Bump version: 0.39.0-beta.5 → 0.39.0-beta.6

chore: update lance dependency to v12.0.0-beta.15 (#4143 )
Updates the Rust workspace Lance dependencies, Cargo lockfile, and Java lance-core from v12.0.0-beta.14 to [v12.0.0-beta.15](https://github.com/lance-format/lance/releases/tag/v12.0.0-beta.15). No compatibility fixes were required; `cargo clippy --quiet --workspace --tests --all-features -- -D warnings`, `cargo fmt --all --quiet`, and `git diff --check` passed. --------- Co-authored-by: Jack Ye <yezhaoqin@gmail.com>
2026-09-10 15:22:25 +00:00 · 2026-09-08 12:14:12 +00:00 · 2026-09-08 05:12:45 -07:00 · 2026-09-08 12:03:46 +00:00 · 2026-09-08 04:38:22 -07:00 · 2026-09-06 20:46:55 +00:00
94 changed files with 9865 additions and 1408 deletions
@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "0.38.0"
+current_version = "0.39.0-beta.6"
 parse = """(?x)
    (?P<major>0|[1-9]\\d*)\\.
    (?P<minor>0|[1-9]\\d*)\\.
@@ -232,7 +232,10 @@ jobs:
          ALL_FEATURES=`cargo metadata --format-version=1 --no-deps \
            | jq -r '.packages[] | .features | keys | .[]' \
            | grep -v s3-test | sort | uniq | paste -s -d "," -`
-          cargo test --profile ci --features $ALL_FEATURES --locked
+          # Run doctests before test binaries fill the runner disk. Examples are
+          # already built by the Linux job, so avoid retaining them here.
+          cargo test --profile ci --features $ALL_FEATURES --locked --doc
+          cargo test --profile ci --features $ALL_FEATURES --locked --lib --tests

  windows:
    strategy:
@@ -332,6 +332,34 @@ dependencies = [
 "num-traits",
 ]

+[[package]]
+name = "arrow-flight"
+version = "58.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b2dbe34824c639e43136af8f106992792ab456540d54b880bc320a3192502d2e"
+dependencies = [
+ "arrow-arith",
+ "arrow-array",
+ "arrow-buffer",
+ "arrow-cast",
+ "arrow-data",
+ "arrow-ipc",
+ "arrow-ord",
+ "arrow-row",
+ "arrow-schema",
+ "arrow-select",
+ "arrow-string",
+ "base64 0.22.1",
+ "bytes",
+ "futures",
+ "once_cell",
+ "paste",
+ "prost",
+ "prost-types",
+ "tonic",
+ "tonic-prost",
+]
+
 [[package]]
 name = "arrow-ipc"
 version = "58.4.0"
@@ -1129,7 +1157,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "edca88bc138befd0323b20752846e6587272d3b03b0343c8ea28a6f819e6e71f"
 dependencies = [
 "async-trait",
- "axum-core",
+ "axum-core 0.4.5",
 "bytes",
 "futures-util",
 "http 1.5.0",
@@ -1138,7 +1166,7 @@ dependencies = [
 "hyper 1.9.0",
 "hyper-util",
 "itoa",
- "matchit",
+ "matchit 0.7.3",
 "memchr",
 "mime",
 "percent-encoding",
@@ -1156,6 +1184,31 @@ dependencies = [
 "tracing",
 ]

+[[package]]
+name = "axum"
+version = "0.8.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "31b698c5f9a010f6573133b09e0de5408834d0c82f8d7475a89fc1867a71cd90"
+dependencies = [
+ "axum-core 0.5.6",
+ "bytes",
+ "futures-util",
+ "http 1.5.0",
+ "http-body 1.1.0",
+ "http-body-util",
+ "itoa",
+ "matchit 0.8.4",
+ "memchr",
+ "mime",
+ "percent-encoding",
+ "pin-project-lite",
+ "serde_core",
+ "sync_wrapper",
+ "tower",
+ "tower-layer",
+ "tower-service",
+]
+
 [[package]]
 name = "axum-core"
 version = "0.4.5"
@@ -1177,6 +1230,24 @@ dependencies = [
 "tracing",
 ]

+[[package]]
+name = "axum-core"
+version = "0.5.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "08c78f31d7b1291f7ee735c1c6780ccde7785daae9a9206026862dab7d8792d1"
+dependencies = [
+ "bytes",
+ "futures-core",
+ "http 1.5.0",
+ "http-body 1.1.0",
+ "http-body-util",
+ "mime",
+ "pin-project-lite",
+ "sync_wrapper",
+ "tower-layer",
+ "tower-service",
+]
+
 [[package]]
 name = "backoff"
 version = "0.4.0"
@@ -3455,8 +3526,8 @@ checksum = "42703706b716c37f96a77aea830392ad231f44c9e9a67872fa5548707e11b11c"

 [[package]]
 name = "fsst"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-array",
 "rand 0.9.5",
@@ -4815,8 +4886,8 @@ checksum = "e037a2e1d8d5fdbd49b16a4ea09d5d6401c1f29eca5ff29d03d3824dba16256a"

 [[package]]
 name = "lance"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arc-swap",
 "arrow",
@@ -4888,8 +4959,8 @@ dependencies = [

 [[package]]
 name = "lance-arrow"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-array",
 "arrow-buffer",
@@ -4911,7 +4982,7 @@ dependencies = [
 [[package]]
 name = "lance-arrow-scalar"
 version = "58.0.0"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-array",
 "arrow-buffer",
@@ -4925,7 +4996,7 @@ dependencies = [
 [[package]]
 name = "lance-arrow-stats"
 version = "58.0.0"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-array",
 "arrow-schema",
@@ -4934,8 +5005,8 @@ dependencies = [

 [[package]]
 name = "lance-bitpacking"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrayref",
 "crunchy",
@@ -4945,8 +5016,8 @@ dependencies = [

 [[package]]
 name = "lance-core"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-array",
 "arrow-buffer",
@@ -4983,8 +5054,8 @@ dependencies = [

 [[package]]
 name = "lance-datafusion"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow",
 "arrow-array",
@@ -5000,6 +5071,7 @@ dependencies = [
 "datafusion-functions",
 "datafusion-physical-expr",
 "futures",
+ "half",
 "jsonb",
 "lance-arrow",
 "lance-core",
@@ -5013,8 +5085,8 @@ dependencies = [

 [[package]]
 name = "lance-datagen"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow",
 "arrow-array",
@@ -5031,8 +5103,8 @@ dependencies = [

 [[package]]
 name = "lance-derive"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -5041,8 +5113,8 @@ dependencies = [

 [[package]]
 name = "lance-encoding"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-arith",
 "arrow-array",
@@ -5075,8 +5147,8 @@ dependencies = [

 [[package]]
 name = "lance-file"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-arith",
 "arrow-array",
@@ -5107,8 +5179,8 @@ dependencies = [

 [[package]]
 name = "lance-index"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arc-swap",
 "arrow",
@@ -5172,8 +5244,8 @@ dependencies = [

 [[package]]
 name = "lance-index-core"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-array",
 "arrow-schema",
@@ -5195,8 +5267,8 @@ dependencies = [

 [[package]]
 name = "lance-io"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow",
 "arrow-array",
@@ -5236,8 +5308,8 @@ dependencies = [

 [[package]]
 name = "lance-linalg"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-array",
 "arrow-schema",
@@ -5251,27 +5323,29 @@ dependencies = [

 [[package]]
 name = "lance-namespace"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow",
 "async-trait",
 "bytes",
 "lance-core",
 "lance-namespace-reqwest-client",
+ "serde",
+ "serde_json",
 "snafu 0.9.0",
 ]

 [[package]]
 name = "lance-namespace-impls"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow",
 "arrow-ipc",
 "arrow-schema",
 "async-trait",
- "axum",
+ "axum 0.7.9",
 "base64 0.22.1",
 "bytes",
 "chrono",
@@ -5304,9 +5378,9 @@ dependencies = [

 [[package]]
 name = "lance-namespace-reqwest-client"
-version = "0.11.0"
+version = "0.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0a030196da1c994b63a96a4f0bf5b0cfa459fe6dadc9e962320246ca328da22a"
+checksum = "d8d23e54b1634d5bbb434f8dd33dc3c05f6e58d876a9a27b3b4aef58ddbe11af"
 dependencies = [
 "reqwest 0.12.28",
 "serde",
@@ -5318,8 +5392,8 @@ dependencies = [

 [[package]]
 name = "lance-select"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-array",
 "arrow-buffer",
@@ -5333,8 +5407,8 @@ dependencies = [

 [[package]]
 name = "lance-table"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow",
 "arrow-array",
@@ -5374,8 +5448,8 @@ dependencies = [

 [[package]]
 name = "lance-testing"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "arrow-array",
 "arrow-schema",
@@ -5388,8 +5462,8 @@ dependencies = [

 [[package]]
 name = "lance-tokenizer"
-version = "12.0.0-beta.5"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.5#556637791d0048c2b4f1342dd84b67c8bbd65259"
+version = "12.0.0-beta.15"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.15#271a155ba505fd8e1094c095d4ce356707e93866"
 dependencies = [
 "frostem",
 "icu_segmenter",
@@ -5402,7 +5476,7 @@ dependencies = [

 [[package]]
 name = "lancedb"
-version = "0.38.0"
+version = "0.39.0-beta.5"
 dependencies = [
 "ahash",
 "anyhow",
@@ -5411,6 +5485,7 @@ dependencies = [
 "arrow-buffer",
 "arrow-cast",
 "arrow-data",
+ "arrow-flight",
 "arrow-ipc",
 "arrow-ord",
 "arrow-schema",
@@ -5466,6 +5541,7 @@ dependencies = [
 "polars",
 "polars-arrow",
 "pprof 0.14.1",
+ "prost",
 "rand 0.9.5",
 "random_word",
 "regex",
@@ -5482,6 +5558,7 @@ dependencies = [
 "test-log",
 "tokenizers",
 "tokio",
+ "tonic",
 "url",
 "urlencoding",
 "uuid",
@@ -5490,7 +5567,7 @@ dependencies = [

 [[package]]
 name = "lancedb-nodejs"
-version = "0.38.0"
+version = "0.39.0-beta.5"
 dependencies = [
 "arrow-array",
 "arrow-buffer",
@@ -5515,7 +5592,7 @@ dependencies = [

 [[package]]
 name = "lancedb-python"
-version = "0.38.0"
+version = "0.39.0-beta.5"
 dependencies = [
 "arrow",
 "async-trait",
@@ -5539,6 +5616,7 @@ dependencies = [
 "serde_json",
 "snafu 0.8.9",
 "tokio",
+ "uuid",
 ]

 [[package]]
@@ -5859,6 +5937,12 @@ version = "0.7.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0e7465ac9959cc2b1404e8e2367b43684a6d13790fe23056cc8c6c5a6b7bcb94"

+[[package]]
+name = "matchit"
+version = "0.8.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "47e1ffaa40ddd1f3ed91f717a33c8c0ee23fff369e3aa8772b9605cc1d22f4c3"
+
 [[package]]
 name = "matrixmultiply"
 version = "0.3.10"
@@ -7734,6 +7818,7 @@ dependencies = [
 "pyo3-build-config",
 "pyo3-ffi",
 "pyo3-macros",
+ "uuid",
 ]

 [[package]]
@@ -10086,6 +10171,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ac2a5518c70fa84342385732db33fb3f44bc4cc748936eb5833d2df34d6445ef"
 dependencies = [
 "async-trait",
+ "axum 0.8.9",
 "base64 0.22.1",
 "bytes",
 "h2 0.4.16",
@@ -10097,9 +10183,11 @@ dependencies = [
 "hyper-util",
 "percent-encoding",
 "pin-project",
+ "rustls-native-certs",
 "socket2 0.6.3",
 "sync_wrapper",
 "tokio",
+ "tokio-rustls 0.26.4",
 "tokio-stream",
 "tower",
 "tower-layer",
@@ -13,20 +13,20 @@ categories = ["database-implementations"]
 rust-version = "1.91.0"

 [workspace.dependencies]
-lance = { "version" = "=12.0.0-beta.5", default-features = false, "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-core = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-datagen = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-file = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-io = { "version" = "=12.0.0-beta.5", default-features = false, "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-index = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-linalg = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-namespace = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-namespace-impls = { "version" = "=12.0.0-beta.5", default-features = false, "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-table = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-testing = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-datafusion = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-encoding = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
-lance-arrow = { "version" = "=12.0.0-beta.5", "tag" = "v12.0.0-beta.5", "git" = "https://github.com/lance-format/lance.git" }
+lance = { "version" = "=12.0.0-beta.15", default-features = false, "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-core = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-datagen = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-file = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-io = { "version" = "=12.0.0-beta.15", default-features = false, "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-index = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-linalg = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace-impls = { "version" = "=12.0.0-beta.15", default-features = false, "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-table = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-testing = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-datafusion = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-encoding = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
+lance-arrow = { "version" = "=12.0.0-beta.15", "tag" = "v12.0.0-beta.15", "git" = "https://github.com/lance-format/lance.git" }
 lancedb = { path = "rust/lancedb", default-features = false }
 ahash = "0.8"
 # Note that this one does not include pyarrow
@@ -39,6 +39,7 @@ arrow-ord = "58.0.0"
 arrow-schema = "58.0.0"
 arrow-select = "58.0.0"
 arrow-cast = "58.0.0"
+arrow-flight = { version = "58.0.0", features = ["flight-sql-experimental"] }
 async-trait = "0"
 bytes = "1"
 datafusion = { version = "54.0.0", default-features = false }
@@ -71,7 +72,8 @@ serde = "1"
 serde_json = "1"
 tempfile = "3.5.0"
 tokio = { version = "1.23", features = ["rt-multi-thread", "sync"] }
-uuid = { version = "1.7.0", features = ["v4"] }
+tonic = { version = "0.14", features = ["tls-native-roots", "tls-ring"] }
+uuid = { version = "1.7.0", features = ["v4", "v7"] }
 chrono = { version = "0.4", default-features = false, features = ["clock"] }

 [profile.ci]
@@ -446,6 +446,15 @@ paths:
              properties:
                column:
                  type: string
+                name:
+                  type: string
+                  description: Optional name for the created index.
+                replace:
+                  type: boolean
+                  default: true
+                  description: |
+                    Whether to replace an existing index with the same resolved
+                    name. Defaults to true.
                metric_type:
                  type: string
                  nullable: false
@@ -14,7 +14,7 @@ Add the following dependency to your `pom.xml`:
 <dependency>
    <groupId>com.lancedb</groupId>
    <artifactId>lancedb-core</artifactId>
-    <version>0.38.0</version>
+    <version>0.39.0-beta.6</version>
 </dependency>
 ```

@@ -448,26 +448,6 @@ on the returned job to know when cleanup has finished.

 ***

-### getJob()
-
-```ts
-abstract getJob(jobId): Promise<null | JobDescription>
-```
-
-Describe a single server-side job by id.
-
-Resolves to `null` when the server has no such job.
-
-#### Parameters
-
-* **jobId**: `string`
-
-#### Returns
-
-`Promise`&lt;`null` \| [`JobDescription`](../interfaces/JobDescription.md)&gt;
-
-***
-
 ### isOpen()

 ```ts
@@ -482,48 +462,6 @@ Return true if the connection has not been closed

 ***

-### job()
-
-```ts
-abstract job(jobId): Job
-```
-
-A [Job](Job.md) handle for a server-side job by id.
-
-The handle is constructed without a server round trip; an unknown id
-surfaces when the handle is used. Dropping the handle has no effect on
-the job itself.
-
-#### Parameters
-
-* **jobId**: `string`
-
-#### Returns
-
-[`Job`](Job.md)
-
-***
-
-### jobHistory()
-
-```ts
-abstract jobHistory(jobId?): Promise<Table<any>>
-```
-
-The lifecycle event history of a server-side job, as an Arrow table.
-
-Lists history across all jobs when `jobId` is omitted.
-
-#### Parameters
-
-* **jobId?**: `string`
-
-#### Returns
-
-`Promise`&lt;`Table`&lt;`any`&gt;&gt;
-
-***
-
 ### listJobs()

 ```ts
@@ -648,6 +586,30 @@ A page of table names and an

 ***

+### openJob()
+
+```ts
+abstract openJob(jobId): Promise<Job>
+```
+
+Open a server-side job by id, returning a handle with its record already
+populated. Rejects when the server has no such job, the way
+[Connection.openTable](Connection.md#opentable) does for a missing table.
+
+The returned [Job](Job.md) answers for its own state, specification,
+result, failure and event history, so there is no separate
+connection-level call for any of them.
+
+#### Parameters
+
+* **jobId**: `string`
+
+#### Returns
+
+`Promise`&lt;[`Job`](Job.md)&gt;
+
+***
+
 ### openMaterializedView()

 ```ts
@@ -8,19 +8,46 @@

 A handle to an operation that may still be running.

-## Constructors
+The operation may already be complete when the handle is created.

-### new Job()
+The detail getters read what the handle last observed. Submitting an
+operation returns only a job id, so populating them eagerly would cost an
+extra round trip on every call:
+
+- [Job.refresh](Job.md#refresh) and [Job.status](Job.md#status) fetch the whole record.
+- [Job.wait](Job.md#wait) records the terminal state it establishes, but not the
+  rest of the record.
+- Everything is null until one of those runs.
+
+## Accessors
+
+### creationMs

 ```ts
-new Job(): Job
+get creationMs(): null | number
 ```

+When the job was created, in milliseconds since the epoch.
+
 #### Returns

-[`Job`](Job.md)
+`null` \| `number`

-## Accessors
+***
+
+### failure
+
+```ts
+get failure(): null | JobFailureInfo
+```
+
+Why the job failed, when it failed and the server reports a reason.
+
+#### Returns
+
+`null` \| [`JobFailureInfo`](../interfaces/JobFailureInfo.md)
+
+***

 ### id

@@ -28,8 +55,69 @@ new Job(): Job
 get id(): null | string
 ```

-Identifies the operation on the server that is running it. Operations
-that run in this process have no server id. The value is opaque.
+Identifies the operation on the server that is running it.
+
+Operations that run in this process have no server id. The value is
+opaque: parsing it or storing it to resume the job later is not supported.
+
+#### Returns
+
+`null` \| `string`
+
+***
+
+### jobType
+
+```ts
+get jobType(): null | string
+```
+
+The job's type, as the server names it. Null for an in-process job, which
+has no server-side record.
+
+#### Returns
+
+`null` \| `string`
+
+***
+
+### result
+
+```ts
+get result(): any
+```
+
+The job-type-specific terminal result. Null until the job succeeds, so a
+job that never terminates reports its progress through [Job.events](Job.md#events)
+instead.
+
+#### Returns
+
+`any`
+
+***
+
+### spec
+
+```ts
+get spec(): any
+```
+
+The job-type-specific specification it was submitted with.
+
+#### Returns
+
+`any`
+
+***
+
+### state
+
+```ts
+get state(): null | string
+```
+
+The last observed lifecycle state, without contacting the backend.

 #### Returns

@@ -51,18 +139,61 @@ Request cancellation. Cancelling a finished operation is a no-op.

 ***

+### events()
+
+```ts
+events(options?): Promise<Table<any>>
+```
+
+This job's recorded lifecycle events.
+
+Where the getters above report a terminal result only once the job reaches
+one, events are written as the job runs and outlive the workers that
+produced them. A distributed job records a `claim`/`claim_complete` pair
+per unit of work, each carrying `rows_processed`, so a job that never
+finishes still accounts for what it did.
+
+The server caps results at 1000 rows by default and 10,000 at most, and
+truncates without saying so, so pass `limit` for a job that emits an event
+per fragment. `filter` is a SQL-like expression over the `state`,
+`updated_by`, `emitted_from`, `emitted_by`, and `claim_entity` columns.
+
+#### Parameters
+
+* **options?**: [`JobEventsOptions`](../interfaces/JobEventsOptions.md)
+
+#### Returns
+
+`Promise`&lt;`Table`&lt;`any`&gt;&gt;
+
+***
+
+### refresh()
+
+```ts
+refresh(): Promise<void>
+```
+
+Ask the backend for this job's current state, and for a server-side job
+its full record, then cache it for the getters above.
+
+#### Returns
+
+`Promise`&lt;`void`&gt;
+
+***
+
 ### status()

 ```ts
 status(): Promise<string>
 ```

-The operation's current lifecycle state: "running", "finished",
-"failed", or "cancelled".
+The operation's current lifecycle state: "running", "finished", "failed",
+or "cancelled".

-A point snapshot; unlike [Job.wait](Job.md#wait) it does not block or reject
-on a terminal failure state. States a newer server reports that this
-client version does not know pass through as-is.
+A point snapshot; unlike [Job.wait](Job.md#wait) it does not block or reject on a
+terminal failure state. Also refreshes the getters above.

 #### Returns

@@ -70,6 +201,22 @@ client version does not know pass through as-is.

 ***

+### toString()
+
+```ts
+toString(): string
+```
+
+Every field the handle currently knows, one per line, with the JSON
+payloads indented -- a refresh job's spec and result are the point of
+printing it.
+
+#### Returns
+
+`string`
+
+***
+
 ### wait()

 ```ts
@@ -676,9 +676,17 @@ List all the versions of the table
 abstract mergeInsert(on): MergeInsertBuilder
 ```

+Create a [MergeInsertBuilder](MergeInsertBuilder.md), which combines new data with the
+existing table in a single transaction — inserting, updating and deleting
+rows depending on how they match.
+
 #### Parameters

 * **on**: `string` \| `string`[]
+    The column, or columns, to match source rows against target
+    rows on. Typically a key or id column. Several columns match on the
+    composite key: a source row updates a target row only when it agrees on
+    every one of them.

 #### Returns

@@ -96,7 +96,7 @@
 - [IvfFlatOptions](interfaces/IvfFlatOptions.md)
 - [IvfPqOptions](interfaces/IvfPqOptions.md)
 - [IvfRqOptions](interfaces/IvfRqOptions.md)
- [JobDescription](interfaces/JobDescription.md)
+- [JobEventsOptions](interfaces/JobEventsOptions.md)
 - [JobFailureInfo](interfaces/JobFailureInfo.md)
 - [JobInfo](interfaces/JobInfo.md)
 - [ListNamespacesOptions](interfaces/ListNamespacesOptions.md)
@@ -1,66 +0,0 @@
-[**@lancedb/lancedb**](../README.md) • **Docs**
-
-***
-
-[@lancedb/lancedb](../globals.md) / JobDescription
-
-# Interface: JobDescription
-
-A described job from `Connection.getJob`.
-
-## Properties
-
-### creationMs
-
-```ts
-creationMs: number;
-```
-
-When the job was created, in milliseconds since the epoch.
-
-***
-
-### failure?
-
-```ts
-optional failure: JobFailureInfo;
-```
-
-Why the job failed, when the job is failed and the server reports a
-reason.
-
-***
-
-### jobId
-
-```ts
-jobId: string;
-```
-
-***
-
-### jobType
-
-```ts
-jobType: string;
-```
-
-***
-
-### specJson?
-
-```ts
-optional specJson: string;
-```
-
-The job-type-specific specification as a JSON string, when present.
-
-***
-
-### state
-
-```ts
-state: string;
-```
-
-Lifecycle state: "running", "finished", "failed", or "cancelled".
@@ -0,0 +1,29 @@
+[**@lancedb/lancedb**](../README.md) • **Docs**
+
+***
+
+[@lancedb/lancedb](../globals.md) / JobEventsOptions
+
+# Interface: JobEventsOptions
+
+Which of a job's events [Job.events](../classes/Job.md#events) returns.
+
+## Properties
+
+### filter?
+
+```ts
+optional filter: string;
+```
+
+SQL-like filter over the event columns.
+
+***
+
+### limit?
+
+```ts
+optional limit: number;
+```
+
+Maximum event rows to return, up to the server maximum of 10,000.
@@ -26,7 +26,7 @@ When the job was created, in milliseconds since the epoch.
 jobId: string;
 ```

-The job id -- what `Connection.getJob` and `Connection.cancelJob`
+The job id -- what `Connection.openJob` and `Connection.cancelJob`
 accept.

