VIRTUAL_FILE_IO_ENGINE: &str = "tokio-epoll-uring";
+
+    #[cfg(not(target_os = "linux"))]
    pub const DEFAULT_VIRTUAL_FILE_IO_ENGINE: &str = "std-fs";

    pub const DEFAULT_GET_VECTORED_IMPL: &str = "sequential";
--- a/pageserver/src/context.rs
+++ b/pageserver/src/context.rs
@@ -88,13 +88,16 @@

 use crate::task_mgr::TaskKind;

+pub(crate) mod optional_counter;
+
 // The main structure of this module, see module-level comment.
-#[derive(Clone, Debug)]
+#[derive(Debug)]
 pub struct RequestContext {
    task_kind: TaskKind,
    download_behavior: DownloadBehavior,
    access_stats_behavior: AccessStatsBehavior,
    page_content_kind: PageContentKind,
+    pub micros_spent_throttled: optional_counter::MicroSecondsCounterU32,
 }

 /// The kind of access to the page cache.
@@ -150,6 +153,7 @@ impl RequestContextBuilder {
                download_behavior: DownloadBehavior::Download,
                access_stats_behavior: AccessStatsBehavior::Update,
                page_content_kind: PageContentKind::Unknown,
+                micros_spent_throttled: Default::default(),
            },
        }
    }
@@ -163,6 +167,7 @@ impl RequestContextBuilder {
                download_behavior: original.download_behavior,
                access_stats_behavior: original.access_stats_behavior,
                page_content_kind: original.page_content_kind,
+                micros_spent_throttled: Default::default(),
            },
        }
    }
--- a/pageserver/src/context/optional_counter.rs
+++ b/pageserver/src/context/optional_counter.rs
@@ -0,0 +1,101 @@
+use std::{
+    sync::atomic::{AtomicU32, Ordering},
+    time::Duration,
+};
+
+#[derive(Debug)]
+pub struct CounterU32 {
+    inner: AtomicU32,
+}
+impl Default for CounterU32 {
+    fn default() -> Self {
+        Self {
+            inner: AtomicU32::new(u32::MAX),
+        }
+    }
+}
+impl CounterU32 {
+    pub fn open(&self) -> Result<(), &'static str> {
+        match self
+            .inner
+            .compare_exchange(u32::MAX, 0, Ordering::Relaxed, Ordering::Relaxed)
+        {
+            Ok(_) => Ok(()),
+            Err(_) => Err("open() called on clsoed state"),
+        }
+    }
+    pub fn close(&self) -> Result<u32, &'static str> {
+        match self.inner.swap(u32::MAX, Ordering::Relaxed) {
+            u32::MAX => Err("close() called on closed state"),
+            x => Ok(x),
+        }
+    }
+
+    pub fn add(&self, count: u32) -> Result<(), &'static str> {
+        if count == 0 {
+            return Ok(());
+        }
+        let mut had_err = None;
+        self.inner
+            .fetch_update(Ordering::Relaxed, Ordering::Relaxed, |cur| match cur {
+                u32::MAX => {
+                    had_err = Some("add() called on closed state");
+                    None
+                }
+                x => {
+                    let (new, overflowed) = x.overflowing_add(count);
+                    if new == u32::MAX || overflowed {
+                        had_err = Some("add() overflowed the counter");
+                        None
+                    } else {
+                        Some(new)
+                    }
+                }
+            })
+            .map_err(|_| had_err.expect("we set it whenever the function returns None"))
+            .map(|_| ())
+    }
+}
+
+#[derive(Default, Debug)]
+pub struct MicroSecondsCounterU32 {
+    inner: CounterU32,
+}
+
+impl MicroSecondsCounterU32 {
+    pub fn open(&self) -> Result<(), &'static str> {
+        self.inner.open()
+    }
+    pub fn add(&self, duration: Duration) -> Result<(), &'static str> {
+        match duration.as_micros().try_into() {
+            Ok(x) => self.inner.add(x),
+            Err(_) => Err("add(): duration conversion error"),
+        }
+    }
+    pub fn close_and_checked_sub_from(&self, from: Duration) -> Result<Duration, &'static str> {
+        let val = self.inner.close()?;
+        let val = Duration::from_micros(val as u64);
+        let subbed = match from.checked_sub(val) {
+            Some(v) => v,
+            None => return Err("Duration::checked_sub"),
+        };
+        Ok(subbed)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+
+    use super::*;
+
+    #[test]
+    fn test_basic() {
+        let counter = MicroSecondsCounterU32::default();
+        counter.open().unwrap();
+        counter.add(Duration::from_micros(23)).unwrap();
+        let res = counter
+            .close_and_checked_sub_from(Duration::from_micros(42))
+            .unwrap();
+        assert_eq!(res, Duration::from_micros(42 - 23));
+    }
+}
--- a/pageserver/src/http/openapi_spec.yml
+++ b/pageserver/src/http/openapi_spec.yml
@@ -932,6 +932,59 @@ paths:
              schema:
                $ref: "#/components/schemas/ServiceUnavailableError"

+  /v1/tenant/{tenant_shard_id}/heatmap_upload:
+    parameters:
+      - name: tenant_shard_id
+        in: path
+        required: true
+        schema:
+          type: string
+    post:
+      description: |
+        If the location is in an attached mode, upload the current state to the remote heatmap
+      responses:
+        "200":
+          description: Success
+        "500":
+          description: Generic operation error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
+  /v1/tenant/{tenant_shard_id}/secondary/download:
+    parameters:
+      - name: tenant_shard_id
+        in: path
+        required: true
+        schema:
+          type: string
+    post:
+      description: |
+        If the location is in secondary mode, download latest heatmap and layers
+      responses:
+        "200":
+          description: Success
+        "500":
+          description: Generic operation error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
+

  /v1/tenant/{tenant_id}/timeline/:
    parameters:
@@ -1339,6 +1392,10 @@ components:
          type: array
          items:
            $ref: "#/components/schemas/TenantShardLocation"
+        stripe_size:
+          description: If multiple shards are present, this field contains the sharding stripe size, else it is null.
+          type: integer
+          nullable: true
    TenantShardLocation:
      type: object
      required:
@@ -1387,7 +1444,7 @@ components:
        trace_read_requests:
          type: boolean
        heatmap_period:
-          type: integer
+          type: string
    TenantConfigResponse:
      type: object
      properties:
--- a/pageserver/src/http/routes.rs
+++ b/pageserver/src/http/routes.rs
@@ -14,6 +14,7 @@ use hyper::header;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
 use metrics::launch_timestamp::LaunchTimestamp;
+use pageserver_api::models::LocationConfig;
 use pageserver_api::models::LocationConfigListResponse;
 use pageserver_api::models::ShardParameters;
 use pageserver_api::models::TenantDetails;
@@ -1451,11 +1452,12 @@ async fn put_tenant_location_config_handler(
        tenant::SpawnMode::Eager
    };

-    let attached = state
+    let tenant = state
        .tenant_manager
        .upsert_location(tenant_shard_id, location_conf, flush, spawn_mode, &ctx)
-        .await?
-        .is_some();
+        .await?;
+    let stripe_size = tenant.as_ref().map(|t| t.get_shard_stripe_size());
+    let attached = tenant.is_some();

    if let Some(_flush_ms) = flush {
        match state
@@ -1477,12 +1479,20 @@ async fn put_tenant_location_config_handler(
    // This API returns a vector of pageservers where the tenant is attached: this is
    // primarily for use in the sharding service.  For compatibilty, we also return this
    // when called directly on a pageserver, but the payload is always zero or one shards.
-    let mut response = TenantLocationConfigResponse { shards: Vec::new() };
+    let mut response = TenantLocationConfigResponse {
+        shards: Vec::new(),
+        stripe_size: None,
+    };
    if attached {
        response.shards.push(TenantShardLocation {
            shard_id: tenant_shard_id,
            node_id: state.conf.id,
-        })
+        });
+        if tenant_shard_id.shard_count.count() > 1 {
+            // Stripe size should be set if we are attached
+            debug_assert!(stripe_size.is_some());
+            response.stripe_size = stripe_size;
+        }
    }

    json_response(StatusCode::OK, response)
@@ -1510,6 +1520,29 @@ async fn list_location_config_handler(
    json_response(StatusCode::OK, result)
 }

+async fn get_location_config_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let state = get_state(&request);
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    let slot = state.tenant_manager.get(tenant_shard_id);
+
+    let Some(slot) = slot else {
+        return Err(ApiError::NotFound(
+            anyhow::anyhow!("Tenant shard not found").into(),
+        ));
+    };
+
+    let result: Option<LocationConfig> = match slot {
+        TenantSlot::Attached(t) => Some(t.get_location_conf()),
+        TenantSlot::Secondary(s) => Some(s.get_location_conf()),
+        TenantSlot::InProgress(_) => None,
+    };
+
+    json_response(StatusCode::OK, result)
+}
+
 // Do a time travel recovery on the given tenant/tenant shard. Tenant needs to be detached
 // (from all pageservers) as it invalidates consistency assumptions.
 async fn tenant_time_travel_remote_storage_handler(
@@ -2214,6 +2247,9 @@ pub fn make_router(
        .get("/v1/location_config", |r| {
            api_handler(r, list_location_config_handler)
        })
+        .get("/v1/location_config/:tenant_id", |r| {
+            api_handler(r, get_location_config_handler)
+        })
        .put(
            "/v1/tenant/:tenant_shard_id/time_travel_remote_storage",
            |r| api_handler(r, tenant_time_travel_remote_storage_handler),
--- a/pageserver/src/metrics.rs
+++ b/pageserver/src/metrics.rs
@@ -1,5 +1,4 @@
 use enum_map::EnumMap;
-use metrics::metric_vec_duration::DurationResultObserver;
 use metrics::{
    register_counter_vec, register_gauge_vec, register_histogram, register_histogram_vec,
    register_int_counter, register_int_counter_pair_vec, register_int_counter_vec,
@@ -11,6 +10,7 @@ use once_cell::sync::Lazy;
 use pageserver_api::shard::TenantShardId;
 use strum::{EnumCount, IntoEnumIterator, VariantNames};
 use strum_macros::{EnumVariantNames, IntoStaticStr};
+use tracing::warn;
 use utils::id::TimelineId;

 /// Prometheus histogram buckets (in seconds) for operations in the critical
@@ -1005,15 +1005,39 @@ impl GlobalAndPerTimelineHistogram {
    }
 }

-struct GlobalAndPerTimelineHistogramTimer<'a> {
+struct GlobalAndPerTimelineHistogramTimer<'a, 'c> {
    h: &'a GlobalAndPerTimelineHistogram,
+    ctx: &'c RequestContext,
    start: std::time::Instant,
+    op: SmgrQueryType,
 }

-impl<'a> Drop for GlobalAndPerTimelineHistogramTimer<'a> {
+impl<'a, 'c> Drop for GlobalAndPerTimelineHistogramTimer<'a, 'c> {
    fn drop(&mut self) {
        let elapsed = self.start.elapsed();
-        self.h.observe(elapsed.as_secs_f64());
+        let ex_throttled = self
+            .ctx
+            .micros_spent_throttled
+            .close_and_checked_sub_from(elapsed);
+        let ex_throttled = match ex_throttled {
+            Ok(res) => res,
+            Err(error) => {
+                use utils::rate_limit::RateLimit;
+                static LOGGED: Lazy<Mutex<enum_map::EnumMap<SmgrQueryType, RateLimit>>> =
+                    Lazy::new(|| {
+                        Mutex::new(enum_map::EnumMap::from_array(std::array::from_fn(|_| {
+                            RateLimit::new(Duration::from_secs(10))
+                        })))
+                    });
+                let mut guard = LOGGED.lock().unwrap();
+                let rate_limit = &mut guard[self.op];
+                rate_limit.call(|| {
+                    warn!(op=?self.op, error, "error deducting time spent throttled; this message is logged at a global rate limit");
+                });
+                elapsed
+            }
+        };
+        self.h.observe(ex_throttled.as_secs_f64());
    }
 }

@@ -1025,6 +1049,7 @@ impl<'a> Drop for GlobalAndPerTimelineHistogramTimer<'a> {
    strum_macros::EnumCount,
    strum_macros::EnumIter,
    strum_macros::FromRepr,
+    enum_map::Enum,
 )]
 #[strum(serialize_all = "snake_case")]
 pub enum SmgrQueryType {
@@ -1130,11 +1155,35 @@ impl SmgrQueryTimePerTimeline {
        });
        Self { metrics }
    }
-    pub(crate) fn start_timer(&self, op: SmgrQueryType) -> impl Drop + '_ {
+    pub(crate) fn start_timer<'c: 'a, 'a>(
+        &'a self,
+        op: SmgrQueryType,
+        ctx: &'c RequestContext,
+    ) -> impl Drop + '_ {
        let metric = &self.metrics[op as usize];
+        let start = Instant::now();
+        match ctx.micros_spent_throttled.open() {
+            Ok(()) => (),
+            Err(error) => {
+                use utils::rate_limit::RateLimit;
+                static LOGGED: Lazy<Mutex<enum_map::EnumMap<SmgrQueryType, RateLimit>>> =
+                    Lazy::new(|| {
+                        Mutex::new(enum_map::EnumMap::from_array(std::array::from_fn(|_| {
+                            RateLimit::new(Duration::from_secs(10))
+                        })))
+                    });
+                let mut guard = LOGGED.lock().unwrap();
+                let rate_limit = &mut guard[op];
+                rate_limit.call(|| {
+                    warn!(?op, error, "error opening micros_spent_throttled; this message is logged at a global rate limit");
+                });
+            }
+        }
        GlobalAndPerTimelineHistogramTimer {
            h: metric,
-            start: std::time::Instant::now(),
+            ctx,
+            start,
+            op,
        }
    }
 }
@@ -1145,6 +1194,11 @@ mod smgr_query_time_tests {
    use strum::IntoEnumIterator;
    use utils::id::{TenantId, TimelineId};

+    use crate::{
+        context::{DownloadBehavior, RequestContext},
+        task_mgr::TaskKind,
+    };
+
    // Regression test, we used hard-coded string constants before using an enum.
    #[test]
    fn op_label_name() {
@@ -1193,7 +1247,8 @@ mod smgr_query_time_tests {
            let (pre_global, pre_per_tenant_timeline) = get_counts();
            assert_eq!(pre_per_tenant_timeline, 0);

-            let timer = metrics.start_timer(*op);
+            let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Download);
+            let timer = metrics.start_timer(*op, &ctx);
            drop(timer);

            let (post_global, post_per_tenant_timeline) = get_counts();
@@ -1227,11 +1282,65 @@ pub(crate) static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|
    })
 });

-impl DurationResultObserver for BasebackupQueryTime {
-    fn observe_result<T, E>(&self, res: &Result<T, E>, duration: std::time::Duration) {
+pub(crate) struct BasebackupQueryTimeOngoingRecording<'a, 'c> {
+    parent: &'a BasebackupQueryTime,
+    ctx: &'c RequestContext,
+    start: std::time::Instant,
+}
+
+impl BasebackupQueryTime {
+    pub(crate) fn start_recording<'c: 'a, 'a>(
+        &'a self,
+        ctx: &'c RequestContext,
+    ) -> BasebackupQueryTimeOngoingRecording<'_, '_> {
+        let start = Instant::now();
+        match ctx.micros_spent_throttled.open() {
+            Ok(()) => (),
+            Err(error) => {
+                use utils::rate_limit::RateLimit;
+                static LOGGED: Lazy<Mutex<RateLimit>> =
+                    Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
+                let mut rate_limit = LOGGED.lock().unwrap();
+                rate_limit.call(|| {
+                    warn!(error, "error opening micros_spent_throttled; this message is logged at a global rate limit");
+                });
+            }
+        }
+        BasebackupQueryTimeOngoingRecording {
+            parent: self,
+            ctx,
+            start,
+        }
+    }
+}
+
+impl<'a, 'c> BasebackupQueryTimeOngoingRecording<'a, 'c> {
+    pub(crate) fn observe<T, E>(self, res: &Result<T, E>) {
+        let elapsed = self.start.elapsed();
+        let ex_throttled = self
+            .ctx
+            .micros_spent_throttled
+            .close_and_checked_sub_from(elapsed);
+        let ex_throttled = match ex_throttled {
+            Ok(ex_throttled) => ex_throttled,
+            Err(error) => {
+                use utils::rate_limit::RateLimit;
+                static LOGGED: Lazy<Mutex<RateLimit>> =
+                    Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
+                let mut rate_limit = LOGGED.lock().unwrap();
+                rate_limit.call(|| {
+                    warn!(error, "error deducting time spent throttled; this message is logged at a global rate limit");
+                });
+                elapsed
+            }
+        };
        let label_value = if res.is_ok() { "ok" } else { "error" };
-        let metric = self.0.get_metric_with_label_values(&[label_value]).unwrap();
-        metric.observe(duration.as_secs_f64());
+        let metric = self
+            .parent
+            .0
+            .get_metric_with_label_values(&[label_value])
+            .unwrap();
+        metric.observe(ex_throttled.as_secs_f64());
    }
 }

@@ -1908,10 +2017,8 @@ impl TimelineMetrics {
    pub(crate) fn resident_physical_size_get(&self) -> u64 {
        self.resident_physical_size_gauge.get()
    }
-}

-impl Drop for TimelineMetrics {
-    fn drop(&mut self) {
+    pub(crate) fn shutdown(&self) {
        let tenant_id = &self.tenant_id;
        let timeline_id = &self.timeline_id;
        let shard_id = &self.shard_id;
@@ -2567,6 +2674,12 @@ pub fn preinitialize_metrics() {
    Lazy::force(&crate::tenant::storage_layer::layer::LAYER_IMPL_METRICS);
    Lazy::force(&disk_usage_based_eviction::METRICS);

+    for state_name in pageserver_api::models::TenantState::VARIANTS {
+        // initialize the metric for all gauges, otherwise the time series might seemingly show
+        // values from last restart.
+        TENANT_STATE_METRIC.with_label_values(&[state_name]).set(0);
+    }
+
    // countervecs
    [&BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT]
        .into_iter()
--- a/pageserver/src/page_service.rs
+++ b/pageserver/src/page_service.rs
@@ -910,7 +910,7 @@ impl PageServerHandler {
        let timeline = self.get_timeline_shard_zero(tenant_id, timeline_id).await?;
        let _timer = timeline
            .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetRelExists);
+            .start_timer(metrics::SmgrQueryType::GetRelExists, ctx);

        let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
        let lsn =
@@ -938,7 +938,7 @@ impl PageServerHandler {

        let _timer = timeline
            .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetRelSize);
+            .start_timer(metrics::SmgrQueryType::GetRelSize, ctx);

        let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
        let lsn =
@@ -966,7 +966,7 @@ impl PageServerHandler {

        let _timer = timeline
            .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetDbSize);
+            .start_timer(metrics::SmgrQueryType::GetDbSize, ctx);

        let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
        let lsn =
@@ -1144,7 +1144,7 @@ impl PageServerHandler {

        let _timer = timeline
            .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetPageAtLsn);
+            .start_timer(metrics::SmgrQueryType::GetPageAtLsn, ctx);

        let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
        let lsn =
@@ -1172,7 +1172,7 @@ impl PageServerHandler {

        let _timer = timeline
            .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetSlruSegment);
+            .start_timer(metrics::SmgrQueryType::GetSlruSegment, ctx);

        let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
        let lsn =
@@ -1199,7 +1199,7 @@ impl PageServerHandler {
        prev_lsn: Option<Lsn>,
        full_backup: bool,
        gzip: bool,
-        ctx: RequestContext,
+        ctx: &RequestContext,
    ) -> Result<(), QueryError>
    where
        IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
@@ -1214,7 +1214,7 @@ impl PageServerHandler {
        if let Some(lsn) = lsn {
            // Backup was requested at a particular LSN. Wait for it to arrive.
            info!("waiting for {}", lsn);
-            timeline.wait_lsn(lsn, &ctx).await?;
+            timeline.wait_lsn(lsn, ctx).await?;
            timeline
                .check_lsn_is_in_scope(lsn, &latest_gc_cutoff_lsn)
                .context("invalid basebackup lsn")?;
@@ -1236,7 +1236,7 @@ impl PageServerHandler {
                lsn,
                prev_lsn,
                full_backup,
-                &ctx,
+                ctx,
            )
            .await?;
        } else {
@@ -1257,7 +1257,7 @@ impl PageServerHandler {
                    lsn,
                    prev_lsn,
                    full_backup,
-                    &ctx,
+                    ctx,
                )
                .await?;
                // shutdown the encoder to ensure the gzip footer is written
@@ -1269,7 +1269,7 @@ impl PageServerHandler {
                    lsn,
                    prev_lsn,
                    full_backup,
-                    &ctx,
+                    ctx,
                )
                .await?;
            }
@@ -1449,25 +1449,25 @@ where
                false
            };