 ***
@@ -50,6 +50,16 @@ projections: [string, string][];

 ***

+### sourceNamespace
+
+```ts
+sourceNamespace: string[];
+```
+
+Namespace holding the source table; empty is the root namespace.
+
+***
+
 ### sourceTable

 ```ts
@@ -28,6 +28,59 @@ is also an [asynchronous API client](#connections-asynchronous).

 ::: lancedb.Session

+## Remote SQL
+
+Submit SQL against a remote LanceDB database through the connection.
+The connected database and `default_namespace_path=["public"]` are used for
+unqualified tables. Fully qualified references can still query other databases
+and namespaces available to the same deployment. `execute_query` returns a
+reader as soon as its initial result stream is available. `execute_query_async`
+returns a query handle immediately; use it to inspect progress, open a reader,
+or cancel the query. The SQL client is initialized by the first query and
+retained for the lifetime of the remote connection. Query ids are random,
+connection-scoped references rather than encoded SQL or durable resume tokens:
+
+```python
+import lancedb
+
+db = lancedb.connect(
+    "db://analytics",
+    api_key="ldb_...",
+    host_override="https://api.example.com",
+    sql_host_override="grpc+tls://sql.example.com:10026",
+)
+reader = db.execute_query(
+    """
+    SELECT events.id, accounts.name
+    FROM analytics.public.events AS events
+    JOIN users.public.accounts AS accounts ON events.user_id = accounts.id
+    """,
+    default_namespace_path=["public"],
+)
+for batch in reader:
+    print(batch.num_rows)
+
+query = db.execute_query_async("SELECT * FROM events")
+print(query.id)
+print(query.describe().status)
+for batch in query.reader():
+    print(batch.num_rows)
+
+# The async connection exposes the same lifecycle without blocking:
+# async_db = await lancedb.connect_async(
+#     "db://analytics",
+#     api_key="ldb_...",
+#     host_override="https://api.example.com",
+#     sql_host_override="grpc+tls://sql.example.com:10026",
+# )
+# reader = await async_db.execute_query("SELECT * FROM events")
+# query = await async_db.execute_query_async("SELECT * FROM events")
+# description = await async_db.describe_query(query.id)
+# async for batch in await query.reader():
+#     print(batch.num_rows)
+# await query.cancel()
+```
+
 ## Namespaces (Synchronous)

 A namespace-backed connection resolves tables through a
@@ -94,6 +147,8 @@ listing a storage directory.

 ::: lancedb.functions.OutputMapping

+::: lancedb.functions.AssignmentMapping
+
 ::: lancedb.functions.FunctionBinding

 ::: lancedb.functions.RefreshColumnResult
@@ -102,6 +157,18 @@ listing a storage directory.

 ::: lancedb.job.AsyncJob

+::: lancedb.job.JobInfo
+
+::: lancedb.job.JobDescription
+
+::: lancedb.job.JobFailureInfo
+
+::: lancedb.sql.Query
+
+::: lancedb.sql.AsyncQuery
+
+::: lancedb.sql.QueryDescription
+
 ## Materialized Views (Synchronous)

 ::: lancedb.materialized_view.MaterializedView
@@ -249,6 +316,12 @@ still work. Queries return descriptors. Call

 ::: lancedb.exceptions.MissingColumnError

+::: lancedb.exceptions.JobNotFoundError
+
+::: lancedb.exceptions.JobFailedError
+
+::: lancedb.exceptions.JobCancelledError
+
 ## Integrations

 ## Pydantic
@@ -8,7 +8,7 @@
    <parent>
      <groupId>com.lancedb</groupId>
      <artifactId>lancedb-parent</artifactId>
-      <version>0.38.0-final.0</version>
+      <version>0.39.0-beta.6</version>
      <relativePath>../pom.xml</relativePath>
    </parent>

@@ -6,7 +6,7 @@

    <groupId>com.lancedb</groupId>
    <artifactId>lancedb-parent</artifactId>
-    <version>0.38.0-final.0</version>
+    <version>0.39.0-beta.6</version>
    <packaging>pom</packaging>
    <name>${project.artifactId}</name>
    <description>LanceDB Java SDK Parent POM</description>
@@ -28,7 +28,7 @@
    <properties>
        <project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
        <arrow.version>15.0.0</arrow.version>
-        <lance-core.version>12.0.0-beta.5</lance-core.version>
+        <lance-core.version>12.0.0-beta.15</lance-core.version>
        <spotless.skip>false</spotless.skip>
        <spotless.version>2.30.0</spotless.version>
        <spotless.java.googlejavaformat.version>1.7</spotless.java.googlejavaformat.version>
@@ -1,7 +1,7 @@
 [package]
 name = "lancedb-nodejs"
 edition.workspace = true
-version = "0.38.0"
+version = "0.39.0-beta.6"
 publish = false
 license.workspace = true
 description.workspace = true
@@ -48,6 +48,28 @@ describe("materialized views", () => {
    expect(definitionFromMetadata(safe, "v").limit).toBe(42);
  });

+  it("reads the namespaced select kind and refuses unknown kinds", () => {
+    // "namespaced_select" is the namespaced form of "select": same shape, a
+    // separate kind so readers that predate it refuse instead of resolving
+    // the source at the root.
+    const namespaced = new Map([
+      [
+        DEFINITION_META_KEY,
+        '{"kind":"namespaced_select","source_table":"people","source_namespace":["ns"]}',
+      ],
+    ]);
+    const definition = definitionFromMetadata(namespaced, "v");
+    expect(definition.sourceTable).toBe("people");
+    expect(definition.sourceNamespace).toEqual(["ns"]);
+
+    const unknown = new Map([
+      [DEFINITION_META_KEY, '{"kind":"select_v3","source_table":"people"}'],
+    ]);
+    expect(() => definitionFromMetadata(unknown, "v")).toThrow(
+      /cannot refresh/,
+    );
+  });
+
  it("creates, refreshes and queries a view", async () => {
    const view = await db.createMaterializedView("adults", "people", {
      select: ["name", ["shout", "upper(name)"]],
@@ -939,6 +939,7 @@ describe("remote connection jobs surface", () => {
    const { tableFromArrays, tableToIPC } = await import("apache-arrow");
    const eventsTable = tableFromArrays({ state: ["created", "succeeded"] });
    const eventsBody = Buffer.from(tableToIPC(eventsTable, "stream"));
+    const queryEventsPayloads: Record<string, unknown>[] = [];

    await withMockDatabase(
      (req, res) => {
@@ -967,6 +968,16 @@ describe("remote connection jobs surface", () => {
                );
            }
          } else if (req.url === "/v1/jobs/describe") {
+            if (payload["job_id"] === "job-2") {
+              res
+                .writeHead(200, { "Content-Type": "application/json" })
+                .end(
+                  '{"job_id": "job-2", "job_type": "refresh_column", ' +
+                    '"job_state": "DONE", "creation_ms": 2000, ' +
+                    '"result": {"rows_assigned": 1000000}}',
+                );
+              return;
+            }
            if (payload["job_id"] !== "job-1") {
              res.writeHead(404).end("no such job");
              return;
@@ -988,6 +999,7 @@ describe("remote connection jobs surface", () => {
              .writeHead(200, { "Content-Type": "application/json" })
              .end('{"job_id": "job-1"}');
          } else if (req.url === "/v1/jobs/query_events") {
+            queryEventsPayloads.push(payload);
            res
              .writeHead(200, {
                "Content-Type": "application/vnd.apache.arrow.stream",
@@ -1004,22 +1016,65 @@ describe("remote connection jobs surface", () => {
        expect(jobs[0].state).toEqual("running");
        expect(jobs[1].state).toEqual("finished");

-        const description = await db.getJob("job-1");
-        expect(description?.state).toEqual("failed");
-        expect(JSON.parse(description?.specJson ?? "")).toEqual({
-          column: "vec",
-        });
-        expect(description?.failure?.message).toEqual("worker died");
-        expect(await db.getJob("missing")).toBeNull();
-
        expect(await db.cancelJob("job-1")).toBe(true);
        expect(await db.cancelJob("missing")).toBe(false);

-        const history = await db.jobHistory("job-1");
-        expect(history.numRows).toEqual(2);
+        // Opening a job hands back a populated handle; a missing one rejects.
+        await expect(db.openJob("missing")).rejects.toThrow("not found");
+        const finished = await db.openJob("job-2");
+        expect(finished.state).toEqual("finished");
+        expect(finished.result).toEqual({
+          // biome-ignore lint/style/useNamingConvention: snake_case mandated by the server wire format
+          rows_assigned: 1000000,
+        });

-        const job = db.job("job-1");
+        const job = await db.openJob("job-1");
        expect(job.id).toEqual("job-1");
+
+        // openJob already populated the handle; refresh() re-reads it.
+        expect(job.state).toEqual("failed");
+        await job.refresh();
+        expect(job.state).toEqual("failed");
+        expect(job.jobType).toEqual("create_index");
+        expect(job.creationMs).toEqual(1000);
+        expect(job.spec).toEqual({ column: "vec" });
+        expect(job.result).toBeNull();
+        expect(job.failure?.message).toEqual("worker died");
+
+        // The handle reaches its own events, supplying its job id.
+        const jobEvents = await job.events({
+          limit: 500,
+          filter: "state = 'claim_complete'",
+        });
+        expect(jobEvents.numRows).toEqual(2);
+        expect(queryEventsPayloads.pop()).toEqual({
+          // biome-ignore lint/style/useNamingConvention: snake_case mandated by the server wire format
+          job_id: "job-1",
+          limit: 500,
+          filter: "state = 'claim_complete'",
+        });
+
+        // Printing lays every known field out on its own line, with the JSON
+        // payloads indented rather than crammed onto one line.
+        expect(`${job}`).toEqual(
+          [
+            "Job(",
+            '    id="job-1",',
+            '    state="failed",',
+            '    jobType="create_index",',
+            "    creationMs=1000,",
+            "    spec={",
+            '        "column": "vec"',
+            "    },",
+            "    failure={",
+            '        "phase": "execute",',
+            '        "message": "worker died",',
+            '        "retryable": true',
+            "    },",
+            ")",
+          ].join("\n"),
+        );
+
        expect(await job.status()).toEqual("failed");
        await expect(job.wait()).rejects.toThrow("worker died");
      },
@@ -281,7 +281,7 @@ describe.each([arrow15, arrow16, arrow17, arrow18])(
        numIndices: 0,
        numRows: 3,
        // Full on-disk size of the two data files, footers and metadata included.
-        totalBytes: 684,
+        totalBytes: 550,
      });

      // Index files count toward totalBytes too (only deletion files and
@@ -289,7 +289,7 @@ describe.each([arrow15, arrow16, arrow17, arrow18])(
      await table.createIndex("id", { config: Index.btree() });
      const statsWithIndex = await table.stats();
      expect(statsWithIndex.numIndices).toBe(1);
-      expect(statsWithIndex.totalBytes).toBeGreaterThan(684);
+      expect(statsWithIndex.totalBytes).toBeGreaterThan(550);
    });

    it("should overwrite data if asked", async () => {
@@ -737,11 +737,12 @@ it("should query documents with LangChain PDF metadata", async () => {

 describe("merge insert", () => {
  let tmpDir: tmp.DirResult;
+  let conn: Connection;
  let table: Table;

  beforeEach(async () => {
    tmpDir = tmp.dirSync({ unsafeCleanup: true });
-    const conn = await connect(tmpDir.name);
+    conn = await connect(tmpDir.name);

    table = await conn.createTable("some_table", [
      { a: 1, b: "a" },
@@ -779,6 +780,38 @@ describe("merge insert", () => {

    expect(result.map((row) => ({ ...row }))).toEqual(expected);
  });
+  test("upsert on a composite key", async () => {
+    const composite = await conn.createTable("composite", [
+      { shard: "a", id: 1, val: "x" },
+      { shard: "a", id: 2, val: "y" },
+      { shard: "b", id: 1, val: "z" },
+    ]);
+
+    // ("a", 1) matches an existing row and updates it. ("b", 2) agrees with an
+    // existing row on each key column separately but on neither pair, so it is
+    // an insert.
+    const mergeInsertRes = await composite
+      .mergeInsert(["shard", "id"])
+      .whenMatchedUpdateAll()
+      .whenNotMatchedInsertAll()
+      .execute([
+        { shard: "a", id: 1, val: "X" },
+        { shard: "b", id: 2, val: "W" },
+      ]);
+    expect(mergeInsertRes.numUpdatedRows).toBe(1);
+    expect(mergeInsertRes.numInsertedRows).toBe(1);
+
+    const result = (await composite.toArrow())
+      .toArray()
+      .sort((a, b) => a.shard.localeCompare(b.shard) || a.id - b.id);
+
+    expect(result.map((row) => ({ ...row }))).toEqual([
+      { shard: "a", id: 1, val: "X" },
+      { shard: "a", id: 2, val: "y" },
+      { shard: "b", id: 1, val: "z" },
+      { shard: "b", id: 2, val: "W" },
+    ]);
+  });
  test("conditional update", async () => {
    const newData = [
      { a: 2, b: "x" },
@@ -1,7 +1,6 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors

-import { tableFromIPC } from "apache-arrow";
 import {
  Data,
  SchemaLike,
@@ -16,6 +15,7 @@ import {
  makeEmptyTable,
 } from "./arrow";
 import { EmbeddingFunctionConfig, getRegistry } from "./embedding/registry";
+import { Job } from "./job";
 import {
  MaterializedView,
  MaterializedViewSelect,
@@ -27,8 +27,6 @@ import type {
  CreateNamespaceResponse,
  DescribeNamespaceResponse,
  DropNamespaceResponse,
-  Job,
-  JobDescription,
  JobInfo,
  ListNamespacesResponse,
  ListTablesResponse,
@@ -557,24 +555,19 @@ export abstract class Connection {
  ): Promise<void>;

  /**
-   * A {@link Job} handle for a server-side job by id.
+   * Open a server-side job by id, returning a handle with its record already
+   * populated. Rejects when the server has no such job, the way
+   * {@link Connection.openTable} does for a missing table.
   *
-   * The handle is constructed without a server round trip; an unknown id
-   * surfaces when the handle is used. Dropping the handle has no effect on
-   * the job itself.
+   * The returned {@link Job} answers for its own state, specification,
+   * result, failure and event history, so there is no separate
+   * connection-level call for any of them.
   */
-  abstract job(jobId: string): Job;
+  abstract openJob(jobId: string): Promise<Job>;

  /** List server-side jobs across the database's tables. */
  abstract listJobs(): Promise<JobInfo[]>;

-  /**
-   * Describe a single server-side job by id.
-   *
-   * Resolves to `null` when the server has no such job.
-   */
-  abstract getJob(jobId: string): Promise<JobDescription | null>;
-
  /**
   * Request cancellation of a server-side job by id.
   *
@@ -582,13 +575,6 @@ export abstract class Connection {
   * such job exists. Cancelling an already-terminal job is a no-op success.
   */
  abstract cancelJob(jobId: string): Promise<boolean>;
-
-  /**
-   * The lifecycle event history of a server-side job, as an Arrow table.
-   *
-   * Lists history across all jobs when `jobId` is omitted.
-   */
-  abstract jobHistory(jobId?: string): Promise<ArrowTable>;
 }

 /** @hideconstructor */
@@ -869,7 +855,7 @@ export class LocalConnection extends Connection {
  }

  async dropTableAsync(name: string, namespacePath?: string[]): Promise<Job> {
-    return this.inner.dropTableAsync(name, namespacePath ?? []);
+    return new Job(await this.inner.dropTableAsync(name, namespacePath ?? []));
  }

  async dropAllTables(namespacePath?: string[]): Promise<void> {
@@ -928,29 +914,17 @@ export class LocalConnection extends Connection {
    );
  }

-  job(jobId: string): Job {
-    return this.inner.job(jobId);
+  async openJob(jobId: string): Promise<Job> {
+    return new Job(await this.inner.openJob(jobId));
  }

  async listJobs(): Promise<JobInfo[]> {
    return this.inner.listJobs();
  }

-  async getJob(jobId: string): Promise<JobDescription | null> {
-    return this.inner.getJob(jobId);
-  }
-
  async cancelJob(jobId: string): Promise<boolean> {
    return this.inner.cancelJob(jobId);
  }
-
-  async jobHistory(jobId?: string): Promise<ArrowTable> {
-    const buf = await this.inner.jobHistory(jobId);
-    if (buf.length === 0) {
-      return new ArrowTable();
-    }
-    return tableFromIPC(buf);
-  }
 }

 /**
@@ -94,13 +94,9 @@ export {
  RenameTableOptions,
 } from "./connection";

-export {
-  Job,
-  JobDescription,
-  JobFailureInfo,
-  JobInfo,
-  Session,
-} from "./native.js";
+export { JobFailureInfo, JobInfo, Session } from "./native.js";
+
+export { Job, JobEventsOptions } from "./job";

 export {
  AutoQuery,
@@ -0,0 +1,188 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+import { Table as ArrowTable, tableFromIPC } from "apache-arrow";
+import { JobFailureInfo, Job as NativeJob } from "./native";
+
+/** Which of a job's events {@link Job.events} returns. */
+export interface JobEventsOptions {
+  /** Maximum event rows to return, up to the server maximum of 10,000. */
+  limit?: number;
+  /** SQL-like filter over the event columns. */
+  filter?: string;
+}
+
+/**
+ * A handle to an operation that may still be running.
+ *
+ * The operation may already be complete when the handle is created.
+ *
+ * The detail getters read what the handle last observed. Submitting an
+ * operation returns only a job id, so populating them eagerly would cost an
+ * extra round trip on every call:
+ *
+ * - {@link Job.refresh} and {@link Job.status} fetch the whole record.
+ * - {@link Job.wait} records the terminal state it establishes, but not the
+ *   rest of the record.
+ * - Everything is null until one of those runs.
+ *
+ * @hideconstructor
+ */
+export class Job {
+  private readonly inner: NativeJob;
+
+  constructor(inner: NativeJob) {
+    this.inner = inner;
+  }
+
+  /**
+   * Identifies the operation on the server that is running it.
+   *
+   * Operations that run in this process have no server id. The value is
+   * opaque: parsing it or storing it to resume the job later is not supported.
+   */
+  get id(): string | null {
+    return this.inner.id ?? null;
+  }
+
+  /** The last observed lifecycle state, without contacting the backend. */
+  get state(): string | null {
+    return this.inner.state ?? null;
+  }
+
+  /**
+   * The job's type, as the server names it. Null for an in-process job, which
+   * has no server-side record.
+   */
+  get jobType(): string | null {
+    return this.inner.jobType ?? null;
+  }
+
+  /** When the job was created, in milliseconds since the epoch. */
+  get creationMs(): number | null {
+    return this.inner.creationMs ?? null;
+  }
+
+  /** The job-type-specific specification it was submitted with. */
+  // biome-ignore lint/suspicious/noExplicitAny: shape varies by job type
+  get spec(): any | null {
+    return parseJson(this.inner.specJson);
+  }
+
+  /**
+   * The job-type-specific terminal result. Null until the job succeeds, so a
+   * job that never terminates reports its progress through {@link Job.events}
+   * instead.
+   */
+  // biome-ignore lint/suspicious/noExplicitAny: shape varies by job type
+  get result(): any | null {
+    return parseJson(this.inner.resultJson);
+  }
+
+  /** Why the job failed, when it failed and the server reports a reason. */
+  get failure(): JobFailureInfo | null {
+    return this.inner.failure ?? null;
+  }
+
+  /**
+   * The operation's current lifecycle state: "running", "finished", "failed",
+   * or "cancelled".
+   *
+   * A point snapshot; unlike {@link Job.wait} it does not block or reject on a
+   * terminal failure state. Also refreshes the getters above.
+   */
+  async status(): Promise<string> {
+    return this.inner.status();
+  }
+
+  /** Wait until the operation reaches a terminal state. */
+  async wait(): Promise<void> {
+    return this.inner.wait();
+  }
+
+  /** Request cancellation. Cancelling a finished operation is a no-op. */
+  async cancel(): Promise<void> {
+    return this.inner.cancel();
+  }
+
+  /**
+   * Ask the backend for this job's current state, and for a server-side job
+   * its full record, then cache it for the getters above.
+   */
+  async refresh(): Promise<void> {
+    return this.inner.refresh();
+  }
+
+  /**
+   * This job's recorded lifecycle events.
+   *
+   * Where the getters above report a terminal result only once the job reaches
+   * one, events are written as the job runs and outlive the workers that
+   * produced them. A distributed job records a `claim`/`claim_complete` pair
+   * per unit of work, each carrying `rows_processed`, so a job that never
+   * finishes still accounts for what it did.
+   *
+   * The server caps results at 1000 rows by default and 10,000 at most, and
+   * truncates without saying so, so pass `limit` for a job that emits an event
+   * per fragment. `filter` is a SQL-like expression over the `state`,
+   * `updated_by`, `emitted_from`, `emitted_by`, and `claim_entity` columns.
+   */
+  async events(options?: JobEventsOptions): Promise<ArrowTable> {
+    const buf = await this.inner.events(options?.limit, options?.filter);
+    if (buf.length === 0) {
+      return new ArrowTable();
+    }
+    return tableFromIPC(buf);
+  }
+
+  /**
+   * Every field the handle currently knows, one per line, with the JSON
+   * payloads indented -- a refresh job's spec and result are the point of
+   * printing it.
+   */
+  toString(): string {
+    if (this.state === null) {
+      const known = this.id === null ? "" : `id=${JSON.stringify(this.id)}, `;
+      return `Job(${known}not refreshed)`;
+    }
+    const fields: string[] = [];
+    if (this.id !== null) {
+      fields.push(`id=${JSON.stringify(this.id)}`);
+    }
+    fields.push(`state=${JSON.stringify(this.state)}`);
+    if (this.jobType !== null) {
+      fields.push(`jobType=${JSON.stringify(this.jobType)}`);
+    }
+    if (this.creationMs !== null) {
+      fields.push(`creationMs=${this.creationMs}`);
+    }
+    for (const [name, value] of [
+      ["spec", this.spec],
+      ["result", this.result],
+    ] as const) {
+      if (value !== null) {
+        fields.push(`${name}=${indentJson(value)}`);
+      }
+    }
+    if (this.failure !== null) {
+      fields.push(`failure=${indentJson(this.failure)}`);
+    }
+    return `Job(${fields.map((field) => `\n${REPR_INDENT}${field},`).join("")}\n)`;
+  }
+
+  [Symbol.for("nodejs.util.inspect.custom")](): string {
+    return this.toString();
+  }
+}
+
+const REPR_INDENT = "    ";
+
+// biome-ignore lint/suspicious/noExplicitAny: shape varies by job type
+function indentJson(value: any): string {
+  return JSON.stringify(value, null, 4).replace(/\n/g, `\n${REPR_INDENT}`);
+}
+
+// biome-ignore lint/suspicious/noExplicitAny: shape varies by job type
+function parseJson(raw: string | null | undefined): any | null {
+  return raw === null || raw === undefined ? null : JSON.parse(raw);
+}
@@ -19,6 +19,8 @@ export interface MaterializedViewDefinition {
  limit?: number;
  /** Source columns the projections and filter read. */
  inputs: string[];
+  /** Namespace holding the source table; empty is the root namespace. */
+  sourceNamespace: string[];
 }

 /**
@@ -78,7 +80,8 @@ export function definitionFromMetadata(
  }
  // biome-ignore lint/suspicious/noExplicitAny: raw JSON
  const value: any = JSON.parse(raw);
-  if (value.kind !== "select") {
+  // "namespaced_select" keeps older readers from resolving the source at root.
+  if (value.kind !== "select" && value.kind !== "namespaced_select") {
    throw new Error(
      `materialized view '${name}' is defined by '${value.kind}', which this ` +
        "version of lancedb cannot refresh",
@@ -103,6 +106,7 @@ export function definitionFromMetadata(
    filter: value.filter ?? undefined,
    limit,
    inputs: value.inputs ?? [],
+    sourceNamespace: value.source_namespace ?? [],
  };
 }

@@ -19,6 +19,7 @@ import {

 import { EmbeddingFunctionConfig, getRegistry } from "./embedding/registry";
 import { IndexOptions } from "./indices";
+import { Job } from "./job";
 import { MergeInsertBuilder } from "./merge";
 import {
  AddColumnsResult,
@@ -30,7 +31,6 @@ import {
  DropColumnsResult,
  IndexConfig,
  IndexStatistics,
-  Job,
  LsmStats,
  Branches as NativeBranches,
  OptimizeStats,
@@ -919,6 +919,16 @@ export abstract class Table {
  /** Return the table as an arrow table */
  abstract toArrow(): Promise<ArrowTable>;

+  /**
+   * Create a {@link MergeInsertBuilder}, which combines new data with the
+   * existing table in a single transaction — inserting, updating and deleting
+   * rows depending on how they match.
+   *
+   * @param on - The column, or columns, to match source rows against target
+   * rows on. Typically a key or id column. Several columns match on the
+   * composite key: a source row updates a target row only when it agrees on
+   * every one of them.
+   */
  abstract mergeInsert(on: string | string[]): MergeInsertBuilder;

  /** List all the stats of a specified index
@@ -1114,13 +1124,15 @@ export class LocalTable extends Table {
  ): Promise<Job> {
    // biome-ignore lint/suspicious/noExplicitAny: skip
    const nativeIndex = (options?.config as any)?.inner;
-    return await this.inner.createIndexAsync(
-      nativeIndex,
-      column,
-      options?.replace,
-      options?.waitTimeoutSeconds,
-      options?.name,
-      options?.train,
+    return new Job(
+      await this.inner.createIndexAsync(
+        nativeIndex,
+        column,
+        options?.replace,
+        options?.waitTimeoutSeconds,
+        options?.name,
+        options?.train,
+      ),
    );
  }

@@ -1303,7 +1315,7 @@ export class LocalTable extends Table {
  }

  async refreshColumnAsync(column: string): Promise<Job> {
-    return await this.inner.refreshColumnAsync(column);
+    return new Job(await this.inner.refreshColumnAsync(column));
  }

  async refreshMaterializedView(
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-darwin-arm64",
-	"version": "0.38.0",
+	"version": "0.39.0-beta.6",
 	"os": ["darwin"],
 	"cpu": ["arm64"],
 	"main": "lancedb.darwin-arm64.node",
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-arm64-gnu",
-	"version": "0.38.0",
+	"version": "0.39.0-beta.6",
 	"os": ["linux"],