-            ::metrics::metric_vec_duration::observe_async_block_duration_by_result(
-                &*metrics::BASEBACKUP_QUERY_TIME,
-                async move {
-                    self.handle_basebackup_request(
-                        pgb,
-                        tenant_id,
-                        timeline_id,
-                        lsn,
-                        None,
-                        false,
-                        gzip,
-                        ctx,
-                    )
-                    .await?;
-                    pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
-                    Result::<(), QueryError>::Ok(())
-                },
-            )
-            .await?;
+            let metric_recording = metrics::BASEBACKUP_QUERY_TIME.start_recording(&ctx);
+            let res = async {
+                self.handle_basebackup_request(
+                    pgb,
+                    tenant_id,
+                    timeline_id,
+                    lsn,
+                    None,
+                    false,
+                    gzip,
+                    &ctx,
+                )
+                .await?;
+                pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
+                Result::<(), QueryError>::Ok(())
+            }
+            .await;
+            metric_recording.observe(&res);
+            res?;
        }
        // return pair of prev_lsn and last_lsn
        else if query_string.starts_with("get_last_record_rlsn ") {
@@ -1563,7 +1563,7 @@ where
                prev_lsn,
                true,
                false,
-                ctx,
+                &ctx,
            )
            .await?;
            pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
--- a/pageserver/src/pgdatadir_mapping.rs
+++ b/pageserver/src/pgdatadir_mapping.rs
@@ -15,6 +15,7 @@ use crate::walrecord::NeonWalRecord;
 use anyhow::{ensure, Context};
 use bytes::{Buf, Bytes, BytesMut};
 use enum_map::Enum;
+use itertools::Itertools;
 use pageserver_api::key::{
    dbdir_key_range, is_rel_block_key, is_slru_block_key, rel_block_to_key, rel_dir_to_key,
    rel_key_range, rel_size_to_key, relmap_file_key, slru_block_to_key, slru_dir_to_key,
@@ -1498,7 +1499,7 @@ impl<'a> DatadirModification<'a> {
            return Ok(());
        }

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

        // Flush relation and  SLRU data blocks, keep metadata.
        let mut retained_pending_updates = HashMap::<_, Vec<_>>::new();
@@ -1537,14 +1538,22 @@ impl<'a> DatadirModification<'a> {
    /// All the modifications in this atomic update are stamped by the specified LSN.
    ///
    pub async fn commit(&mut self, ctx: &RequestContext) -> anyhow::Result<()> {
-        let writer = self.tline.writer().await;
+        let mut writer = self.tline.writer().await;

        let pending_nblocks = self.pending_nblocks;
        self.pending_nblocks = 0;

        if !self.pending_updates.is_empty() {
-            writer.put_batch(&self.pending_updates, ctx).await?;
-            self.pending_updates.clear();
+            // The put_batch call below expects expects the inputs to be sorted by Lsn,
+            // so we do that first.
+            let lsn_ordered_batch: Vec<(Key, Lsn, Value)> = self
+                .pending_updates
+                .drain()
+                .map(|(key, vals)| vals.into_iter().map(move |(lsn, val)| (key, lsn, val)))
+                .kmerge_by(|lhs, rhs| lhs.1 .0 < rhs.1 .0)
+                .collect();
+
+            writer.put_batch(lsn_ordered_batch, ctx).await?;
        }

        if !self.pending_deletions.is_empty() {
@@ -1677,7 +1686,7 @@ struct RelDirectory {
    rels: HashSet<(Oid, u8)>,
 }

-#[derive(Debug, Serialize, Deserialize, Default)]
+#[derive(Debug, Serialize, Deserialize, Default, PartialEq)]
 pub(crate) struct AuxFilesDirectory {
    pub(crate) files: HashMap<String, Bytes>,
 }
--- a/pageserver/src/task_mgr.rs
+++ b/pageserver/src/task_mgr.rs
@@ -272,9 +272,6 @@ pub enum TaskKind {
    // Task that uploads a file to remote storage
    RemoteUploadTask,

-    // Task that downloads a file from remote storage
-    RemoteDownloadTask,
-
    // task that handles the initial downloading of all tenants
    InitialLoad,

--- a/pageserver/src/tenant.rs
+++ b/pageserver/src/tenant.rs
@@ -22,6 +22,7 @@ use pageserver_api::models;
 use pageserver_api::models::TimelineState;
 use pageserver_api::models::WalRedoManagerStatus;
 use pageserver_api::shard::ShardIdentity;
+use pageserver_api::shard::ShardStripeSize;
 use pageserver_api::shard::TenantShardId;
 use remote_storage::DownloadError;
 use remote_storage::GenericRemoteStorage;
@@ -1845,6 +1846,8 @@ impl Tenant {
        // Wait for any in-flight operations to complete
        self.gate.close().await;

+        remove_tenant_metrics(&self.tenant_shard_id);
+
        Ok(())
    }

@@ -2086,6 +2089,10 @@ impl Tenant {
        &self.tenant_shard_id
    }

+    pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
+        self.shard_identity.stripe_size
+    }
+
    pub(crate) fn get_generation(&self) -> Generation {
        self.generation
    }
@@ -3552,11 +3559,6 @@ async fn run_initdb(
    Ok(())
 }

-impl Drop for Tenant {
-    fn drop(&mut self) {
-        remove_tenant_metrics(&self.tenant_shard_id);
-    }
-}
 /// Dump contents of a layer file to stdout.
 pub async fn dump_layerfile_from_path(
    path: &Utf8Path,
@@ -3674,7 +3676,10 @@ pub(crate) mod harness {
    }

    impl TenantHarness {
-        pub fn create(test_name: &'static str) -> anyhow::Result<Self> {
+        pub fn create_custom(
+            test_name: &'static str,
+            tenant_conf: TenantConf,
+        ) -> anyhow::Result<Self> {
            setup_logging();

            let repo_dir = PageServerConf::test_repo_dir(test_name);
@@ -3686,14 +3691,6 @@ pub(crate) mod harness {
            // OK in a test.
            let conf: &'static PageServerConf = Box::leak(Box::new(conf));

-            // Disable automatic GC and compaction to make the unit tests more deterministic.
-            // The tests perform them manually if needed.
-            let tenant_conf = TenantConf {
-                gc_period: Duration::ZERO,
-                compaction_period: Duration::ZERO,
-                ..TenantConf::default()
-            };
-
            let tenant_id = TenantId::generate();
            let tenant_shard_id = TenantShardId::unsharded(tenant_id);
            fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
@@ -3721,6 +3718,18 @@ pub(crate) mod harness {
            })
        }

+        pub fn create(test_name: &'static str) -> anyhow::Result<Self> {
+            // Disable automatic GC and compaction to make the unit tests more deterministic.
+            // The tests perform them manually if needed.
+            let tenant_conf = TenantConf {
+                gc_period: Duration::ZERO,
+                compaction_period: Duration::ZERO,
+                ..TenantConf::default()
+            };
+
+            Self::create_custom(test_name, tenant_conf)
+        }
+
        pub fn span(&self) -> tracing::Span {
            info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
        }
@@ -3828,6 +3837,7 @@ mod tests {
    use crate::keyspace::KeySpaceAccum;
    use crate::repository::{Key, Value};
    use crate::tenant::harness::*;
+    use crate::tenant::timeline::CompactFlags;
    use crate::DEFAULT_PG_VERSION;
    use bytes::BytesMut;
    use hex_literal::hex;
@@ -3844,7 +3854,7 @@ mod tests {
            .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
            .await?;

-        let writer = tline.writer().await;
+        let mut writer = tline.writer().await;
        writer
            .put(
                *TEST_KEY,
@@ -3856,7 +3866,7 @@ mod tests {
        writer.finish_write(Lsn(0x10));
        drop(writer);

-        let writer = tline.writer().await;
+        let mut writer = tline.writer().await;
        writer
            .put(
                *TEST_KEY,
@@ -3922,7 +3932,7 @@ mod tests {
        let tline = tenant
            .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
            .await?;
-        let writer = tline.writer().await;
+        let mut writer = tline.writer().await;

        #[allow(non_snake_case)]
        let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
@@ -3956,7 +3966,7 @@ mod tests {
        let newtline = tenant
            .get_timeline(NEW_TIMELINE_ID, true)
            .expect("Should have a local timeline");
-        let new_writer = newtline.writer().await;
+        let mut new_writer = newtline.writer().await;
        new_writer
            .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
            .await?;
@@ -3988,7 +3998,7 @@ mod tests {
    ) -> anyhow::Result<()> {
        let mut lsn = start_lsn;
        {
-            let writer = tline.writer().await;
+            let mut writer = tline.writer().await;
            // Create a relation on the timeline
            writer
                .put(
@@ -4013,7 +4023,7 @@ mod tests {
        }
        tline.freeze_and_flush().await?;
        {
-            let writer = tline.writer().await;
+            let mut writer = tline.writer().await;
            writer
                .put(
                    *TEST_KEY,
@@ -4376,7 +4386,7 @@ mod tests {
            .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
            .await?;

-        let writer = tline.writer().await;
+        let mut writer = tline.writer().await;
        writer
            .put(
                *TEST_KEY,
@@ -4393,7 +4403,7 @@ mod tests {
            .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
            .await?;

-        let writer = tline.writer().await;
+        let mut writer = tline.writer().await;
        writer
            .put(
                *TEST_KEY,
@@ -4410,7 +4420,7 @@ mod tests {
            .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
            .await?;

-        let writer = tline.writer().await;
+        let mut writer = tline.writer().await;
        writer
            .put(
                *TEST_KEY,
@@ -4427,7 +4437,7 @@ mod tests {
            .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
            .await?;

-        let writer = tline.writer().await;
+        let mut writer = tline.writer().await;
        writer
            .put(
                *TEST_KEY,
@@ -4484,7 +4494,7 @@ mod tests {
        for _ in 0..repeat {
            for _ in 0..key_count {
                test_key.field6 = blknum;
-                let writer = timeline.writer().await;
+                let mut writer = timeline.writer().await;
                writer
                    .put(
                        test_key,
@@ -4632,6 +4642,145 @@ mod tests {
        Ok(())
    }

+    // Test that vectored get handles layer gaps correctly
+    // by advancing into the next ancestor timeline if required.
+    //
+    // The test generates timelines that look like the diagram below.
+    // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
+    // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
+    //
+    // ```
+    //-------------------------------+
+    //                          ...  |
+    //               [   L1   ]      |
+    //     [ / L1   ]                | Child Timeline
+    // ...                           |
+    // ------------------------------+
+    //     [ X L1   ]                | Parent Timeline
+    // ------------------------------+
+    // ```
+    #[tokio::test]
+    async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
+        let tenant_conf = TenantConf {
+            // Make compaction deterministic
+            gc_period: Duration::ZERO,
+            compaction_period: Duration::ZERO,
+            // Encourage creation of L1 layers
+            checkpoint_distance: 16 * 1024,
+            compaction_target_size: 8 * 1024,
+            ..TenantConf::default()
+        };
+
+        let harness = TenantHarness::create_custom("test_get_vectored_key_gap", tenant_conf)?;
+        let (tenant, ctx) = harness.load().await;
+
+        let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
+        let gap_at_key = current_key.add(100);
+        let mut current_lsn = Lsn(0x10);
+
+        const KEY_COUNT: usize = 10_000;
+
+        let timeline_id = TimelineId::generate();
+        let current_timeline = tenant
+            .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
+            .await?;
+
+        current_lsn += 0x100;
+
+        let mut writer = current_timeline.writer().await;
+        writer
+            .put(
+                gap_at_key,
+                current_lsn,
+                &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
+                &ctx,
+            )
+            .await?;
+        writer.finish_write(current_lsn);
+        drop(writer);
+
+        let mut latest_lsns = HashMap::new();
+        latest_lsns.insert(gap_at_key, current_lsn);
+
+        current_timeline.freeze_and_flush().await?;
+
+        let child_timeline_id = TimelineId::generate();
+
+        tenant
+            .branch_timeline_test(
+                &current_timeline,
+                child_timeline_id,
+                Some(current_lsn),
+                &ctx,
+            )
+            .await?;
+        let child_timeline = tenant
+            .get_timeline(child_timeline_id, true)
+            .expect("Should have the branched timeline");
+
+        for i in 0..KEY_COUNT {
+            if current_key == gap_at_key {
+                current_key = current_key.next();
+                continue;
+            }
+
+            current_lsn += 0x10;
+
+            let mut writer = child_timeline.writer().await;
+            writer
+                .put(
+                    current_key,
+                    current_lsn,
+                    &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
+                    &ctx,
+                )
+                .await?;
+            writer.finish_write(current_lsn);
+            drop(writer);
+
+            latest_lsns.insert(current_key, current_lsn);
+            current_key = current_key.next();
+
+            // Flush every now and then to encourage layer file creation.
+            if i % 500 == 0 {
+                child_timeline.freeze_and_flush().await?;
+            }
+        }
+
+        child_timeline.freeze_and_flush().await?;
+        let mut flags = EnumSet::new();
+        flags.insert(CompactFlags::ForceRepartition);
+        child_timeline
+            .compact(&CancellationToken::new(), flags, &ctx)
+            .await?;
+
+        let key_near_end = {
+            let mut tmp = current_key;
+            tmp.field6 -= 10;
+            tmp
+        };
+
+        let key_near_gap = {
+            let mut tmp = gap_at_key;
+            tmp.field6 -= 10;
+            tmp
+        };
+
+        let read = KeySpace {
+            ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
+        };
+        let results = child_timeline
+            .get_vectored_impl(read.clone(), current_lsn, &ctx)
+            .await?;
+
+        for (key, img_res) in results {
+            let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
+            assert_eq!(img_res?, expected);
+        }
+
+        Ok(())
+    }
+
    #[tokio::test]
    async fn test_random_updates() -> anyhow::Result<()> {
        let harness = TenantHarness::create("test_random_updates")?;
@@ -4655,7 +4804,7 @@ mod tests {
        for blknum in 0..NUM_KEYS {
            lsn = Lsn(lsn.0 + 0x10);
            test_key.field6 = blknum as u32;
-            let writer = tline.writer().await;
+            let mut writer = tline.writer().await;
            writer
                .put(
                    test_key,
@@ -4676,7 +4825,7 @@ mod tests {
                lsn = Lsn(lsn.0 + 0x10);
                let blknum = thread_rng().gen_range(0..NUM_KEYS);
                test_key.field6 = blknum as u32;
-                let writer = tline.writer().await;
+                let mut writer = tline.writer().await;
                writer
                    .put(
                        test_key,
@@ -4744,7 +4893,7 @@ mod tests {
        for blknum in 0..NUM_KEYS {
            lsn = Lsn(lsn.0 + 0x10);
            test_key.field6 = blknum as u32;
-            let writer = tline.writer().await;
+            let mut writer = tline.writer().await;
            writer
                .put(
                    test_key,
@@ -4773,7 +4922,7 @@ mod tests {
                lsn = Lsn(lsn.0 + 0x10);
                let blknum = thread_rng().gen_range(0..NUM_KEYS);
                test_key.field6 = blknum as u32;
-                let writer = tline.writer().await;
+                let mut writer = tline.writer().await;
                writer
                    .put(
                        test_key,
@@ -4850,7 +4999,7 @@ mod tests {
                lsn = Lsn(lsn.0 + 0x10);
                let blknum = thread_rng().gen_range(0..NUM_KEYS);
                test_key.field6 = blknum as u32;
-                let writer = tline.writer().await;
+                let mut writer = tline.writer().await;
                writer
                    .put(
                        test_key,
--- a/pageserver/src/tenant/blob_io.rs
+++ b/pageserver/src/tenant/blob_io.rs
@@ -12,7 +12,7 @@
 //! len >= 128: 1XXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
 //!
 use bytes::{BufMut, BytesMut};
-use tokio_epoll_uring::{BoundedBuf, Slice};
+use tokio_epoll_uring::{BoundedBuf, IoBuf, Slice};

 use crate::context::RequestContext;
 use crate::page_cache::PAGE_SZ;
@@ -127,7 +127,7 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
    /// You need to make sure that the internal buffer is empty, otherwise
    /// data will be written in wrong order.
    #[inline(always)]
-    async fn write_all_unbuffered<B: BoundedBuf>(
+    async fn write_all_unbuffered<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
        &mut self,
        src_buf: B,
    ) -> (B::Buf, Result<(), Error>) {
@@ -162,7 +162,10 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
    }

    /// Internal, possibly buffered, write function
-    async fn write_all<B: BoundedBuf>(&mut self, src_buf: B) -> (B::Buf, Result<(), Error>) {
+    async fn write_all<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
+        &mut self,
+        src_buf: B,
+    ) -> (B::Buf, Result<(), Error>) {
        if !BUFFERED {
            assert!(self.buf.is_empty());
            return self.write_all_unbuffered(src_buf).await;
@@ -210,7 +213,10 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {

    /// Write a blob of data. Returns the offset that it was written to,
    /// which can be used to retrieve the data later.
-    pub async fn write_blob<B: BoundedBuf>(&mut self, srcbuf: B) -> (B::Buf, Result<u64, Error>) {
+    pub async fn write_blob<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
+        &mut self,
+        srcbuf: B,
+    ) -> (B::Buf, Result<u64, Error>) {
        let offset = self.offset;

        let len = srcbuf.bytes_init();
--- a/pageserver/src/tenant/config.rs
+++ b/pageserver/src/tenant/config.rs
@@ -354,6 +354,7 @@ pub struct TenantConf {
    /// If non-zero, the period between uploads of a heatmap from attached tenants.  This
    /// may be disabled if a Tenant will not have secondary locations: only secondary
    /// locations will use the heatmap uploaded by attached locations.
+    #[serde(with = "humantime_serde")]
    pub heatmap_period: Duration,

    /// If true then SLRU segments are dowloaded on demand, if false SLRU segments are included in basebackup
--- a/pageserver/src/tenant/disk_btree.rs
+++ b/pageserver/src/tenant/disk_btree.rs
@@ -18,10 +18,19 @@
 //! - An Iterator interface would be more convenient for the callers than the
 //!   'visit' function
 //!
+use async_stream::try_stream;
 use byteorder::{ReadBytesExt, BE};
 use bytes::{BufMut, Bytes, BytesMut};
 use either::Either;
-use std::{cmp::Ordering, io, result};
+use futures::Stream;
+use hex;
+use std::{
+    cmp::Ordering,
+    io,
+    iter::Rev,
+    ops::{Range, RangeInclusive},
+    result,
+};
 use thiserror::Error;
 use tracing::error;

@@ -250,6 +259,90 @@ where
        Ok(result)
    }

+    /// Return a stream which yields all key, value pairs from the index
+    /// starting from the first key greater or equal to `start_key`.
+    ///
+    /// Note that this is a copy of [`Self::visit`].
+    /// TODO: Once the sequential read path is removed this will become
+    /// the only index traversal method.
+    pub fn get_stream_from<'a>(
+        &'a self,
+        start_key: &'a [u8; L],
+        ctx: &'a RequestContext,
+    ) -> impl Stream<Item = std::result::Result<(Vec<u8>, u64), DiskBtreeError>> + 'a {
+        try_stream! {
+            let mut stack = Vec::new();
+            stack.push((self.root_blk, None));
+            let block_cursor = self.reader.block_cursor();
+            while let Some((node_blknum, opt_iter)) = stack.pop() {
+                // Locate the node.
+                let node_buf = block_cursor
+                    .read_blk(self.start_blk + node_blknum, ctx)
+                    .await?;
+
+                let node = OnDiskNode::deparse(node_buf.as_ref())?;
+                let prefix_len = node.prefix_len as usize;
+                let suffix_len = node.suffix_len as usize;
+
+                assert!(node.num_children > 0);
+
+                let mut keybuf = Vec::new();
+                keybuf.extend(node.prefix);
+                keybuf.resize(prefix_len + suffix_len, 0);
+
+                let mut iter: Either<Range<usize>, Rev<RangeInclusive<usize>>> = if let Some(iter) = opt_iter {
+                    iter
+                } else {
+                    // Locate the first match
+                    let idx = match node.binary_search(start_key, keybuf.as_mut_slice()) {
+                        Ok(idx) => idx,
+                        Err(idx) => {
+                            if node.level == 0 {
+                                // Imagine that the node contains the following keys:
+                                //
+                                // 1
+                                // 3  <-- idx
+                                // 5
+                                //
+                                // If the search key is '2' and there is exact match,
+                                // the binary search would return the index of key
+                                // '3'. That's cool, '3' is the first key to return.
+                                idx
+                            } else {
+                                // This is an internal page, so each key represents a lower
+                                // bound for what's in the child page. If there is no exact
+                                // match, we have to return the *previous* entry.
+                                //
+                                // 1  <-- return this
+                                // 3  <-- idx
+                                // 5
+                                idx.saturating_sub(1)
+                            }
+                        }
+                    };
+                    Either::Left(idx..node.num_children.into())
+                };
+
+                // idx points to the first match now. Keep going from there
+                while let Some(idx) = iter.next() {
+                    let key_off = idx * suffix_len;
+                    let suffix = &node.keys[key_off..key_off + suffix_len];
+                    keybuf[prefix_len..].copy_from_slice(suffix);
+                    let value = node.value(idx);
+                    #[allow(clippy::collapsible_if)]
+                    if node.level == 0 {
+                        // leaf
+                        yield (keybuf.clone(), value.to_u64());
+                    } else {
+                        stack.push((node_blknum, Some(iter)));
+                        stack.push((value.to_blknum(), None));
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
    ///
    /// Scan the tree, starting from 'search_key', in the given direction. 'visitor'
    /// will be called for every key >= 'search_key' (or <= 'search_key', if scanning
--- a/pageserver/src/tenant/layer_map.rs
+++ b/pageserver/src/tenant/layer_map.rs
@@ -460,15 +460,22 @@ impl LayerMap {
        }
    }

-    pub fn range_search(&self, key_range: Range<Key>, end_lsn: Lsn) -> Option<RangeSearchResult> {
-        let version = self.historic.get().unwrap().get_version(end_lsn.0 - 1)?;
+    pub fn range_search(&self, key_range: Range<Key>, end_lsn: Lsn) -> RangeSearchResult {
+        let version = match self.historic.get().unwrap().get_version(end_lsn.0 - 1) {
+            Some(version) => version,
+            None => {
+                let mut result = RangeSearchResult::new();
+                result.not_found.add_range(key_range);
+                return result;
+            }
+        };

        let raw_range = key_range.start.to_i128()..key_range.end.to_i128();
        let delta_changes = version.delta_coverage.range_overlaps(&raw_range);
        let image_changes = version.image_coverage.range_overlaps(&raw_range);

        let collector = RangeSearchCollector::new(key_range, end_lsn, delta_changes, image_changes);
-        Some(collector.collect())
+        collector.collect()
    }

    /// Start a batch of updates, applied on drop
@@ -995,8 +1002,13 @@ mod tests {
        let layer_map = LayerMap::default();
        let range = Key::from_i128(100)..Key::from_i128(200);

-        let res = layer_map.range_search(range, Lsn(100));
-        assert!(res.is_none());
+        let res = layer_map.range_search(range.clone(), Lsn(100));
+        assert_eq!(
+            res.not_found.to_keyspace(),
+            KeySpace {
+                ranges: vec![range]
+            }
+        );
    }

    #[test]
@@ -1033,7 +1045,7 @@ mod tests {
        for start in 0..60 {
            for end in (start + 1)..60 {
                let range = Key::from_i128(start)..Key::from_i128(end);
-                let result = layer_map.range_search(range.clone(), Lsn(100)).unwrap();
+                let result = layer_map.range_search(range.clone(), Lsn(100));
                let expected = brute_force_range_search(&layer_map, range, Lsn(100));

                assert_range_search_result_eq(result, expected);
--- a/pageserver/src/tenant/mgr.rs
+++ b/pageserver/src/tenant/mgr.rs
@@ -1358,6 +1358,16 @@ impl TenantManager {
        }
    }

+    pub(crate) fn get(&self, tenant_shard_id: TenantShardId) -> Option<TenantSlot> {
+        let locked = self.tenants.read().unwrap();
+        match &*locked {
+            TenantsMap::Initializing => None,
+            TenantsMap::Open(map) | TenantsMap::ShuttingDown(map) => {
+                map.get(&tenant_shard_id).cloned()
+            }
+        }
+    }
+
    pub(crate) async fn delete_tenant(
        &self,
        tenant_shard_id: TenantShardId,
--- a/pageserver/src/tenant/remote_timeline_client/download.rs
+++ b/pageserver/src/tenant/remote_timeline_client/download.rs
@@ -14,14 +14,14 @@ use tokio::io::{AsyncSeekExt, AsyncWriteExt};
 use tokio_util::io::StreamReader;
 use tokio_util::sync::CancellationToken;
 use tracing::warn;
-use utils::{backoff, crashsafe};
+use utils::backoff;

 use crate::config::PageServerConf;
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::Generation;
-use crate::virtual_file::on_fatal_io_error;
+use crate::virtual_file::{on_fatal_io_error, MaybeFatalIo, VirtualFile};
 use crate::TEMP_FILE_SUFFIX;
 use remote_storage::{DownloadError, GenericRemoteStorage, ListingMode};
 use utils::crashsafe::path_with_suffix_extension;
@@ -50,9 +50,8 @@ pub async fn download_layer_file<'a>(
 ) -> Result<u64, DownloadError> {
    debug_assert_current_span_has_tenant_and_timeline_id();

-    let local_path = conf
-        .timeline_path(&tenant_shard_id, &timeline_id)
-        .join(layer_file_name.file_name());
+    let timeline_path = conf.timeline_path(&tenant_shard_id, &timeline_id);
+    let local_path = timeline_path.join(layer_file_name.file_name());

    let remote_path = remote_layer_path(
        &tenant_shard_id.tenant_id,
@@ -149,10 +148,21 @@ pub async fn download_layer_file<'a>(
        .with_context(|| format!("rename download layer file to {local_path}"))
        .map_err(DownloadError::Other)?;

-    crashsafe::fsync_async(&local_path)
-        .await
-        .with_context(|| format!("fsync layer file {local_path}"))
-        .map_err(DownloadError::Other)?;
+    // We use fatal_err() below because the after the rename above,
+    // the in-memory state of the filesystem already has the layer file in its final place,
+    // and subsequent pageserver code could think it's durable while it really isn't.
+    let work = async move {
+        let timeline_dir = VirtualFile::open(&timeline_path)
+            .await
+            .fatal_err("VirtualFile::open for timeline dir fsync");
+        timeline_dir
+            .sync_all()
+            .await
+            .fatal_err("VirtualFile::sync_all timeline dir");
+    };
+    crate::virtual_file::io_engine::get()
+        .spawn_blocking_and_block_on_if_std(work)
+        .await;

    tracing::debug!("download complete: {local_path}");

--- a/pageserver/src/tenant/storage_layer/delta_layer.rs
+++ b/pageserver/src/tenant/storage_layer/delta_layer.rs
@@ -46,6 +46,7 @@ use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
 use anyhow::{anyhow, bail, ensure, Context, Result};
 use bytes::BytesMut;
 use camino::{Utf8Path, Utf8PathBuf};
+use futures::StreamExt;
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::LayerAccessKind;
 use pageserver_api::shard::TenantShardId;
@@ -847,10 +848,33 @@ impl DeltaLayerInner {
        reconstruct_state: &mut ValuesReconstructState,
        ctx: &RequestContext,
    ) -> Result<(), GetVectoredError> {
-        let reads = self
-            .plan_reads(keyspace, lsn_range, reconstruct_state, ctx)
-            .await
-            .map_err(GetVectoredError::Other)?;
+        let block_reader = FileBlockReader::new(&self.file, self.file_id);
+        let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
+            self.index_start_blk,
+            self.index_root_blk,
+            block_reader,
+        );
+
+        let planner = VectoredReadPlanner::new(
+            self.max_vectored_read_bytes
+                .expect("Layer is loaded with max vectored bytes config")
+                .0
+                .into(),
+        );
+
+        let data_end_offset = self.index_start_blk as u64 * PAGE_SZ as u64;
+
+        let reads = Self::plan_reads(
+            keyspace,
+            lsn_range,
+            data_end_offset,
+            index_reader,
+            planner,
+            reconstruct_state,
+            ctx,
+        )
+        .await
+        .map_err(GetVectoredError::Other)?;

        self.do_reads_and_update_state(reads, reconstruct_state)
            .await;
@@ -858,73 +882,64 @@ impl DeltaLayerInner {
        Ok(())
    }

-    async fn plan_reads(
-        &self,
+    async fn plan_reads<Reader>(
        keyspace: KeySpace,
        lsn_range: Range<Lsn>,
+        data_end_offset: u64,
+        index_reader: DiskBtreeReader<Reader, DELTA_KEY_SIZE>,
+        mut planner: VectoredReadPlanner,
        reconstruct_state: &mut ValuesReconstructState,
        ctx: &RequestContext,
-    ) -> anyhow::Result<Vec<VectoredRead>> {
-        let mut planner = VectoredReadPlanner::new(
-            self.max_vectored_read_bytes
-                .expect("Layer is loaded with max vectored bytes config")
-                .0
-                .into(),
-        );
-
-        let block_reader = FileBlockReader::new(&self.file, self.file_id);
-        let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
-            self.index_start_blk,
-            self.index_root_blk,
-            block_reader,
-        );
+    ) -> anyhow::Result<Vec<VectoredRead>>
+    where
+        Reader: BlockReader,
+    {
+        let ctx = RequestContextBuilder::extend(ctx)
+            .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
+            .build();

        for range in keyspace.ranges.iter() {
            let mut range_end_handled = false;

            let start_key = DeltaKey::from_key_lsn(&range.start, lsn_range.start);
-            tree_reader
-                .visit(
-                    &start_key.0,
-                    VisitDirection::Forwards,
-                    |raw_key, value| {
-                        let key = Key::from_slice(&raw_key[..KEY_SIZE]);
-                        let lsn = DeltaKey::extract_lsn_from_buf(raw_key);
-                        let blob_ref = BlobRef(value);
+            let index_stream = index_reader.get_stream_from(&start_key.0, &ctx);
+            let mut index_stream = std::pin::pin!(index_stream);

-                        assert!(key >= range.start && lsn >= lsn_range.start);
+            while let Some(index_entry) = index_stream.next().await {
+                let (raw_key, value) = index_entry?;
+                let key = Key::from_slice(&raw_key[..KEY_SIZE]);
+                let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
+                let blob_ref = BlobRef(value);

-                        let cached_lsn = reconstruct_state.get_cached_lsn(&key);
-                        let flag = {
-                            if cached_lsn >= Some(lsn) {
-                                BlobFlag::Ignore
-                            } else if blob_ref.will_init() {
-                                BlobFlag::Replaces
-                            } else {
-                                BlobFlag::None
-                            }
-                        };
+                // Lsns are not monotonically increasing across keys, so we don't assert on them.
+                assert!(key >= range.start);

-                        if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
-                            planner.handle_range_end(blob_ref.pos());
-                            range_end_handled = true;
-                            false
-                        } else {
-                            planner.handle(key, lsn, blob_ref.pos(), flag);
-                            true
-                        }
-                    },
-                    &RequestContextBuilder::extend(ctx)
-                        .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
-                        .build(),
-                )
-                .await
-                .map_err(|err| anyhow!(err))?;
+                let outside_lsn_range = !lsn_range.contains(&lsn);
+                let below_cached_lsn = reconstruct_state.get_cached_lsn(&key) >= Some(lsn);
+
+                let flag = {
+                    if outside_lsn_range || below_cached_lsn {
+                        BlobFlag::Ignore
+                    } else if blob_ref.will_init() {
+                        BlobFlag::ReplaceAll
+                    } else {
+                        // Usual path: add blob to the read
+                        BlobFlag::None
+                    }
+                };
+
+                if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
+                    planner.handle_range_end(blob_ref.pos());
+                    range_end_handled = true;
+                    break;
+                } else {
+                    planner.handle(key, lsn, blob_ref.pos(), flag);
+                }
+            }

            if !range_end_handled {
-                let payload_end = self.index_start_blk as u64 * PAGE_SZ as u64;
-                tracing::info!("Handling range end fallback at {}", payload_end);
-                planner.handle_range_end(payload_end);
+                tracing::info!("Handling range end fallback at {}", data_end_offset);
+                planner.handle_range_end(data_end_offset);
            }
        }

@@ -1190,3 +1205,131 @@ impl<'a> pageserver_compaction::interface::CompactionDeltaEntry<'a, Key> for Del
        self.size
    }
 }
+
+#[cfg(test)]
+mod test {
+    use std::collections::BTreeMap;
+
+    use super::*;
+    use crate::{
+        context::DownloadBehavior, task_mgr::TaskKind, tenant::disk_btree::tests::TestDisk,
+    };
+
+    /// Construct an index for a fictional delta layer and and then
+    /// traverse in order to plan vectored reads for a query. Finally,
+    /// verify that the traversal fed the right index key and value
+    /// pairs into the planner.
+    #[tokio::test]
+    async fn test_delta_layer_index_traversal() {
+        let base_key = Key {
+            field1: 0,
+            field2: 1663,
+            field3: 12972,
+            field4: 16396,
+            field5: 0,
+            field6: 246080,
+        };
+
+        // Populate the index with some entries
+        let entries: BTreeMap<Key, Vec<Lsn>> = BTreeMap::from([
+            (base_key, vec![Lsn(1), Lsn(5), Lsn(25), Lsn(26), Lsn(28)]),
+            (base_key.add(1), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
+            (base_key.add(2), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
+            (base_key.add(5), vec![Lsn(10), Lsn(15), Lsn(16), Lsn(20)]),
+        ]);
+
+        let mut disk = TestDisk::default();
+        let mut writer = DiskBtreeBuilder::<_, DELTA_KEY_SIZE>::new(&mut disk);
+
+        let mut disk_offset = 0;
+        for (key, lsns) in &entries {
+            for lsn in lsns {
+                let index_key = DeltaKey::from_key_lsn(key, *lsn);
+                let blob_ref = BlobRef::new(disk_offset, false);
+                writer
+                    .append(&index_key.0, blob_ref.0)
+                    .expect("In memory disk append should never fail");
+
+                disk_offset += 1;
+            }
+        }
+
+        // Prepare all the arguments for the call into `plan_reads` below
+        let (root_offset, _writer) = writer
+            .finish()
+            .expect("In memory disk finish should never fail");
+        let reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(0, root_offset, disk);
+        let planner = VectoredReadPlanner::new(100);
+        let mut reconstruct_state = ValuesReconstructState::new();
+        let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
+
+        let keyspace = KeySpace {
+            ranges: vec![
+                base_key..base_key.add(3),
+                base_key.add(3)..base_key.add(100),
+            ],
+        };
+        let lsn_range = Lsn(2)..Lsn(40);
+
+        // Plan and validate
+        let vectored_reads = DeltaLayerInner::plan_reads(
+            keyspace.clone(),
+            lsn_range.clone(),
+            disk_offset,
+            reader,
+            planner,
+            &mut reconstruct_state,
+            &ctx,
+        )
+        .await
+        .expect("Read planning should not fail");
+
+        validate(keyspace, lsn_range, vectored_reads, entries);
+    }
+
+    fn validate(
+        keyspace: KeySpace,
+        lsn_range: Range<Lsn>,
+        vectored_reads: Vec<VectoredRead>,
+        index_entries: BTreeMap<Key, Vec<Lsn>>,
+    ) {
+        #[derive(Debug, PartialEq, Eq)]
+        struct BlobSpec {
+            key: Key,
+            lsn: Lsn,
+            at: u64,
+        }
+
+        let mut planned_blobs = Vec::new();
+        for read in vectored_reads {
+            for (at, meta) in read.blobs_at.as_slice() {
+                planned_blobs.push(BlobSpec {
+                    key: meta.key,
+                    lsn: meta.lsn,
+                    at: *at,
+                });
+            }
+        }
+
+        let mut expected_blobs = Vec::new();
+        let mut disk_offset = 0;
+        for (key, lsns) in index_entries {
+            for lsn in lsns {
+                let key_included = keyspace.ranges.iter().any(|range| range.contains(&key));
+                let lsn_included = lsn_range.contains(&lsn);
+
+                if key_included && lsn_included {
+                    expected_blobs.push(BlobSpec {
+                        key,
+                        lsn,
+                        at: disk_offset,
+                    });
+                }
+
+                disk_offset += 1;
+            }
+        }
+
+        assert_eq!(planned_blobs, expected_blobs);
+    }
+}
--- a/pageserver/src/tenant/storage_layer/image_layer.rs
+++ b/pageserver/src/tenant/storage_layer/image_layer.rs
@@ -43,6 +43,7 @@ use crate::{IMAGE_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
 use anyhow::{anyhow, bail, ensure, Context, Result};
 use bytes::{Bytes, BytesMut};
 use camino::{Utf8Path, Utf8PathBuf};
+use hex;
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::LayerAccessKind;
 use pageserver_api::shard::TenantShardId;
@@ -54,6 +55,7 @@ use std::ops::Range;
 use std::os::unix::prelude::FileExt;
 use std::sync::Arc;
 use tokio::sync::OnceCell;
+use tokio_stream::StreamExt;
 use tracing::*;

 use utils::{
@@ -488,35 +490,33 @@ impl ImageLayerInner {
        let tree_reader =
            DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);

+        let ctx = RequestContextBuilder::extend(ctx)
+            .page_content_kind(PageContentKind::ImageLayerBtreeNode)
+            .build();
+
        for range in keyspace.ranges.iter() {
            let mut range_end_handled = false;

            let mut search_key: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
            range.start.write_to_byte_slice(&mut search_key);

-            tree_reader
-                .visit(
-                    &search_key,
-                    VisitDirection::Forwards,
-                    |raw_key, offset| {
-                        let key = Key::from_slice(&raw_key[..KEY_SIZE]);
-                        assert!(key >= range.start);
+            let index_stream = tree_reader.get_stream_from(&search_key, &ctx);
+            let mut index_stream = std::pin::pin!(index_stream);

-                        if key >= range.end {
-                            planner.handle_range_end(offset);
-                            range_end_handled = true;
-                            false
-                        } else {
-                            planner.handle(key, self.lsn, offset, BlobFlag::None);
-                            true
-                        }
-                    },
-                    &RequestContextBuilder::extend(ctx)
-                        .page_content_kind(PageContentKind::ImageLayerBtreeNode)
-                        .build(),
-                )
-                .await
-                .map_err(|err| GetVectoredError::Other(anyhow!(err)))?;
+            while let Some(index_entry) = index_stream.next().await {
+                let (raw_key, offset) = index_entry?;
+
+                let key = Key::from_slice(&raw_key[..KEY_SIZE]);
+                assert!(key >= range.start);
+
+                if key >= range.end {
+                    planner.handle_range_end(offset);
+                    range_end_handled = true;
+                    break;
+                } else {
+                    planner.handle(key, self.lsn, offset, BlobFlag::None);
+                }
+            }

            if !range_end_handled {
                let payload_end = self.index_start_blk as u64 * PAGE_SZ as u64;
--- a/pageserver/src/tenant/storage_layer/inmemory_layer.rs
+++ b/pageserver/src/tenant/storage_layer/inmemory_layer.rs
@@ -336,32 +336,17 @@ impl InMemoryLayer {

    /// Common subroutine of the public put_wal_record() and put_page_image() functions.
    /// Adds the page version to the in-memory tree
+
    pub(crate) async fn put_value(
        &self,
        key: Key,
        lsn: Lsn,
-        val: &Value,
+        buf: &[u8],
        ctx: &RequestContext,
    ) -> Result<()> {
        let mut inner = self.inner.write().await;
        self.assert_writable();
-        self.put_value_locked(&mut inner, key, lsn, val, ctx).await
-    }
-
-    pub(crate) async fn put_values(
-        &self,
-        values: &HashMap<Key, Vec<(Lsn, Value)>>,
-        ctx: &RequestContext,
-    ) -> Result<()> {
-        let mut inner = self.inner.write().await;
-        self.assert_writable();
-        for (key, vals) in values {
-            for (lsn, val) in vals {
-                self.put_value_locked(&mut inner, *key, *lsn, val, ctx)
-                    .await?;
-            }
-        }
-        Ok(())
+        self.put_value_locked(&mut inner, key, lsn, buf, ctx).await
    }

    async fn put_value_locked(
@@ -369,22 +354,16 @@ impl InMemoryLayer {
        locked_inner: &mut RwLockWriteGuard<'_, InMemoryLayerInner>,
        key: Key,
        lsn: Lsn,
-        val: &Value,
+        buf: &[u8],
        ctx: &RequestContext,
    ) -> Result<()> {
        trace!("put_value key {} at {}/{}", key, self.timeline_id, lsn);

        let off = {
-            // Avoid doing allocations for "small" values.
-            // In the regression test suite, the limit of 256 avoided allocations in 95% of cases:
-            // https://github.com/neondatabase/neon/pull/5056#discussion_r1301975061
-            let mut buf = smallvec::SmallVec::<[u8; 256]>::new();
-            buf.clear();
-            val.ser_into(&mut buf)?;
            locked_inner
                .file
                .write_blob(
-                    &buf,
+                    buf,
                    &RequestContextBuilder::extend(ctx)
                        .page_content_kind(PageContentKind::InMemoryLayer)
                        .build(),
@@ -412,7 +391,12 @@ impl InMemoryLayer {
    pub async fn freeze(&self, end_lsn: Lsn) {
        let inner = self.inner.write().await;