 	"cpu": ["arm64"],
 	"main": "lancedb.linux-arm64-gnu.node",
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-arm64-musl",
-	"version": "0.38.0",
+	"version": "0.39.0-beta.6",
 	"os": ["linux"],
 	"cpu": ["arm64"],
 	"main": "lancedb.linux-arm64-musl.node",
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-x64-gnu",
-	"version": "0.38.0",
+	"version": "0.39.0-beta.6",
 	"os": ["linux"],
 	"cpu": ["x64"],
 	"main": "lancedb.linux-x64-gnu.node",
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-x64-musl",
-	"version": "0.38.0",
+	"version": "0.39.0-beta.6",
 	"os": ["linux"],
 	"cpu": ["x64"],
 	"main": "lancedb.linux-x64-musl.node",
@@ -1,6 +1,6 @@
 {
  "name": "@lancedb/lancedb-win32-arm64-msvc",
-  "version": "0.38.0",
+  "version": "0.39.0-beta.6",
  "os": [
    "win32"
  ],
@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-win32-x64-msvc",
-	"version": "0.38.0",
+	"version": "0.39.0-beta.6",
 	"os": ["win32"],
 	"cpu": ["x64"],
 	"main": "lancedb.win32-x64-msvc.node",
@@ -11,7 +11,7 @@
    "ann"
  ],
  "private": false,
-  "version": "0.38.0",
+  "version": "0.39.0-beta.6",
  "main": "dist/index.js",
  "exports": {
    ".": "./dist/index.js",
@@ -44,7 +44,7 @@
    "@biomejs/biome": "^1.7.3",
    "@jest/globals": "^29.7.0",
    "@napi-rs/cli": "3.7.0",
-    "@opentelemetry/sdk-metrics": "^1.30.0",
+    "@opentelemetry/sdk-metrics": "^2.10.0",
    "@types/axios": "^0.14.0",
    "@types/jest": "^29.1.2",
    "@types/node": "22.7.4",
@@ -56,7 +56,7 @@
    "eslint": "^8.57.0",
    "jest": "^29.7.0",
    "shx": "^0.3.4",
-    "tmp": "^0.2.3",
+    "tmp": "^0.2.7",
    "ts-jest": "^29.1.2",
    "typedoc": "0.26.4",
    "typedoc-plugin-markdown": "4.2.1",
@@ -16,3 +16,41 @@ allowBuilds:
  onnxruntime-node: true
  protobufjs: true
  sharp: true
+
+minimumReleaseAgeExclude:
+  - protobufjs@7.5.8
+  - tmp@0.2.6
+  - form-data@4.0.6
+  - tar@7.5.16
+  - markdown-it@14.1.2
+  - linkify-it@5.0.1
+  - js-yaml@3.15.0
+  - js-yaml@4.1.2
+  - protobufjs@7.6.1
+  - protobufjs@7.6.3
+  - '@babel/core@7.29.1'
+  - axios@1.18.0
+  - brace-expansion@2.1.2
+  - brace-expansion@1.1.16
+  - js-yaml@4.3.0
+  - tar@7.5.18
+  - tar@7.5.19
+  - tar@7.5.17
+  - protobufjs@7.6.5
+  - linkify-it@5.0.2
+  - sharp@0.35.0
+  - brace-expansion@1.1.17
+  - brace-expansion@2.1.3
+  - brace-expansion@2.1.4
+  - brace-expansion@1.1.18
+  - js-yaml@3.15.1
+  - js-yaml@4.3.1
+  - tar@7.5.21
+  - '@opentelemetry/core@2.8.0'
+
+# @huggingface/transformers pins sharp ^0.33.5 and no released version has moved
+# past ^0.34.5, all of which inherit the libvips CVEs in GHSA-f88m-g3jw-g9cj.
+# Force the patched line. sharp is only reached by transformers' image pipeline,
+# which LanceDB's text embedding function never uses.
+overrides:
+  sharp: ^0.35.4
@@ -442,13 +442,15 @@ impl Connection {
        self.get_inner()?.drop_all_tables(&ns).await.default_error()
    }

-    /// A `Job` handle for a server-side job by id.
+    /// Open a server-side job by id, returning a handle with its record
+    /// already populated. Rejects when the server has no such job.
    ///
-    /// The handle is constructed without a server round trip; an unknown id
-    /// surfaces when the handle is used.
-    #[napi]
-    pub fn job(&self, job_id: String) -> napi::Result<crate::job::Job> {
-        let job = self.get_inner()?.job(job_id).default_error()?;
+    /// The returned handle answers for its own state, specification, result,
+    /// failure and event history, so there is no separate connection-level
+    /// call for any of them.
+    #[napi(catch_unwind)]
+    pub async fn open_job(&self, job_id: String) -> napi::Result<crate::job::Job> {
+        let job = self.get_inner()?.open_job(&job_id).await.default_error()?;
        Ok(crate::job::Job::new(job))
    }

@@ -459,17 +461,6 @@ impl Connection {
        Ok(jobs.into_iter().map(Into::into).collect())
    }

-    /// Describe a single server-side job by id. `null` when the server has
-    /// no such job.
-    #[napi(catch_unwind)]
-    pub async fn get_job(
-        &self,
-        job_id: String,
-    ) -> napi::Result<Option<crate::job::JobDescription>> {
-        let description = self.get_inner()?.get_job(&job_id).await.default_error()?;
-        Ok(description.map(Into::into))
-    }
-
    /// Request cancellation of a server-side job by id. Returns true if the
    /// server accepted the cancellation, false if no such job exists.
    #[napi(catch_unwind)]
@@ -477,34 +468,6 @@ impl Connection {
        self.get_inner()?.cancel_job(&job_id).await.default_error()
    }

-    /// The lifecycle event history of a server-side job (all jobs when
-    /// `job_id` is null), as an Arrow IPC stream buffer. Empty when there is
-    /// no history.
-    #[napi(catch_unwind)]
-    pub async fn job_history(&self, job_id: Option<String>) -> napi::Result<Buffer> {
-        let batches = self
-            .get_inner()?
-            .job_history(job_id.as_deref())
-            .await
-            .default_error()?;
-        let Some(first) = batches.first() else {
-            return Ok(Buffer::from(Vec::<u8>::new()));
-        };
-        let mut out = Vec::new();
-        let mut writer = arrow_ipc::writer::StreamWriter::try_new(&mut out, &first.schema())
-            .map_err(|e| napi::Error::from_reason(e.to_string()))?;
-        for batch in &batches {
-            writer
-                .write(batch)
-                .map_err(|e| napi::Error::from_reason(e.to_string()))?;
-        }
-        writer
-            .finish()
-            .map_err(|e| napi::Error::from_reason(e.to_string()))?;
-        drop(writer);
-        Ok(Buffer::from(out))
-    }
-
    #[napi(catch_unwind)]
    /// Describe a namespace and return its properties.
    pub async fn describe_namespace(
@@ -3,6 +3,9 @@

 use std::sync::Arc;

+use arrow_array::RecordBatch;
+use lancedb::job::JobEventsRequest;
+use napi::bindgen_prelude::Buffer;
 use napi_derive::napi;

 use crate::error::NapiErrorExt;
@@ -55,12 +58,98 @@ impl Job {
    pub async fn cancel(&self) -> napi::Result<()> {
        self.inner.cancel().await.default_error()
    }
+
+    /// Ask the backend for this job's current state, and for a server-side job
+    /// its full record, then cache it for the getters below.
+    ///
+    /// They are all null until this runs, because submitting an operation
+    /// returns only a job id. {@link Job.status} fetches the whole record too;
+    /// {@link Job.wait} records only the terminal state it establishes.
+    #[napi(catch_unwind)]
+    pub async fn refresh(&self) -> napi::Result<()> {
+        self.inner.refresh().await.default_error()
+    }
+
+    /// The last observed lifecycle state, without contacting the backend.
+    #[napi(getter)]
+    pub fn state(&self) -> Option<String> {
+        self.inner.state()
+    }
+
+    /// The job's type, as the server names it. Null for an in-process job,
+    /// which has no server-side record.
+    #[napi(getter)]
+    pub fn job_type(&self) -> Option<String> {
+        self.inner.job_type()
+    }
+
+    /// When the job was created, in milliseconds since the epoch.
+    #[napi(getter)]
+    pub fn creation_ms(&self) -> Option<i64> {
+        self.inner.creation_ms()
+    }
+
+    /// The job-type-specific specification as a JSON string, when present.
+    #[napi(getter)]
+    pub fn spec_json(&self) -> Option<String> {
+        self.inner.spec().map(|spec| spec.to_string())
+    }
+
+    /// The job-type-specific terminal result as a JSON string. Null until the
+    /// job succeeds, so a job that never terminates reports its progress
+    /// through {@link Job.events} instead.
+    #[napi(getter)]
+    pub fn result_json(&self) -> Option<String> {
+        self.inner.result().map(|result| result.to_string())
+    }
+
+    /// Why the job failed, when it failed and the server reports a reason.
+    #[napi(getter)]
+    pub fn failure(&self) -> Option<JobFailureInfo> {
+        self.inner.failure().map(|failure| JobFailureInfo {
+            phase: failure.phase,
+            message: failure.message,
+            retryable: failure.retryable,
+        })
+    }
+
+    /// This job's recorded lifecycle events, as an Arrow IPC stream buffer.
+    /// The TypeScript wrapper turns it into an Arrow table.
+    #[napi(catch_unwind)]
+    pub async fn events(&self, limit: Option<u32>, filter: Option<String>) -> napi::Result<Buffer> {
+        let batches = self
+            .inner
+            .events(JobEventsRequest { limit, filter })
+            .await
+            .default_error()?;
+        batches_to_ipc_buffer(&batches)
+    }
+}
+
+/// Serialise Arrow batches as a single IPC stream for the TypeScript layer.
+fn batches_to_ipc_buffer(batches: &[RecordBatch]) -> napi::Result<Buffer> {
+    let Some(first) = batches.first() else {
+        return Ok(Buffer::from(Vec::<u8>::new()));
+    };
+    let mut out = Vec::new();
+    let mut writer = arrow_ipc::writer::StreamWriter::try_new(&mut out, &first.schema())
+        .map_err(|e| napi::Error::from_reason(e.to_string()))?;
+    for batch in batches {
+        writer
+            .write(batch)
+            .map_err(|e| napi::Error::from_reason(e.to_string()))?;
+    }
+    writer
+        .finish()
+        .map_err(|e| napi::Error::from_reason(e.to_string()))?;
+    drop(writer);
+    Ok(Buffer::from(out))
 }

 /// A row from `Connection.listJobs`: one server-side job.
 #[napi(object)]
 pub struct JobInfo {
-    /// The job id -- what `Connection.getJob` and `Connection.cancelJob`
+    /// The job id -- what `Connection.openJob` and `Connection.cancelJob`
    /// accept.
    pub job_id: String,
    /// The table the job runs against, without URI or namespace.
@@ -91,36 +180,3 @@ pub struct JobFailureInfo {
    pub message: Option<String>,
    pub retryable: Option<bool>,
 }
-
-/// A described job from `Connection.getJob`.
-#[napi(object)]
-pub struct JobDescription {
-    pub job_id: String,
-    pub job_type: String,
-    /// Lifecycle state: "running", "finished", "failed", or "cancelled".
-    pub state: String,
-    /// When the job was created, in milliseconds since the epoch.
-    pub creation_ms: i64,
-    /// The job-type-specific specification as a JSON string, when present.
-    pub spec_json: Option<String>,
-    /// Why the job failed, when the job is failed and the server reports a
-    /// reason.
-    pub failure: Option<JobFailureInfo>,
-}
-
-impl From<lancedb::database::JobDescription> for JobDescription {
-    fn from(description: lancedb::database::JobDescription) -> Self {
-        Self {
-            job_id: description.job_id,
-            job_type: description.job_type,
-            state: description.state,
-            creation_ms: description.creation_ms,
-            spec_json: (!description.spec.is_null()).then(|| description.spec.to_string()),
-            failure: description.failure.map(|failure| JobFailureInfo {
-                phase: failure.phase,
-                message: failure.message,
-                retryable: failure.retryable,
-            }),
-        }
-    }
-}
@@ -1,6 +1,6 @@
 [package]
 name = "lancedb-python"
-version = "0.38.0"
+version = "0.39.0-beta.6"
 publish = false
 edition.workspace = true
 description = "Python bindings for LanceDB"
@@ -28,7 +28,7 @@ env_logger.workspace = true
 log.workspace = true
 # Maturin enables extension-module mode for Python builds. Keeping it out of
 # Cargo features lets Rust unit tests link against libpython.
-pyo3 = { version = "0.28", features = ["abi3-py310", "chrono"] }
+pyo3 = { version = "0.28", features = ["abi3-py310", "chrono", "uuid"] }
 chrono.workspace = true
 pyo3-async-runtimes = { version = "0.28", features = [
    "attributes",
@@ -40,6 +40,7 @@ serde.workspace = true
 serde_json.workspace = true
 snafu.workspace = true
 tokio.workspace = true
+uuid.workspace = true
 libc = "0.2"

 [build-dependencies]
@@ -139,6 +139,7 @@ include = [
    "python/lancedb/exceptions.py",
    "python/lancedb/background_loop.py",
    "python/lancedb/schema.py",
+    "python/lancedb/sql.py",
    "python/lancedb/remote/__init__.py",
    "python/lancedb/remote/errors.py",
    "python/lancedb/embeddings/__init__.py",
@@ -22,7 +22,11 @@ from .remote.db import RemoteDBConnection
 from .expr import Expr, col, lit, func
 from .schema import blob, vector
 from .job import AsyncJob, Job
+from .sql import AsyncQuery as AsyncSqlQuery
+from .sql import Query as SqlQuery
+from .sql import QueryDescription
 from .functions import (
+    AssignmentMapping as AssignmentMapping,
    FunctionArtifactRequest as FunctionArtifactRequest,
    FunctionApplication as FunctionApplication,
    FunctionBinding as FunctionBinding,
@@ -101,6 +105,7 @@ def connect(
    api_key: Optional[str] = None,
    region: str = "us-east-1",
    host_override: Optional[str] = None,
+    sql_host_override: Optional[str] = None,
    read_consistency_interval: Optional[timedelta] = None,
    request_thread_pool: Optional[Union[int, ThreadPoolExecutor]] = None,
    client_config: Union[ClientConfig, Dict[str, Any], None] = None,
@@ -129,6 +134,9 @@ def connect(
        The region to use for LanceDB Cloud.
    host_override: str, optional
        The override url for LanceDB Cloud.
+    sql_host_override: str, optional
+        The remote SQL service endpoint override. The client connects lazily when SQL
+        is first executed and retains that connection.
    read_consistency_interval: timedelta, default None
        The interval at which to check for updates to the table from other
        processes. If None, then consistency is not checked. For performance
@@ -270,6 +278,7 @@ def connect(
            api_key,
            region,
            host_override,
+            sql_host_override=sql_host_override,
            # TODO: remove this (deprecation warning downstream)
            request_thread_pool=request_thread_pool,
            client_config=client_config,
@@ -412,6 +421,7 @@ def deserialize_conn(
            parsed["api_key"],
            parsed.get("region", "us-east-1"),
            host_override=parsed.get("host_override"),
+            sql_host_override=parsed.get("sql_host_override"),
            client_config=parsed.get("client_config"),
            storage_options=storage_options,
        )
@@ -425,6 +435,7 @@ async def connect_async(
    api_key: Optional[str] = None,
    region: str = "us-east-1",
    host_override: Optional[str] = None,
+    sql_host_override: Optional[str] = None,
    read_consistency_interval: Optional[timedelta] = None,
    client_config: Optional[Union[ClientConfig, Dict[str, Any]]] = None,
    storage_options: Optional[Dict[str, str]] = None,
@@ -447,6 +458,9 @@ async def connect_async(
        The region to use for LanceDB Cloud.
    host_override: str, optional
        The override url for LanceDB Cloud.
+    sql_host_override: str, optional
+        The remote SQL service endpoint override. The client connects lazily when SQL
+        is first executed and retains that connection.
    read_consistency_interval: timedelta, default None
        The interval at which to check for updates to the table from other
        processes. If None, then consistency is not checked. For performance
@@ -534,6 +548,7 @@ async def connect_async(
            api_key,
            region,
            host_override,
+            sql_host_override,
            read_consistency_interval_secs,
            client_config,
            storage_options,
@@ -556,6 +571,7 @@ __all__ = [
    "connect_namespace_async",
    "AsyncConnection",
    "AsyncJob",
+    "AsyncSqlQuery",
    "AsyncLanceNamespaceDBConnection",
    "AsyncTable",
    "FtsToken",
@@ -570,6 +586,8 @@ __all__ = [
    "vector",
    "DBConnection",
    "Job",
+    "QueryDescription",
+    "SqlQuery",
    "LanceDBConnection",
    "LanceNamespaceDBConnection",
    "LsmWriteSpec",
@@ -1,6 +1,7 @@
 from datetime import date, datetime, timedelta
 from decimal import Decimal
 from typing import Dict, List, Optional, Tuple, Any, TypedDict, Union, Literal
+from uuid import UUID

 import pyarrow as pa

@@ -147,16 +148,20 @@ class Connection(object):
        start_after: Optional[str],
        limit: Optional[int],
    ) -> list[str]: ...  # Deprecated: Use list_tables instead
-    def job(self, job_id: str) -> Job: ...
+    async def open_job(self, job_id: str) -> Job: ...
    async def create_function_async(self, request_json: str) -> Job: ...
    async def get_function(self, name: str, version: str) -> str: ...
+    async def list_functions(self) -> List[str]: ...
    async def drop_function(self, name: str, version: str) -> bool: ...
    async def list_jobs(self) -> List[JobInfo]: ...
-    async def get_job(self, job_id: str) -> Optional[JobDescription]: ...
    async def cancel_job(self, job_id: str) -> bool: ...
-    async def job_history(
-        self, job_id: Optional[str] = None
-    ) -> List[pa.RecordBatch]: ...
+    async def execute_query_async(
+        self,
+        query: str,
+        *,
+        default_namespace_path: Optional[List[str]] = None,
+    ) -> SqlQuery: ...
+    async def describe_query(self, query_id: UUID) -> QueryDescription: ...
    async def create_table(
        self,
        name: str,
@@ -235,9 +240,20 @@ class BlobFile:
 class Job:
    @property
    def id(self) -> Optional[str]: ...
+    @property
+    def _state(self) -> Optional[str]: ...
+    @property
+    def _description(self) -> Optional[JobDescription]: ...
    async def status(self) -> str: ...
    async def wait(self) -> Optional[str]: ...
    async def cancel(self) -> None: ...
+    async def refresh(self) -> None: ...
+    async def events(
+        self,
+        *,
+        limit: Optional[int] = None,
+        filter: Optional[str] = None,
+    ) -> pa.Table: ...

 class JobInfo:
    @property
@@ -269,10 +285,33 @@ class JobDescription:
    @property
    def creation_ms(self) -> int: ...
    @property
-    def spec_json(self) -> Optional[str]: ...
+    def _spec_json(self) -> Optional[str]: ...
+    @property
+    def _result_json(self) -> Optional[str]: ...
+    @property
+    def spec(self) -> Optional[Any]: ...
+    @property
+    def result(self) -> Optional[Any]: ...
    @property
    def failure(self) -> Optional[JobFailureInfo]: ...

+class SqlQuery:
+    @property
+    def id(self) -> UUID: ...
+    async def describe(self) -> QueryDescription: ...
+    async def reader(self) -> RecordBatchStream: ...
+    async def cancel(self) -> None: ...
+
+class QueryDescription:
+    @property
+    def id(self) -> UUID: ...
+    @property
+    def status(self) -> str: ...
+    @property
+    def progress(self) -> Optional[float]: ...
+    @property
+    def expires_at(self) -> Optional[datetime]: ...
+
 class Table:
    def name(self) -> str: ...
    def __repr__(self) -> str: ...
@@ -451,6 +490,7 @@ async def connect(
    api_key: Optional[str],
    region: Optional[str],
    host_override: Optional[str],
+    sql_host_override: Optional[str],
    read_consistency_interval: Optional[float],
    client_config: Optional[Union[ClientConfig, Dict[str, Any]]],
    storage_options: Optional[Dict[str, str]],
@@ -19,6 +19,7 @@ from typing import (
    Optional,
    Union,
 )
+from uuid import UUID

 if sys.version_info >= (3, 12):
    from typing import override
@@ -47,6 +48,9 @@ from . import __version__
 from ._lancedb import connect as lancedb_connect  # type: ignore
 from .functions import FunctionVersion, UdfDefinition
 from .job import AsyncJob, Job, _typed_job
+from .sql import AsyncQuery as AsyncSqlQuery
+from .sql import Query as SqlQuery
+from .sql import QueryDescription
 from .materialized_view import (
    AsyncMaterializedView,
    MaterializedView,
@@ -68,10 +72,11 @@ import deprecation

 if TYPE_CHECKING:
    import pyarrow as pa
+    from .arrow import AsyncRecordBatchReader
    from .pydantic import LanceModel

    from ._lancedb import Connection as LanceDbConnection
-    from ._lancedb import JobDescription, JobInfo
+    from ._lancedb import JobInfo
    from .common import DATA, URI
    from .embeddings import EmbeddingFunctionConfig
    from ._lancedb import Session
@@ -712,6 +717,24 @@ class DBConnection(EnforceOverrides):
            "Function catalog operations are not supported for this connection type"
        )

+    def list_functions(self) -> List[FunctionVersion]:
+        """List every published immutable Function version.
+
+        Results are ordered by Function name then version. Local connections
+        raise ``NotImplementedError``.
+
+        Examples
+        --------
+        List the identities available to use in Function-backed columns:
+
+        ```python
+        [(function.name, function.version) for function in db.list_functions()]
+        ```
+        """
+        raise NotImplementedError(
+            "Function catalog operations are not supported for this connection type"
+        )
+
    def drop_function(self, name: str, *, version: str) -> bool:
        """Drop one exact immutable Function version from the remote catalog.

@@ -722,26 +745,23 @@ class DBConnection(EnforceOverrides):
            "Function catalog operations are not supported for this connection type"
        )

-    def job(self, job_id: str) -> Job:
-        """A [Job][lancedb.job.Job] handle for a server-side job by id.
+    def open_job(self, job_id: str) -> Job:
+        """Open a server-side job by id, returning a handle with its record
+        already populated.

-        The handle is constructed without a server round trip; an unknown id
-        surfaces when the handle is used. Dropping the handle has no effect
-        on the job itself.
+        The returned [Job][lancedb.job.Job] answers for its own state,
+        specification, result, failure and event history, so there is no
+        separate connection-level call for any of them.
+
+        Raises `JobNotFoundError` when the server has no such job, the way
+        `open_table` does for a missing table.
        """
-        raise NotImplementedError("job is not supported for this connection type")
+        raise NotImplementedError("open_job is not supported for this connection type")

    def list_jobs(self) -> List[JobInfo]:
        """List server-side jobs across the database's tables."""
        raise NotImplementedError("list_jobs is not supported for this connection type")

-    def get_job(self, job_id: str) -> Optional[JobDescription]:
-        """Describe a single server-side job by id.
-
-        Returns None when the server has no such job.
-        """
-        raise NotImplementedError("get_job is not supported for this connection type")
-
    def cancel_job(self, job_id: str) -> bool:
        """Request cancellation of a server-side job by id.

@@ -753,14 +773,38 @@ class DBConnection(EnforceOverrides):
            "cancel_job is not supported for this connection type"
        )

-    def job_history(self, job_id: Optional[str] = None) -> List[pa.RecordBatch]:
-        """The lifecycle event history of a server-side job, as Arrow batches.
+    def execute_query(
+        self,
+        query: str,
+        *,
+        default_namespace_path: Optional[List[str]] = None,
+    ) -> pa.RecordBatchReader:
+        """Execute SQL and return a blocking Arrow reader.

-        Lists history across all jobs when `job_id` is None.
+        This submits through :meth:`execute_query_async` and waits until the
+        initial result stream is readable. It does not wait for the full query
+        to finish.
        """
-        raise NotImplementedError(
-            "job_history is not supported for this connection type"
-        )
+        return self.execute_query_async(
+            query,
+            default_namespace_path=default_namespace_path,
+        ).reader()
+
+    def execute_query_async(
+        self,
+        query: str,
+        *,
+        default_namespace_path: Optional[List[str]] = None,
+    ) -> SqlQuery:
+        """Start executing SQL and return its query handle.
+
+        Local connections do not support SQL.
+        """
+        raise NotImplementedError("SQL is not supported for this connection type")
+
+    def describe_query(self, query_id: UUID) -> QueryDescription:
+        """Describe a submitted SQL query by its connection-scoped id."""
+        raise NotImplementedError("SQL is not supported for this connection type")


 class LanceDBConnection(DBConnection):
@@ -857,6 +901,7 @@ class LanceDBConnection(DBConnection):
                None,
                None,
                None,
+                None,
                read_consistency_interval_secs,
                None,
                storage_options,
@@ -1405,14 +1450,11 @@ class LanceDBConnection(DBConnection):
        )

    @override
-    def job(self, job_id: str) -> Job:
-        """A [Job][lancedb.job.Job] handle for a server-side job by id.
-
-        The handle is constructed without a server round trip; an unknown id
-        surfaces when the handle is used. Dropping the handle has no effect
-        on the job itself.
+    def open_job(self, job_id: str) -> Job:
+        """Open a server-side job by id. See
+        [DBConnection.open_job][lancedb.db.DBConnection.open_job].
        """
-        return Job(self._conn.job(job_id))
+        return Job(LOOP.run(self._conn.open_job(job_id)))

    @override
    def create_function_async(self, definition: UdfDefinition) -> Job[FunctionVersion]:
@@ -1423,6 +1465,10 @@ class LanceDBConnection(DBConnection):
    def get_function(self, name: str, *, version: str) -> FunctionVersion:
        return LOOP.run(self._conn.get_function(name, version=version))

+    @override
+    def list_functions(self) -> List[FunctionVersion]:
+        return LOOP.run(self._conn.list_functions())
+
    @override
    def drop_function(self, name: str, *, version: str) -> bool:
        return LOOP.run(self._conn.drop_function(name, version=version))
@@ -1432,14 +1478,6 @@ class LanceDBConnection(DBConnection):
        """List server-side jobs across the database's tables."""
        return LOOP.run(self._conn.list_jobs())

-    @override
-    def get_job(self, job_id: str) -> Optional[JobDescription]:
-        """Describe a single server-side job by id.
-
-        Returns None when the server has no such job.
-        """
-        return LOOP.run(self._conn.get_job(job_id))
-
    @override
    def cancel_job(self, job_id: str) -> bool:
        """Request cancellation of a server-side job by id.
@@ -1450,14 +1488,6 @@ class LanceDBConnection(DBConnection):
        """
        return LOOP.run(self._conn.cancel_job(job_id))

-    @override
-    def job_history(self, job_id: Optional[str] = None) -> List[pa.RecordBatch]:
-        """The lifecycle event history of a server-side job, as Arrow batches.
-
-        Lists history across all jobs when `job_id` is None.
-        """
-        return LOOP.run(self._conn.job_history(job_id))
-
    @override
    def namespace_client(self) -> LanceNamespace:
        """Get the equivalent namespace client for this connection.
@@ -2228,15 +2258,11 @@ class AsyncConnection(object):
            namespace_path = []
        await self._inner.drop_all_tables(namespace_path=namespace_path)

-    def job(self, job_id: str) -> AsyncJob:
-        """An [AsyncJob][lancedb.job.AsyncJob] handle for a server-side job
-        by id.
-
-        The handle is constructed without a server round trip; an unknown id
-        surfaces when the handle is used. Dropping the handle has no effect
-        on the job itself.
+    async def open_job(self, job_id: str) -> AsyncJob:
+        """Open a server-side job by id. See
+        [DBConnection.open_job][lancedb.db.DBConnection.open_job].
        """
-        return AsyncJob(self._inner.job(job_id))
+        return AsyncJob(await self._inner.open_job(job_id))

    async def create_function_async(
        self, definition: UdfDefinition
@@ -2257,6 +2283,17 @@ class AsyncConnection(object):
        """Open one exact immutable Function version from the remote catalog."""
        return FunctionVersion.from_json(await self._inner.get_function(name, version))

+    async def list_functions(self) -> List[FunctionVersion]:
+        """List every published immutable Function version.
+
+        Results are ordered by Function name then version. Local connections
+        raise ``NotImplementedError``.
+        """
+        return [
+            FunctionVersion.from_json(value)
+            for value in await self._inner.list_functions()
+        ]
+
    async def drop_function(self, name: str, *, version: str) -> bool:
        """Drop one exact immutable Function version from the remote catalog."""
        return await self._inner.drop_function(name, version)
@@ -2265,13 +2302,6 @@ class AsyncConnection(object):
        """List server-side jobs across the database's tables."""
        return await self._inner.list_jobs()

-    async def get_job(self, job_id: str) -> Optional[JobDescription]:
-        """Describe a single server-side job by id.
-
-        Returns None when the server has no such job.
-        """
-        return await self._inner.get_job(job_id)
-
    async def cancel_job(self, job_id: str) -> bool:
        """Request cancellation of a server-side job by id.

@@ -2281,12 +2311,46 @@ class AsyncConnection(object):
        """
        return await self._inner.cancel_job(job_id)

-    async def job_history(self, job_id: Optional[str] = None) -> List[pa.RecordBatch]:
-        """The lifecycle event history of a server-side job, as Arrow batches.
+    async def execute_query(
+        self,
+        query: str,
+        *,
+        default_namespace_path: Optional[List[str]] = None,
+    ) -> AsyncRecordBatchReader:
+        """Execute SQL and return an asynchronous Arrow reader.

-        Lists history across all jobs when `job_id` is None.
+        This submits through :meth:`execute_query_async` and waits until the
+        initial result stream is readable. It does not wait for the full query
+        to finish.
        """
-        return await self._inner.job_history(job_id)
+        submitted = await self.execute_query_async(
+            query,
+            default_namespace_path=default_namespace_path,
+        )
+        return await submitted.reader()
+
+    async def execute_query_async(
+        self,
+        query: str,
+        *,
+        default_namespace_path: Optional[List[str]] = None,
+    ) -> AsyncSqlQuery:
+        """Start executing SQL and return its query handle.
+
+        The database from ``connect_async`` is used for unqualified database
+        references. The namespace defaults to ``["public"]``. Local
+        connections raise ``NotImplementedError``.
+        """
+        return AsyncSqlQuery(
+            await self._inner.execute_query_async(
+                query,
+                default_namespace_path=default_namespace_path,
+            )
+        )
+
+    async def describe_query(self, query_id: UUID) -> QueryDescription:
+        """Describe a submitted SQL query by its connection-scoped id."""
+        return await self._inner.describe_query(query_id)

    async def namespace_client(self) -> LanceNamespace:
        """Get the equivalent namespace client for this connection.
@@ -35,3 +35,9 @@ class JobCancelledError(RuntimeError):
    """Exception raised when an asynchronous job was cancelled."""

    pass
+
+
+class JobNotFoundError(ValueError):
+    """Exception raised when opening a job the server does not have."""
+
+    pass
@@ -470,11 +470,7 @@ class InputBinding(_RemoteValue):


 class OutputMapping(_RemoteValue):
-    """One stable result-field mapping.
-
-    Assignment state is outside the Slice 1 client contract. During the NULL
-    transition Lance exposes no public cell-flag identifier to persist here.
-    """
+    """One stable result-field mapping."""

    result_field: str
    output_name: str
@@ -484,6 +480,13 @@ class OutputMapping(_RemoteValue):
    nullable: bool


+class AssignmentMapping(_RemoteValue):
+    """Internal physical column preserving flattened struct validity."""
+
+    output_name: str
+    output_field_id: _Int32
+
+
 class FunctionBinding(_RemoteValue):
    """Immutable Function binding persisted by the Enterprise table service."""

@@ -491,6 +494,7 @@ class FunctionBinding(_RemoteValue):
    function: FunctionVersionRef
    inputs: tuple[InputBinding, ...]
    outputs: tuple[OutputMapping, ...]
+    assignment: Optional[AssignmentMapping] = None
    input_schema: Optional[Mapping[str, Any]] = None
    output_schema: Optional[Mapping[str, Any]] = None

@@ -521,6 +525,12 @@ class RefreshColumnResult(_RemoteValue):

 _FUNCTION_NAME = re.compile(r"^[A-Za-z_][A-Za-z0-9_.-]*$")
 _FUNCTION_BLOB_V2_TYPE = "blob_v2"
+_ARROW_EXTENSION_NAME_KEY = "ARROW:extension:name"
+_BLOB_V2_EXTENSION_NAME = "lance.blob.v2"
+_NESTED_BLOB_COLLECTION_ERROR = (
+    "unsupported Arrow type for Function signature: Blob v2 fields nested under "
+    "collection types are not supported"
+)


 _GRAMMAR_PRIMITIVES = (
@@ -591,6 +601,19 @@ def _validate_exact_arrow_field(field: pa.Field) -> None:
                "unsupported Arrow type for Function signature: lance.blob.v2 "
                f"requires a supported Blob storage layout, got {field}"
            )
+        metadata = {
+            (key.decode() if isinstance(key, bytes) else key): (
+                value.decode() if isinstance(value, bytes) else value
+            )
+            for key, value in (field.metadata or {}).items()
+        }
+        if metadata and metadata != {
+            _ARROW_EXTENSION_NAME_KEY: _BLOB_V2_EXTENSION_NAME
+        }:
+            raise TypeError(
+                "unsupported Arrow type for Function signature: lance.blob.v2 "
+                "field metadata must contain only its canonical extension marker"
+            )
    elif field.metadata:
        raise TypeError(
            "unsupported Arrow type for Function signature: field metadata "
@@ -655,23 +678,84 @@ def _canonical_arrow_field(field: pa.Field) -> str:
    return _canonical_arrow_type(field.type)


-def _exact_arrow_field(field: pa.Field) -> dict[str, Any]:
+def _blob_storage_type(field: pa.Field) -> pa.DataType:
+    data_type = field.type
+    if isinstance(data_type, pa.ExtensionType):
+        return data_type.storage_type
+    return data_type
+
+
+def _exact_blob_storage_type(field: pa.Field) -> dict[str, Any]:
+    storage = _blob_storage_type(field)
+    if not pa.types.is_struct(storage):
+        raise TypeError(
+            "unsupported Arrow type for Function signature: lance.blob.v2 "
+            "requires struct storage"
+        )
+    return {
+        "type": "struct",
+        "fields": [
+            {
+                "name": child.name,
+                "nullable": child.nullable,
+                "type": (
+                    {"type": "large_binary"}
+                    if pa.types.is_large_binary(child.type)
+                    else _exact_arrow_type(child.type)
+                ),
+            }
+            for child in storage
+        ],
+    }
+
+
+def _data_type_has_blob_v2(data_type: pa.DataType) -> bool:
+    if pa.types.is_struct(data_type):
+        return any(
+            _is_blob_v2_field(field) or _data_type_has_blob_v2(field.type)
+            for field in data_type
+        )
+    if (
+        pa.types.is_list(data_type)
+        or pa.types.is_large_list(data_type)
+        or pa.types.is_fixed_size_list(data_type)
+    ):
+        field = data_type.value_field
+        return _is_blob_v2_field(field) or _data_type_has_blob_v2(field.type)
+    if pa.types.is_map(data_type):
+        return any(
+            _is_blob_v2_field(field) or _data_type_has_blob_v2(field.type)
+            for field in (data_type.key_field, data_type.item_field)
+        )
+    return False
+
+
+def _exact_arrow_field(
+    field: pa.Field, *, inside_collection: bool = False
+) -> dict[str, Any]:
    _validate_exact_arrow_field(field)
    if _is_blob_v2_field(field):
-        raise TypeError(
-            "unsupported Arrow type for Function signature: nested Blob v2 "
-            "fields are not supported; declare Blob parameters or named result "
-            "fields directly"
-        )
+        if inside_collection:
+            raise TypeError(_NESTED_BLOB_COLLECTION_ERROR)
+        return {
+            "name": field.name,
+            "nullable": field.nullable,
+            "type": _exact_blob_storage_type(field),
+            "metadata": {
+                _ARROW_EXTENSION_NAME_KEY: _BLOB_V2_EXTENSION_NAME,
+            },
+        }
    value = {
        "name": field.name,
        "nullable": field.nullable,
-        "type": _exact_arrow_type(field.type),
+        "type": _exact_arrow_type(field.type, inside_collection=inside_collection),
    }
    return value


-def _exact_arrow_type(data_type: pa.DataType) -> dict[str, Any]:
+def _exact_arrow_type(
+    data_type: pa.DataType, *, inside_collection: bool = False
+) -> dict[str, Any]:
    for candidate, name in _GRAMMAR_PRIMITIVES:
        if data_type == candidate:
            return {"type": name}
@@ -685,7 +769,10 @@ def _exact_arrow_type(data_type: pa.DataType) -> dict[str, Any]:
            )
        return {
            "type": "struct",
-            "fields": [_exact_arrow_field(field) for field in fields],
+            "fields": [
+                _exact_arrow_field(field, inside_collection=inside_collection)
+                for field in fields
+            ],
        }
    if (
        pa.types.is_list(data_type)
@@ -710,11 +797,15 @@ def _exact_arrow_type(data_type: pa.DataType) -> dict[str, Any]:
                if pa.types.is_large_list(data_type)
                else "fixed_size_list"
            ),
-            "fields": [_exact_arrow_field(data_type.value_field)],
+            "fields": [
+                _exact_arrow_field(data_type.value_field, inside_collection=True)
+            ],
        }
        if pa.types.is_fixed_size_list(data_type):
            value["length"] = data_type.list_size
        return value
+    if pa.types.is_map(data_type) and _data_type_has_blob_v2(data_type):
+        raise TypeError(_NESTED_BLOB_COLLECTION_ERROR)
    raise TypeError(f"unsupported Arrow type for Function signature: {data_type}")


@@ -824,8 +915,6 @@ def _function_output(output: pa.DataType | pa.Field | pa.Schema) -> FunctionOutp

    if not fields:
        raise ValueError("named-struct Function output must contain at least one field")
-    if any(field.nullable for field in fields):
-        raise ValueError("Function output fields must be non-nullable")
    for field in fields:
        _validate_exact_arrow_field(field)
    names = [field.name for field in fields]
@@ -837,7 +926,7 @@ def _function_output(output: pa.DataType | pa.Field | pa.Schema) -> FunctionOutp
            FunctionResultField(
                name=field.name,
                arrow_type=_canonical_arrow_field(field),
-                nullable=False,
+                nullable=field.nullable,
            )
            for field in fields
        ),
@@ -1220,8 +1309,9 @@ def udf(

    Input and output signatures are inferred from supported annotations. For
    Arrow types annotations cannot express precisely, pass ``input_schema``
-    and ``output_schema`` together. Nullable outputs are rejected because V1
-    uses physical NULL to represent unassigned computed-column rows.
+    and ``output_schema`` together. Scalar outputs must be non-nullable. Every
+    named-struct field may be nullable; Enterprise preserves the struct's
+    validity when the result is expanded into sibling columns.

    Parameters
    ----------
@@ -1233,8 +1323,8 @@ def udf(
        Explicit input fields in the exact order of the callable parameters.
        Must be provided together with ``output_schema``.
    output_schema : pyarrow.DataType, pyarrow.Field, or pyarrow.Schema, optional
-        Explicit scalar or named-struct output. Must be non-nullable and be
-        provided together with ``input_schema``.
+        Explicit scalar or named-struct output. Scalar outputs must be
+        non-nullable. Must be provided together with ``input_schema``.
    pip : sequence of str, optional
        Pip requirements for the remote environment.
    conda : sequence of str, optional
@@ -1300,6 +1390,7 @@ def udf(


 __all__ = [
+    "AssignmentMapping",
    "ApplicationInput",
    "FunctionApplication",
    "FunctionArtifact",
@@ -4,15 +4,27 @@
 """Handles to operations a server may run asynchronously."""

 import asyncio
+import json
 from datetime import timedelta
 from typing import Any, Callable, Generic, Optional, TypeVar, cast

+import pyarrow as pa
+
 from lancedb.background_loop import LOOP

 from . import _lancedb
+from ._lancedb import JobDescription, JobFailureInfo, JobInfo

 T = TypeVar("T")

+__all__ = [
+    "AsyncJob",
+    "Job",
+    "JobDescription",
+    "JobFailureInfo",
+    "JobInfo",
+]
+

 class AsyncJob(Generic[T]):
    """A handle to an operation that may still be running.
@@ -78,6 +90,149 @@ class AsyncJob(Generic[T]):
            return
        await self._inner.cancel()

+    async def refresh(self) -> None:
+        """Ask the backend for this job's current state, and for a server-side
+        job its full record, then cache it for the properties below.
+
+        The properties are all `None` until this runs, because submitting an
+        operation returns only a job id. `status` fetches the whole record too;
+        `wait` records only the terminal state it establishes.
+        """
+        if self._inner is None:
+            return
+        await self._inner.refresh()
+
+    @property
+    def state(self) -> Optional[str]:
+        """The last observed lifecycle state, without contacting the backend.
+
+        `None` until the handle has talked to it. See :meth:`AsyncJob.refresh`.
+        """
+        if self._inner is None:
+            return "finished"
+        return self._inner._state
+
+    @property
+    def job_type(self) -> Optional[str]:
+        """The job's type, as the server names it.
+
+        `None` for an in-process job, which has no server-side record.
+        """
+        return self._field("job_type")
+
+    @property
+    def creation_ms(self) -> Optional[int]:
+        """When the job was created, in milliseconds since the epoch."""
+        return self._field("creation_ms")
+
+    @property
+    def spec(self) -> Optional[Any]:
+        """The job-type-specific specification it was submitted with."""
+        return self._field("spec")
+
+    @property
+    def result(self) -> Optional[Any]:
+        """The job-type-specific terminal result, as reported data rather than
+        the typed model :meth:`AsyncJob.wait` returns.
+
+        `None` until the job succeeds, so a job that never terminates reports
+        its progress through :meth:`AsyncJob.events` instead.
+        """
+        return self._field("result")
+
+    @property
+    def failure(self) -> Optional[JobFailureInfo]:
+        """Why the job failed, when it failed and the server reports a reason."""
+        return self._field("failure")
+
+    @property
+    def _spec_json(self) -> Optional[str]:
+        return self._field("_spec_json")
+
+    @property
+    def _result_json(self) -> Optional[str]:
+        return self._field("_result_json")
+
+    def _field(self, name: str) -> Optional[Any]:
+        description = self._inner._description if self._inner is not None else None
+        return getattr(description, name) if description is not None else None
+
+    async def events(
+        self,
+        *,
+        limit: Optional[int] = None,
+        filter: Optional[str] = None,
+    ) -> "pa.Table":
+        """This job's recorded lifecycle events.
+
+        Where the properties above report a terminal result only once the job
+        reaches one, events are written as the job runs and outlive the workers
+        that produced them. A distributed job records a `claim`/`claim_complete`
+        pair per unit of work, each carrying `rows_processed`, so a job that
+        never finishes still accounts for what it did.
+
+        Parameters
+        ----------
+        limit: int, optional
+            Maximum event rows to return. The server caps results at 1000 by
+            default and 10,000 at most, and truncates without saying so, so
+            pass this for a job that emits an event per fragment.
+        filter: str, optional
+            SQL-like expression over the `state`, `updated_by`, `emitted_from`,
+            `emitted_by`, and `claim_entity` columns, such as
+            ``state = 'claim_complete'``.
+        """
+        if self._inner is None:
+            raise NotImplementedError(
+                "job event history is only available for server-side jobs"
+            )
+        return await self._inner.events(limit=limit, filter=filter)
+
+    def __repr__(self) -> str:
+        return _job_repr("AsyncJob", self)
+
+
+_REPR_INDENT = " " * 4
+
+
+def _repr_payload(value: Any) -> str:
+    """Render a job payload as indented JSON, aligned under its field."""
+    try:
+        rendered = json.dumps(value, indent=4)
+    except TypeError:
+        return repr(value)
+    return rendered.replace("\n", "\n" + _REPR_INDENT)
+
+
+def _job_repr(kind: str, job: Any) -> str:
+    """Render every field the handle currently knows, omitting the rest.
+
+    One field per line, with the JSON payloads indented, because a refresh
+    job's spec and result are the point of printing it.
+    """
+    state = job.state
+    if state is None:
+        # Nothing has been fetched yet, so there is nothing to lay out.
+        known = f"id={job.id!r}, " if job.id is not None else ""
+        return f"{kind}({known}not refreshed)"
+
+    fields = []
+    if job.id is not None:
+        fields.append(f"id={job.id!r}")
+    fields.append(f"state={state!r}")
+    for name in ("job_type", "creation_ms"):
+        value = getattr(job, name)
+        if value is not None:
+            fields.append(f"{name}={value!r}")
+    for name in ("spec", "result"):
+        value = getattr(job, name)
+        if value is not None:
+            fields.append(f"{name}={_repr_payload(value)}")
+    if job.failure is not None:
+        fields.append(f"failure={job.failure!r}")
+    body = "".join(f"\n{_REPR_INDENT}{field}," for field in fields)
+    return f"{kind}({body}\n)"
+

 class Job(Generic[T]):
    """Synchronous counterpart of `AsyncJob` with the same result type."""
@@ -122,6 +277,75 @@ class Job(Generic[T]):
            return
        LOOP.run(self._inner.cancel())

+    def refresh(self) -> None:
+        """Ask the backend for this job's current state and record.
+
+        See :meth:`AsyncJob.refresh`.
+        """
+        if self._inner is None:
+            return
+        LOOP.run(self._inner.refresh())
+
+    @property
+    def state(self) -> Optional[str]:
+        """The last observed lifecycle state. See :attr:`AsyncJob.state`."""
+        return self._inner.state if self._inner is not None else "finished"
+
+    @property
+    def job_type(self) -> Optional[str]:
+        """The job's type. See :attr:`AsyncJob.job_type`."""
+        return self._field("job_type")
+
+    @property
+    def creation_ms(self) -> Optional[int]:
+        """When the job was created. See :attr:`AsyncJob.creation_ms`."""
+        return self._field("creation_ms")
+
+    @property
+    def spec(self) -> Optional[Any]:
+        """The job's specification. See :attr:`AsyncJob.spec`."""
+        return self._field("spec")
+
+    @property
+    def result(self) -> Optional[Any]:
+        """The job's terminal result. See :attr:`AsyncJob.result`."""
+        return self._field("result")
+
+    @property
+    def failure(self) -> Optional[JobFailureInfo]:
+        """Why the job failed. See :attr:`AsyncJob.failure`."""
+        return self._field("failure")
+
+    @property
+    def _spec_json(self) -> Optional[str]:
+        return self._field("_spec_json")
+
+    @property
+    def _result_json(self) -> Optional[str]:
+        return self._field("_result_json")
+
+    def _field(self, name: str) -> Optional[Any]:
+        return getattr(self._inner, name) if self._inner is not None else None
+
+    def events(
+        self,
+        *,
+        limit: Optional[int] = None,
+        filter: Optional[str] = None,
+    ) -> "pa.Table":
+        """This job's recorded lifecycle events.
+
+        See :meth:`AsyncJob.events`.
+        """
+        if self._inner is None:
+            raise NotImplementedError(
+                "job event history is only available for server-side jobs"
+            )
+        return LOOP.run(self._inner.events(limit=limit, filter=filter))
+
+    def __repr__(self) -> str:
+        return _job_repr("Job", self)
+

 def _typed_job(
    inner: "_lancedb.Job", result_decoder: Callable[[str], T]
@@ -42,6 +42,8 @@ class MaterializedViewDefinition:
    """Cap on the number of rows the view holds."""
    inputs: List[str] = field(default_factory=list)
    """Source columns the projections and filter read."""
+    source_namespace: List[str] = field(default_factory=list)
+    """Namespace holding the source table; empty is the root namespace."""


 def _definition_from_schema(
@@ -53,7 +55,8 @@ def _definition_from_schema(
        raise ValueError(f"Table '{name}' is not a materialized view")
    value = json.loads(raw)
    kind = value.get("kind")
-    if kind != "select":
+    # "namespaced_select" keeps older readers from resolving the source at root.
+    if kind not in ("select", "namespaced_select"):
        raise NotImplementedError(
            f"materialized view '{name}' is defined by '{kind}', which this "
            "version of lancedb cannot refresh"
@@ -66,6 +69,7 @@ def _definition_from_schema(
        filter=value.get("filter"),
        limit=value.get("limit"),
        inputs=value.get("inputs", []),
+        source_namespace=value.get("source_namespace", []),
    )


@@ -12,6 +12,7 @@ from __future__ import annotations

 import sys
 from typing import TYPE_CHECKING, Any, Dict, Iterable, List, Optional, Union
+from uuid import UUID

 if sys.version_info >= (3, 12):
    from typing import override
@@ -48,8 +49,11 @@ from lancedb._lancedb import (
    connect_namespace_client as _connect_namespace_client,
 )
 from lancedb.background_loop import LOOP
+from lancedb.arrow import AsyncRecordBatchReader
 from lancedb.db import AsyncConnection, DBConnection
 from lancedb.job import AsyncJob, Job
+from lancedb.sql import AsyncQuery as AsyncSqlQuery
+from lancedb.sql import QueryDescription
 from lance_namespace import (
    LanceNamespace,
    connect as namespace_connect,
@@ -1447,6 +1451,37 @@ class AsyncLanceNamespaceDBConnection:
            namespace_path=namespace_path, page_token=page_token, limit=limit
        )

+    async def execute_query(
+        self,
+        query: str,
+        *,
+        default_namespace_path: Optional[List[str]] = None,
+    ) -> AsyncRecordBatchReader:
+        """Execute SQL when supported by the underlying connection."""
+        return await self._inner.execute_query(
+            query,
+            default_namespace_path=default_namespace_path,
+        )
+
+    async def execute_query_async(
+        self,
+        query: str,
+        *,
+        default_namespace_path: Optional[List[str]] = None,
+    ) -> AsyncSqlQuery:
+        """Start executing SQL when supported by the underlying connection.
+
+        Namespace-backed local connections do not support SQL.
+        """
+        return await self._inner.execute_query_async(
+            query,
+            default_namespace_path=default_namespace_path,
+        )
+
+    async def describe_query(self, query_id: UUID) -> QueryDescription:
+        """Describe a submitted SQL query when supported."""
+        return await self._inner.describe_query(query_id)
+
    async def namespace_client(self) -> LanceNamespace:
        """Get the namespace client for this connection.