-        assert!(self.start_lsn < end_lsn);
+        assert!(
+            self.start_lsn < end_lsn,
+            "{} >= {}",
+            self.start_lsn,
+            end_lsn
+        );
        self.end_lsn.set(end_lsn).expect("end_lsn set only once");

        for vec_map in inner.index.values() {
--- a/pageserver/src/tenant/storage_layer/layer.rs
+++ b/pageserver/src/tenant/storage_layer/layer.rs
@@ -195,6 +195,7 @@ impl Layer {
        let downloaded = resident.expect("just initialized");

        // if the rename works, the path is as expected
+        // TODO: sync system call
        std::fs::rename(temp_path, owner.local_path())
            .with_context(|| format!("rename temporary file as correct path for {owner}"))?;

@@ -535,6 +536,18 @@ impl Drop for LayerInner {
            // carry this until we are finished for [`Layer::wait_drop`] support
            let _status = status;

+            let Some(timeline) = timeline.upgrade() else {
+                // no need to nag that timeline is gone: under normal situation on
+                // task_mgr::remove_tenant_from_memory the timeline is gone before we get dropped.
+                LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);
+                return;
+            };
+
+            let Ok(_guard) = timeline.gate.enter() else {
+                LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);
+                return;
+            };
+
            let removed = match std::fs::remove_file(path) {
                Ok(()) => true,
                Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
@@ -553,32 +566,26 @@ impl Drop for LayerInner {
                }
            };

-            if let Some(timeline) = timeline.upgrade() {
-                if removed {
-                    timeline.metrics.resident_physical_size_sub(file_size);
-                }
-                if let Some(remote_client) = timeline.remote_client.as_ref() {
-                    let res = remote_client.schedule_deletion_of_unlinked(vec![(file_name, meta)]);
+            if removed {
+                timeline.metrics.resident_physical_size_sub(file_size);
+            }
+            if let Some(remote_client) = timeline.remote_client.as_ref() {
+                let res = remote_client.schedule_deletion_of_unlinked(vec![(file_name, meta)]);

-                    if let Err(e) = res {
-                        // test_timeline_deletion_with_files_stuck_in_upload_queue is good at
-                        // demonstrating this deadlock (without spawn_blocking): stop will drop
-                        // queued items, which will have ResidentLayer's, and those drops would try
-                        // to re-entrantly lock the RemoteTimelineClient inner state.
-                        if !timeline.is_active() {
-                            tracing::info!("scheduling deletion on drop failed: {e:#}");
-                        } else {
-                            tracing::warn!("scheduling deletion on drop failed: {e:#}");
-                        }
-                        LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::DeleteSchedulingFailed);
+                if let Err(e) = res {
+                    // test_timeline_deletion_with_files_stuck_in_upload_queue is good at
+                    // demonstrating this deadlock (without spawn_blocking): stop will drop
+                    // queued items, which will have ResidentLayer's, and those drops would try
+                    // to re-entrantly lock the RemoteTimelineClient inner state.
+                    if !timeline.is_active() {
+                        tracing::info!("scheduling deletion on drop failed: {e:#}");
                    } else {
-                        LAYER_IMPL_METRICS.inc_completed_deletes();
+                        tracing::warn!("scheduling deletion on drop failed: {e:#}");
                    }
+                    LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::DeleteSchedulingFailed);
+                } else {
+                    LAYER_IMPL_METRICS.inc_completed_deletes();
                }
-            } else {
-                // no need to nag that timeline is gone: under normal situation on
-                // task_mgr::remove_tenant_from_memory the timeline is gone before we get dropped.
-                LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);
            }
        });
    }
@@ -879,23 +886,18 @@ impl LayerInner {
    ) -> Result<heavier_once_cell::InitPermit, DownloadError> {
        debug_assert_current_span_has_tenant_and_timeline_id();

-        let task_name = format!("download layer {}", self);
-
        let (tx, rx) = tokio::sync::oneshot::channel();

-        // this is sadly needed because of task_mgr::shutdown_tasks, otherwise we cannot
-        // block tenant::mgr::remove_tenant_from_memory.
-
        let this: Arc<Self> = self.clone();

-        crate::task_mgr::spawn(
-            &tokio::runtime::Handle::current(),
-            crate::task_mgr::TaskKind::RemoteDownloadTask,
-            Some(self.desc.tenant_shard_id),
-            Some(self.desc.timeline_id),
-            &task_name,
-            false,
-            async move {
+        let guard = timeline
+            .gate
+            .enter()
+            .map_err(|_| DownloadError::DownloadCancelled)?;
+
+        tokio::task::spawn(async move {
+
+                let _guard = guard;

                let client = timeline
                    .remote_client
@@ -905,7 +907,7 @@ impl LayerInner {
                let result = client.download_layer_file(
                    &this.desc.filename(),
                    &this.metadata(),
-                    &crate::task_mgr::shutdown_token()
+                    &timeline.cancel
                )
                .await;

@@ -928,7 +930,6 @@ impl LayerInner {

                        tokio::select! {
                            _ = tokio::time::sleep(backoff) => {},
-                            _ = crate::task_mgr::shutdown_token().cancelled_owned() => {},
                            _ = timeline.cancel.cancelled() => {},
                        };

@@ -958,11 +959,10 @@ impl LayerInner {
                        }
                    }
                }
-
-                Ok(())
            }
            .in_current_span(),
        );
+
        match rx.await {
            Ok((Ok(()), permit)) => {
                if let Some(reason) = self
@@ -975,7 +975,7 @@ impl LayerInner {
                }

                self.consecutive_failures.store(0, Ordering::Relaxed);
-                tracing::info!("on-demand download successful");
+                tracing::info!(size=%self.desc.file_size, "on-demand download successful");

                Ok(permit)
            }
@@ -1101,6 +1101,10 @@ impl LayerInner {
            return Err(EvictionCancelled::TimelineGone);
        };

+        let Ok(_gate) = timeline.gate.enter() else {
+            return Err(EvictionCancelled::TimelineGone);
+        };
+
        // to avoid starting a new download while we evict, keep holding on to the
        // permit.
        let _permit = {
--- a/pageserver/src/tenant/tasks.rs
+++ b/pageserver/src/tenant/tasks.rs
@@ -101,6 +101,7 @@ pub fn start_background_loops(
                    _ = completion::Barrier::maybe_wait(background_jobs_can_start) => {}
                };
                compaction_loop(tenant, cancel)
+                    // If you rename this span, change the RUST_LOG env variable in test_runner/performance/test_branch_creation.py
                    .instrument(info_span!("compaction_loop", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug()))
                    .await;
                Ok(())
@@ -198,7 +199,11 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                }
            };

-            warn_when_period_overrun(started_at.elapsed(), period, BackgroundLoopKind::Compaction);
+            let elapsed = started_at.elapsed();
+            warn_when_period_overrun(elapsed, period, BackgroundLoopKind::Compaction);
+
+            // the duration is recorded by performance tests by enabling debug in this function
+            tracing::debug!(elapsed_ms=elapsed.as_millis(), "compaction iteration complete");

            // Perhaps we did no work and the walredo process has been idle for some time:
            // give it a chance to shut down to avoid leaving walredo process running indefinitely.
@@ -217,7 +222,7 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                }
                let allowed_rps = tenant.timeline_get_throttle.steady_rps();
                let delta = now - prev;
-                warn!(
+                info!(
                    n_seconds=%format_args!("{:.3}",
                    delta.as_secs_f64()),
                    count_accounted,
--- a/pageserver/src/tenant/throttle.rs
+++ b/pageserver/src/tenant/throttle.rs
@@ -2,14 +2,14 @@ use std::{
    str::FromStr,
    sync::{
        atomic::{AtomicU64, Ordering},
-        Arc,
+        Arc, Mutex,
    },
    time::{Duration, Instant},
 };

 use arc_swap::ArcSwap;
 use enumset::EnumSet;
-use tracing::error;
+use tracing::{error, warn};

 use crate::{context::RequestContext, task_mgr::TaskKind};

@@ -157,6 +157,19 @@ where
                .fetch_add(wait_time.as_micros() as u64, Ordering::Relaxed);
            let observation = Observation { wait_time };
            self.metric.observe_throttling(&observation);
+            match ctx.micros_spent_throttled.add(wait_time) {
+                Ok(res) => res,
+                Err(error) => {
+                    use once_cell::sync::Lazy;
+                    use utils::rate_limit::RateLimit;
+                    static WARN_RATE_LIMIT: Lazy<Mutex<RateLimit>> =
+                        Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
+                    let mut guard = WARN_RATE_LIMIT.lock().unwrap();
+                    guard.call(move || {
+                        warn!(error, "error adding time spent throttled; this message is logged at a global rate limit");
+                    });
+                }
+            }
        }
    }
 }
--- a/pageserver/src/tenant/timeline.rs
+++ b/pageserver/src/tenant/timeline.rs
--- a/pageserver/src/tenant/timeline/compaction.rs
+++ b/pageserver/src/tenant/timeline/compaction.rs
@@ -4,24 +4,32 @@
 //!
 //! The old legacy algorithm is implemented directly in `timeline.rs`.

+use std::collections::BinaryHeap;
 use std::ops::{Deref, Range};
 use std::sync::Arc;

-use super::Timeline;
+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;
+use pageserver_api::shard::TenantShardId;
 use tokio_util::sync::CancellationToken;
-use tracing::{debug, trace, warn};
+use tracing::{debug, info, info_span, trace, warn, Instrument};
+use utils::id::TimelineId;

-use crate::context::RequestContext;
+use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
 use crate::tenant::storage_layer::{AsLayerDesc, PersistentLayerDesc};
-use crate::tenant::timeline::{is_rel_fsm_block_key, is_rel_vm_block_key};
+use crate::tenant::timeline::{drop_rlock, is_rel_fsm_block_key, is_rel_vm_block_key, Hole};
 use crate::tenant::timeline::{DeltaLayerWriter, ImageLayerWriter};
 use crate::tenant::timeline::{Layer, ResidentLayer};
 use crate::tenant::DeltaLayer;
 use crate::tenant::PageReconstructError;
-use crate::ZERO_PAGE;
+use crate::virtual_file::{MaybeFatalIo, VirtualFile};
+use crate::{page_cache, ZERO_PAGE};

 use crate::keyspace::KeySpace;
 use crate::repository::Key;
@@ -33,6 +41,694 @@ use pageserver_compaction::interface::*;

 use super::CompactionError;