@@ -9,6 +9,7 @@ from concurrent.futures import ThreadPoolExecutor
 import sys
 from typing import TYPE_CHECKING, Any, Dict, Iterable, List, Optional, Union
 from urllib.parse import urlparse
+from uuid import UUID
 import warnings

 if sys.version_info >= (3, 12):
@@ -25,10 +26,12 @@ from ..common import DATA
 from ..db import DBConnection, LOOP
 from ..functions import FunctionVersion, UdfDefinition
 from ..job import AsyncJob, Job
+from ..sql import Query as SqlQuery
+from ..sql import QueryDescription
 from ..materialized_view import MaterializedView, SelectArg

 if TYPE_CHECKING:
-    from .._lancedb import JobDescription, JobInfo
+    from .._lancedb import JobInfo
 from ..embeddings import EmbeddingFunctionConfig
 from lance_namespace import (
    LanceNamespace,
@@ -116,6 +119,7 @@ class RemoteDBConnection(DBConnection):
        read_timeout: Optional[float] = None,
        storage_options: Optional[Dict[str, str]] = None,
        read_consistency_interval: Optional[timedelta] = None,
+        sql_host_override: Optional[str] = None,
    ):
        """Connect to a remote LanceDB database."""
        if isinstance(client_config, dict):
@@ -161,6 +165,7 @@ class RemoteDBConnection(DBConnection):
        self.api_key = api_key
        self.region = region
        self.host_override = host_override
+        self.sql_host_override = sql_host_override
        self.storage_options = storage_options
        self.db_name = parsed.netloc

@@ -175,6 +180,7 @@ class RemoteDBConnection(DBConnection):
                api_key=api_key,
                region=region,
                host_override=host_override,
+                sql_host_override=sql_host_override,
                client_config=client_config,
                storage_options=storage_options,
                read_consistency_interval=read_consistency_interval,
@@ -193,6 +199,7 @@ class RemoteDBConnection(DBConnection):
                "api_key": self.api_key,
                "region": self.region,
                "host_override": self.host_override,
+                "sql_host_override": self.sql_host_override,
                "client_config": _client_config_to_dict(self.client_config),
                "storage_options": self.storage_options,
            }
@@ -732,14 +739,11 @@ class RemoteDBConnection(DBConnection):
        )

    @override
-    def job(self, job_id: str) -> Job:
-        """A [Job][lancedb.job.Job] handle for a server-side job by id.
-
-        The handle is constructed without a server round trip; an unknown id
-        surfaces when the handle is used. Dropping the handle has no effect
-        on the job itself.
+    def open_job(self, job_id: str) -> Job:
+        """Open a server-side job by id. See
+        [DBConnection.open_job][lancedb.db.DBConnection.open_job].
        """
-        return Job(self._conn.job(job_id))
+        return Job(LOOP.run(self._conn.open_job(job_id)))

    @override
    def create_function_async(self, definition: UdfDefinition) -> Job[FunctionVersion]:
@@ -749,6 +753,10 @@ class RemoteDBConnection(DBConnection):
    def get_function(self, name: str, *, version: str) -> FunctionVersion:
        return LOOP.run(self._conn.get_function(name, version=version))

+    @override
+    def list_functions(self) -> List[FunctionVersion]:
+        return LOOP.run(self._conn.list_functions())
+
    @override
    def drop_function(self, name: str, *, version: str) -> bool:
        return LOOP.run(self._conn.drop_function(name, version=version))
@@ -758,14 +766,6 @@ class RemoteDBConnection(DBConnection):
        """List server-side jobs across the database's tables."""
        return LOOP.run(self._conn.list_jobs())

-    @override
-    def get_job(self, job_id: str) -> Optional["JobDescription"]:
-        """Describe a single server-side job by id.
-
-        Returns None when the server has no such job.
-        """
-        return LOOP.run(self._conn.get_job(job_id))
-
    @override
    def cancel_job(self, job_id: str) -> bool:
        """Request cancellation of a server-side job by id.
@@ -777,12 +777,35 @@ class RemoteDBConnection(DBConnection):
        return LOOP.run(self._conn.cancel_job(job_id))

    @override
-    def job_history(self, job_id: Optional[str] = None) -> List[pa.RecordBatch]:
-        """The lifecycle event history of a server-side job, as Arrow batches.
+    def execute_query_async(
+        self,
+        query: str,
+        *,
+        default_namespace_path: Optional[List[str]] = None,
+    ) -> SqlQuery:
+        """Start executing SQL through this remote connection.

-        Lists history across all jobs when `job_id` is None.
+        Unqualified tables use this connection's database and the
+        ``["public"]`` namespace by default. Fully qualified table names may
+        reference other databases available to the same deployment.
        """
-        return LOOP.run(self._conn.job_history(job_id))
+        return SqlQuery(
+            LOOP.run(
+                self._conn.execute_query_async(
+                    query,
+                    default_namespace_path=default_namespace_path,
+                )
+            )
+        )
+
+    @override
+    def describe_query(self, query_id: UUID) -> QueryDescription:
+        """Describe a submitted SQL query by its connection-scoped id."""
+        return LOOP.run(
+            self._conn.describe_query(
+                query_id,
+            )
+        )

    @override
    def namespace_client(self) -> LanceNamespace:
@@ -177,4 +177,7 @@ class OAuthProvider(HeaderProvider):
        if not self._current_token:
            raise RuntimeError("Failed to obtain OAuth token")

-        return {"Authorization": f"Bearer {self._current_token}"}
+        return {
+            "Authorization": f"Bearer {self._current_token}",
+            "x-lancedb-credential-type": "oidc",
+        }
@@ -548,6 +548,7 @@ class RemoteTable(Table):
        LOOP.run(
            self._table.create_index(
                column,
+                replace=replace,
                config=config,
                wait_timeout=wait_timeout,
                name=name,
@@ -0,0 +1,88 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+"""Handles to SQL queries running on a remote database."""
+
+from uuid import UUID
+
+import pyarrow as pa
+
+from lancedb.background_loop import LOOP
+
+from . import _lancedb
+from .arrow import AsyncRecordBatchReader
+
+QueryDescription = _lancedb.QueryDescription
+
+
+class AsyncQuery:
+    """A handle to a submitted SQL query on an asynchronous connection."""
+
+    def __init__(self, inner: "_lancedb.SqlQuery"):
+        self._inner = inner
+
+    @property
+    def id(self) -> UUID:
+        """The stable identifier scoped to the connection that submitted it."""
+        return self._inner.id
+
+    async def describe(self) -> QueryDescription:
+        """Get a point-in-time description of the query."""
+        return await self._inner.describe()
+
+    async def reader(self) -> AsyncRecordBatchReader:
+        """Wait for the initial result stream and return its Arrow reader.
+
+        Results are single-consumer. Calling this method more than once on the
+        same query raises an error. Later batches are streamed as they become
+        available without waiting for the full query to finish.
+        """
+        return AsyncRecordBatchReader(await self._inner.reader())
+
+    async def cancel(self) -> None:
+        """Request cancellation of the query."""
+        await self._inner.cancel()
+
+
+class Query:
+    """Synchronous counterpart of :class:`AsyncQuery`."""
+
+    def __init__(self, inner: AsyncQuery):
+        self._inner = inner
+
+    @property
+    def id(self) -> UUID:
+        """The stable identifier scoped to the connection that submitted it."""
+        return self._inner.id
+
+    def describe(self) -> QueryDescription:
+        """Get a point-in-time description of the query."""
+        return LOOP.run(self._inner.describe())
+
+    def reader(self) -> pa.RecordBatchReader:
+        """Wait for the initial result stream and return a blocking reader.
+
+        Results are single-consumer. Calling this method more than once on the
+        same query raises an error. Later batches block only until they become
+        available, without waiting for the full query to finish.
+        """
+        reader = LOOP.run(self._inner.reader())
+
+        def next_batch():
+            try:
+                return LOOP.run(reader.__anext__())
+            except StopAsyncIteration:
+                return None
+
+        def batches():
+            while (batch := next_batch()) is not None:
+                yield batch
+
+        return pa.RecordBatchReader.from_batches(reader.schema, batches())
+
+    def cancel(self) -> None:
+        """Request cancellation of the query."""
+        LOOP.run(self._inner.cancel())
+
+
+__all__ = ["AsyncQuery", "Query", "QueryDescription"]
@@ -1547,7 +1547,9 @@ class Table(ABC):
        on: Union[str, Iterable[str]]
            A column (or columns) to join on.  This is how records from the
            source table and target table are matched.  Typically this is some
-            kind of key or id column.
+            kind of key or id column.  Passing several columns matches on the
+            composite key: a source row updates a target row only when it
+            agrees on every one of them.

        Examples
        --------
@@ -1791,6 +1793,9 @@ class Table(ABC):
        The result has the same length and order as ``row_ids``. Null blobs
        produce null slots; valid empty blobs produce ``b""``.

+        ``_rowid`` values stay valid after compaction when the table has stable
+        row ids.
+
        Convenience for small payloads. For large values use
        :meth:`fetch_blob_files`.
        """
@@ -1808,6 +1813,9 @@ class Table(ABC):
        The result has the same length and order as ``requests``; null blobs
        produce null slots and empty ranges on non-null blobs produce ``b""``.

+        ``_rowid`` values stay valid after compaction when the table has stable
+        row ids.
+
        Row IDs can be obtained from a query with ``with_row_id(True)``. This
        API is currently supported only by local tables.
        """
@@ -1823,6 +1831,9 @@ class Table(ABC):
        ``_rowid`` or a ``_lance_row_id`` field on the blob descriptor. Null
        rows are ``None``. Remote tables require LanceDB Cloud server 0.5.0 or
        newer.
+
+        ``_rowid`` values stay valid after compaction when the table has stable
+        row ids.
        """

    @abstractmethod
@@ -5701,7 +5712,9 @@ class AsyncTable:
        on: Union[str, Iterable[str]]
            A column (or columns) to join on.  This is how records from the
            source table and target table are matched.  Typically this is some
-            kind of key or id column.
+            kind of key or id column.  Passing several columns matches on the
+            composite key: a source row updates a target row only when it
+            agrees on every one of them.

        Examples
        --------
@@ -66,6 +66,25 @@ def _row_ids_by_id(table):
    return dict(zip(hits["id"].to_pylist(), hits["_rowid"].to_pylist()))


+def _assert_missing_blob_row_ids(exc_info):
+    message = str(exc_info.value)
+    assert "row ids" in message
+    assert "rowaddr" not in message
+    assert "fragment" not in message
+
+
+def _assert_fetch_apis_reject_missing_row_ids(table, row_ids):
+    with pytest.raises(ValueError) as exc_info:
+        table.fetch_blobs("image", row_ids)
+    _assert_missing_blob_row_ids(exc_info)
+    with pytest.raises(ValueError) as exc_info:
+        table.fetch_blob_files("image", row_ids)
+    _assert_missing_blob_row_ids(exc_info)
+    with pytest.raises(ValueError) as exc_info:
+        table.fetch_blob_ranges("image", [(row_id, 0, 1) for row_id in row_ids])
+    _assert_missing_blob_row_ids(exc_info)
+
+
 def test_blob_factory_declares_v2_field():
    field = lancedb.blob("image")
    assert isinstance(field.type, pa.ExtensionType)
@@ -278,7 +297,10 @@ def test_blob_v2_projection_sources_use_typed_column_name():


 def _legacy_v1_table(name):
-    db = lancedb.connect("memory:///")
+    # Legacy v1 blob columns are only writable at file version <= 2.1.
+    db = lancedb.connect(
+        "memory:///", storage_options={"new_table_data_storage_version": "2.1"}
+    )
    schema = pa.schema(
        [
            pa.field("id", pa.int64()),
@@ -691,6 +713,25 @@ def test_fetch_blobs_accepts_query_result():
    assert {blobs[i].as_py() for i in range(len(blobs))} == {b"gamma"}


+def test_fetch_blobs_after_compact_with_stable_row_ids(tmp_path):
+    db = lancedb.connect(
+        tmp_path, storage_options={"new_table_enable_stable_row_ids": "true"}
+    )
+    schema = pa.schema([pa.field("id", pa.int64()), lancedb.blob("image")])
+    table = db.create_table("t", schema=schema)
+    table.add([{"id": 1, "image": b"frag-one"}])
+    table.add([{"id": 2, "image": b"frag-two"}])
+    by_id = _row_ids_by_id(table)
+    ids = [by_id[1], by_id[2]]
+
+    table.optimize()
+
+    blobs = table.fetch_blobs("image", ids)
+    assert blobs.to_pylist() == [b"frag-one", b"frag-two"]
+    ranges = table.fetch_blob_ranges("image", [(ids[0], 5, 3), (ids[1], 5, 3)])
+    assert ranges.to_pylist() == [b"one", b"two"]
+
+
 def test_fetch_blobs_preserves_null_and_empty_values():
    table = _blob_table(
        "nulls",
@@ -739,8 +780,25 @@ def test_fetch_blob_ranges_validates_requests():
    with pytest.raises(ValueError, match="offset \\+ length overflowed"):
        table.fetch_blob_ranges("image", [(row_id, 2**64 - 1, 1)])

-    with pytest.raises(ValueError, match="row IDs"):
+    with pytest.raises(ValueError) as exc_info:
        table.fetch_blob_ranges("image", [(2**64 - 1, 0, 1)])
+    _assert_missing_blob_row_ids(exc_info)
+
+
+def test_fetch_blob_apis_reject_missing_fragment_row_addr():
+    table = _blob_table("missing_frag", [{"id": 1, "image": b"x"}])
+    live = _row_ids_by_id(table)[1]
+    _assert_fetch_apis_reject_missing_row_ids(table, [1 << 32, live])
+
+
+def test_fetch_blob_apis_reject_deleted_row_ids():
+    table = _blob_table(
+        "deleted_rows",
+        [{"id": 1, "image": b"one"}, {"id": 2, "image": b"two"}],
+    )
+    by_id = _row_ids_by_id(table)
+    table.delete("id = 2")
+    _assert_fetch_apis_reject_missing_row_ids(table, [by_id[2], by_id[1]])


 def test_fetch_blob_ranges_empty_requests_returns_empty_array():
@@ -668,6 +668,167 @@ def test_blob_fields_use_the_scalar_function_semantic_type():
    assert signature.output.arrow_type == "blob_v2"


+def test_whole_named_struct_function_can_include_a_blob_result_field():
+    @udf(
+        input_schema=pa.schema([lancedb.blob("image", nullable=False)]),
+        output_schema=pa.field(
+            "payload",
+            pa.struct(
+                [
+                    pa.field("mime_type", pa.string(), nullable=False),
+                    lancedb.blob("image", nullable=False),
+                ]
+            ),
+            nullable=False,
+        ),
+    )
+    def inspect_blob(image):
+        return {"mime_type": "image/png", "image": image}
+
+    output = inspect_blob.registration_request.signature.output
+    assert output.kind == "named_struct"
+    assert [(field.name, field.arrow_type) for field in output.fields] == [
+        ("mime_type", "utf8"),
+        ("image", "blob_v2"),
+    ]
+
+
+def test_struct_blob_signature_fields_preserve_exact_metadata_and_nullability():
+    nested_input = pa.field(
+        "payload",
+        pa.struct(
+            [
+                pa.field("mime_type", pa.string(), nullable=False),
+                pa.field(
+                    "nested",
+                    pa.struct([lancedb.blob("image", nullable=True)]),
+                    nullable=True,
+                ),
+            ]
+        ),
+        nullable=True,
+    )
+    nested_output = pa.field(
+        "result",
+        pa.struct(
+            [
+                pa.field("mime_type", pa.string(), nullable=False),
+                pa.field(
+                    "nested",
+                    pa.struct([lancedb.blob("image", nullable=True)]),
+                    nullable=False,
+                ),
+            ]
+        ),
+        nullable=False,
+    )
+
+    @udf(input_schema=pa.schema([nested_input]), output_schema=nested_output)
+    def copy_payload(payload):
+        return payload
+
+    signature = copy_payload.registration_request.signature
+    input_type = json.loads(signature.inputs[0].arrow_type)
+    assert input_type["fields"][1]["nullable"] is True
+    input_blob = input_type["fields"][1]["type"]["fields"][0]
+    assert input_blob["nullable"] is True
+    assert input_blob["metadata"] == {"ARROW:extension:name": "lance.blob.v2"}
+
+    assert signature.output.kind == "named_struct"
+    nested_result = next(
+        field for field in signature.output.fields if field.name == "nested"
+    )
+    output_type = json.loads(nested_result.arrow_type)
+    output_blob = output_type["fields"][0]
+    assert output_blob["nullable"] is True
+    assert output_blob["metadata"] == {"ARROW:extension:name": "lance.blob.v2"}
+
+
+def test_struct_blob_signature_supports_multiple_struct_levels():
+    recursive = pa.field(
+        "value",
+        pa.struct(
+            [
+                pa.field(
+                    "level_1",
+                    pa.struct(
+                        [
+                            pa.field(
+                                "level_2",
+                                pa.struct([lancedb.blob("image", nullable=False)]),
+                                nullable=False,
+                            )
+                        ]
+                    ),
+                    nullable=False,
+                )
+            ]
+        ),
+        nullable=False,
+    )
+
+    @udf(
+        input_schema=pa.schema([recursive]),
+        output_schema=pa.field("size", pa.int64(), nullable=False),
+    )
+    def blob_size(value):
+        return len(value["level_1"]["level_2"]["image"])
+
+    encoded = json.loads(blob_size.registration_request.signature.inputs[0].arrow_type)
+    blob = encoded["fields"][0]["type"]["fields"][0]["type"]["fields"][0]
+    assert blob["metadata"]["ARROW:extension:name"] == "lance.blob.v2"
+
+
+@pytest.mark.parametrize(
+    "data_type",
+    [
+        pa.list_(lancedb.blob("item", nullable=False)),
+        pa.large_list(lancedb.blob("item", nullable=False)),
+        pa.list_(lancedb.blob("item", nullable=False), 2),
+        pa.map_(pa.string(), lancedb.blob("value", nullable=False).type),
+    ],
+)
+def test_blob_signature_rejects_collection_ancestors(data_type):
+    with pytest.raises(
+        TypeError,
+        match="Blob v2 fields nested under collection types are not supported",
+    ):
+
+        @udf(
+            input_schema=pa.schema([pa.field("value", data_type, nullable=False)]),
+            output_schema=pa.field("size", pa.int64(), nullable=False),
+        )
+        def blob_size(value):
+            return len(value)
+
+
+def test_blob_signature_rejects_collection_below_a_struct():
+    nested = pa.field(
+        "value",
+        pa.struct(
+            [
+                pa.field(
+                    "images",
+                    pa.list_(lancedb.blob("item", nullable=False)),
+                    nullable=False,
+                )
+            ]
+        ),
+        nullable=False,
+    )
+    with pytest.raises(
+        TypeError,
+        match="Blob v2 fields nested under collection types are not supported",
+    ):
+
+        @udf(
+            input_schema=pa.schema([nested]),
+            output_schema=pa.field("size", pa.int64(), nullable=False),
+        )
+        def blob_size(value):
+            return len(value["images"])
+
+
 def test_named_struct_function_can_include_a_blob_result_field():
    @udf(
        input_schema=pa.schema([lancedb.blob("image", nullable=False)]),
@@ -689,6 +850,43 @@ def test_named_struct_function_can_include_a_blob_result_field():
    ]


+def test_named_struct_function_preserves_nullable_result_fields():
+    @udf(
+        input_schema=pa.schema([pa.field("value", pa.int64(), nullable=False)]),
+        output_schema=pa.schema(
+            [
+                pa.field("result", pa.int64(), nullable=True),
+                pa.field("failure_code", pa.int32(), nullable=False),
+            ]
+        ),
+    )
+    def nullable_result(value):
+        return {"result": value, "failure_code": 0}
+
+    output = nullable_result.registration_request.signature.output
+    assert [(field.name, field.nullable) for field in output.fields] == [
+        ("result", True),
+        ("failure_code", False),
+    ]
+
+    @udf(
+        input_schema=pa.schema([pa.field("value", pa.int64(), nullable=False)]),
+        output_schema=pa.schema(
+            [
+                pa.field("result", pa.int64(), nullable=True),
+                pa.field("failure_code", pa.int32(), nullable=True),
+            ]
+        ),
+    )
+    def all_nullable(value):
+        return {"result": value, "failure_code": None}
+
+    assert all(
+        field.nullable
+        for field in all_nullable.registration_request.signature.output.fields
+    )
+
+
 def test_metadata_marked_blob_field_uses_the_semantic_type():
    extension = lancedb.blob("image", nullable=False).type
    storage = (
@@ -729,22 +927,6 @@ def test_blob_marker_rejects_invalid_storage_layout():
            return len(image)


-def test_nested_blob_signature_field_has_a_clear_error():
-    nested = pa.field(
-        "value",
-        pa.struct([lancedb.blob("image", nullable=False)]),
-        nullable=False,
-    )
-    with pytest.raises(TypeError, match="nested Blob v2 fields are not supported"):
-
-        @udf(
-            input_schema=pa.schema([nested]),
-            output_schema=pa.field("size", pa.int64(), nullable=False),
-        )
-        def blob_size(value):
-            return len(value["image"])
-
-
 def test_nested_non_blob_extension_is_not_silently_unwrapped():
    class TestExtension(pa.ExtensionType):
        def __init__(self):
@@ -1017,6 +1199,8 @@ def test_local_function_catalog_operations_are_not_supported(tmp_path):
        db.create_function_async(normalize_score)
    with pytest.raises(NotImplementedError, match=message):
        db.get_function("normalize_score", version="fv_exact")
+    with pytest.raises(NotImplementedError, match=message):
+        db.list_functions()
    with pytest.raises(NotImplementedError, match=message):
        db.drop_function("normalize_score", version="fv_exact")

@@ -1064,6 +1248,22 @@ def _mock_remote_function_catalog():
                    "version": "fv_exact",
                }
                response = state["version"]
+            elif self.path == "/v1/functions/list":
+                assert body["include_definition"] is True
+                if "page_token" not in body:
+                    response = {
+                        "functions": [
+                            {
+                                "name": "normalize_score",
+                                "version": "fv_exact",
+                                "definition": state["version"],
+                            }
+                        ],
+                        "page_token": "next",
+                    }
+                else:
+                    assert body["page_token"] == "next"
+                    response = {"functions": []}
            elif self.path == "/v1/functions/drop":
                assert body == {
                    "name": "normalize_score",
@@ -1130,6 +1330,49 @@ def test_blocking_remote_registration_returns_function_version():
    ]


+def test_remote_list_functions_paginates_and_returns_typed_versions():
+    with _mock_remote_function_catalog() as (host, state):
+        db = lancedb.connect(
+            "db://dev",
+            api_key="fake",
+            host_override=host,
+            client_config={"retry_config": {"retries": 0}},
+        )
+        created = db.create_function(normalize_score)
+        state["requests"].clear()
+        functions = db.list_functions()
+
+    assert functions == [created]
+    assert state["requests"] == [
+        ("/v1/functions/list", {"include_definition": True}),
+        (
+            "/v1/functions/list",
+            {"include_definition": True, "page_token": "next"},
+        ),
+    ]
+
+
+@pytest.mark.asyncio
+async def test_async_remote_list_functions_returns_typed_versions():
+    with _mock_remote_function_catalog() as (host, state):
+        db = await lancedb.connect_async(
+            "db://dev",
+            api_key="fake",
+            host_override=host,
+            client_config={"retry_config": {"retries": 0}},
+        )
+        registration = await db.create_function_async(normalize_score)
+        created = await registration.wait()
+        state["requests"].clear()
+        functions = await db.list_functions()
+
+    assert functions == [created]
+    assert [path for path, _ in state["requests"]] == [
+        "/v1/functions/list",
+        "/v1/functions/list",
+    ]
+
+
 def test_remote_drop_function_sends_exact_version():
    with _mock_remote_function_catalog() as (host, state):
        db = lancedb.connect(
@@ -54,7 +54,10 @@ class TestOAuthProvider:
        provider = OAuthProvider(fetcher)
        headers = provider.get_headers()

-        assert headers == {"Authorization": "Bearer token123"}
+        assert headers == {
+            "Authorization": "Bearer token123",
+            "x-lancedb-credential-type": "oidc",
+        }
        assert provider._current_token == "token123"
        assert provider._token_expires_at is not None

@@ -73,14 +76,20 @@ class TestOAuthProvider:

        # First call
        headers1 = provider.get_headers()
-        assert headers1 == {"Authorization": "Bearer token1"}
+        assert headers1 == {
+            "Authorization": "Bearer token1",