+impl Timeline {
+    /// TODO: cancellation
+    pub(crate) async fn compact_legacy(
+        self: &Arc<Self>,
+        _cancel: &CancellationToken,
+        flags: EnumSet<CompactFlags>,
+        ctx: &RequestContext,
+    ) -> Result<(), CompactionError> {
+        // High level strategy for compaction / image creation:
+        //
+        // 1. First, calculate the desired "partitioning" of the
+        // currently in-use key space. The goal is to partition the
+        // key space into roughly fixed-size chunks, but also take into
+        // account any existing image layers, and try to align the
+        // chunk boundaries with the existing image layers to avoid
+        // too much churn. Also try to align chunk boundaries with
+        // relation boundaries.  In principle, we don't know about
+        // relation boundaries here, we just deal with key-value
+        // pairs, and the code in pgdatadir_mapping.rs knows how to
+        // map relations into key-value pairs. But in practice we know
+        // that 'field6' is the block number, and the fields 1-5
+        // identify a relation. This is just an optimization,
+        // though.
+        //
+        // 2. Once we know the partitioning, for each partition,
+        // decide if it's time to create a new image layer. The
+        // criteria is: there has been too much "churn" since the last
+        // image layer? The "churn" is fuzzy concept, it's a
+        // combination of too many delta files, or too much WAL in
+        // total in the delta file. Or perhaps: if creating an image
+        // file would allow to delete some older files.
+        //
+        // 3. After that, we compact all level0 delta files if there
+        // are too many of them.  While compacting, we also garbage
+        // collect any page versions that are no longer needed because
+        // of the new image layers we created in step 2.
+        //
+        // TODO: This high level strategy hasn't been implemented yet.
+        // Below are functions compact_level0() and create_image_layers()
+        // but they are a bit ad hoc and don't quite work like it's explained
+        // above. Rewrite it.
+
+        // Is the timeline being deleted?
+        if self.is_stopping() {
+            trace!("Dropping out of compaction on timeline shutdown");
+            return Err(CompactionError::ShuttingDown);
+        }
+
+        let target_file_size = self.get_checkpoint_distance();
+
+        // Define partitioning schema if needed
+
+        // FIXME: the match should only cover repartitioning, not the next steps
+        match self
+            .repartition(
+                self.get_last_record_lsn(),
+                self.get_compaction_target_size(),
+                flags,
+                ctx,
+            )
+            .await
+        {
+            Ok((partitioning, lsn)) => {
+                // Disables access_stats updates, so that the files we read remain candidates for eviction after we're done with them
+                let image_ctx = RequestContextBuilder::extend(ctx)
+                    .access_stats_behavior(AccessStatsBehavior::Skip)
+                    .build();
+
+                // 2. Compact
+                let timer = self.metrics.compact_time_histo.start_timer();
+                self.compact_level0(target_file_size, ctx).await?;
+                timer.stop_and_record();
+
+                // 3. Create new image layers for partitions that have been modified
+                // "enough".
+                let layers = self
+                    .create_image_layers(
+                        &partitioning,
+                        lsn,
+                        flags.contains(CompactFlags::ForceImageLayerCreation),
+                        &image_ctx,
+                    )
+                    .await
+                    .map_err(anyhow::Error::from)?;
+                if let Some(remote_client) = &self.remote_client {
+                    for layer in layers {
+                        remote_client.schedule_layer_file_upload(layer)?;
+                    }
+                }
+
+                if let Some(remote_client) = &self.remote_client {
+                    // should any new image layer been created, not uploading index_part will
+                    // result in a mismatch between remote_physical_size and layermap calculated
+                    // size, which will fail some tests, but should not be an issue otherwise.
+                    remote_client.schedule_index_upload_for_file_changes()?;
+                }
+            }
+            Err(err) => {
+                // no partitioning? This is normal, if the timeline was just created
+                // as an empty timeline. Also in unit tests, when we use the timeline
+                // as a simple key-value store, ignoring the datadir layout. Log the
+                // error but continue.
+                //
+                // Suppress error when it's due to cancellation
+                if !self.cancel.is_cancelled() {
+                    tracing::error!("could not compact, repartitioning keyspace failed: {err:?}");
+                }
+            }
+        };
+
+        Ok(())
+    }
+
+    /// Collect a bunch of Level 0 layer files, and compact and reshuffle them as
+    /// as Level 1 files.
+    async fn compact_level0(
+        self: &Arc<Self>,
+        target_file_size: u64,
+        ctx: &RequestContext,
+    ) -> Result<(), CompactionError> {
+        let CompactLevel0Phase1Result {
+            new_layers,
+            deltas_to_compact,
+        } = {
+            let phase1_span = info_span!("compact_level0_phase1");
+            let ctx = ctx.attached_child();
+            let mut stats = CompactLevel0Phase1StatsBuilder {
+                version: Some(2),
+                tenant_id: Some(self.tenant_shard_id),
+                timeline_id: Some(self.timeline_id),
+                ..Default::default()
+            };
+
+            let begin = tokio::time::Instant::now();
+            let phase1_layers_locked = Arc::clone(&self.layers).read_owned().await;
+            let now = tokio::time::Instant::now();
+            stats.read_lock_acquisition_micros =
+                DurationRecorder::Recorded(RecordedDuration(now - begin), now);
+            self.compact_level0_phase1(phase1_layers_locked, stats, target_file_size, &ctx)
+                .instrument(phase1_span)
+                .await?
+        };
+
+        if new_layers.is_empty() && deltas_to_compact.is_empty() {
+            // nothing to do
+            return Ok(());
+        }
+
+        self.finish_compact_batch(&new_layers, &Vec::new(), &deltas_to_compact)
+            .await?;
+        Ok(())
+    }
+
+    /// Level0 files first phase of compaction, explained in the [`Self::compact_legacy`] comment.
+    async fn compact_level0_phase1(
+        self: &Arc<Self>,
+        guard: tokio::sync::OwnedRwLockReadGuard<LayerManager>,
+        mut stats: CompactLevel0Phase1StatsBuilder,
+        target_file_size: u64,
+        ctx: &RequestContext,
+    ) -> Result<CompactLevel0Phase1Result, CompactionError> {
+        stats.read_lock_held_spawn_blocking_startup_micros =
+            stats.read_lock_acquisition_micros.till_now(); // set by caller
+        let layers = guard.layer_map();
+        let level0_deltas = layers.get_level0_deltas()?;
+        let mut level0_deltas = level0_deltas
+            .into_iter()
+            .map(|x| guard.get_from_desc(&x))
+            .collect_vec();
+        stats.level0_deltas_count = Some(level0_deltas.len());
+        // Only compact if enough layers have accumulated.
+        let threshold = self.get_compaction_threshold();
+        if level0_deltas.is_empty() || level0_deltas.len() < threshold {
+            debug!(
+                level0_deltas = level0_deltas.len(),
+                threshold, "too few deltas to compact"
+            );
+            return Ok(CompactLevel0Phase1Result::default());
+        }
+
+        // This failpoint is used together with `test_duplicate_layers` integration test.
+        // It returns the compaction result exactly the same layers as input to compaction.
+        // We want to ensure that this will not cause any problem when updating the layer map
+        // after the compaction is finished.
+        //
+        // Currently, there are two rare edge cases that will cause duplicated layers being
+        // inserted.
+        // 1. The compaction job is inturrupted / did not finish successfully. Assume we have file 1, 2, 3, 4, which
+        //    is compacted to 5, but the page server is shut down, next time we start page server we will get a layer
+        //    map containing 1, 2, 3, 4, and 5, whereas 5 has the same content as 4. If we trigger L0 compation at this
+        //    point again, it is likely that we will get a file 6 which has the same content and the key range as 5,
+        //    and this causes an overwrite. This is acceptable because the content is the same, and we should do a
+        //    layer replace instead of the normal remove / upload process.
+        // 2. The input workload pattern creates exactly n files that are sorted, non-overlapping and is of target file
+        //    size length. Compaction will likely create the same set of n files afterwards.
+        //
+        // This failpoint is a superset of both of the cases.
+        if cfg!(feature = "testing") {
+            let active = (|| {
+                ::fail::fail_point!("compact-level0-phase1-return-same", |_| true);
+                false
+            })();
+
+            if active {
+                let mut new_layers = Vec::with_capacity(level0_deltas.len());
+                for delta in &level0_deltas {
+                    // we are just faking these layers as being produced again for this failpoint
+                    new_layers.push(
+                        delta
+                            .download_and_keep_resident()
+                            .await
+                            .context("download layer for failpoint")?,
+                    );
+                }
+                tracing::info!("compact-level0-phase1-return-same"); // so that we can check if we hit the failpoint
+                return Ok(CompactLevel0Phase1Result {
+                    new_layers,
+                    deltas_to_compact: level0_deltas,
+                });
+            }
+        }
+
+        // Gather the files to compact in this iteration.
+        //
+        // Start with the oldest Level 0 delta file, and collect any other
+        // level 0 files that form a contiguous sequence, such that the end
+        // LSN of previous file matches the start LSN of the next file.
+        //
+        // Note that if the files don't form such a sequence, we might
+        // "compact" just a single file. That's a bit pointless, but it allows
+        // us to get rid of the level 0 file, and compact the other files on
+        // the next iteration. This could probably made smarter, but such
+        // "gaps" in the sequence of level 0 files should only happen in case
+        // of a crash, partial download from cloud storage, or something like
+        // that, so it's not a big deal in practice.
+        level0_deltas.sort_by_key(|l| l.layer_desc().lsn_range.start);
+        let mut level0_deltas_iter = level0_deltas.iter();
+
+        let first_level0_delta = level0_deltas_iter.next().unwrap();
+        let mut prev_lsn_end = first_level0_delta.layer_desc().lsn_range.end;
+        let mut deltas_to_compact = Vec::with_capacity(level0_deltas.len());
+
+        deltas_to_compact.push(first_level0_delta.download_and_keep_resident().await?);
+        for l in level0_deltas_iter {
+            let lsn_range = &l.layer_desc().lsn_range;
+
+            if lsn_range.start != prev_lsn_end {
+                break;
+            }
+            deltas_to_compact.push(l.download_and_keep_resident().await?);
+            prev_lsn_end = lsn_range.end;
+        }
+        let lsn_range = Range {
+            start: deltas_to_compact
+                .first()
+                .unwrap()
+                .layer_desc()
+                .lsn_range
+                .start,
+            end: deltas_to_compact.last().unwrap().layer_desc().lsn_range.end,
+        };
+
+        info!(
+            "Starting Level0 compaction in LSN range {}-{} for {} layers ({} deltas in total)",
+            lsn_range.start,
+            lsn_range.end,
+            deltas_to_compact.len(),
+            level0_deltas.len()
+        );
+
+        for l in deltas_to_compact.iter() {
+            info!("compact includes {l}");
+        }
+
+        // We don't need the original list of layers anymore. Drop it so that
+        // we don't accidentally use it later in the function.
+        drop(level0_deltas);
+
+        stats.read_lock_held_prerequisites_micros = stats
+            .read_lock_held_spawn_blocking_startup_micros
+            .till_now();
+
+        // Determine N largest holes where N is number of compacted layers.
+        let max_holes = deltas_to_compact.len();
+        let last_record_lsn = self.get_last_record_lsn();
+        let min_hole_range = (target_file_size / page_cache::PAGE_SZ as u64) as i128;
+        let min_hole_coverage_size = 3; // TODO: something more flexible?
+
+        // min-heap (reserve space for one more element added before eviction)
+        let mut heap: BinaryHeap<Hole> = BinaryHeap::with_capacity(max_holes + 1);
+        let mut prev: Option<Key> = None;
+
+        let mut all_keys = Vec::new();
+
+        for l in deltas_to_compact.iter() {
+            all_keys.extend(l.load_keys(ctx).await?);
+        }
+
+        // FIXME: should spawn_blocking the rest of this function
+
+        // The current stdlib sorting implementation is designed in a way where it is
+        // particularly fast where the slice is made up of sorted sub-ranges.
+        all_keys.sort_by_key(|DeltaEntry { key, lsn, .. }| (*key, *lsn));
+
+        stats.read_lock_held_key_sort_micros = stats.read_lock_held_prerequisites_micros.till_now();
+
+        for &DeltaEntry { key: next_key, .. } in all_keys.iter() {
+            if let Some(prev_key) = prev {
+                // just first fast filter
+                if next_key.to_i128() - prev_key.to_i128() >= min_hole_range {
+                    let key_range = prev_key..next_key;
+                    // Measuring hole by just subtraction of i128 representation of key range boundaries
+                    // has not so much sense, because largest holes will corresponds field1/field2 changes.
+                    // But we are mostly interested to eliminate holes which cause generation of excessive image layers.
+                    // That is why it is better to measure size of hole as number of covering image layers.
+                    let coverage_size = layers.image_coverage(&key_range, last_record_lsn).len();
+                    if coverage_size >= min_hole_coverage_size {
+                        heap.push(Hole {
+                            key_range,
+                            coverage_size,
+                        });
+                        if heap.len() > max_holes {
+                            heap.pop(); // remove smallest hole
+                        }
+                    }
+                }
+            }
+            prev = Some(next_key.next());
+        }
+        stats.read_lock_held_compute_holes_micros = stats.read_lock_held_key_sort_micros.till_now();
+        drop_rlock(guard);
+        stats.read_lock_drop_micros = stats.read_lock_held_compute_holes_micros.till_now();
+        let mut holes = heap.into_vec();
+        holes.sort_unstable_by_key(|hole| hole.key_range.start);
+        let mut next_hole = 0; // index of next hole in holes vector
+
+        // This iterator walks through all key-value pairs from all the layers
+        // we're compacting, in key, LSN order.
+        let all_values_iter = all_keys.iter();
+
+        // This iterator walks through all keys and is needed to calculate size used by each key
+        let mut all_keys_iter = all_keys
+            .iter()
+            .map(|DeltaEntry { key, lsn, size, .. }| (*key, *lsn, *size))
+            .coalesce(|mut prev, cur| {
+                // Coalesce keys that belong to the same key pair.
+                // This ensures that compaction doesn't put them
+                // into different layer files.
+                // Still limit this by the target file size,
+                // so that we keep the size of the files in
+                // check.
+                if prev.0 == cur.0 && prev.2 < target_file_size {
+                    prev.2 += cur.2;
+                    Ok(prev)
+                } else {
+                    Err((prev, cur))
+                }
+            });
+
+        // Merge the contents of all the input delta layers into a new set
+        // of delta layers, based on the current partitioning.
+        //
+        // We split the new delta layers on the key dimension. We iterate through the key space, and for each key, check if including the next key to the current output layer we're building would cause the layer to become too large. If so, dump the current output layer and start new one.
+        // It's possible that there is a single key with so many page versions that storing all of them in a single layer file
+        // would be too large. In that case, we also split on the LSN dimension.
+        //
+        // LSN
+        //  ^
+        //  |
+        //  | +-----------+            +--+--+--+--+
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+            |  |  |  |  |
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+     ==>    |  |  |  |  |
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+            |  |  |  |  |
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+            +--+--+--+--+
+        //  |
+        //  +--------------> key
+        //
+        //
+        // If one key (X) has a lot of page versions:
+        //
+        // LSN
+        //  ^
+        //  |                                 (X)
+        //  | +-----------+            +--+--+--+--+
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+            |  |  +--+  |
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+     ==>    |  |  |  |  |
+        //  | |           |            |  |  +--+  |
+        //  | +-----------+            |  |  |  |  |
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+            +--+--+--+--+
+        //  |
+        //  +--------------> key
+        // TODO: this actually divides the layers into fixed-size chunks, not
+        // based on the partitioning.
+        //
+        // TODO: we should also opportunistically materialize and
+        // garbage collect what we can.
+        let mut new_layers = Vec::new();
+        let mut prev_key: Option<Key> = None;
+        let mut writer: Option<DeltaLayerWriter> = None;
+        let mut key_values_total_size = 0u64;
+        let mut dup_start_lsn: Lsn = Lsn::INVALID; // start LSN of layer containing values of the single key
+        let mut dup_end_lsn: Lsn = Lsn::INVALID; // end LSN of layer containing values of the single key
+
+        for &DeltaEntry {
+            key, lsn, ref val, ..
+        } in all_values_iter
+        {
+            let value = val.load(ctx).await?;
+            let same_key = prev_key.map_or(false, |prev_key| prev_key == key);
+            // We need to check key boundaries once we reach next key or end of layer with the same key
+            if !same_key || lsn == dup_end_lsn {
+                let mut next_key_size = 0u64;
+                let is_dup_layer = dup_end_lsn.is_valid();
+                dup_start_lsn = Lsn::INVALID;
+                if !same_key {
+                    dup_end_lsn = Lsn::INVALID;
+                }
+                // Determine size occupied by this key. We stop at next key or when size becomes larger than target_file_size
+                for (next_key, next_lsn, next_size) in all_keys_iter.by_ref() {
+                    next_key_size = next_size;
+                    if key != next_key {
+                        if dup_end_lsn.is_valid() {
+                            // We are writting segment with duplicates:
+                            // place all remaining values of this key in separate segment
+                            dup_start_lsn = dup_end_lsn; // new segments starts where old stops
+                            dup_end_lsn = lsn_range.end; // there are no more values of this key till end of LSN range
+                        }
+                        break;
+                    }
+                    key_values_total_size += next_size;
+                    // Check if it is time to split segment: if total keys size is larger than target file size.
+                    // We need to avoid generation of empty segments if next_size > target_file_size.
+                    if key_values_total_size > target_file_size && lsn != next_lsn {
+                        // Split key between multiple layers: such layer can contain only single key
+                        dup_start_lsn = if dup_end_lsn.is_valid() {
+                            dup_end_lsn // new segment with duplicates starts where old one stops
+                        } else {
+                            lsn // start with the first LSN for this key
+                        };
+                        dup_end_lsn = next_lsn; // upper LSN boundary is exclusive
+                        break;
+                    }
+                }
+                // handle case when loop reaches last key: in this case dup_end is non-zero but dup_start is not set.
+                if dup_end_lsn.is_valid() && !dup_start_lsn.is_valid() {
+                    dup_start_lsn = dup_end_lsn;
+                    dup_end_lsn = lsn_range.end;
+                }
+                if writer.is_some() {
+                    let written_size = writer.as_mut().unwrap().size();
+                    let contains_hole =
+                        next_hole < holes.len() && key >= holes[next_hole].key_range.end;
+                    // check if key cause layer overflow or contains hole...
+                    if is_dup_layer
+                        || dup_end_lsn.is_valid()
+                        || written_size + key_values_total_size > target_file_size
+                        || contains_hole
+                    {
+                        // ... if so, flush previous layer and prepare to write new one
+                        new_layers.push(
+                            writer
+                                .take()
+                                .unwrap()
+                                .finish(prev_key.unwrap().next(), self)
+                                .await?,
+                        );
+                        writer = None;
+
+                        if contains_hole {
+                            // skip hole
+                            next_hole += 1;
+                        }
+                    }
+                }
+                // Remember size of key value because at next iteration we will access next item
+                key_values_total_size = next_key_size;
+            }
+            fail_point!("delta-layer-writer-fail-before-finish", |_| {
+                Err(CompactionError::Other(anyhow::anyhow!(
+                    "failpoint delta-layer-writer-fail-before-finish"
+                )))
+            });
+
+            if !self.shard_identity.is_key_disposable(&key) {
+                if writer.is_none() {
+                    // Create writer if not initiaized yet
+                    writer = Some(
+                        DeltaLayerWriter::new(
+                            self.conf,
+                            self.timeline_id,
+                            self.tenant_shard_id,
+                            key,
+                            if dup_end_lsn.is_valid() {
+                                // this is a layer containing slice of values of the same key
+                                debug!("Create new dup layer {}..{}", dup_start_lsn, dup_end_lsn);
+                                dup_start_lsn..dup_end_lsn
+                            } else {
+                                debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
+                                lsn_range.clone()
+                            },
+                        )
+                        .await?,
+                    );
+                }
+
+                writer.as_mut().unwrap().put_value(key, lsn, value).await?;
+            } else {
+                debug!(
+                    "Dropping key {} during compaction (it belongs on shard {:?})",
+                    key,
+                    self.shard_identity.get_shard_number(&key)
+                );
+            }
+
+            if !new_layers.is_empty() {
+                fail_point!("after-timeline-compacted-first-L1");
+            }
+
+            prev_key = Some(key);
+        }
+        if let Some(writer) = writer {
+            new_layers.push(writer.finish(prev_key.unwrap().next(), self).await?);
+        }
+
+        // Sync layers
+        if !new_layers.is_empty() {
+            // Print a warning if the created layer is larger than double the target size
+            // Add two pages for potential overhead. This should in theory be already
+            // accounted for in the target calculation, but for very small targets,
+            // we still might easily hit the limit otherwise.
+            let warn_limit = target_file_size * 2 + page_cache::PAGE_SZ as u64 * 2;
+            for layer in new_layers.iter() {
+                if layer.layer_desc().file_size > warn_limit {
+                    warn!(
+                        %layer,
+                        "created delta file of size {} larger than double of target of {target_file_size}", layer.layer_desc().file_size
+                    );
+                }
+            }
+
+            // The writer.finish() above already did the fsync of the inodes.
+            // We just need to fsync the directory in which these inodes are linked,
+            // which we know to be the timeline directory.
+            //
+            // We use fatal_err() below because the after writer.finish() returns with success,
+            // the in-memory state of the filesystem already has the layer file in its final place,
+            // and subsequent pageserver code could think it's durable while it really isn't.
+            let timeline_dir = VirtualFile::open(
+                &self
+                    .conf
+                    .timeline_path(&self.tenant_shard_id, &self.timeline_id),
+            )
+            .await
+            .fatal_err("VirtualFile::open for timeline dir fsync");
+            timeline_dir
+                .sync_all()
+                .await
+                .fatal_err("VirtualFile::sync_all timeline dir");
+        }
+
+        stats.write_layer_files_micros = stats.read_lock_drop_micros.till_now();
+        stats.new_deltas_count = Some(new_layers.len());
+        stats.new_deltas_size = Some(new_layers.iter().map(|l| l.layer_desc().file_size).sum());
+
+        match TryInto::<CompactLevel0Phase1Stats>::try_into(stats)
+            .and_then(|stats| serde_json::to_string(&stats).context("serde_json::to_string"))
+        {
+            Ok(stats_json) => {
+                info!(
+                    stats_json = stats_json.as_str(),
+                    "compact_level0_phase1 stats available"
+                )
+            }
+            Err(e) => {
+                warn!("compact_level0_phase1 stats failed to serialize: {:#}", e);
+            }
+        }
+
+        Ok(CompactLevel0Phase1Result {
+            new_layers,
+            deltas_to_compact: deltas_to_compact
+                .into_iter()
+                .map(|x| x.drop_eviction_guard())
+                .collect::<Vec<_>>(),
+        })
+    }
+}
+
+#[derive(Default)]
+struct CompactLevel0Phase1Result {
+    new_layers: Vec<ResidentLayer>,
+    deltas_to_compact: Vec<Layer>,
+}
+
+#[derive(Default)]
+struct CompactLevel0Phase1StatsBuilder {
+    version: Option<u64>,
+    tenant_id: Option<TenantShardId>,
+    timeline_id: Option<TimelineId>,
+    read_lock_acquisition_micros: DurationRecorder,
+    read_lock_held_spawn_blocking_startup_micros: DurationRecorder,
+    read_lock_held_key_sort_micros: DurationRecorder,
+    read_lock_held_prerequisites_micros: DurationRecorder,
+    read_lock_held_compute_holes_micros: DurationRecorder,
+    read_lock_drop_micros: DurationRecorder,
+    write_layer_files_micros: DurationRecorder,
+    level0_deltas_count: Option<usize>,
+    new_deltas_count: Option<usize>,
+    new_deltas_size: Option<u64>,
+}
+
+#[derive(serde::Serialize)]
+struct CompactLevel0Phase1Stats {
+    version: u64,
+    tenant_id: TenantShardId,
+    timeline_id: TimelineId,
+    read_lock_acquisition_micros: RecordedDuration,
+    read_lock_held_spawn_blocking_startup_micros: RecordedDuration,
+    read_lock_held_key_sort_micros: RecordedDuration,
+    read_lock_held_prerequisites_micros: RecordedDuration,
+    read_lock_held_compute_holes_micros: RecordedDuration,
+    read_lock_drop_micros: RecordedDuration,
+    write_layer_files_micros: RecordedDuration,
+    level0_deltas_count: usize,
+    new_deltas_count: usize,
+    new_deltas_size: u64,
+}
+
+impl TryFrom<CompactLevel0Phase1StatsBuilder> for CompactLevel0Phase1Stats {
+    type Error = anyhow::Error;
+
+    fn try_from(value: CompactLevel0Phase1StatsBuilder) -> Result<Self, Self::Error> {
+        Ok(Self {
+            version: value.version.ok_or_else(|| anyhow!("version not set"))?,
+            tenant_id: value
+                .tenant_id
+                .ok_or_else(|| anyhow!("tenant_id not set"))?,
+            timeline_id: value
+                .timeline_id
+                .ok_or_else(|| anyhow!("timeline_id not set"))?,
+            read_lock_acquisition_micros: value
+                .read_lock_acquisition_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_acquisition_micros not set"))?,
+            read_lock_held_spawn_blocking_startup_micros: value
+                .read_lock_held_spawn_blocking_startup_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_held_spawn_blocking_startup_micros not set"))?,
+            read_lock_held_key_sort_micros: value
+                .read_lock_held_key_sort_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_held_key_sort_micros not set"))?,
+            read_lock_held_prerequisites_micros: value
+                .read_lock_held_prerequisites_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_held_prerequisites_micros not set"))?,
+            read_lock_held_compute_holes_micros: value
+                .read_lock_held_compute_holes_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_held_compute_holes_micros not set"))?,
+            read_lock_drop_micros: value
+                .read_lock_drop_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_drop_micros not set"))?,
+            write_layer_files_micros: value
+                .write_layer_files_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("write_layer_files_micros not set"))?,
+            level0_deltas_count: value
+                .level0_deltas_count
+                .ok_or_else(|| anyhow!("level0_deltas_count not set"))?,
+            new_deltas_count: value
+                .new_deltas_count
+                .ok_or_else(|| anyhow!("new_deltas_count not set"))?,
+            new_deltas_size: value
+                .new_deltas_size
+                .ok_or_else(|| anyhow!("new_deltas_size not set"))?,
+        })
+    }
+}
+
 impl Timeline {
    /// Entry point for new tiered compaction algorithm.
    ///
@@ -134,7 +830,6 @@ struct ResidentDeltaLayer(ResidentLayer);
 #[derive(Clone)]
 struct ResidentImageLayer(ResidentLayer);

-#[async_trait]
 impl CompactionJobExecutor for TimelineAdaptor {
    type Key = crate::repository::Key;

--- a/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs
+++ b/pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs
@@ -343,23 +343,6 @@ pub(super) async fn handle_walreceiver_connection(
                            modification.commit(&ctx).await?;
                            uncommitted_records = 0;
                            filtered_records = 0;
-
-                            //
-                            // We should check checkpoint distance after appending each ingest_batch_size bytes because otherwise
-                            // layer size can become much larger than `checkpoint_distance`.
-                            // It can append because wal-sender is sending WAL using 125kb chucks and some WAL records can cause writing large
-                            // amount of data to key-value storage. So performing this check only after processing
-                            // all WAL records in the chunk, can cause huge L0 layer files.
-                            //
-                            timeline
-                                .check_checkpoint_distance()
-                                .await
-                                .with_context(|| {
-                                    format!(
-                                        "Failed to check checkpoint distance for timeline {}",
-                                        timeline.timeline_id
-                                    )
-                                })?;
                        }
                    }

@@ -406,15 +389,16 @@ pub(super) async fn handle_walreceiver_connection(
            }
        }

-        timeline
-            .check_checkpoint_distance()
-            .await
-            .with_context(|| {
-                format!(
-                    "Failed to check checkpoint distance for timeline {}",
-                    timeline.timeline_id
-                )
-            })?;
+        {
+            // This is a hack. It piggybacks on the keepalive messages sent by the
+            // safekeeper in order to enforce `checkpoint_timeout` on the currently
+            // open layer. This hack doesn't provide a bound on the total size of
+            // in-memory layers on a pageserver. See https://github.com/neondatabase/neon/issues/6916.
+            let mut writer = timeline.writer().await;
+            if let Err(err) = writer.tick().await {
+                warn!("Timeline writer tick failed: {err}");
+            }
+        }

        if let Some(last_lsn) = status_update {
            let timeline_remote_consistent_lsn = timeline
--- a/pageserver/src/tenant/vectored_blob_io.rs
+++ b/pageserver/src/tenant/vectored_blob_io.rs
@@ -128,7 +128,7 @@ impl VectoredReadBuilder {
 pub enum BlobFlag {
    None,
    Ignore,
-    Replaces,
+    ReplaceAll,
 }

 /// Planner for vectored blob reads.
@@ -170,7 +170,7 @@ impl VectoredReadPlanner {
    /// incorrect data to the user.
    ///
    /// The `flag` argument has two interesting values:
-    /// * [`BlobFlag::Replaces`]: The blob for this key should replace all existing blobs.
+    /// * [`BlobFlag::ReplaceAll`]: The blob for this key should replace all existing blobs.
    /// This is used for WAL records that `will_init`.
    /// * [`BlobFlag::Ignore`]: This blob should not be included in the read. This happens
    /// if the blob is cached.
@@ -204,7 +204,7 @@ impl VectoredReadPlanner {
                let blobs_for_key = self.blobs.entry(key).or_default();
                blobs_for_key.push((lsn, start_offset, end_offset));
            }
-            BlobFlag::Replaces => {
+            BlobFlag::ReplaceAll => {
                let blobs_for_key = self.blobs.entry(key).or_default();
                blobs_for_key.clear();
                blobs_for_key.push((lsn, start_offset, end_offset));
@@ -411,10 +411,10 @@ mod tests {
        let blob_descriptions = vec![
            (first_key, lsn, 0, BlobFlag::None),    // First in read 1
            (first_key, lsn, 1024, BlobFlag::None), // Last in read 1
-            (second_key, lsn, 2 * 1024, BlobFlag::Replaces),
+            (second_key, lsn, 2 * 1024, BlobFlag::ReplaceAll),
            (second_key, lsn, 3 * 1024, BlobFlag::None),
-            (second_key, lsn, 4 * 1024, BlobFlag::Replaces), // First in read 2
-            (second_key, lsn, 5 * 1024, BlobFlag::None),     // Last in read 2
+            (second_key, lsn, 4 * 1024, BlobFlag::ReplaceAll), // First in read 2
+            (second_key, lsn, 5 * 1024, BlobFlag::None),       // Last in read 2
        ];

        let ranges = [&blob_descriptions[0..2], &blob_descriptions[4..]];
--- a/pageserver/src/virtual_file.rs
+++ b/pageserver/src/virtual_file.rs
@@ -17,20 +17,21 @@ use crate::tenant::TENANTS_SEGMENT_NAME;
 use camino::{Utf8Path, Utf8PathBuf};
 use once_cell::sync::OnceCell;
 use pageserver_api::shard::TenantShardId;
-use std::fs::{self, File};
+use std::fs::File;
 use std::io::{Error, ErrorKind, Seek, SeekFrom};
 use tokio_epoll_uring::{BoundedBuf, IoBuf, IoBufMut, Slice};

 use std::os::fd::{AsRawFd, FromRawFd, IntoRawFd, OwnedFd, RawFd};
-use std::os::unix::fs::FileExt;
 use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
 use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
 use tokio::time::Instant;

 pub use pageserver_api::models::virtual_file as api;
 pub(crate) mod io_engine;
+mod metadata;
 mod open_options;
 pub(crate) use io_engine::IoEngineKind;
+pub(crate) use metadata::Metadata;
 pub(crate) use open_options::*;

 ///
@@ -435,13 +436,25 @@ impl VirtualFile {

    /// Call File::sync_all() on the underlying File.
    pub async fn sync_all(&self) -> Result<(), Error> {
-        with_file!(self, StorageIoOperation::Fsync, |file_guard| file_guard
-            .with_std_file(|std_file| std_file.sync_all()))
+        with_file!(self, StorageIoOperation::Fsync, |file_guard| {
+            let (_file_guard, res) = io_engine::get().sync_all(file_guard).await;
+            res
+        })
    }

-    pub async fn metadata(&self) -> Result<fs::Metadata, Error> {
-        with_file!(self, StorageIoOperation::Metadata, |file_guard| file_guard
-            .with_std_file(|std_file| std_file.metadata()))
+    /// Call File::sync_data() on the underlying File.
+    pub async fn sync_data(&self) -> Result<(), Error> {
+        with_file!(self, StorageIoOperation::Fsync, |file_guard| {
+            let (_file_guard, res) = io_engine::get().sync_data(file_guard).await;
+            res
+        })
+    }
+
+    pub async fn metadata(&self) -> Result<Metadata, Error> {
+        with_file!(self, StorageIoOperation::Metadata, |file_guard| {
+            let (_file_guard, res) = io_engine::get().metadata(file_guard).await;
+            res
+        })
    }

    /// Helper function internal to `VirtualFile` that looks up the underlying File,
@@ -579,7 +592,7 @@ impl VirtualFile {
    }

    // Copied from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#219-235
-    pub async fn write_all_at<B: BoundedBuf>(
+    pub async fn write_all_at<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
        &self,
        buf: B,
        mut offset: u64,
@@ -590,8 +603,9 @@ impl VirtualFile {
        }
        let mut buf = buf.slice(0..buf_len);
        while !buf.is_empty() {
-            // TODO: push `buf` further down
-            match self.write_at(&buf, offset).await {
+            let res;
+            (buf, res) = self.write_at(buf, offset).await;
+            match res {
                Ok(0) => {
                    return (
                        Slice::into_inner(buf),
@@ -605,7 +619,7 @@ impl VirtualFile {
                    buf = buf.slice(n..);
                    offset += n as u64;
                }
-                Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
+                Err(e) if e.kind() == std::io::ErrorKind::Interrupted => {}
                Err(e) => return (Slice::into_inner(buf), Err(e)),
            }
        }
@@ -616,15 +630,19 @@ impl VirtualFile {
    /// Returns the IoBuf that is underlying the BoundedBuf `buf`.
    /// I.e., the returned value's `bytes_init()` method returns something different than the `bytes_init()` that was passed in.
    /// It's quite brittle and easy to mis-use, so, we return the size in the Ok() variant.
-    pub async fn write_all<B: BoundedBuf>(&mut self, buf: B) -> (B::Buf, Result<usize, Error>) {
+    pub async fn write_all<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
+        &mut self,
+        buf: B,
+    ) -> (B::Buf, Result<usize, Error>) {
        let nbytes = buf.bytes_init();
        if nbytes == 0 {
            return (Slice::into_inner(buf.slice_full()), Ok(0));
        }
        let mut buf = buf.slice(0..nbytes);
        while !buf.is_empty() {
-            // TODO: push `Slice` further down
-            match self.write(&buf).await {
+            let res;
+            (buf, res) = self.write(buf).await;
+            match res {
                Ok(0) => {
                    return (
                        Slice::into_inner(buf),
@@ -644,11 +662,18 @@ impl VirtualFile {
        (Slice::into_inner(buf), Ok(nbytes))
    }

-    async fn write(&mut self, buf: &[u8]) -> Result<usize, std::io::Error> {
+    async fn write<B: IoBuf + Send>(
+        &mut self,
+        buf: Slice<B>,
+    ) -> (Slice<B>, Result<usize, std::io::Error>) {
        let pos = self.pos;
-        let n = self.write_at(buf, pos).await?;
+        let (buf, res) = self.write_at(buf, pos).await;
+        let n = match res {
+            Ok(n) => n,
+            Err(e) => return (buf, Err(e)),
+        };
        self.pos += n as u64;
-        Ok(n)
+        (buf, Ok(n))
    }

    pub(crate) async fn read_at<B>(&self, buf: B, offset: u64) -> (B, Result<usize, Error>)
@@ -676,16 +701,30 @@ impl VirtualFile {
        })
    }

-    async fn write_at(&self, buf: &[u8], offset: u64) -> Result<usize, Error> {
-        let result = with_file!(self, StorageIoOperation::Write, |file_guard| {
-            file_guard.with_std_file(|std_file| std_file.write_at(buf, offset))
-        });
-        if let Ok(size) = result {
-            STORAGE_IO_SIZE
-                .with_label_values(&["write", &self.tenant_id, &self.shard_id, &self.timeline_id])
-                .add(size as i64);
-        }
-        result
+    async fn write_at<B: IoBuf + Send>(
+        &self,
+        buf: Slice<B>,
+        offset: u64,
+    ) -> (Slice<B>, Result<usize, Error>) {
+        let file_guard = match self.lock_file().await {
+            Ok(file_guard) => file_guard,
+            Err(e) => return (buf, Err(e)),
+        };
+        observe_duration!(StorageIoOperation::Write, {
+            let ((_file_guard, buf), result) =
+                io_engine::get().write_at(file_guard, offset, buf).await;
+            if let Ok(size) = result {
+                STORAGE_IO_SIZE
+                    .with_label_values(&[
+                        "write",
+                        &self.tenant_id,
+                        &self.shard_id,
+                        &self.timeline_id,
+                    ])
+                    .add(size as i64);
+            }
+            (buf, result)
+        })
    }
 }

@@ -1083,6 +1122,7 @@ mod tests {
    use rand::Rng;
    use std::future::Future;
    use std::io::Write;
+    use std::os::unix::fs::FileExt;
    use std::sync::Arc;

    enum MaybeVirtualFile {
@@ -1103,7 +1143,11 @@ mod tests {
                MaybeVirtualFile::File(file) => file.read_exact_at(&mut buf, offset).map(|()| buf),
            }
        }
-        async fn write_all_at<B: BoundedBuf>(&self, buf: B, offset: u64) -> Result<(), Error> {
+        async fn write_all_at<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
+            &self,
+            buf: B,
+            offset: u64,
+        ) -> Result<(), Error> {
            match self {
                MaybeVirtualFile::VirtualFile(file) => {
                    let (_buf, res) = file.write_all_at(buf, offset).await;
@@ -1124,7 +1168,10 @@ mod tests {
                MaybeVirtualFile::File(file) => file.seek(pos),
            }
        }
-        async fn write_all<B: BoundedBuf>(&mut self, buf: B) -> Result<(), Error> {
+        async fn write_all<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
+            &mut self,
+            buf: B,
+        ) -> Result<(), Error> {
            match self {
                MaybeVirtualFile::VirtualFile(file) => {
                    let (_buf, res) = file.write_all(buf).await;
--- a/pageserver/src/virtual_file/io_engine.rs
+++ b/pageserver/src/virtual_file/io_engine.rs
@@ -7,6 +7,9 @@
 //!
 //! Then use [`get`] and  [`super::OpenOptions`].

+use tokio_epoll_uring::{IoBuf, Slice};
+use tracing::Instrument;
+
 pub(crate) use super::api::IoEngineKind;
 #[derive(Clone, Copy)]
 #[repr(u8)]
@@ -61,7 +64,8 @@ pub(super) fn init(engine_kind: IoEngineKind) {
    set(engine_kind);
 }

-pub(super) fn get() -> IoEngine {
+/// Longer-term, this API should only be used by [`super::VirtualFile`].
+pub(crate) fn get() -> IoEngine {
    let cur = IoEngine::try_from(IO_ENGINE.load(Ordering::Relaxed)).unwrap();
    if cfg!(test) {
        let env_var_name = "NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE";
@@ -98,7 +102,17 @@ use std::{
    sync::atomic::{AtomicU8, Ordering},
 };

-use super::FileGuard;
+use super::{FileGuard, Metadata};
+
+#[cfg(target_os = "linux")]
+fn epoll_uring_error_to_std(e: tokio_epoll_uring::Error<std::io::Error>) -> std::io::Error {
+    match e {
+        tokio_epoll_uring::Error::Op(e) => e,
+        tokio_epoll_uring::Error::System(system) => {
+            std::io::Error::new(std::io::ErrorKind::Other, system)
+        }
+    }
+}

 impl IoEngine {
    pub(super) async fn read_at<B>(
@@ -133,16 +147,109 @@ impl IoEngine {
            IoEngine::TokioEpollUring => {
                let system = tokio_epoll_uring::thread_local_system().await;
                let (resources, res) = system.read(file_guard, offset, buf).await;
+                (resources, res.map_err(epoll_uring_error_to_std))
+            }
+        }
+    }
+    pub(super) async fn sync_all(&self, file_guard: FileGuard) -> (FileGuard, std::io::Result<()>) {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            IoEngine::StdFs => {
+                let res = file_guard.with_std_file(|std_file| std_file.sync_all());
+                (file_guard, res)
+            }
+            #[cfg(target_os = "linux")]
+            IoEngine::TokioEpollUring => {
+                let system = tokio_epoll_uring::thread_local_system().await;
+                let (resources, res) = system.fsync(file_guard).await;
+                (resources, res.map_err(epoll_uring_error_to_std))
+            }
+        }
+    }
+    pub(super) async fn sync_data(
+        &self,
+        file_guard: FileGuard,
+    ) -> (FileGuard, std::io::Result<()>) {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            IoEngine::StdFs => {
+                let res = file_guard.with_std_file(|std_file| std_file.sync_data());
+                (file_guard, res)
+            }
+            #[cfg(target_os = "linux")]
+            IoEngine::TokioEpollUring => {
+                let system = tokio_epoll_uring::thread_local_system().await;
+                let (resources, res) = system.fdatasync(file_guard).await;
+                (resources, res.map_err(epoll_uring_error_to_std))
+            }
+        }
+    }
+    pub(super) async fn metadata(
+        &self,
+        file_guard: FileGuard,
+    ) -> (FileGuard, std::io::Result<Metadata>) {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            IoEngine::StdFs => {
+                let res =
+                    file_guard.with_std_file(|std_file| std_file.metadata().map(Metadata::from));
+                (file_guard, res)
+            }
+            #[cfg(target_os = "linux")]
+            IoEngine::TokioEpollUring => {
+                let system = tokio_epoll_uring::thread_local_system().await;
+                let (resources, res) = system.statx(file_guard).await;
                (
                    resources,
-                    res.map_err(|e| match e {
-                        tokio_epoll_uring::Error::Op(e) => e,
-                        tokio_epoll_uring::Error::System(system) => {
-                            std::io::Error::new(std::io::ErrorKind::Other, system)
-                        }
-                    }),
+                    res.map_err(epoll_uring_error_to_std).map(Metadata::from),
                )
            }
        }
    }
+    pub(super) async fn write_at<B: IoBuf + Send>(
+        &self,
+        file_guard: FileGuard,
+        offset: u64,
+        buf: Slice<B>,
+    ) -> ((FileGuard, Slice<B>), std::io::Result<usize>) {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            IoEngine::StdFs => {
+                let result = file_guard.with_std_file(|std_file| std_file.write_at(&buf, offset));
+                ((file_guard, buf), result)
+            }
+            #[cfg(target_os = "linux")]
+            IoEngine::TokioEpollUring => {
+                let system = tokio_epoll_uring::thread_local_system().await;
+                let (resources, res) = system.write(file_guard, offset, buf).await;
+                (resources, res.map_err(epoll_uring_error_to_std))
+            }
+        }
+    }
+
+    /// If we switch a user of [`tokio::fs`] to use [`super::io_engine`],
+    /// they'd start blocking the executor thread if [`IoEngine::StdFs`] is configured
+    /// whereas before the switch to [`super::io_engine`], that wasn't the case.
+    /// This method helps avoid such a regression.
+    ///
+    /// Panics if the `spawn_blocking` fails, see [`tokio::task::JoinError`] for reasons why that can happen.
+    pub(crate) async fn spawn_blocking_and_block_on_if_std<Fut, R>(&self, work: Fut) -> R
+    where
+        Fut: 'static + Send + std::future::Future<Output = R>,
+        R: 'static + Send,
+    {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            IoEngine::StdFs => {
+                let span = tracing::info_span!("spawn_blocking_block_on_if_std");
+                tokio::task::spawn_blocking({
+                    move || tokio::runtime::Handle::current().block_on(work.instrument(span))
+                })
+                .await
+                .expect("failed to join blocking code most likely it panicked, panicking as well")
+            }
+            #[cfg(target_os = "linux")]
+            IoEngine::TokioEpollUring => work.await,
+        }
+    }
 }
--- a/pageserver/src/virtual_file/metadata.rs
+++ b/pageserver/src/virtual_file/metadata.rs
@@ -0,0 +1,30 @@
+use std::fs;
+
+pub enum Metadata {
+    StdFs(fs::Metadata),
+    #[cfg(target_os = "linux")]
+    TokioEpollUring(Box<tokio_epoll_uring::ops::statx::statx>),
+}
+
+#[cfg(target_os = "linux")]
+impl From<Box<tokio_epoll_uring::ops::statx::statx>> for Metadata {
+    fn from(value: Box<tokio_epoll_uring::ops::statx::statx>) -> Self {
+        Metadata::TokioEpollUring(value)
+    }
+}
+
+impl From<std::fs::Metadata> for Metadata {
+    fn from(value: std::fs::Metadata) -> Self {
+        Metadata::StdFs(value)
+    }
+}
+
+impl Metadata {
+    pub fn len(&self) -> u64 {
+        match self {
+            Metadata::StdFs(metadata) => metadata.len(),
+            #[cfg(target_os = "linux")]
+            Metadata::TokioEpollUring(statx) => statx.stx_size,
+        }
+    }
+}
--- a/pgxn/hnsw/hnsw.c
+++ b/pgxn/hnsw/hnsw.c
@@ -149,7 +149,7 @@ hnsw_check_available_memory(Size requested)
 	struct sysinfo si;
 	Size total;
 	if (sysinfo(&si) < 0)
-		elog(ERROR, "Failed to get amount of RAM: %n");
+		elog(ERROR, "Failed to get amount of RAM: %m");

 	total = si.totalram*si.mem_unit;
 	if ((Size)NBuffers*BLCKSZ + requested >= total)
--- a/pgxn/neon/Makefile
+++ b/pgxn/neon/Makefile
@@ -21,7 +21,7 @@ SHLIB_LINK_INTERNAL = $(libpq)
 SHLIB_LINK = -lcurl

 EXTENSION = neon
-DATA = neon--1.0.sql neon--1.0--1.1.sql neon--1.1--1.2.sql neon--1.2--1.3.sql
+DATA = neon--1.0.sql neon--1.0--1.1.sql neon--1.1--1.2.sql neon--1.2--1.3.sql neon--1.3--1.2.sql neon--1.2--1.1.sql neon--1.1--1.0.sql
 PGFILEDESC = "neon - cloud storage for PostgreSQL"

 EXTRA_CLEAN = \
--- a/pgxn/neon/control_plane_connector.c
+++ b/pgxn/neon/control_plane_connector.c
@@ -35,6 +35,7 @@
 #include "utils/memutils.h"
 #include "utils/jsonb.h"

+#include "control_plane_connector.h"
 #include "neon_utils.h"

 static ProcessUtility_hook_type PreviousProcessUtilityHook = NULL;
@@ -113,6 +114,8 @@ ConstructDeltaMessage()
 	if (RootTable.db_table)
 	{
 		JsonbValue	dbs;
+		HASH_SEQ_STATUS status;
+		DbEntry    *entry;

 		dbs.type = jbvString;
 		dbs.val.string.val = "dbs";
@@ -120,9 +123,6 @@ ConstructDeltaMessage()
 		pushJsonbValue(&state, WJB_KEY, &dbs);
 		pushJsonbValue(&state, WJB_BEGIN_ARRAY, NULL);

-		HASH_SEQ_STATUS status;
-		DbEntry    *entry;
-
 		hash_seq_init(&status, RootTable.db_table);
 		while ((entry = hash_seq_search(&status)) != NULL)
 		{
@@ -168,8 +168,9 @@ ConstructDeltaMessage()
 #else
 				const char *logdetail;
 #endif
+				char	   *encrypted_password;
 				PushKeyValue(&state, "password", (char *) entry->password);
-				char	   *encrypted_password = get_role_password(entry->name, &logdetail);
+				encrypted_password = get_role_password(entry->name, &logdetail);

 				if (encrypted_password)
 				{
@@ -831,7 +832,7 @@ NeonProcessUtility(
 	}
 }

-extern void
+void
 InitControlPlaneConnector()
 {
 	PreviousProcessUtilityHook = ProcessUtility_hook;
--- a/pgxn/neon/control_plane_connector.h
+++ b/pgxn/neon/control_plane_connector.h
@@ -1,6 +1,6 @@
 #ifndef CONTROL_PLANE_CONNECTOR_H
 #define CONTROL_PLANE_CONNECTOR_H

-void		InitControlPlaneConnector();
+void		InitControlPlaneConnector(void);

 #endif
--- a/pgxn/neon/extension_server.c
+++ b/pgxn/neon/extension_server.c
@@ -14,6 +14,7 @@

 #include "utils/guc.h"

+#include "extension_server.h" 
 #include "neon_utils.h"

 static int	extension_server_port = 0;
@@ -37,7 +38,6 @@ neon_download_extension_file_http(const char *filename, bool is_library)

 	CURLcode	res;
 	char	   *compute_ctl_url;
-	char	   *postdata;
 	bool		ret = false;

 	if (handle == NULL)
--- a/pgxn/neon/extension_server.h
+++ b/pgxn/neon/extension_server.h
@@ -0,0 +1,17 @@
+/*-------------------------------------------------------------------------
+ *
+ * extension_server.h
+ *	  Request compute_ctl to download extension files.
+ *
+ * IDENTIFICATION
+ *	 contrib/neon/extension_server.h
+ *
+ *-------------------------------------------------------------------------
+ */
+
+#ifndef EXTENSION_SERVER_H
+#define EXTENSION_SERVER_H
+
+void pg_init_extension_server(void);
+
+#endif							/* EXTENSION_SERVER_H */
--- a/pgxn/neon/libpagestore.c
+++ b/pgxn/neon/libpagestore.c
@@ -316,6 +316,7 @@ pageserver_connect(shardno_t shard_no, int elevel)
 	static uint64_t delay_us = MIN_RECONNECT_INTERVAL_USEC;
 	TimestampTz now;
 	uint64_t	us_since_last_connect;
+	bool	broke_from_loop = false;

 	Assert(page_servers[shard_no].conn == NULL);