+            "x-lancedb-credential-type": "oidc",
+        }

        # Wait for token to expire
        time.sleep(1.1)

        # Second call should refresh
        headers2 = provider.get_headers()
-        assert headers2 == {"Authorization": "Bearer token2"}
+        assert headers2 == {
+            "Authorization": "Bearer token2",
+            "x-lancedb-credential-type": "oidc",
+        }
        assert call_count == 2

    def test_no_expiry_info(self):
@@ -92,12 +101,18 @@ class TestOAuthProvider:
        provider = OAuthProvider(fetcher)
        headers = provider.get_headers()

-        assert headers == {"Authorization": "Bearer permanent_token"}
+        assert headers == {
+            "Authorization": "Bearer permanent_token",
+            "x-lancedb-credential-type": "oidc",
+        }
        assert provider._token_expires_at is None

        # Should not refresh on second call
        headers2 = provider.get_headers()
-        assert headers2 == {"Authorization": "Bearer permanent_token"}
+        assert headers2 == {
+            "Authorization": "Bearer permanent_token",
+            "x-lancedb-credential-type": "oidc",
+        }

    def test_missing_access_token(self):
        """Test error handling when access_token is missing."""
@@ -121,7 +136,10 @@ class TestOAuthProvider:
        provider = OAuthProvider(fetcher)
        headers = provider.get_headers()

-        assert headers == {"Authorization": "Bearer sync_token"}
+        assert headers == {
+            "Authorization": "Bearer sync_token",
+            "x-lancedb-credential-type": "oidc",
+        }


 class TestClientConfigIntegration:
@@ -266,3 +266,38 @@ async def test_async_namespace_connection_materialized_views(tmp_path):
            handle._route_pushdown_to_rust == through_namespace._route_pushdown_to_rust
        )
        assert handle._namespace_path == through_namespace._namespace_path
+
+
+def test_namespaced_select_kind_is_read_and_unknown_kinds_are_refused():
+    import json
+
+    import pyarrow as pa
+
+    from lancedb.materialized_view import _definition_from_schema
+
+    def schema_with(definition: dict) -> pa.Schema:
+        return pa.schema([pa.field("id", pa.int32())]).with_metadata(
+            {b"mv.definition": json.dumps(definition).encode()}
+        )
+
+    # "namespaced_select" is the namespaced form of "select": same shape,
+    # a separate kind so readers that predate it refuse instead of
+    # resolving the source at the root.
+    definition = _definition_from_schema(
+        schema_with(
+            {
+                "kind": "namespaced_select",
+                "source_table": "people",
+                "source_namespace": ["ns"],
+                "projections": [{"output": "name", "expression": "name"}],
+            }
+        ),
+        "v",
+    )
+    assert definition.source_table == "people"
+    assert definition.source_namespace == ["ns"]
+
+    with pytest.raises(NotImplementedError, match="cannot refresh"):
+        _definition_from_schema(
+            schema_with({"kind": "select_v3", "source_table": "people"}), "v"
+        )
@@ -193,7 +193,13 @@ class TestNamespaceConnection:
            ),
        )

-        table = db.create_table("blob_table", data, namespace_path=["test_ns"])
+        # Legacy v1 blob columns are only writable at file version <= 2.1.
+        table = db.create_table(
+            "blob_table",
+            data,
+            namespace_path=["test_ns"],
+            storage_options={"new_table_data_storage_version": "2.1"},
+        )
        df = table.to_pandas(blob_mode="lazy").sort_values("id")

        blob = df["blob"].iloc[0]
@@ -40,6 +40,10 @@ from utils import exception_output
 from importlib.util import find_spec


+# Legacy v1 blob columns are only writable at file version <= 2.1.
+LEGACY_BLOB_STORAGE_OPTIONS = {"new_table_data_storage_version": "2.1"}
+
+
 def _blob_query_data():
    return pa.table(
        {
@@ -119,13 +123,17 @@ def _assert_blob_bytes_projection(df):

 def _blob_query_table(db, name, blob_schema):
    if blob_schema == "v1":
-        return db.create_table(name, _blob_query_data())
+        return db.create_table(
+            name, _blob_query_data(), storage_options=LEGACY_BLOB_STORAGE_OPTIONS
+        )
    return _create_blob_v2_query_table(db, name)


 async def _blob_query_table_async(db, name, blob_schema):
    if blob_schema == "v1":
-        return await db.create_table(name, _blob_query_data())
+        return await db.create_table(
+            name, _blob_query_data(), storage_options=LEGACY_BLOB_STORAGE_OPTIONS
+        )
    return await _create_blob_v2_query_table_async(db, name)


@@ -275,7 +283,9 @@ async def test_query_to_pandas_kwargs(table, table_async):
 def test_plain_scan_query_to_pandas_blob_modes(tmp_db, blob_mode):
    pytest.importorskip("lance")
    table = tmp_db.create_table(
-        f"test_query_to_pandas_blob_{blob_mode}", _blob_query_data()
+        f"test_query_to_pandas_blob_{blob_mode}",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
    )

    df = (
@@ -322,7 +332,9 @@ def test_plain_scan_query_to_pandas_blob_mode_does_not_collect_arrow(
 ):
    pytest.importorskip("lance")
    table = tmp_db.create_table(
-        "test_query_to_pandas_blob_no_arrow_collect", _blob_query_data()
+        "test_query_to_pandas_blob_no_arrow_collect",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
    )
    query = table.search().where("id = 1").select(["id", "blob"])

@@ -347,7 +359,9 @@ def test_plain_scan_query_to_pandas_blob_descriptions_flatten_uses_scanner(
 ):
    pytest.importorskip("lance")
    table = tmp_db.create_table(
-        "test_query_to_pandas_blob_desc_flatten", _blob_query_data()
+        "test_query_to_pandas_blob_desc_flatten",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
    )
    query = table.search().where("id = 1").select(["id", "blob"])

@@ -365,7 +379,11 @@ def test_plain_scan_query_to_pandas_blob_descriptions_flatten_uses_scanner(
 def test_plain_scan_query_to_pandas_scanner_state(tmp_db):
    pytest.importorskip("lance")
    data = _blob_query_data()
-    table = tmp_db.create_table("test_query_to_pandas_scanner_state", data.slice(0, 2))
+    table = tmp_db.create_table(
+        "test_query_to_pandas_scanner_state",
+        data.slice(0, 2),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
+    )
    table.add(data.slice(2, 2))

    fragments = table.to_lance().get_fragments()
@@ -400,7 +418,9 @@ def test_plain_scan_query_to_pandas_scanner_state(tmp_db):
 async def test_async_plain_scan_query_to_pandas_blob_projection(tmp_db_async):
    pytest.importorskip("lance")
    table = await tmp_db_async.create_table(
-        "test_async_query_to_pandas_blob_projection", _blob_query_data()
+        "test_async_query_to_pandas_blob_projection",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
    )

    lazy_df = await (
@@ -452,7 +472,9 @@ async def test_async_plain_scan_query_to_pandas_blob_mode_does_not_collect_arrow
 ):
    pytest.importorskip("lance")
    table = await tmp_db_async.create_table(
-        "test_async_query_to_pandas_blob_no_arrow_collect", _blob_query_data()
+        "test_async_query_to_pandas_blob_no_arrow_collect",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
    )
    query = table.query().where("id = 1").select(["id", "blob"])

@@ -474,7 +496,11 @@ async def test_async_plain_scan_query_to_pandas_blob_mode_does_not_collect_arrow

 def test_vector_query_to_pandas_blob_mode_requires_native_path(tmp_db):
    pytest.importorskip("lance")
-    table = tmp_db.create_table("test_vector_query_blob_mode", _blob_query_data())
+    table = tmp_db.create_table(
+        "test_vector_query_blob_mode",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
+    )

    with pytest.raises(RuntimeError, match="Lance native pandas conversion"):
        table.search([1.0, 0.0]).select(["blob", "vector"]).limit(1).to_pandas(
@@ -485,7 +511,9 @@ def test_vector_query_to_pandas_blob_mode_requires_native_path(tmp_db):
 def test_vector_query_to_pandas_blob_descriptions_requires_plain_scan(tmp_db):
    pytest.importorskip("lance")
    table = tmp_db.create_table(
-        "test_vector_query_blob_descriptions", _blob_query_data()
+        "test_vector_query_blob_descriptions",
+        _blob_query_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
    )

    with pytest.raises(RuntimeError, match="plain scan query"):
@@ -820,11 +820,13 @@ def test_table_create_indices():
        scalar_req = received_requests[0]
        assert "name" in scalar_req
        assert scalar_req["name"] == "custom_scalar_idx"
+        assert scalar_req["replace"] is False

        # Check FTS index request has custom name
        fts_req = received_requests[1]
        assert "name" in fts_req
        assert fts_req["name"] == "custom_fts_idx"
+        assert fts_req["replace"] is False
        assert fts_req["block_size"] == 256
        assert fts_req["custom_stop_words"] == ["cloud"]

@@ -832,6 +834,7 @@ def test_table_create_indices():
        vector_req = received_requests[2]
        assert "name" in vector_req
        assert vector_req["name"] == "custom_vector_idx"
+        assert "replace" not in vector_req

        table.wait_for_index(["custom_scalar_idx"], timedelta(seconds=2))
        table.wait_for_index(
@@ -1104,6 +1107,9 @@ def test_remote_create_index_new_api():
            table.create_index("text", config=FTS(block_size=256))
            # IvfRq via new API
            table.create_index("vector", config=IvfRq(distance_type="l2"))
+            table.create_index(
+                "vector", config=IvfPq(distance_type="l2"), replace=False
+            )

        # Legacy index_type="IVF_RQ" routes to IvfRq config under the hood.
        with pytest.warns(DeprecationWarning, match="create_index"):
@@ -1113,15 +1119,17 @@ def test_remote_create_index_new_api():
                num_partitions=8,
            )

-        assert len(received_requests) == 5
+        assert len(received_requests) == 6
        assert [req["column"] for req in received_requests] == [
            "vector",
            "category",
            "text",
            "vector",
            "vector",
+            "vector",
        ]
        assert received_requests[2]["block_size"] == 256
+        assert received_requests[4]["replace"] is False


 def test_table_wait_for_index_timeout():
@@ -2459,7 +2467,7 @@ def test_remote_blob_byte_apis_not_supported_on_old_server():


 def test_remote_connection_jobs_surface():
-    from lancedb.exceptions import JobFailedError
+    from lancedb.exceptions import JobFailedError, JobNotFoundError

    schema = pa.schema([("state", pa.string())])
    batch = pa.record_batch([pa.array(["created", "done"])], schema=schema)
@@ -2467,6 +2475,7 @@ def test_remote_connection_jobs_surface():
    with pa.ipc.new_stream(sink, schema) as writer:
        writer.write_batch(batch)
    events_body = sink.getvalue().to_pybytes()
+    query_events_payloads = []

    def handler(request):
        content_len = int(request.headers.get("Content-Length", 0))
@@ -2504,6 +2513,22 @@ def test_remote_connection_jobs_surface():
            request.end_headers()
            request.wfile.write(json.dumps(rsp).encode())
        elif request.path == "/v1/jobs/describe":
+            if payload["job_id"] == "job-2":
+                request.send_response(200)
+                request.send_header("Content-Type", "application/json")
+                request.end_headers()
+                request.wfile.write(
+                    json.dumps(
+                        dict(
+                            job_id="job-2",
+                            job_type="refresh_column",
+                            job_state="DONE",
+                            creation_ms=2000,
+                            result=dict(rows_assigned=1000000, rows_failed=0),
+                        )
+                    ).encode()
+                )
+                return
            if payload["job_id"] != "job-1":
                request.send_response(404)
                request.end_headers()
@@ -2535,7 +2560,7 @@ def test_remote_connection_jobs_surface():
            request.end_headers()
            request.wfile.write(b'{"job_id": "job-1"}')
        elif request.path == "/v1/jobs/query_events":
-            assert payload["job_id"] == "job-1"
+            query_events_payloads.append(payload)
            request.send_response(200)
            request.send_header("Content-Type", "application/vnd.apache.arrow.stream")
            request.end_headers()
@@ -2551,24 +2576,109 @@ def test_remote_connection_jobs_surface():
        assert jobs[0].table == "t1"
        assert jobs[1].state == "finished"

-        description = db.get_job("job-1")
-        assert description.job_type == "create_index"
-        assert description.state == "failed"
-        assert json.loads(description.spec_json) == {"column": "vec"}
-        assert description.failure.message == "worker died"
-        assert description.failure.retryable is True
-        assert db.get_job("missing") is None
-
        assert db.cancel_job("job-1") is True
        assert db.cancel_job("missing") is False

-        batches = db.job_history("job-1")
-        assert len(batches) == 1
-        assert batches[0].num_rows == 2
-        assert batches[0].column("state").to_pylist() == ["created", "done"]
+        # Opening a job hands back a populated handle; a missing one fails.
+        with pytest.raises(JobNotFoundError, match="missing"):
+            db.open_job("missing")
+        finished = db.open_job("job-2")
+        assert finished.state == "finished"
+        assert finished.result == {"rows_assigned": 1000000, "rows_failed": 0}

-        job = db.job("job-1")
+        job = db.open_job("job-1")
        assert job.id == "job-1"
+        # Opening already populated the handle.
+        assert job.state == "failed"
+        assert job.spec == {"column": "vec"}
+        assert job.failure.message == "worker died"
        assert job.status() == "failed"
        with pytest.raises(JobFailedError, match="worker died"):
            job.wait(timeout=timedelta(seconds=5))
+
+
+def test_remote_job_handle_reports_its_own_detail():
+    schema = pa.schema([("state", pa.string())])
+    batch = pa.record_batch([pa.array(["claim_complete"])], schema=schema)
+    sink = pa.BufferOutputStream()
+    with pa.ipc.new_stream(sink, schema) as writer:
+        writer.write_batch(batch)
+    events_body = sink.getvalue().to_pybytes()
+    event_payloads = []
+
+    def handler(request):
+        content_len = int(request.headers.get("Content-Length", 0))
+        body = request.rfile.read(content_len) if content_len > 0 else b""
+        payload = json.loads(body) if body else {}
+        if request.path == "/v1/jobs/describe":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            request.wfile.write(
+                json.dumps(
+                    dict(
+                        job_id="job-1",
+                        job_type="refresh_column",
+                        job_state="DONE",
+                        creation_ms=2000,
+                        spec=dict(column="vec"),
+                        result=dict(rows_assigned=1000000),
+                    )
+                ).encode()
+            )
+        elif request.path == "/v1/jobs/query_events":
+            event_payloads.append(payload)
+            request.send_response(200)
+            request.send_header("Content-Type", "application/vnd.apache.arrow.stream")
+            request.end_headers()
+            request.wfile.write(events_body)
+        else:
+            request.send_response(404)
+            request.end_headers()
+
+    with mock_lancedb_connection(handler) as db:
+        job = db.open_job("job-1")
+
+        # Opening populates the handle in the same round trip.
+        assert job.state == "finished"
+        job.refresh()
+        assert job.job_type == "refresh_column"
+        assert job.creation_ms == 2000
+        assert job.spec == {"column": "vec"}
+        assert job.result == {"rows_assigned": 1000000}
+        assert job.failure is None
+        # The JSON payloads stay reachable, but as internal APIs.
+        assert json.loads(job._spec_json) == {"column": "vec"}
+        assert json.loads(job._result_json) == {"rows_assigned": 1000000}
+
+        # print() shows everything the handle knows and nothing it does not.
+        # print() lays every known field out on its own line, with the JSON
+        # payloads indented rather than crammed onto one line.
+        assert repr(job) == "\n".join(
+            [
+                "Job(",
+                "    id='job-1',",
+                "    state='finished',",
+                "    job_type='refresh_column',",
+                "    creation_ms=2000,",
+                "    spec={",
+                '        "column": "vec"',
+                "    },",
+                "    result={",
+                '        "rows_assigned": 1000000',
+                "    },",
+                ")",
+            ]
+        )
+        # Nothing it does not know shows up.
+        assert "failure" not in repr(job)
+
+        events = job.events(filter="state = 'claim_complete'", limit=500)
+        assert isinstance(events, pa.Table)
+        assert events.column("state").to_pylist() == ["claim_complete"]
+        # The handle supplies job_id; the caller only narrows the query.
+        assert event_payloads[-1] == {
+            "job_id": "job-1",
+            "limit": 500,
+            "filter": "state = 'claim_complete'",
+        }
@@ -0,0 +1,162 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+from uuid import UUID
+
+import pytest
+import pyarrow as pa
+
+import lancedb
+from lancedb import _lancedb
+from lancedb.arrow import AsyncRecordBatchReader
+from lancedb.db import AsyncConnection
+from lancedb.remote.db import RemoteDBConnection
+from lancedb.sql import AsyncQuery, Query
+
+NIL_QUERY_ID = UUID(int=0)
+
+
+class FakeNativeQuery:
+    id = UUID("0198f1b2-c3d4-7e5f-8123-456789abcdef")
+
+    async def reader(self):
+        return pa.table({"value": [1, 2]})
+
+
+class FakeNativeConnection:
+    async def execute_query_async(self, query, *, default_namespace_path=None):
+        return FakeNativeQuery()
+
+
+class FakeAsyncConnection:
+    async def execute_query_async(self, query, *, default_namespace_path=None):
+        return AsyncQuery(FakeNativeQuery())
+
+
+def remote_connection(sql_host_override=None):
+    return lancedb.connect(
+        "db://analytics",
+        api_key="test-key",
+        host_override="http://localhost:10024",
+        sql_host_override=sql_host_override,
+    )
+
+
+def test_sql_is_connection_scoped():
+    assert hasattr(lancedb, "sql")
+    assert not callable(lancedb.sql)
+    assert not hasattr(_lancedb, "sql")
+    assert not hasattr(remote_connection(), "sql")
+    assert hasattr(remote_connection(), "execute_query")
+    assert hasattr(remote_connection(), "execute_query_async")
+    assert hasattr(remote_connection(), "describe_query")
+
+
+def test_query_id_is_uuid():
+    query = AsyncQuery(FakeNativeQuery())
+    assert isinstance(query.id, UUID)
+    assert Query(query).id == query.id
+
+
+def test_connection_serializes_sql_host_override():
+    endpoint = "grpc+tls://sql.example.com:10026"
+    restored = lancedb.deserialize_conn(
+        remote_connection(sql_host_override=endpoint).serialize()
+    )
+    assert restored.sql_host_override == endpoint
+
+
+@pytest.mark.asyncio
+async def test_async_sql_reader_is_record_batch_stream():
+    reader = await AsyncQuery(FakeNativeQuery()).reader()
+    assert isinstance(reader, AsyncRecordBatchReader)
+    assert (await reader.read_all())[0].column(0).to_pylist() == [1, 2]
+
+
+def test_sync_sql_reader_is_record_batch_reader():
+    reader = Query(AsyncQuery(FakeNativeQuery())).reader()
+    assert isinstance(reader, pa.RecordBatchReader)
+    assert reader.read_all().column(0).to_pylist() == [1, 2]
+
+
+def test_execute_query_returns_blocking_reader():
+    connection = RemoteDBConnection.__new__(RemoteDBConnection)
+    connection._conn = FakeAsyncConnection()
+    reader = connection.execute_query("SELECT 1")
+    assert isinstance(reader, pa.RecordBatchReader)
+    assert reader.read_all().column(0).to_pylist() == [1, 2]
+
+
+@pytest.mark.asyncio
+async def test_async_execute_query_returns_async_reader():
+    connection = AsyncConnection(FakeNativeConnection())
+    reader = await connection.execute_query("SELECT 1")
+    assert isinstance(reader, AsyncRecordBatchReader)
+    assert (await reader.read_all())[0].column(0).to_pylist() == [1, 2]
+
+
+def test_local_connection_rejects_sql(tmp_path):
+    connection = lancedb.connect(tmp_path)
+    with pytest.raises(NotImplementedError, match="SQL"):
+        connection.execute_query("SELECT 1")
+    with pytest.raises(NotImplementedError, match="SQL"):
+        connection.execute_query_async("SELECT 1")
+    with pytest.raises(NotImplementedError, match="SQL"):
+        connection.describe_query(NIL_QUERY_ID)
+
+
+@pytest.mark.asyncio
+async def test_local_async_connection_rejects_sql(tmp_path):
+    connection = await lancedb.connect_async(tmp_path)
+    with pytest.raises(NotImplementedError, match="SQL"):
+        await connection.execute_query("SELECT 1")
+    with pytest.raises(NotImplementedError, match="SQL"):
+        await connection.execute_query_async("SELECT 1")
+    with pytest.raises(NotImplementedError, match="SQL"):
+        await connection.describe_query(NIL_QUERY_ID)
+
+
+@pytest.mark.asyncio
+async def test_async_namespace_connection_rejects_sql(tmp_path):
+    connection = lancedb.connect_namespace_async("dir", {"root": str(tmp_path)})
+    with pytest.raises(NotImplementedError, match="SQL"):
+        await connection.execute_query("SELECT 1")
+    with pytest.raises(NotImplementedError, match="SQL"):
+        await connection.execute_query_async("SELECT 1")
+    with pytest.raises(NotImplementedError, match="SQL"):
+        await connection.describe_query(NIL_QUERY_ID)
+
+
+def test_describe_query_requires_uuid():
+    with pytest.raises(TypeError, match="UUID"):
+        remote_connection().describe_query(str(NIL_QUERY_ID))
+
+
+@pytest.mark.parametrize(
+    "default_namespace_path",
+    ["public", ("public",), [1]],
+)
+def test_execute_query_async_requires_namespace_path_list(default_namespace_path):
+    with pytest.raises(ValueError, match="default_namespace_path"):
+        remote_connection().execute_query_async(
+            "SELECT 1", default_namespace_path=default_namespace_path
+        )
+
+
+def test_execute_query_async_rejects_invalid_endpoint():
+    connection = remote_connection(sql_host_override="invalid://localhost")
+    with pytest.raises(ValueError, match="sql_host_override"):
+        connection.execute_query_async("SELECT 1")
+
+
+@pytest.mark.parametrize(
+    "default_namespace_path",
+    [[""], ["café"], ["pub\tlic"], ["events$raw"]],
+)
+def test_execute_query_async_rejects_invalid_namespace_components(
+    default_namespace_path,
+):
+    with pytest.raises(ValueError, match="default_namespace_path"):
+        remote_connection().execute_query_async(
+            "SELECT 1", default_namespace_path=default_namespace_path
+        )
@@ -64,15 +64,23 @@ async def _blob_v2_table_async(db: AsyncConnection, name: str):
    return table


+# Legacy v1 blob columns are only writable at file version <= 2.1.
+LEGACY_BLOB_STORAGE_OPTIONS = {"new_table_data_storage_version": "2.1"}
+
+
 def _blob_table(db: DBConnection, name: str, blob_schema: str):
    if blob_schema == "v1":
-        return db.create_table(name, data=_blob_test_data())
+        return db.create_table(
+            name, data=_blob_test_data(), storage_options=LEGACY_BLOB_STORAGE_OPTIONS
+        )
    return _blob_v2_table(db, name)


 async def _blob_table_async(db: AsyncConnection, name: str, blob_schema: str):
    if blob_schema == "v1":
-        return await db.create_table(name, data=_blob_test_data())
+        return await db.create_table(
+            name, data=_blob_test_data(), storage_options=LEGACY_BLOB_STORAGE_OPTIONS
+        )
    return await _blob_v2_table_async(db, name)


@@ -147,7 +155,11 @@ def test_table_to_pandas_invalid_blob_mode_non_blob_table(tmp_db: DBConnection):
@pytest.mark.parametrize("blob_mode", ["lazy", "bytes", "descriptions"])
 def test_table_to_pandas_blob_modes(tmp_db: DBConnection, blob_mode):
    pytest.importorskip("lance")
-    table = tmp_db.create_table(f"test_to_pandas_blob_{blob_mode}", _blob_test_data())
+    table = tmp_db.create_table(
+        f"test_to_pandas_blob_{blob_mode}",
+        _blob_test_data(),
+        storage_options=LEGACY_BLOB_STORAGE_OPTIONS,
+    )

    df = table.to_pandas(blob_mode=blob_mode)

@@ -2682,6 +2694,43 @@ def test_merge_insert(mem_db: DBConnection):
        )


+def test_merge_insert_composite_key(mem_db: DBConnection):
+    table = mem_db.create_table(
+        "my_table",
+        data=pa.table(
+            {
+                "shard": ["a", "a", "b"],
+                "id": [1, 2, 1],
+                "val": ["x", "y", "z"],
+            }
+        ),
+    )
+
+    # ("a", 1) matches an existing row and updates it. ("b", 2) agrees with an
+    # existing row on each key column separately but on neither pair, so it is
+    # an insert.
+    new_data = pa.table({"shard": ["a", "b"], "id": [1, 2], "val": ["X", "W"]})
+    res = (
+        table.merge_insert(["shard", "id"])
+        .when_matched_update_all()
+        .when_not_matched_insert_all()
+        .execute(new_data)
+    )
+    assert res.num_updated_rows == 1
+    assert res.num_inserted_rows == 1
+
+    expected = pa.table(
+        {
+            "shard": ["a", "a", "b", "b"],
+            "id": [1, 2, 1, 2],
+            "val": ["X", "y", "z", "W"],
+        }
+    )
+    assert table.to_arrow().sort_by([("shard", "ascending"), ("id", "ascending")]) == (