@@ -418,7 +419,9 @@ pageserver_connect(shardno_t shard_no, int elevel)

 					neon_shard_log(shard_no, elevel, "could not complete handshake with pageserver: %s",
 								   msg);
-					return false;
+					/* Returning from inside PG_TRY is bad, so we break/return later */
+					broke_from_loop = true;
+					break;
 				}
 			}
 		}
@@ -431,6 +434,11 @@ pageserver_connect(shardno_t shard_no, int elevel)
 	}
 	PG_END_TRY();

+	if (broke_from_loop)
+	{
+		return false;
+	}
+
 	neon_shard_log(shard_no, LOG, "libpagestore: connected to '%s'", connstr);
 	page_servers[shard_no].conn = conn;
 	page_servers[shard_no].wes = wes;
--- a/pgxn/neon/neon--1.1--1.0.sql
+++ b/pgxn/neon/neon--1.1--1.0.sql
@@ -0,0 +1,6 @@
+-- the order of operations is important here
+-- because the view depends on the function
+
+DROP VIEW IF EXISTS neon_lfc_stats CASCADE;
+
+DROP FUNCTION IF EXISTS neon_get_lfc_stats CASCADE;
--- a/pgxn/neon/neon--1.2--1.1.sql
+++ b/pgxn/neon/neon--1.2--1.1.sql
@@ -0,0 +1 @@
+DROP VIEW IF EXISTS NEON_STAT_FILE_CACHE CASCADE;
--- a/pgxn/neon/neon--1.3--1.2.sql
+++ b/pgxn/neon/neon--1.3--1.2.sql
@@ -0,0 +1 @@
+DROP FUNCTION IF EXISTS approximate_working_set_size(bool) CASCADE;
--- a/pgxn/neon/neon.c
+++ b/pgxn/neon/neon.c
@@ -29,6 +29,7 @@
 #include "utils/guc.h"
 #include "utils/wait_event.h"

+#include "extension_server.h"
 #include "neon.h"
 #include "walproposer.h"
 #include "pagestore_client.h"
@@ -94,7 +95,6 @@ get_num_snap_files_lsn_threshold(void)
 	DIR		   *dirdesc;
 	struct dirent *de;
 	char	   *snap_path = "pg_logical/snapshots/";
-	int			cnt = 0;
 	int			lsns_allocated = 1024;
 	int			lsns_num = 0;
 	XLogRecPtr *lsns;
@@ -160,9 +160,6 @@ get_num_snap_files_lsn_threshold(void)
 PGDLLEXPORT void
 LogicalSlotsMonitorMain(Datum main_arg)
 {
-	TimestampTz now,
-				last_checked;
-
 	/* Establish signal handlers. */
 	pqsignal(SIGUSR1, procsignal_sigusr1_handler);
 	pqsignal(SIGHUP, SignalHandlerForConfigReload);
--- a/pgxn/neon/neon.h
+++ b/pgxn/neon/neon.h
@@ -25,12 +25,11 @@ extern int	wal_acceptor_connection_timeout;
 extern void pg_init_libpagestore(void);
 extern void pg_init_walproposer(void);

-extern void pg_init_extension_server(void);
-
 extern uint64 BackpressureThrottlingTime(void);
 extern void replication_feedback_get_lsns(XLogRecPtr *writeLsn, XLogRecPtr *flushLsn, XLogRecPtr *applyLsn);

 extern void PGDLLEXPORT WalProposerSync(int argc, char *argv[]);
 extern void PGDLLEXPORT WalProposerMain(Datum main_arg);
+PGDLLEXPORT void LogicalSlotsMonitorMain(Datum main_arg);

 #endif							/* NEON_H */
--- a/pgxn/neon/neon_utils.c
+++ b/pgxn/neon/neon_utils.c
@@ -6,6 +6,7 @@

 #include "postgres.h"

+#include "neon_utils.h"
 #include "lib/stringinfo.h"
 #include "libpq/pqformat.h"

@@ -14,7 +15,7 @@
 *
 * Returns -1 if the character is not a hexadecimal digit.
 */
-int
+static int
 HexDecodeChar(char c)
 {
 	if (c >= '0' && c <= '9')
--- a/pgxn/neon/neon_utils.h
+++ b/pgxn/neon/neon_utils.h
@@ -12,7 +12,7 @@ uint32		pq_getmsgint32_le(StringInfo msg);
 uint64		pq_getmsgint64_le(StringInfo msg);
 void		pq_sendint32_le(StringInfo buf, uint32 i);
 void		pq_sendint64_le(StringInfo buf, uint64 i);
-extern void disable_core_dump();
+void        disable_core_dump(void);

 #ifndef WALPROPOSER_LIB

--- a/pgxn/neon/pagestore_smgr.c
+++ b/pgxn/neon/pagestore_smgr.c
@@ -1888,7 +1888,6 @@ neon_zeroextend(SMgrRelation reln, ForkNumber forkNum, BlockNumber blocknum,
 				int nblocks, bool skipFsync)
 {
 	const PGAlignedBlock buffer = {0};
-	BlockNumber curblocknum = blocknum;
 	int			remblocks = nblocks;
 	XLogRecPtr	lsn = 0;

--- a/pgxn/neon/walproposer.c
+++ b/pgxn/neon/walproposer.c
@@ -1220,7 +1220,7 @@ PrepareAppendRequest(WalProposer *wp, AppendRequestHeader *req, XLogRecPtr begin
 	req->epochStartLsn = wp->propEpochStartLsn;
 	req->beginLsn = beginLsn;
 	req->endLsn = endLsn;
-	req->commitLsn = GetAcknowledgedByQuorumWALPosition(wp);
+	req->commitLsn = wp->commitLsn;
 	req->truncateLsn = wp->truncateLsn;
 	req->proposerId = wp->greetRequest.proposerId;
 }
@@ -1405,7 +1405,7 @@ static bool
 RecvAppendResponses(Safekeeper *sk)
 {
 	WalProposer *wp = sk->wp;
-	XLogRecPtr	minQuorumLsn;
+	XLogRecPtr	newCommitLsn;
 	bool		readAnything = false;

 	while (true)
@@ -1444,23 +1444,24 @@ RecvAppendResponses(Safekeeper *sk)
 	if (!readAnything)
 		return sk->state == SS_ACTIVE;

-	HandleSafekeeperResponse(wp);
-
+	/* update commit_lsn */
+	newCommitLsn = GetAcknowledgedByQuorumWALPosition(wp);
 	/*
-	 * Also send the new commit lsn to all the safekeepers.
+	 * Send the new value to all safekeepers.
 	 */
-	minQuorumLsn = GetAcknowledgedByQuorumWALPosition(wp);
-	if (minQuorumLsn > wp->lastSentCommitLsn)
+	if (newCommitLsn > wp->commitLsn)
 	{
+		wp->commitLsn = newCommitLsn;
 		BroadcastAppendRequest(wp);
-		wp->lastSentCommitLsn = minQuorumLsn;
 	}

+	HandleSafekeeperResponse(wp);
+
 	return sk->state == SS_ACTIVE;
 }

 /* Parse a PageserverFeedback message, or the PageserverFeedback part of an AppendResponse */
-void
+static void
 ParsePageserverFeedbackMessage(WalProposer *wp, StringInfo reply_message, PageserverFeedback *rf)
 {
 	uint8		nkeys;
@@ -1590,9 +1591,9 @@ GetAcknowledgedByQuorumWALPosition(WalProposer *wp)
 Safekeeper *
 GetDonor(WalProposer *wp, XLogRecPtr *donor_lsn)
 {
-	*donor_lsn = InvalidXLogRecPtr;
 	Safekeeper *donor = NULL;
 	int			i;
+	*donor_lsn = InvalidXLogRecPtr;

 	if (wp->n_votes < wp->quorum)
 	{
@@ -1632,11 +1633,9 @@ GetDonor(WalProposer *wp, XLogRecPtr *donor_lsn)
 static void
 HandleSafekeeperResponse(WalProposer *wp)
 {
-	XLogRecPtr	minQuorumLsn;
 	XLogRecPtr	candidateTruncateLsn;

-	minQuorumLsn = GetAcknowledgedByQuorumWALPosition(wp);
-	wp->api.process_safekeeper_feedback(wp, minQuorumLsn);
+	wp->api.process_safekeeper_feedback(wp);

 	/*
 	 * Try to advance truncateLsn -- the last record flushed to all
@@ -1649,7 +1648,7 @@ HandleSafekeeperResponse(WalProposer *wp)
 	 * can't commit entries from previous term' in Raft); 2)
 	 */
 	candidateTruncateLsn = CalculateMinFlushLsn(wp);
-	candidateTruncateLsn = Min(candidateTruncateLsn, minQuorumLsn);
+	candidateTruncateLsn = Min(candidateTruncateLsn, wp->commitLsn);
 	if (candidateTruncateLsn > wp->truncateLsn)
 	{
 		wp->truncateLsn = candidateTruncateLsn;
--- a/pgxn/neon/walproposer.h
+++ b/pgxn/neon/walproposer.h
@@ -564,7 +564,7 @@ typedef struct walproposer_api
 	 * backpressure feedback and to confirm WAL persistence (has been commited
 	 * on the quorum of safekeepers).
 	 */
-	void		(*process_safekeeper_feedback) (WalProposer *wp, XLogRecPtr commitLsn);
+	void		(*process_safekeeper_feedback) (WalProposer *wp);

 	/*
 	 * Write a log message to the internal log processor. This is used only
@@ -646,8 +646,8 @@ typedef struct WalProposer
 	/* WAL has been generated up to this point */
 	XLogRecPtr	availableLsn;

-	/* last commitLsn broadcasted to safekeepers */
-	XLogRecPtr	lastSentCommitLsn;
+	/* cached GetAcknowledgedByQuorumWALPosition result */
+	XLogRecPtr	commitLsn;

 	ProposerGreeting greetRequest;

--- a/pgxn/neon/walproposer_pg.c
+++ b/pgxn/neon/walproposer_pg.c
@@ -68,6 +68,8 @@ static WalproposerShmemState *walprop_shared;
 static WalProposerConfig walprop_config;
 static XLogRecPtr sentPtr = InvalidXLogRecPtr;
 static const walproposer_api walprop_pg;
+static volatile sig_atomic_t got_SIGUSR2 = false;
+static bool reported_sigusr2 = false;

 static void nwp_shmem_startup_hook(void);
 static void nwp_register_gucs(void);
@@ -101,6 +103,8 @@ static void add_nwr_event_set(Safekeeper *sk, uint32 events);
 static void update_nwr_event_set(Safekeeper *sk, uint32 events);
 static void rm_safekeeper_event_set(Safekeeper *to_remove, bool is_sk);

+static void CheckGracefulShutdown(WalProposer *wp);
+
 static XLogRecPtr GetLogRepRestartLSN(WalProposer *wp);

 static void
@@ -398,7 +402,7 @@ walprop_pg_get_shmem_state(WalProposer *wp)
 	return walprop_shared;
 }

-void
+static void
 replication_feedback_set(PageserverFeedback *rf)
 {
 	SpinLockAcquire(&walprop_shared->mutex);
@@ -492,6 +496,24 @@ walprop_pg_init_standalone_sync_safekeepers(void)
 	BackgroundWorkerUnblockSignals();
 }

+/*
+ * We pretend to be a walsender process, and the lifecycle of a walsender is
+ * slightly different than other procesess. At shutdown, walsender processes
+ * stay alive until the very end, after the checkpointer has written the
+ * shutdown checkpoint. When the checkpointer exits, the postmaster sends all
+ * remaining walsender processes SIGUSR2. On receiving SIGUSR2, we try to send
+ * the remaining WAL, and then exit. This ensures that the checkpoint record
+ * reaches durable storage (in safekeepers), before the server shuts down
+ * completely.
+ */
+static void
+walprop_sigusr2(SIGNAL_ARGS)
+{
+	got_SIGUSR2 = true;
+
+	SetLatch(MyLatch);
+}
+
 static void
 walprop_pg_init_bgworker(void)
 {
@@ -503,6 +525,7 @@ walprop_pg_init_bgworker(void)
 	pqsignal(SIGUSR1, procsignal_sigusr1_handler);
 	pqsignal(SIGHUP, SignalHandlerForConfigReload);
 	pqsignal(SIGTERM, die);
+	pqsignal(SIGUSR2, walprop_sigusr2);

 	BackgroundWorkerUnblockSignals();

@@ -1026,7 +1049,7 @@ static void
 StartProposerReplication(WalProposer *wp, StartReplicationCmd *cmd)
 {
 	XLogRecPtr	FlushPtr;
-	TimeLineID	currTLI;
+	 __attribute__((unused)) TimeLineID	currTLI;

 #if PG_VERSION_NUM < 150000
 	if (ThisTimeLineID == 0)
@@ -1075,14 +1098,26 @@ StartProposerReplication(WalProposer *wp, StartReplicationCmd *cmd)
 #endif

 	/*
-	 * When we first start replication the standby will be behind the primary.
-	 * For some applications, for example synchronous replication, it is
-	 * important to have a clear state for this initial catchup mode, so we
-	 * can trigger actions when we change streaming state later. We may stay
-	 * in this state for a long time, which is exactly why we want to be able
-	 * to monitor whether or not we are still here.
+	 * XXX: Move straight to STOPPING state, skipping the STREAMING state.
+	 *
+	 * This is a bit weird. Normal walsenders stay in STREAMING state, until
+	 * the checkpointer signals them that it is about to start writing the
+	 * shutdown checkpoint. The walsenders acknowledge that they have received
+	 * that signal by switching to STOPPING state. That tells the walsenders
+	 * that they must not write any new WAL.
+	 *
+	 * However, we cannot easily intercept that signal from the checkpointer.
+	 * It's sent by WalSndInitStopping(), using
+	 * SendProcSignal(PROCSIGNAL_WALSND_INIT_STOPPING). It's received by
+	 * HandleWalSndInitStopping, which sets a process-local got_STOPPING flag.
+	 * However, that's all private to walsender.c.
+	 *
+	 * We don't need to do anything special upon receiving the signal, the
+	 * walproposer doesn't write any WAL anyway, so we skip the STREAMING
+	 * state and go directly to STOPPING mode. That way, the checkpointer
+	 * won't wait for us.
 	 */
-	WalSndSetState(WALSNDSTATE_CATCHUP);
+	WalSndSetState(WALSNDSTATE_STOPPING);

 	/*
 	 * Don't allow a request to stream from a future point in WAL that hasn't
@@ -1122,6 +1157,8 @@ StartProposerReplication(WalProposer *wp, StartReplicationCmd *cmd)
 static void
 WalSndLoop(WalProposer *wp)
 {
+	XLogRecPtr	flushPtr;
+
 	/* Clear any already-pending wakeups */
 	ResetLatch(MyLatch);

@@ -1130,9 +1167,6 @@ WalSndLoop(WalProposer *wp)
 		CHECK_FOR_INTERRUPTS();

 		XLogBroadcastWalProposer(wp);
-
-		if (MyWalSnd->state == WALSNDSTATE_CATCHUP)
-			WalSndSetState(WALSNDSTATE_STREAMING);
 		WalProposerPoll(wp);
 	}
 }
@@ -1230,7 +1264,6 @@ WalProposerRecovery(WalProposer *wp, Safekeeper *sk)
 	TimeLineID	timeline;
 	XLogRecPtr	startpos;
 	XLogRecPtr	endpos;
-	uint64		download_range_mb;

 	startpos = GetLogRepRestartLSN(wp);
 	if (startpos == InvalidXLogRecPtr)
@@ -1745,6 +1778,9 @@ walprop_pg_wait_event_set(WalProposer *wp, long timeout, Safekeeper **sk, uint32
 	{
 		ConditionVariableCancelSleep();
 		ResetLatch(MyLatch);
+
+		CheckGracefulShutdown(wp);
+
 		*events = WL_LATCH_SET;
 		return 1;
 	}
@@ -1798,6 +1834,41 @@ walprop_pg_finish_sync_safekeepers(WalProposer *wp, XLogRecPtr lsn)
 	exit(0);
 }

+/*
+ * Like vanilla walsender, on sigusr2 send all remaining WAL and exit.
+ *
+ * Note that unlike sync-safekeepers waiting here is not reliable: we
+ * don't check that majority of safekeepers received and persisted
+ * commit_lsn -- only that walproposer reached it (which immediately
+ * broadcasts new value). Doing that without incurring redundant control
+ * file syncing would need wp -> sk protocol change. OTOH unlike
+ * sync-safekeepers which must bump commit_lsn or basebackup will fail,
+ * this catchup is important only for tests where safekeepers/network
+ * don't crash on their own.
+ */
+static void
+CheckGracefulShutdown(WalProposer *wp)
+{
+	if (got_SIGUSR2)
+	{
+		if (!reported_sigusr2)
+		{
+			XLogRecPtr	flushPtr = walprop_pg_get_flush_rec_ptr(wp);
+
+			wpg_log(LOG, "walproposer will send and wait for remaining WAL between %X/%X and %X/%X",
+					LSN_FORMAT_ARGS(wp->commitLsn), LSN_FORMAT_ARGS(flushPtr));
+			reported_sigusr2 = true;
+		}
+
+		if (wp->commitLsn >= walprop_pg_get_flush_rec_ptr(wp))
+		{
+			wpg_log(LOG, "walproposer sent all WAL up to %X/%X, exiting",
+					LSN_FORMAT_ARGS(wp->commitLsn));
+			proc_exit(0);
+		}
+	}
+}
+
 /*
 * Choose most advanced PageserverFeedback and set it to *rf.
 */
@@ -1878,7 +1949,7 @@ CombineHotStanbyFeedbacks(HotStandbyFeedback *hs, WalProposer *wp)
 * None of that is functional in sync-safekeepers.
 */
 static void
-walprop_pg_process_safekeeper_feedback(WalProposer *wp, XLogRecPtr commitLsn)
+walprop_pg_process_safekeeper_feedback(WalProposer *wp)
 {
 	HotStandbyFeedback hsFeedback;
 	XLogRecPtr	oldDiskConsistentLsn;
@@ -1893,10 +1964,10 @@ walprop_pg_process_safekeeper_feedback(WalProposer *wp, XLogRecPtr commitLsn)
 	replication_feedback_set(&quorumFeedback.rf);
 	SetZenithCurrentClusterSize(quorumFeedback.rf.currentClusterSize);

-	if (commitLsn > quorumFeedback.flushLsn || oldDiskConsistentLsn != quorumFeedback.rf.disk_consistent_lsn)
+	if (wp->commitLsn > quorumFeedback.flushLsn || oldDiskConsistentLsn != quorumFeedback.rf.disk_consistent_lsn)
 	{
-		if (commitLsn > quorumFeedback.flushLsn)
-			quorumFeedback.flushLsn = commitLsn;
+		if (wp->commitLsn > quorumFeedback.flushLsn)
+			quorumFeedback.flushLsn = wp->commitLsn;

 		/*
 		 * Advance the replication slot to commitLsn. WAL before it is
@@ -1929,6 +2000,8 @@ walprop_pg_process_safekeeper_feedback(WalProposer *wp, XLogRecPtr commitLsn)
 								 XidFromFullTransactionId(hsFeedback.catalog_xmin),
 								 EpochFromFullTransactionId(hsFeedback.catalog_xmin));
 	}
+
+	CheckGracefulShutdown(wp);
 }

 static XLogRecPtr
--- a/pgxn/neon_test_utils/neontest.c
+++ b/pgxn/neon_test_utils/neontest.c
@@ -182,8 +182,6 @@ test_consume_memory(PG_FUNCTION_ARGS)
 Datum
 test_release_memory(PG_FUNCTION_ARGS)
 {
-	TimestampTz start;
-
 	if (PG_ARGISNULL(0))
 	{
 		if (consume_cxt)
--- a/pgxn/neon_walredo/walredoproc.c
+++ b/pgxn/neon_walredo/walredoproc.c
@@ -220,6 +220,9 @@ enter_seccomp_mode(void)
 }
 #endif /* HAVE_LIBSECCOMP */

+PGDLLEXPORT void
+WalRedoMain(int argc, char *argv[]);
+
 /*
 * Entry point for the WAL redo process.
 *
--- a/proxy/Cargo.toml
+++ b/proxy/Cargo.toml
@@ -68,7 +68,6 @@ task-local-extensions.workspace = true
 thiserror.workspace = true
 tikv-jemallocator.workspace = true
 tikv-jemalloc-ctl = { workspace = true, features = ["use_std"] }
-tls-listener.workspace = true
 tokio-postgres.workspace = true
 tokio-rustls.workspace = true
 tokio-util.workspace = true
--- a/proxy/src/auth/backend/link.rs
+++ b/proxy/src/auth/backend/link.rs
@@ -102,6 +102,8 @@ 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");