+        expected
+    )
+
+
 def test_merge_insert_nullable_pandas_into_pydantic_schema(mem_db: DBConnection):
    # Regression test for https://github.com/lancedb/lancedb/issues/2366
    pd = pytest.importorskip("pandas")
@@ -3922,7 +3971,7 @@ def test_stats(mem_db: DBConnection):
    print(f"{stats=}")
    assert stats == {
        # Full on-disk size of the data file, footer and metadata included.
-        "total_bytes": 633,
+        "total_bytes": 637,
        "num_rows": 2,
        "num_indices": 0,
        "fragment_stats": {
@@ -13,11 +13,7 @@ use crate::{
    runtime::future_into_py,
    table::Table,
 };
-use arrow::{
-    datatypes::Schema,
-    ffi_stream::ArrowArrayStreamReader,
-    pyarrow::{FromPyArrow, ToPyArrow},
-};
+use arrow::{datatypes::Schema, ffi_stream::ArrowArrayStreamReader, pyarrow::FromPyArrow};
 use lancedb::{
    connection::Connection as LanceConnection,
    connection::NamespaceClientPushdownOperation,
@@ -28,7 +24,7 @@ use pyo3::{
    Bound, FromPyObject, Py, PyAny, PyRef, PyResult, Python,
    exceptions::{PyRuntimeError, PyValueError},
    pyclass, pyfunction, pymethods,
-    types::{PyDict, PyDictMethods, PyList, PyListMethods},
+    types::{PyAnyMethods, PyDict, PyDictMethods, PyList},
 };

 #[pyclass]
@@ -86,6 +82,24 @@ impl Connection {
    }
 }

+fn parse_default_namespace_path(path: Option<Bound<'_, PyAny>>) -> PyResult<Vec<String>> {
+    match path {
+        Some(path) => {
+            if !path.is_instance_of::<PyList>() {
+                return Err(PyValueError::new_err(
+                    "Connection.execute_query_async default_namespace_path must be a list",
+                ));
+            }
+            path.extract::<Vec<String>>().map_err(|_| {
+                PyValueError::new_err(
+                    "Connection.execute_query_async default_namespace_path components must be strings",
+                )
+            })
+        }
+        None => Ok(vec!["public".to_string()]),
+    }
+}
+
 #[pymethods]
 impl Connection {
    fn __repr__(&self) -> String {
@@ -108,6 +122,40 @@ impl Connection {
        self.get_inner().map(|inner| inner.uri().to_string())
    }

+    #[pyo3(signature = (query, *, default_namespace_path=None))]
+    pub fn execute_query_async<'a>(
+        self_: PyRef<'a, Self>,
+        query: String,
+        default_namespace_path: Option<Bound<'_, PyAny>>,
+    ) -> PyResult<Bound<'a, PyAny>> {
+        let inner = self_.get_inner()?.clone();
+        let default_namespace_path = parse_default_namespace_path(default_namespace_path)?;
+        future_into_py(self_.py(), async move {
+            let operation = inner
+                .execute_query_async(query)
+                .default_namespace_path(default_namespace_path);
+            operation
+                .execute()
+                .await
+                .map(crate::sql::Query::new)
+                .infer_error()
+        })
+    }
+
+    pub fn describe_query<'a>(
+        self_: PyRef<'a, Self>,
+        query_id: uuid::Uuid,
+    ) -> PyResult<Bound<'a, PyAny>> {
+        let inner = self_.get_inner()?.clone();
+        future_into_py(self_.py(), async move {
+            inner
+                .describe_query(query_id)
+                .await
+                .map(crate::sql::QueryDescription::from)
+                .infer_error()
+        })
+    }
+
    #[pyo3(signature = ())]
    pub fn get_read_consistency_interval(self_: PyRef<'_, Self>) -> PyResult<Bound<'_, PyAny>> {
        let inner = self_.get_inner()?.clone();
@@ -592,9 +640,12 @@ impl Connection {
        })
    }

-    pub fn job(&self, job_id: String) -> PyResult<crate::job::Job> {
-        let inner = self.get_inner()?.clone();
-        Ok(crate::job::Job::new(inner.job(job_id).infer_error()?))
+    pub fn open_job(self_: PyRef<'_, Self>, job_id: String) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.get_inner()?.clone();
+        future_into_py(self_.py(), async move {
+            let job = inner.open_job(&job_id).await.infer_error()?;
+            Ok(crate::job::Job::new(job))
+        })
    }

    pub fn create_function_async(
@@ -629,6 +680,19 @@ impl Connection {
        })
    }

+    pub fn list_functions(self_: PyRef<'_, Self>) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.get_inner()?.clone();
+        future_into_py(self_.py(), async move {
+            inner
+                .list_functions()
+                .await
+                .infer_error()?
+                .into_iter()
+                .map(|function| function.to_canonical_json().infer_error())
+                .collect::<PyResult<Vec<_>>>()
+        })
+    }
+
    pub fn drop_function(
        self_: PyRef<'_, Self>,
        name: String,
@@ -651,42 +715,16 @@ impl Connection {
        })
    }

-    pub fn get_job(self_: PyRef<'_, Self>, job_id: String) -> PyResult<Bound<'_, PyAny>> {
-        let inner = self_.get_inner()?.clone();
-        future_into_py(self_.py(), async move {
-            let description = inner.get_job(&job_id).await.infer_error()?;
-            Ok(description.map(crate::job::JobDescription::from))
-        })
-    }
-
    pub fn cancel_job(self_: PyRef<'_, Self>, job_id: String) -> PyResult<Bound<'_, PyAny>> {
        let inner = self_.get_inner()?.clone();
        future_into_py(self_.py(), async move {
            inner.cancel_job(&job_id).await.infer_error()
        })
    }
-
-    #[pyo3(signature = (job_id=None))]
-    pub fn job_history(
-        self_: PyRef<'_, Self>,
-        job_id: Option<String>,
-    ) -> PyResult<Bound<'_, PyAny>> {
-        let inner = self_.get_inner()?.clone();
-        future_into_py(self_.py(), async move {
-            let batches = inner.job_history(job_id.as_deref()).await.infer_error()?;
-            Python::attach(|py| {
-                let list = PyList::empty(py);
-                for batch in batches {
-                    list.append(batch.to_pyarrow(py)?)?;
-                }
-                Ok(list.unbind())
-            })
-        })
-    }
 }

 #[pyfunction]
-#[pyo3(signature = (uri, api_key=None, region=None, host_override=None, read_consistency_interval=None, client_config=None, storage_options=None, session=None, manifest_enabled=false, namespace_client_properties=None, oauth_config=None))]
+#[pyo3(signature = (uri, api_key=None, region=None, host_override=None, sql_host_override=None, read_consistency_interval=None, client_config=None, storage_options=None, session=None, manifest_enabled=false, namespace_client_properties=None, oauth_config=None))]
 #[allow(clippy::too_many_arguments)]
 pub fn connect(
    py: Python<'_>,
@@ -694,6 +732,7 @@ pub fn connect(
    api_key: Option<String>,
    region: Option<String>,
    host_override: Option<String>,
+    sql_host_override: Option<String>,
    read_consistency_interval: Option<f64>,
    client_config: Option<PyClientConfig>,
    storage_options: Option<HashMap<String, String>>,
@@ -713,6 +752,12 @@ pub fn connect(
        if let Some(host_override) = host_override {
            builder = builder.host_override(&host_override);
        }
+        #[cfg(feature = "remote")]
+        if let Some(sql_host_override) = sql_host_override {
+            builder = builder.sql_host_override(&sql_host_override);
+        }
+        #[cfg(not(feature = "remote"))]
+        let _ = sql_host_override;
        if let Some(read_consistency_interval) = read_consistency_interval {
            let read_consistency_interval = Duration::from_secs_f64(read_consistency_interval);
            builder = builder.read_consistency_interval(read_consistency_interval);
@@ -114,6 +114,12 @@ impl<T> PythonErrorExt<T> for std::result::Result<T, LanceError> {
                        .getattr(intern!(py, "JobCancelledError"))?;
                    Err(PyErr::from_value(cls.call1((err.to_string(),))?))
                }),
+                LanceError::JobNotFound { .. } => Python::attach(|py| {
+                    let cls = py
+                        .import(intern!(py, "lancedb.exceptions"))?
+                        .getattr(intern!(py, "JobNotFoundError"))?;
+                    Err(PyErr::from_value(cls.call1((err.to_string(),))?))
+                }),
                _ => self.runtime_error(),
            },
        }
@@ -4,11 +4,50 @@
 use std::sync::Arc;

 use crate::runtime::future_into_py;
-use pyo3::{Bound, PyAny, PyRef, PyResult, pyclass, pymethods};
+use arrow::{
+    datatypes::Schema,
+    pyarrow::{IntoPyArrow, Table as PyArrowTable},
+};
+use lancedb::job::JobEventsRequest;
+use pyo3::{
+    Bound, PyAny, PyRef, PyResult, Python,
+    exceptions::PyValueError,
+    pyclass, pymethods,
+    types::{PyAnyMethods, PyDict, PyDictMethods},
+};
 use serde::Serialize;

 use crate::error::PythonErrorExt;

+const REPR_INDENT: &str = "    ";
+
+/// Parse a stored JSON payload into Python data. The bindings carry these as
+/// strings because that is what crosses the boundary cheaply; the public
+/// Python surface is the parsed form.
+fn parse_json_payload<'py>(
+    py: Python<'py>,
+    raw: Option<&str>,
+) -> PyResult<Option<Bound<'py, PyAny>>> {
+    match raw {
+        None => Ok(None),
+        Some(raw) => Ok(Some(py.import("json")?.call_method1("loads", (raw,))?)),
+    }
+}
+
+/// A payload rendered as indented JSON, aligned under the field that holds it.
+fn pretty_json_payload(py: Python<'_>, raw: Option<&str>) -> PyResult<Option<String>> {
+    let Some(parsed) = parse_json_payload(py, raw)? else {
+        return Ok(None);
+    };
+    let kwargs = PyDict::new(py);
+    kwargs.set_item("indent", 4)?;
+    let rendered: String = py
+        .import("json")?
+        .call_method("dumps", (parsed,), Some(&kwargs))?
+        .extract()?;
+    Ok(Some(rendered.replace('\n', &format!("\n{REPR_INDENT}"))))
+}
+
 #[pyclass]
 pub struct Job {
    inner: Arc<lancedb::Job<std::result::Result<Option<String>, String>>>,
@@ -67,6 +106,48 @@ impl Job {
            Ok(())
        })
    }
+
+    pub fn refresh(self_: PyRef<'_, Self>) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.inner.clone();
+        future_into_py(self_.py(), async move {
+            inner.refresh().await.infer_error()?;
+            Ok(())
+        })
+    }
+
+    /// The last observed lifecycle state, without contacting the backend.
+    #[getter]
+    pub fn _state(&self) -> Option<String> {
+        self.inner.state()
+    }
+
+    /// The last observed server-side record. `None` for an in-process job.
+    #[getter]
+    pub fn _description(&self) -> Option<JobDescription> {
+        self.inner.description().map(JobDescription::from)
+    }
+
+    #[pyo3(signature = (*, limit=None, filter=None))]
+    pub fn events(
+        self_: PyRef<'_, Self>,
+        limit: Option<u32>,
+        filter: Option<String>,
+    ) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.inner.clone();
+        let request = JobEventsRequest { limit, filter };
+        future_into_py(self_.py(), async move {
+            let batches = inner.events(request).await.infer_error()?;
+            Python::attach(|py| {
+                let schema = batches
+                    .first()
+                    .map(|batch| batch.schema())
+                    .unwrap_or_else(|| Arc::new(Schema::empty()));
+                let table = PyArrowTable::try_new(batches, schema)
+                    .map_err(|err| PyValueError::new_err(err.to_string()))?;
+                table.into_pyarrow(py).map(|table| table.unbind())
+            })
+        })
+    }
 }

 /// A row from `Connection.list_jobs`: one server-side job.
@@ -121,7 +202,7 @@ impl JobFailureInfo {
    }
 }

-/// A described job from `Connection.get_job`.
+/// The server-side record behind a `Job` handle.
 #[pyclass(get_all, skip_from_py_object)]
 #[derive(Clone)]
 pub struct JobDescription {
@@ -129,17 +210,49 @@ pub struct JobDescription {
    job_type: String,
    state: String,
    creation_ms: i64,
-    spec_json: Option<String>,
+    /// Internal: the wire form behind the `spec` property.
+    _spec_json: Option<String>,
+    /// Internal: the wire form behind the `result` property.
+    _result_json: Option<String>,
    failure: Option<JobFailureInfo>,
 }

 #[pymethods]
 impl JobDescription {
-    fn __repr__(&self) -> String {
-        format!(
-            "JobDescription(job_id={:?}, job_type={:?}, state={:?}, creation_ms={})",
-            self.job_id, self.job_type, self.state, self.creation_ms
-        )
+    /// The job-type-specific specification it was submitted with.
+    #[getter]
+    fn spec<'py>(&self, py: Python<'py>) -> PyResult<Option<Bound<'py, PyAny>>> {
+        parse_json_payload(py, self._spec_json.as_deref())
+    }
+
+    /// The job-type-specific terminal result. `None` until the job succeeds.
+    #[getter]
+    fn result<'py>(&self, py: Python<'py>) -> PyResult<Option<Bound<'py, PyAny>>> {
+        parse_json_payload(py, self._result_json.as_deref())
+    }
+
+    fn __repr__(&self, py: Python<'_>) -> PyResult<String> {
+        let mut fields = vec![
+            format!("job_id={:?}", self.job_id),
+            format!("job_type={:?}", self.job_type),
+            format!("state={:?}", self.state),
+            format!("creation_ms={}", self.creation_ms),
+        ];
+        // Lay the payloads out as indented JSON, the same way the `Job` repr
+        // does, so the two agree on how the same data looks.
+        for (name, payload) in [("spec", &self._spec_json), ("result", &self._result_json)] {
+            if let Some(rendered) = pretty_json_payload(py, payload.as_deref())? {
+                fields.push(format!("{name}={rendered}"));
+            }
+        }
+        if let Some(failure) = &self.failure {
+            fields.push(format!("failure={}", failure.__repr__()));
+        }
+        let body = fields
+            .iter()
+            .map(|field| format!("\n{REPR_INDENT}{field},"))
+            .collect::<String>();
+        Ok(format!("JobDescription({body}\n)"))
    }
 }

@@ -150,7 +263,11 @@ impl From<lancedb::database::JobDescription> for JobDescription {
            job_type: description.job_type,
            state: description.state,
            creation_ms: description.creation_ms,
-            spec_json: (!description.spec.is_null()).then(|| description.spec.to_string()),
+            _spec_json: (!description.spec.is_null()).then(|| description.spec.to_string()),
+            _result_json: description
+                .result
+                .filter(|result| !result.is_null())
+                .map(|result| result.to_string()),
            failure: description.failure.map(|failure| JobFailureInfo {
                phase: failure.phase,
                message: failure.message,
@@ -34,6 +34,7 @@ pub mod permutation;
 pub mod query;
 pub mod runtime;
 pub mod session;
+pub mod sql;
 pub mod table;
 pub mod util;

@@ -50,6 +51,8 @@ pub fn _lancedb(_py: Python, m: &Bound<'_, PyModule>) -> PyResult<()> {
    m.add_class::<crate::job::JobInfo>()?;
    m.add_class::<crate::job::JobDescription>()?;
    m.add_class::<crate::job::JobFailureInfo>()?;
+    m.add_class::<crate::sql::Query>()?;
+    m.add_class::<crate::sql::QueryDescription>()?;
    m.add_class::<PyBlobFile>()?;
    m.add_class::<IndexConfig>()?;
    m.add_class::<Query>()?;
@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+use std::sync::Arc;
+
+use chrono::{DateTime, Utc};
+use pyo3::{Bound, PyAny, PyRef, PyResult, pyclass, pymethods};
+use uuid::Uuid;
+
+use crate::arrow::RecordBatchStream;
+use crate::error::PythonErrorExt;
+use crate::runtime::future_into_py;
+
+#[pyclass(name = "SqlQuery")]
+pub struct Query {
+    inner: Arc<lancedb::sql::Query>,
+}
+
+impl Query {
+    pub(crate) fn new(inner: lancedb::sql::Query) -> Self {
+        Self {
+            inner: Arc::new(inner),
+        }
+    }
+}
+
+#[pymethods]
+impl Query {
+    #[getter]
+    pub fn id(&self) -> Uuid {
+        self.inner.id()
+    }
+
+    pub fn describe(self_: PyRef<'_, Self>) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.inner.clone();
+        future_into_py(self_.py(), async move {
+            inner
+                .describe()
+                .await
+                .map(QueryDescription::from)
+                .infer_error()
+        })
+    }
+
+    pub fn reader(self_: PyRef<'_, Self>) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.inner.clone();
+        future_into_py(self_.py(), async move {
+            let stream = inner.reader().await.infer_error()?;
+            Ok(RecordBatchStream::new(stream))
+        })
+    }
+
+    pub fn cancel(self_: PyRef<'_, Self>) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.inner.clone();
+        future_into_py(self_.py(), async move {
+            inner.cancel().await.infer_error()?;
+            Ok(())
+        })
+    }
+}
+
+#[pyclass(get_all, skip_from_py_object)]
+#[derive(Clone)]
+pub struct QueryDescription {
+    id: Uuid,
+    status: String,
+    progress: Option<f64>,
+    expires_at: Option<DateTime<Utc>>,
+}
+
+#[pymethods]
+impl QueryDescription {
+    fn __repr__(&self) -> String {
+        format!(
+            "QueryDescription(id={:?}, status={:?}, progress={:?}, expires_at={:?})",
+            self.id, self.status, self.progress, self.expires_at
+        )
+    }
+}
+
+impl From<lancedb::sql::QueryDescription> for QueryDescription {
+    fn from(description: lancedb::sql::QueryDescription) -> Self {
+        Self {
+            id: description.id,
+            status: description.status.to_string(),
+            progress: description.progress,
+            expires_at: description.expires_at,
+        }
+    }
+}
@@ -1,6 +1,6 @@
 [package]
 name = "lancedb"
-version = "0.38.0"
+version = "0.39.0-beta.6"
 edition.workspace = true
 description = "LanceDB: A serverless, low-latency vector database for AI applications"
 license.workspace = true
@@ -21,6 +21,8 @@ arrow-select = { workspace = true }
 arrow-ord = { workspace = true }
 arrow-cast = { workspace = true }
 arrow-ipc.workspace = true
+arrow-flight = { workspace = true, optional = true }
+prost = { version = "0.14", optional = true }
 chrono = { workspace = true }
 datafusion-catalog.workspace = true
 datafusion-common.workspace = true
@@ -77,6 +79,7 @@ reqwest = { version = "0.12.0", default-features = false, features = [
    "rustls-tls-native-roots",
    "stream",
 ], optional = true }
+tonic = { workspace = true, optional = true }
 http = { version = "1", optional = true } # Matching what is in reqwest
 urlencoding = { version = "2", optional = true }
 uuid = { workspace = true, features = ["v5"] }
@@ -145,8 +148,11 @@ huggingface = [
 ]
 dynamodb = ["lance/dynamodb", "aws"]
 remote = [
+    "dep:arrow-flight",
+    "dep:prost",
    "dep:reqwest",
    "dep:http",
+    "dep:tonic",
    "dep:urlencoding",
    "lance-namespace-impls/rest",
    "lance-namespace-impls/rest-adapter",
@@ -7,7 +7,9 @@
 //! raw `Binary` / `LargeBinary` into the blob struct layout. Queries return
 //! small descriptors, not bytes.
 //!
-//! Blob tables require Lance file format >= 2.2 and stable row ids at create.
+//! Blob tables require Lance file format >= 2.2. `_rowid` values stay valid
+//! after compaction when the table has stable row ids. Overwrite is a new
+//! create and does not keep the previous table's stable row id setting.

 use std::ops::Range;
 use std::sync::Arc;
@@ -324,6 +326,7 @@ pub(crate) fn blob_column_names(schema: &Schema) -> Vec<String> {
 }

 /// Bumps storage format to at least [`LanceFileVersion::V2_2`] for blob schemas.
+/// Leaves `enable_stable_row_ids` unchanged.
 pub(crate) fn ensure_blob_storage_version(schema: &Schema, params: &mut WriteParams) {
    if !has_blob_columns(schema) {
        return;
@@ -385,6 +388,30 @@ fn ensure_all_row_ids_resolved(column: &str, requested: usize, resolved: usize)
    }
 }

+/// Lance take reports a missing physical row address as NotSupported or InvalidInput.
+fn map_blob_take_error(column: &str, requested: usize, err: lance::Error) -> Error {
+    let missing_row_addr = match &err {
+        lance::Error::NotSupported { source, .. } => {
+            source.to_string().contains("must not target deleted rows")
+        }
+        lance::Error::InvalidInput { source, .. } => source
+            .to_string()
+            .contains("belongs to non-existent fragment"),
+        _ => false,
+    };
+
+    if missing_row_addr {
+        Error::InvalidInput {
+            message: format!(
+                "blob read for column '{column}' requested {requested} row ids but some \
+                 do not exist in the table; pass row ids collected from this table"
+            ),
+        }
+    } else {
+        err.into()
+    }
+}
+
 /// Materialize blob-local ranges (same length and order as `requests`, nulls preserved).
 pub(crate) async fn take_blob_ranges_aligned(
    dataset: &Arc<Dataset>,
@@ -405,7 +432,8 @@ pub(crate) async fn take_blob_ranges_aligned(
        .with_row_ids(lance_requests)
        .preserve_order(true)
        .execute()
-        .await?;
+        .await
+        .map_err(|err| map_blob_take_error(column, requests.len(), err))?;
    ensure_all_row_ids_resolved(column, requests.len(), payloads.len())?;

    let mut builder = LargeBinaryBuilder::new();
@@ -434,7 +462,8 @@ pub(crate) async fn take_blobs_aligned(
        .with_row_ids(row_ids.to_vec())
        .preserve_order(true)
        .execute()
-        .await?;
+        .await
+        .map_err(|err| map_blob_take_error(column, row_ids.len(), err))?;
    ensure_all_row_ids_resolved(column, row_ids.len(), payloads.len())?;

    let mut builder = LargeBinaryBuilder::new();
@@ -458,7 +487,10 @@ pub(crate) async fn take_blob_files_aligned(
        return Ok(Vec::new());
    }

-    let handles = dataset.take_blobs(row_ids, column).await?;
+    let handles = dataset
+        .take_blobs(row_ids, column)
+        .await
+        .map_err(|err| map_blob_take_error(column, row_ids.len(), err))?;
    ensure_all_row_ids_resolved(column, row_ids.len(), handles.len())?;
    Ok(handles
        .into_iter()
@@ -500,10 +532,27 @@ mod tests {
    fn storage_version_bumps_to_v2_2() {
        let mut params = WriteParams::default();
        ensure_blob_storage_version(&blob_schema(), &mut params);
-        assert_eq!(
-            params.data_storage_version.unwrap().resolve(),
-            ConcreteFileVersion::V2_2
-        );
+        let resolved = params
+            .data_storage_version
+            .unwrap_or(LanceFileVersion::Stable)
+            .resolve();
+        assert_eq!(resolved, ConcreteFileVersion::V2_2);
+        assert!(!params.enable_stable_row_ids);
+    }
+
+    #[test]
+    fn storage_version_leaves_stable_row_ids_enabled() {
+        let mut params = WriteParams {
+            enable_stable_row_ids: true,
+            ..Default::default()
+        };
+        ensure_blob_storage_version(&blob_schema(), &mut params);
+        assert!(params.enable_stable_row_ids);
+        let resolved = params
+            .data_storage_version
+            .unwrap_or(LanceFileVersion::Stable)
+            .resolve();
+        assert_eq!(resolved, ConcreteFileVersion::V2_2);
    }

    #[test]
@@ -576,5 +625,6 @@ mod tests {
        let mut params = WriteParams::default();
        ensure_blob_storage_version(&schema, &mut params);
        assert!(params.data_storage_version.is_none());
+        assert!(!params.enable_stable_row_ids);
    }
 }
@@ -23,15 +23,18 @@ use crate::connection::create_table::CreateTableBuilder;
 use crate::data::scannable::Scannable;
 use crate::database::listing::ListingDatabase;
 use crate::database::{
-    CloneTableRequest, Database, DatabaseOptions, JobDescription, JobInfo, OpenTableRequest,
-    ReadConsistency, TableNamesRequest,
+    CloneTableRequest, Database, DatabaseOptions, JobInfo, OpenTableRequest, ReadConsistency,
+    TableNamesRequest,
 };
 use crate::embeddings::{EmbeddingRegistry, MemoryRegistry};
 use crate::error::{Error, Result};
 #[cfg(feature = "remote")]
 use crate::remote::{
    client::ClientConfig,
-    db::{OPT_REMOTE_API_KEY, OPT_REMOTE_HOST_OVERRIDE, OPT_REMOTE_REGION},
+    db::{
+        OPT_REMOTE_API_KEY, OPT_REMOTE_HOST_OVERRIDE, OPT_REMOTE_REGION,
+        OPT_REMOTE_SQL_HOST_OVERRIDE,
+    },
 };
 use lance::io::ObjectStoreParams;
 pub use lance_file::version::LanceFileVersion;
@@ -322,6 +325,43 @@ pub struct CloneTableBuilder {
    request: CloneTableRequest,
 }

+/// Builder for asynchronously executing a SQL statement on a remote database.
+pub struct ExecuteQueryAsyncBuilder {
+    parent: Arc<dyn Database>,
+    query: String,
+    default_namespace_path: Vec<String>,
+}
+
+impl ExecuteQueryAsyncBuilder {
+    fn new(parent: Arc<dyn Database>, query: String) -> Self {
+        Self {
+            parent,
+            query,
+            default_namespace_path: vec!["public".to_string()],
+        }
+    }
+
+    /// Set the namespace used for unqualified table names.
+    ///
+    /// An empty path is treated as `public`, which is the SQL name for the
+    /// root Lance namespace.
+    pub fn default_namespace_path<I, S>(mut self, path: I) -> Self
+    where
+        I: IntoIterator<Item = S>,
+        S: Into<String>,
+    {
+        self.default_namespace_path = path.into_iter().map(Into::into).collect();
+        self
+    }
+
+    /// Start the statement and return its asynchronous query handle.
+    pub async fn execute(self) -> Result<crate::sql::Query> {
+        self.parent
+            .execute_query_async(&self.query, &self.default_namespace_path)
+            .await
+    }
+}
+
 impl CloneTableBuilder {
    fn new(parent: Arc<dyn Database>, target_table_name: String, source_uri: String) -> Self {
        Self {
@@ -405,6 +445,51 @@ impl Connection {
        &self.internal
    }

+    /// Start executing SQL on a remote LanceDB database.
+    ///
+    /// The query can reference tables in other databases with SQL dot notation.
+    /// Use [`ExecuteQueryAsyncBuilder::default_namespace_path`] to avoid qualifying
+    /// tables in the default namespace. Local connections return