    // Backwards compatibility. pg_sni_proxy uses "--" in domain names
    // while direct connections do not. Once we migrate to pg_sni_proxy
--- a/proxy/src/bin/pg_sni_router.rs
+++ b/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 rustls::pki_types::PrivateKeyDer;
 use tokio::net::TcpListener;

 use anyhow::{anyhow, bail, ensure, Context};
@@ -76,37 +77,40 @@ async fn main() -> anyhow::Result<()> {
        (Some(key_path), Some(cert_path)) => {
            let key = {
                let key_bytes = std::fs::read(key_path).context("TLS key file")?;
-                let mut keys = rustls_pemfile::pkcs8_private_keys(&mut &key_bytes[..])
-                    .context(format!("Failed to read TLS keys at '{key_path}'"))?;
+
+                let mut keys =
+                    rustls_pemfile::pkcs8_private_keys(&mut &key_bytes[..]).collect_vec();

                ensure!(keys.len() == 1, "keys.len() = {} (should be 1)", keys.len());
-                keys.pop().map(rustls::PrivateKey).unwrap()
+                PrivateKeyDer::Pkcs8(
+                    keys.pop()
+                        .unwrap()
+                        .context(format!("Failed to read TLS keys at '{key_path}'"))?,
+                )
            };

            let cert_chain_bytes = std::fs::read(cert_path)
                .context(format!("Failed to read TLS cert file at '{cert_path}.'"))?;

-            let cert_chain = {
+            let cert_chain: Vec<_> = {
                rustls_pemfile::certs(&mut &cert_chain_bytes[..])
-                    .context(format!(
-                        "Failed to read TLS certificate chain from bytes from file at '{cert_path}'."
-                    ))?
-                    .into_iter()
-                    .map(rustls::Certificate)
-                    .collect_vec()
+                .try_collect()
+                .with_context(|| {
+                    format!("Failed to read TLS certificate chain from bytes from file at '{cert_path}'.")
+                })?
            };

            // needed for channel bindings
            let first_cert = cert_chain.first().context("missing certificate")?;
            let tls_server_end_point = TlsServerEndPoint::new(first_cert)?;

-            let tls_config = rustls::ServerConfig::builder()
-                .with_safe_default_cipher_suites()
-                .with_safe_default_kx_groups()
-                .with_protocol_versions(&[&rustls::version::TLS13, &rustls::version::TLS12])?
-                .with_no_client_auth()
-                .with_single_cert(cert_chain, key)?
-                .into();
+            let tls_config = rustls::ServerConfig::builder_with_protocol_versions(&[
+                &rustls::version::TLS13,
+                &rustls::version::TLS12,
+            ])
+            .with_no_client_auth()
+            .with_single_cert(cert_chain, key)?
+            .into();

            (tls_config, tls_server_end_point)
        }
--- a/proxy/src/config.rs
+++ b/proxy/src/config.rs
@@ -1,6 +1,10 @@
 use crate::{auth, rate_limiter::RateBucketInfo, serverless::GlobalConnPoolOptions};
 use anyhow::{bail, ensure, Context, Ok};
-use rustls::{sign, Certificate, PrivateKey};
+use itertools::Itertools;
+use rustls::{
+    crypto::ring::sign,
+    pki_types::{CertificateDer, PrivateKeyDer},
+};
 use sha2::{Digest, Sha256};
 use std::{
    collections::{HashMap, HashSet},
@@ -88,14 +92,14 @@ pub fn configure_tls(

    let cert_resolver = Arc::new(cert_resolver);

-    let config = rustls::ServerConfig::builder()
-        .with_safe_default_cipher_suites()
-        .with_safe_default_kx_groups()
-        // allow TLS 1.2 to be compatible with older client libraries
-        .with_protocol_versions(&[&rustls::version::TLS13, &rustls::version::TLS12])?
-        .with_no_client_auth()
-        .with_cert_resolver(cert_resolver.clone())
-        .into();
+    // allow TLS 1.2 to be compatible with older client libraries
+    let config = rustls::ServerConfig::builder_with_protocol_versions(&[
+        &rustls::version::TLS13,
+        &rustls::version::TLS12,
+    ])
+    .with_no_client_auth()
+    .with_cert_resolver(cert_resolver.clone())
+    .into();

    Ok(TlsConfig {
        config,
@@ -133,14 +137,14 @@ pub enum TlsServerEndPoint {
 }

 impl TlsServerEndPoint {
-    pub fn new(cert: &Certificate) -> anyhow::Result<Self> {
+    pub fn new(cert: &CertificateDer) -> anyhow::Result<Self> {
        let sha256_oids = [
            // I'm explicitly not adding MD5 or SHA1 here... They're bad.
            oid_registry::OID_SIG_ECDSA_WITH_SHA256,
            oid_registry::OID_PKCS1_SHA256WITHRSA,
        ];

-        let pem = x509_parser::parse_x509_certificate(&cert.0)
+        let pem = x509_parser::parse_x509_certificate(cert)
            .context("Failed to parse PEM object from cerficiate")?
            .1;

@@ -150,8 +154,7 @@ impl TlsServerEndPoint {
        let oid = pem.signature_algorithm.oid();
        let alg = reg.get(oid);
        if sha256_oids.contains(oid) {
-            let tls_server_end_point: [u8; 32] =
-                Sha256::new().chain_update(&cert.0).finalize().into();
+            let tls_server_end_point: [u8; 32] = Sha256::new().chain_update(cert).finalize().into();
            info!(subject = %pem.subject, signature_algorithm = alg.map(|a| a.description()), tls_server_end_point = %base64::encode(tls_server_end_point), "determined channel binding");
            Ok(Self::Sha256(tls_server_end_point))
        } else {
@@ -165,7 +168,7 @@ impl TlsServerEndPoint {
    }
 }

-#[derive(Default)]
+#[derive(Default, Debug)]
 pub struct CertResolver {
    certs: HashMap<String, (Arc<rustls::sign::CertifiedKey>, TlsServerEndPoint)>,
    default: Option<(Arc<rustls::sign::CertifiedKey>, TlsServerEndPoint)>,
@@ -185,11 +188,14 @@ impl CertResolver {
        let priv_key = {
            let key_bytes = std::fs::read(key_path)
                .context(format!("Failed to read TLS keys at '{key_path}'"))?;
-            let mut keys = rustls_pemfile::pkcs8_private_keys(&mut &key_bytes[..])
-                .context(format!("Failed to parse TLS keys at '{key_path}'"))?;
+            let mut keys = rustls_pemfile::pkcs8_private_keys(&mut &key_bytes[..]).collect_vec();

            ensure!(keys.len() == 1, "keys.len() = {} (should be 1)", keys.len());
-            keys.pop().map(rustls::PrivateKey).unwrap()
+            PrivateKeyDer::Pkcs8(
+                keys.pop()
+                    .unwrap()
+                    .context(format!("Failed to parse TLS keys at '{key_path}'"))?,
+            )
        };

        let cert_chain_bytes = std::fs::read(cert_path)
@@ -197,14 +203,10 @@ impl CertResolver {

        let cert_chain = {
            rustls_pemfile::certs(&mut &cert_chain_bytes[..])
+                .try_collect()
                .with_context(|| {
-                    format!(
-                    "Failed to read TLS certificate chain from bytes from file at '{cert_path}'."
-                )
+                    format!("Failed to read TLS certificate chain from bytes from file at '{cert_path}'.")
                })?
-                .into_iter()
-                .map(rustls::Certificate)
-                .collect()
        };

        self.add_cert(priv_key, cert_chain, is_default)
@@ -212,15 +214,15 @@ impl CertResolver {

    pub fn add_cert(
        &mut self,
-        priv_key: PrivateKey,
-        cert_chain: Vec<Certificate>,
+        priv_key: PrivateKeyDer<'static>,
+        cert_chain: Vec<CertificateDer<'static>>,
        is_default: bool,
    ) -> anyhow::Result<()> {
        let key = sign::any_supported_type(&priv_key).context("invalid private key")?;

        let first_cert = &cert_chain[0];
        let tls_server_end_point = TlsServerEndPoint::new(first_cert)?;
-        let pem = x509_parser::parse_x509_certificate(&first_cert.0)
+        let pem = x509_parser::parse_x509_certificate(first_cert)
            .context("Failed to parse PEM object from cerficiate")?
            .1;

--- a/proxy/src/console/messages.rs
+++ b/proxy/src/console/messages.rs
@@ -101,9 +101,10 @@ pub struct MetricsAuxInfo {
    pub cold_start_info: Option<ColdStartInfo>,
 }

-#[derive(Debug, Serialize, Deserialize, Clone)]
+#[derive(Debug, Default, Serialize, Deserialize, Clone)]
 #[serde(rename_all = "snake_case")]
 pub enum ColdStartInfo {
+    #[default]
    Unknown = 0,
    Warm = 1,
    PoolHit = 2,
--- a/proxy/src/console/provider.rs
+++ b/proxy/src/console/provider.rs
@@ -73,7 +73,7 @@ pub mod errors {
                        // Status 406: endpoint is disabled (we don't allow connections).
                        format!("{REQUEST_FAILED}: endpoint is disabled")
                    }
-                    http::StatusCode::LOCKED => {
+                    http::StatusCode::LOCKED | http::StatusCode::UNPROCESSABLE_ENTITY => {
                        // Status 423: project might be in maintenance mode (or bad state), or quotas exceeded.
                        format!("{REQUEST_FAILED}: endpoint is temporary unavailable. check your quotas and/or contact our support")
                    }
@@ -91,6 +91,12 @@ pub mod errors {
                    status: http::StatusCode::NOT_FOUND | http::StatusCode::NOT_ACCEPTABLE,
                    ..
                } => crate::error::ErrorKind::User,
+                ApiError::Console {
+                    status: http::StatusCode::UNPROCESSABLE_ENTITY,
+                    text,
+                } if text.contains("compute time quota of non-primary branches is exceeded") => {
+                    crate::error::ErrorKind::User
+                }
                ApiError::Console {
                    status: http::StatusCode::LOCKED,
                    text,
@@ -120,6 +126,11 @@ pub mod errors {
                    status: http::StatusCode::BAD_REQUEST,
                    ..
                } => true,
+                // don't retry when quotas are exceeded
+                Self::Console {
+                    status: http::StatusCode::UNPROCESSABLE_ENTITY,
+                    ref text,
+                } => !text.contains("compute time quota of non-primary branches is exceeded"),
                // locked can be returned when the endpoint was in transition
                // or when quotas are exceeded. don't retry when quotas are exceeded
                Self::Console {
--- a/proxy/src/console/provider/neon.rs
+++ b/proxy/src/console/provider/neon.rs
@@ -259,6 +259,9 @@ impl super::Api for Api {
        }

        let 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);
        info!(key = &*key, "created a cache entry for compute node info");

--- a/proxy/src/metrics.rs
+++ b/proxy/src/metrics.rs
@@ -4,7 +4,7 @@ use ::metrics::{
    register_int_gauge_vec, Histogram, HistogramVec, HyperLogLogVec, IntCounterPairVec,
    IntCounterVec, IntGauge, IntGaugeVec,
 };
-use metrics::{register_int_counter_pair, IntCounterPair};
+use metrics::{register_int_counter, register_int_counter_pair, IntCounter, IntCounterPair};

 use once_cell::sync::Lazy;
 use tokio::time;
@@ -303,3 +303,20 @@ pub static ENDPOINT_ERRORS_BY_KIND: Lazy<HyperLogLogVec<32>> = Lazy::new(|| {
    )
    .unwrap()
 });
+
+pub static REDIS_BROKEN_MESSAGES: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "proxy_redis_errors_total",
+        "Number of errors by a given classification",
+        &["channel"],
+    )
+    .unwrap()
+});
+
+pub static TLS_HANDSHAKE_FAILURES: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "proxy_tls_handshake_failures",
+        "Number of TLS handshake failures",
+    )
+    .unwrap()
+});
--- a/proxy/src/protocol2.rs
+++ b/proxy/src/protocol2.rs
@@ -1,22 +1,27 @@
 //! Proxy Protocol V2 implementation

 use std::{
-    future::poll_fn,
-    future::Future,
+    future::{poll_fn, Future},
    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 pin_project_lite::pin_project;
-use tls_listener::AsyncAccept;
 use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, ReadBuf};
+use uuid::Uuid;
+
+use crate::{metrics::NUM_CLIENT_CONNECTION_GAUGE, serverless::tls_listener::AsyncAccept};

 pub struct ProxyProtocolAccept {
    pub incoming: AddrIncoming,
+    pub protocol: &'static str,
 }

 pin_project! {
@@ -327,7 +332,7 @@ impl<T: AsyncRead> AsyncRead for WithClientIp<T> {
 }

 impl AsyncAccept for ProxyProtocolAccept {
-    type Connection = WithClientIp<AddrStream>;
+    type Connection = WithConnectionGuard<WithClientIp<AddrStream>>;

    type Error = io::Error;

@@ -336,11 +341,74 @@ impl AsyncAccept for ProxyProtocolAccept {
        cx: &mut Context<'_>,
    ) -> Poll<Option<Result<Self::Connection, Self::Error>>> {
        let conn = ready!(Pin::new(&mut self.incoming).poll_accept(cx)?);
+        tracing::info!(protocol = self.protocol, "accepted new TCP connection");
        let Some(conn) = conn else {
            return Poll::Ready(None);
        };

-        Poll::Ready(Some(Ok(WithClientIp::new(conn))))
+        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(),
+            )),
+        })))
+    }
+}
+
+pin_project! {
+    pub struct WithConnectionGuard<T> {
+        #[pin]
+        pub inner: T,
+        pub connection_id: Uuid,
+        pub gauge: Mutex<Option<IntCounterPairGuard>>,
+    }
+}
+
+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)
    }
 }

--- a/proxy/src/proxy.rs
+++ b/proxy/src/proxy.rs
@@ -24,6 +24,7 @@ use crate::{
 };
 use futures::TryFutureExt;
 use itertools::Itertools;
+use metrics::IntCounterPairGuard;
 use once_cell::sync::OnceCell;
 use pq_proto::{BeMessage as Be, StartupMessageParams};
 use regex::Regex;
@@ -78,10 +79,16 @@ pub async fn task_main(
    {
        let (socket, peer_addr) = accept_result?;

+        let conn_gauge = NUM_CLIENT_CONNECTION_GAUGE
+            .with_label_values(&["tcp"])
+            .guard();
+
        let session_id = uuid::Uuid::new_v4();
        let cancellation_handler = Arc::clone(&cancellation_handler);
        let endpoint_rate_limiter = endpoint_rate_limiter.clone();

+        tracing::info!(protocol = "tcp", %session_id, "accepted new TCP connection");
+
        connections.spawn(async move {
            let mut socket = WithClientIp::new(socket);
            let mut peer_addr = peer_addr.ip();
@@ -116,6 +123,7 @@ pub async fn task_main(
                socket,
                ClientMode::Tcp,
                endpoint_rate_limiter,
+                conn_gauge,
            )
            .instrument(span.clone())
            .await;
@@ -229,13 +237,11 @@ pub async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
    stream: S,
    mode: ClientMode,
    endpoint_rate_limiter: Arc<EndpointRateLimiter>,
+    conn_gauge: IntCounterPairGuard,
 ) -> Result<Option<ProxyPassthrough<S>>, ClientRequestError> {
    info!("handling interactive connection from client");

    let proto = ctx.protocol;
-    let _client_gauge = NUM_CLIENT_CONNECTION_GAUGE
-        .with_label_values(&[proto])
-        .guard();
    let _request_gauge = NUM_CONNECTION_REQUESTS_GAUGE
        .with_label_values(&[proto])
        .guard();
@@ -325,7 +331,7 @@ pub async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
        aux: node.aux.clone(),
        compute: node,
        req: _request_gauge,
-        conn: _client_gauge,
+        conn: conn_gauge,
        cancel: session,
    }))
 }
@@ -374,6 +380,11 @@ impl NeonOptions {
        Self::parse_from_iter(StartupMessageParams::parse_options_raw(options))
    }

+    pub fn is_ephemeral(&self) -> bool {
+        // Currently, neon endpoint options are all reserved for ephemeral endpoints.
+        !self.0.is_empty()
+    }
+
    fn parse_from_iter<'a>(options: impl Iterator<Item = &'a str>) -> Self {
        let mut options = options
            .filter_map(neon_option)
--- a/proxy/src/proxy/tests.rs
+++ b/proxy/src/proxy/tests.rs
@@ -20,6 +20,7 @@ use crate::{http, sasl, scram};
 use anyhow::{bail, Context};
 use async_trait::async_trait;
 use rstest::rstest;
+use rustls::pki_types;
 use tokio_postgres::config::SslMode;
 use tokio_postgres::tls::{MakeTlsConnect, NoTls};
 use tokio_postgres_rustls::{MakeRustlsConnect, RustlsStream};
@@ -28,7 +29,11 @@ use tokio_postgres_rustls::{MakeRustlsConnect, RustlsStream};
 fn generate_certs(
    hostname: &str,
    common_name: &str,
-) -> anyhow::Result<(rustls::Certificate, rustls::Certificate, rustls::PrivateKey)> {
+) -> anyhow::Result<(
+    pki_types::CertificateDer<'static>,
+    pki_types::CertificateDer<'static>,
+    pki_types::PrivateKeyDer<'static>,
+)> {
    let ca = rcgen::Certificate::from_params({
        let mut params = rcgen::CertificateParams::default();
        params.is_ca = rcgen::IsCa::Ca(rcgen::BasicConstraints::Unconstrained);
@@ -45,9 +50,9 @@ fn generate_certs(
    })?;

    Ok((
-        rustls::Certificate(ca.serialize_der()?),
-        rustls::Certificate(cert.serialize_der_with_signer(&ca)?),
-        rustls::PrivateKey(cert.serialize_private_key_der()),
+        pki_types::CertificateDer::from(ca.serialize_der()?),
+        pki_types::CertificateDer::from(cert.serialize_der_with_signer(&ca)?),
+        pki_types::PrivateKeyDer::Pkcs8(cert.serialize_private_key_der().into()),
    ))
 }

@@ -82,9 +87,8 @@ fn generate_tls_config<'a>(

    let tls_config = {
        let config = rustls::ServerConfig::builder()
-            .with_safe_defaults()
            .with_no_client_auth()
-            .with_single_cert(vec![cert.clone()], key.clone())?
+            .with_single_cert(vec![cert.clone()], key.clone_key())?
            .into();

        let mut cert_resolver = CertResolver::new();
@@ -101,10 +105,9 @@ fn generate_tls_config<'a>(

    let client_config = {
        let config = rustls::ClientConfig::builder()
-            .with_safe_defaults()
            .with_root_certificates({
                let mut store = rustls::RootCertStore::empty();
-                store.add(&ca)?;
+                store.add(ca)?;
                store
            })
            .with_no_client_auth();
--- a/proxy/src/proxy/wake_compute.rs
+++ b/proxy/src/proxy/wake_compute.rs
@@ -69,6 +69,12 @@ fn report_error(e: &WakeComputeError, retry: bool) {
        {
            "quota_exceeded"
        }
+        WakeComputeError::ApiError(ApiError::Console {
+            status: StatusCode::UNPROCESSABLE_ENTITY,
+            ref text,
+        }) if text.contains("compute time quota of non-primary branches is exceeded") => {
+            "quota_exceeded"
+        }
        WakeComputeError::ApiError(ApiError::Console {
            status: StatusCode::LOCKED,
            ..
--- a/proxy/src/redis/notifications.rs
+++ b/proxy/src/redis/notifications.rs
@@ -10,6 +10,7 @@ use crate::{
    cache::project_info::ProjectInfoCache,
    cancellation::{CancelMap, CancellationHandler, NotificationsCancellationHandler},
    intern::{ProjectIdInt, RoleNameInt},
+    metrics::REDIS_BROKEN_MESSAGES,
 };

 const CPLANE_CHANNEL_NAME: &str = "neondb-proxy-ws-updates";
@@ -115,6 +116,9 @@ impl<
        let msg: Notification = match serde_json::from_str(&payload) {
            Ok(msg) => msg,
            Err(e) => {
+                REDIS_BROKEN_MESSAGES
+                    .with_label_values(&[msg.get_channel_name()])
+                    .inc();
                tracing::error!("broken message: {e}");
                return Ok(());
            }
--- a/Show More
+++ b/Show More
				`@@ -0,0 +1 @@`
				`DROP VIEW IF EXISTS NEON_STAT_FILE_CACHE CASCADE;`
				`@@ -0,0 +1 @@`
				`DROP FUNCTION IF EXISTS approximate_working_set_size(bool) CASCADE;`