+    /// [`Error::NotSupported`].
+    ///
+    /// # Example
+    ///
+    /// ```no_run
+    /// # async fn query(db: &lancedb::Connection) -> lancedb::Result<()> {
+    /// use futures::TryStreamExt;
+    ///
+    /// let query = db
+    ///     .execute_query_async("SELECT * FROM events LIMIT 10")
+    ///     .default_namespace_path(["public"])
+    ///     .execute()
+    ///     .await?;
+    /// println!("query id: {}", query.id());
+    /// let mut batches = query.reader().await?;
+    /// while let Some(batch) = batches.try_next().await? {
+    ///     println!("received {} rows", batch.num_rows());
+    /// }
+    /// # Ok(())
+    /// # }
+    /// ```
+    pub fn execute_query_async(&self, query: impl Into<String>) -> ExecuteQueryAsyncBuilder {
+        ExecuteQueryAsyncBuilder::new(self.internal.clone(), query.into())
+    }
+
+    /// Describe a submitted SQL query by its connection-scoped id.
+    ///
+    /// This performs one bounded status poll using state retained by this
+    /// connection. Running state with a live query handle is not evicted;
+    /// abandoned state has bounded retention, and server expiration is
+    /// honored. Terminal state is retained briefly.
+    /// Query ids are not portable to another connection. Local connections
+    /// return [`Error::NotSupported`].
+    pub async fn describe_query(
+        &self,
+        query_id: uuid::Uuid,
+    ) -> Result<crate::sql::QueryDescription> {
+        self.internal.describe_query(query_id).await
+    }
+
    /// Get the names of all tables in the database
    ///
    /// The names will be returned in lexicographical order (ascending)
@@ -523,6 +608,28 @@ impl Connection {
            .await
    }

+    /// List every published immutable Function version in the remote catalog.
+    ///
+    /// Results are ordered by Function name then version. The client walks all
+    /// server pages before returning. Local databases return
+    /// [`Error::NotSupported`].
+    ///
+    /// # Example
+    ///
+    /// ```no_run
+    /// # async fn list_functions(
+    /// #     connection: &lancedb::Connection,
+    /// # ) -> Result<(), Box<dyn std::error::Error>> {
+    /// for function in connection.list_functions().await? {
+    ///     println!("{} {}", function.name(), function.version());
+    /// }
+    /// # Ok(())
+    /// # }
+    /// ```
+    pub async fn list_functions(&self) -> Result<Vec<crate::function::FunctionVersion>> {
+        self.internal.list_functions().await
+    }
+
    /// Drop one exact immutable Function version from the remote catalog.
    ///
    /// Returns `true` when the server appended a Dropped transition and
@@ -563,14 +670,34 @@ impl Connection {
        self.internal.read_consistency().await
    }

-    /// A [`crate::job::Job`] handle for a server-side job by id, suitable for
-    /// waiting on or cancelling the job.
+    /// Open a server-side job by id, returning a handle with its record
+    /// already populated. Fails with [`crate::Error::JobNotFound`] when the
+    /// server has no such job, the way [`Connection::open_table`] does for a
+    /// missing table.
    ///
-    /// The handle is constructed without a server round trip; an unknown id
-    /// surfaces when the handle is used. Only server-backed databases support
-    /// job handles by id.
-    pub fn job(&self, job_id: impl AsRef<str>) -> Result<crate::job::Job> {
-        self.internal.job(job_id.as_ref())
+    /// This is the one way in: the returned [`crate::job::Job`] answers for
+    /// its own state, specification, result, failure and event history, so
+    /// there is no separate connection-level call for any of them.
+    ///
+    /// # Example
+    ///
+    /// ```no_run
+    /// # use lancedb::job::JobEventsRequest;
+    /// # async fn open_job(
+    /// #     connection: &lancedb::Connection,
+    /// #     job_id: &str,
+    /// # ) -> Result<(), Box<dyn std::error::Error>> {
+    /// let job = connection.open_job(job_id).await?;
+    /// println!("{:?} {:?}", job.state(), job.result());
+    /// let done = job
+    ///     .events(JobEventsRequest::default().filter("state = 'claim_complete'"))
+    ///     .await?;
+    /// println!("{} completions", done.iter().map(|b| b.num_rows()).sum::<usize>());
+    /// # Ok(())
+    /// # }
+    /// ```
+    pub async fn open_job(&self, job_id: impl AsRef<str>) -> Result<crate::job::Job> {
+        self.internal.open_job(job_id.as_ref()).await
    }

    /// List server-side jobs across the database's tables.
@@ -578,24 +705,12 @@ impl Connection {
        self.internal.list_jobs().await
    }

-    /// Describe a single server-side job by id. `None` when the server has no
-    /// such job.
-    pub async fn get_job(&self, job_id: impl AsRef<str>) -> Result<Option<JobDescription>> {
-        self.internal.get_job(job_id.as_ref()).await
-    }
-
    /// Request cancellation of a server-side job by id. Returns true if the
    /// server accepted the cancellation, false if no such job exists.
    pub async fn cancel_job(&self, job_id: impl AsRef<str>) -> Result<bool> {
        self.internal.cancel_job(job_id.as_ref()).await
    }

-    /// The lifecycle event history of a server-side job (all jobs when
-    /// `job_id` is `None`), as recorded Arrow batches.
-    pub async fn job_history(&self, job_id: Option<&str>) -> Result<Vec<RecordBatch>> {
-        self.internal.job_history(job_id).await
-    }
-
    /// Drop a table in the database.
    ///
    /// # Arguments
@@ -842,6 +957,19 @@ impl ConnectBuilder {
        self
    }

+    /// Set the SQL service host override for a remote connection.
+    ///
+    /// The SQL client is initialized lazily when the connection first executes
+    /// SQL and is retained for the connection's lifetime.
+    #[cfg(feature = "remote")]
+    pub fn sql_host_override(mut self, sql_host_override: &str) -> Self {
+        self.request.options.insert(
+            OPT_REMOTE_SQL_HOST_OVERRIDE.to_string(),
+            sql_host_override.to_string(),
+        );
+        self
+    }
+
    /// Set the database specific options
    ///
    /// See [crate::database::listing::ListingDatabaseOptions] for the options available for
@@ -1031,6 +1159,7 @@ impl ConnectBuilder {

        let mut merged_options = self.request.options.clone();
        Self::apply_env_defaults(&ENV_VARS_TO_STORAGE_OPTS, &mut merged_options);
+        let sql_host_override = merged_options.get(OPT_REMOTE_SQL_HOST_OVERRIDE).cloned();
        let options = RemoteDatabaseOptions::parse_from_map(&merged_options)?;

        let region = options.region.ok_or_else(|| Error::InvalidInput {
@@ -1072,11 +1201,15 @@ impl ConnectBuilder {
        }

        let storage_options = StorageOptions(options.storage_options.clone());
+        let host_overrides = crate::remote::db::RemoteHostOverrides {
+            rest: options.host_override,
+            sql: sql_host_override,
+        };
        let internal = Arc::new(crate::remote::db::RemoteDatabase::try_new(
            &self.request.uri,
            &api_key,
            &region,
-            options.host_override,
+            host_overrides,
            client_config,
            storage_options.into(),
            self.request.read_consistency_interval,
@@ -1370,6 +1503,23 @@ mod tests {
        assert_eq!(tc.connection.uri(), tc.uri);
    }

+    #[tokio::test]
+    async fn test_local_connection_rejects_sql_queries() {
+        let directory = tempdir().unwrap();
+        let connection = connect(directory.path().to_str().unwrap())
+            .execute()
+            .await
+            .unwrap();
+        assert!(matches!(
+            connection.execute_query_async("SELECT 1").execute().await,
+            Err(Error::NotSupported { .. })
+        ));
+        assert!(matches!(
+            connection.describe_query(uuid::Uuid::nil()).await,
+            Err(Error::NotSupported { .. })
+        ));
+    }
+
    #[cfg(feature = "remote")]
    #[test]
    fn test_apply_env_defaults() {
@@ -18,8 +18,6 @@ use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::Duration;

-use arrow_array::RecordBatch;
-
 use lance::dataset::ReadParams;
 use lance_namespace::LanceNamespace;
 use lance_namespace::models::{
@@ -206,8 +204,8 @@ pub enum ReadConsistency {
 /// compaction, column refresh, ...).
 #[derive(Debug, Clone)]
 pub struct JobInfo {
-    /// The job id -- what [`Database::get_job`] and [`Database::cancel_job`]
-    /// accept.
+    /// The job id -- what [`Database::open_job`] and
+    /// [`Database::cancel_job`] accept.
    pub job_id: String,
    /// The table the job runs against, without URI or namespace.
    pub table: String,
@@ -218,8 +216,8 @@ pub struct JobInfo {
    pub created_at_millis: i64,
 }

-/// A described job from [`Database::get_job`]: lifecycle state plus the
-/// job-type-specific specification.
+/// The server-side record behind a [`crate::job::Job`] handle: lifecycle
+/// state plus the job-type-specific specification and result.
 #[derive(Debug, Clone)]
 pub struct JobDescription {
    pub job_id: String,
@@ -230,6 +228,10 @@ pub struct JobDescription {
    pub creation_ms: i64,
    /// The job-type-specific specification. Null when the server omits it.
    pub spec: serde_json::Value,
+    /// The job-type-specific terminal result, for job types that define one.
+    /// `None` until the job succeeds, so a job that never terminates reports
+    /// its progress through [`crate::job::Job::events`] instead.
+    pub result: Option<serde_json::Value>,
    /// Why the job failed, when the job is failed and the server reports a
    /// reason.
    pub failure: Option<crate::error::JobFailure>,
@@ -307,34 +309,45 @@ pub trait Database:
    ) -> Result<crate::function::FunctionVersion> {
        function_catalog_not_supported()
    }
+    /// List every published immutable Function version in the remote catalog.
+    async fn list_functions(&self) -> Result<Vec<crate::function::FunctionVersion>> {
+        function_catalog_not_supported()
+    }
    /// Drop one exact immutable Function version from the remote catalog.
    async fn drop_function(&self, _name: &str, _version: &str) -> Result<bool> {
        function_catalog_not_supported()
    }
-    /// A [`crate::job::Job`] handle for a server-side job by id, suitable for
-    /// waiting on or cancelling the job. The handle is constructed without a
-    /// server round trip; an unknown id surfaces when the handle is used.
-    fn job(&self, _job_id: &str) -> Result<crate::job::Job> {
-        job_op_not_supported("job")
+    /// Open a job by id, returning a handle with its record already
+    /// populated. Fails with [`crate::Error::JobNotFound`] when the server has
+    /// no such job.
+    async fn open_job(&self, _job_id: &str) -> Result<crate::job::Job> {
+        job_op_not_supported("open_job")
    }
    /// List server-side jobs across the database's tables.
    async fn list_jobs(&self) -> Result<Vec<JobInfo>> {
        job_op_not_supported("list_jobs")
    }
-    /// Describe a single job by id. `None` when the server has no such job.
-    async fn get_job(&self, _job_id: &str) -> Result<Option<JobDescription>> {
-        job_op_not_supported("get_job")
-    }
    /// Request cancellation of a job by id. Returns true if the server
    /// accepted the cancellation, false if no such job exists. Cancelling an
    /// already-terminal job is a no-op success.
    async fn cancel_job(&self, _job_id: &str) -> Result<bool> {
        job_op_not_supported("cancel_job")
    }
-    /// The lifecycle event history of a job (all jobs when `job_id` is
-    /// `None`), as recorded Arrow batches.
-    async fn job_history(&self, _job_id: Option<&str>) -> Result<Vec<RecordBatch>> {
-        job_op_not_supported("job_history")
+    /// Start executing a SQL statement on a remote database.
+    async fn execute_query_async(
+        &self,
+        _query: &str,
+        _default_namespace_path: &[String],
+    ) -> Result<crate::sql::Query> {
+        Err(crate::error::Error::NotSupported {
+            message: "SQL is not supported by this database".to_string(),
+        })
+    }
+    /// Describe a submitted SQL query by its connection-scoped id.
+    async fn describe_query(&self, _query_id: uuid::Uuid) -> Result<crate::sql::QueryDescription> {
+        Err(crate::error::Error::NotSupported {
+            message: "SQL is not supported by this database".to_string(),
+        })
    }
    /// Open a table in the database
    async fn open_table(&self, request: OpenTableRequest) -> Result<Arc<dyn BaseTable>>;
@@ -18,7 +18,7 @@ use lance_table::io::commit::commit_handler_from_url;
 use object_store::local::LocalFileSystem;
 use snafu::ResultExt;

-use crate::blob::{ensure_blob_storage_version, has_blob_columns};
+use crate::blob::ensure_blob_storage_version;
 use crate::connection::ConnectRequest;
 use crate::database::ReadConsistency;
 use crate::database::namespace::LanceNamespaceDatabase;
@@ -827,7 +827,6 @@ impl ListingDatabase {
        if let Some(enable_stable_row_ids) = overrides
            .enable_stable_row_ids
            .or(self.new_table_config.enable_stable_row_ids)
-            .or(has_blob_columns(&data_schema).then_some(true))
        {
            write_params.enable_stable_row_ids = enable_stable_row_ids;
        }
@@ -3,6 +3,7 @@

 //! Namespace-based database implementation that delegates table management to lance-namespace

+use lance_datafusion::utils::StreamingWriteSource;
 use std::collections::{HashMap, HashSet};
 use std::sync::{Arc, Mutex};

@@ -23,7 +24,7 @@ use lance_namespace_impls::ConnectBuilder;
 use lance_table::io::commit::CommitHandler;
 use lance_table::io::commit::external_manifest::ExternalManifestCommitHandler;

-use crate::blob::{ensure_blob_storage_version, has_blob_columns};
+use crate::blob::ensure_blob_storage_version;
 use crate::connection::NamespaceClientPushdownOperation;
 use crate::database::ReadConsistency;
 use crate::database::listing::{NewTableConfig, take_request_creation_overrides};
@@ -217,7 +218,6 @@ impl LanceNamespaceDatabase {
        if let Some(enable_stable_row_ids) = overrides
            .enable_stable_row_ids
            .or(self.new_table_config.enable_stable_row_ids)
-            .or(has_blob_columns(data_schema.as_ref()).then_some(true))
        {
            params.enable_stable_row_ids = enable_stable_row_ids;
        }
@@ -305,6 +305,10 @@ impl Database for LanceNamespaceDatabase {
    }

    async fn create_table(&self, request: DbCreateTableRequest) -> Result<Arc<dyn BaseTable>> {
+        // Refuse a bad declaration before the namespace records a table.
+        crate::table::computed_columns::ensure_declarations_are_planned(
+            &request.data.arrow_schema(),
+        )?;
        let mut table_id = request.namespace_path.clone();
        table_id.push(request.name.clone());
        let mut existing_table = None;
@@ -102,6 +102,8 @@ pub enum Error {
    },
    #[snafu(display("Job{} was cancelled", job_id.as_ref().map(|id| format!(" {id}")).unwrap_or_default()))]
    JobCancelled { job_id: Option<String> },
+    #[snafu(display("Job '{job_id}' was not found"))]
+    JobNotFound { job_id: String },

    // 3rd party / external errors
    #[snafu(display("object_store error: {source}"))]
@@ -582,8 +582,8 @@ pub struct InputBinding {

 /// Ordered result-field to table-field mapping for a Function binding.
 ///
-/// Assignment state is not part of the Slice 1 client contract. During the
-/// NULL transition there is no public Lance cell-flag identifier to persist.
+/// `nullable` describes the logical Function result. Physical computed-column
+/// fields remain nullable while unassigned.
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct OutputMapping {
    pub result_field: String,
@@ -594,6 +594,14 @@ pub struct OutputMapping {
    pub nullable: bool,
 }

+/// Internal physical column preserving the parent validity of a flattened
+/// named-struct result.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct AssignmentMapping {
+    pub output_name: String,
+    pub output_field_id: i32,
+}
+
 /// Immutable Function binding persisted by the Enterprise table service.
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct FunctionBinding {
@@ -601,6 +609,8 @@ pub struct FunctionBinding {
    function: FunctionVersionRef,
    inputs: Vec<InputBinding>,
    outputs: Vec<OutputMapping>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    assignment: Option<AssignmentMapping>,
    /// Exact Arrow schema presented to the Function, encoded with the Lance
    /// Namespace Arrow JSON representation.
    #[serde(default, skip_serializing_if = "Option::is_none")]
@@ -627,6 +637,10 @@ impl FunctionBinding {
        &self.outputs
    }

+    pub fn assignment(&self) -> Option<&AssignmentMapping> {
+        self.assignment.as_ref()
+    }
+
    pub fn input_schema(&self) -> Option<&Value> {
        self.input_schema.as_ref()
    }
@@ -3,16 +3,53 @@

 //! Handles to operations a server may run asynchronously.

-use std::sync::Arc;
+use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};

+use arrow_array::RecordBatch;
 use async_trait::async_trait;
 use serde::{Serialize, de::DeserializeOwned};
 use serde_json::Value;
 use tokio::sync::watch;
 use tokio::task::{AbortHandle, JoinHandle};

+use crate::database::JobDescription;
 use crate::error::{Error, JobFailure, Result};

+/// Which of a job's events [`Job::events`] returns.
+///
+/// The handle already knows which job to ask about, so this narrows the
+/// query rather than naming one.
+#[derive(Debug, Clone, Default)]
+pub struct JobEventsRequest {
+    /// Maximum event rows to return. The server applies its own default
+    /// (1000 rows) and maximum (10,000 rows) when this is `None`, and
+    /// truncates without saying so, which matters for a job with one event
+    /// per fragment.
+    pub limit: Option<u32>,
+    /// SQL-like filter over the event columns `state`, `updated_by`,
+    /// `emitted_from`, `emitted_by`, and `claim_entity`. For example
+    /// `state = 'claim_complete'` selects only per-claim completions.
+    pub filter: Option<String>,
+}
+
+impl JobEventsRequest {
+    pub fn limit(mut self, limit: u32) -> Self {
+        self.limit = Some(limit);
+        self
+    }
+
+    pub fn filter(mut self, filter: impl Into<String>) -> Self {
+        self.filter = Some(filter.into());
+        self
+    }
+}
+
+fn job_detail_not_supported<T>(what: &str) -> Result<T> {
+    Err(Error::NotSupported {
+        message: format!("{what} is only available for server-side jobs"),
+    })
+}
+
 /// Backend-specific tracking for an asynchronous operation.
 #[async_trait]
 pub(crate) trait JobHandle: Send + Sync {
@@ -23,6 +60,15 @@ pub(crate) trait JobHandle: Send + Sync {
    async fn status(&self) -> Result<String>;
    async fn wait(&self) -> Result<TerminalResult>;
    async fn cancel(&self) -> Result<()>;
+    /// The job's full server-side record. Backends that run the operation in
+    /// this process have none and keep the default.
+    async fn describe(&self) -> Result<JobDescription> {
+        job_detail_not_supported("describing a job")
+    }
+    /// The job's recorded lifecycle events.
+    async fn events(&self, _request: JobEventsRequest) -> Result<Vec<RecordBatch>> {
+        job_detail_not_supported("job event history")
+    }
 }

 /// A backend-neutral successful terminal result.
@@ -85,16 +131,34 @@ enum JobInner<T> {
    Completed(T),
 }

+/// What a handle last learned about its job. `state` is separate because an
+/// in-process job can report one but has no server-side record behind it.
+#[derive(Default)]
+struct JobCache {
+    state: Option<String>,
+    description: Option<JobDescription>,
+}
+
 /// A handle to an operation that may still be running.
 ///
 /// The operation may already be complete when the handle is created. `T` is
 /// the endpoint's successful terminal result; unit-result operations use the
 /// default `Job<()>`.
+///
+/// The detail accessors ([`Job::state`], [`Job::job_type`], ...) read what the
+/// handle last observed. Submitting an operation returns only a job id, so
+/// populating them eagerly would cost an extra round trip on every call:
+///
+/// - [`Job::refresh`] and [`Job::status`] fetch the whole record.
+/// - [`Job::wait`] records the terminal state it establishes, but not the rest
+///   of the record; call [`Job::refresh`] for that.
+/// - Everything is `None` until one of those runs.
 pub struct Job<T = ()>
 where
    T: Clone + Send + Sync + 'static,
 {
    inner: JobInner<T>,
+    cache: RwLock<JobCache>,
 }

 impl<T> std::fmt::Debug for Job<T>
@@ -102,18 +166,40 @@ where
    T: Clone + Send + Sync + 'static,
 {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.debug_struct("Job")
-            .field("id", &self.id())
-            .field("done", &matches!(self.inner, JobInner::Completed(_)))
-            .finish()
+        let cache = self.cache_read();
+        let mut out = f.debug_struct("Job");
+        out.field("id", &self.id())
+            .field("done", &matches!(self.inner, JobInner::Completed(_)));
+        if let Some(state) = &cache.state {
+            out.field("state", state);
+        }
+        if let Some(description) = &cache.description {
+            out.field("job_type", &description.job_type)
+                .field("creation_ms", &description.creation_ms);
+            if !description.spec.is_null() {
+                out.field("spec", &description.spec);
+            }
+            if let Some(result) = &description.result {
+                out.field("result", result);
+            }
+            if let Some(failure) = &description.failure {
+                out.field("failure", failure);
+            }
+        }
+        out.finish()
    }
 }

 impl Job<()> {
-    /// A job whose operation finished before the handle was created.
+    /// A job whose operation finished before the handle was created. Its
+    /// state is known without asking anyone, so the cache starts populated.
    pub(crate) fn new_done() -> Self {
        Self {
            inner: JobInner::Completed(()),
+            cache: RwLock::new(JobCache {
+                state: Some("finished".to_string()),
+                description: None,
+            }),
        }
    }

@@ -123,8 +209,21 @@ impl Job<()> {
                handle,
                decode: Arc::new(|_| Ok(())),
            },
+            cache: RwLock::default(),
        }
    }
+
+    /// A handle whose record the caller has already fetched, so the detail
+    /// accessors answer without a second round trip.
+    pub(crate) fn opened(handle: Box<dyn JobHandle>, description: JobDescription) -> Self {
+        let job = Self::new(handle);
+        {
+            let mut cache = job.cache_write();
+            cache.state = Some(description.state.clone());
+            cache.description = Some(description);
+        }
+        job
+    }
 }

 impl<T> Job<T>
@@ -138,6 +237,7 @@ where
                handle,
                decode: Arc::new(TerminalResult::decode::<T>),
            },
+            cache: RwLock::default(),
        }
    }
 }
@@ -169,16 +269,124 @@ where
        }
    }

+    fn cache_read(&self) -> RwLockReadGuard<'_, JobCache> {
+        self.cache.read().unwrap_or_else(|err| err.into_inner())
+    }
+
+    fn cache_write(&self) -> RwLockWriteGuard<'_, JobCache> {
+        self.cache.write().unwrap_or_else(|err| err.into_inner())
+    }
+
+    /// Asks the backend for this job's current state, and for a server-side
+    /// job its full record, then caches the answer for the detail accessors.
+    ///
+    /// In-process operations have no server-side record, so only
+    /// [`Job::state`] is populated for them.
+    pub async fn refresh(&self) -> Result<()> {
+        self.refresh_state().await.map(|_| ())
+    }
+
+    /// Refreshes and reports the state, which every backend can answer.
+    async fn refresh_state(&self) -> Result<String> {
+        let JobInner::Handle { handle, .. } = &self.inner else {
+            let state = "finished".to_string();
+            self.cache_write().state = Some(state.clone());
+            return Ok(state);
+        };
+        match handle.describe().await {
+            Ok(description) => {
+                let state = description.state.clone();
+                let mut cache = self.cache_write();
+                cache.state = Some(state.clone());
+                cache.description = Some(description);
+                Ok(state)
+            }
+            // An in-process job knows its own state and nothing more.
+            Err(Error::NotSupported { .. }) => {
+                let state = handle.status().await?;
+                self.cache_write().state = Some(state.clone());
+                Ok(state)
+            }
+            Err(err) => Err(err),
+        }
+    }
+
    /// The operation's current lifecycle state: "running", "finished",
    /// "failed", or "cancelled".
    ///
    /// A point snapshot; unlike [`Job::wait`] it does not block, raise on a
    /// terminal failure state, or retry. States a newer server reports that
-    /// this client version does not know pass through as-is.
+    /// this client version does not know pass through as-is. Also refreshes
+    /// the detail accessors.
    pub async fn status(&self) -> Result<String> {
+        self.refresh_state().await
+    }
+
+    /// The last lifecycle state this handle observed, without contacting the
+    /// backend. `None` until the handle has.
+    pub fn state(&self) -> Option<String> {
+        self.cache_read().state.clone()
+    }
+
+    /// The whole server-side record this handle last observed. The accessors
+    /// below read individual fields out of it. `None` for an in-process job,
+    /// which has no such record.
+    pub fn description(&self) -> Option<JobDescription> {
+        self.cache_read().description.clone()
+    }
+
+    /// The job's type, as the server names it. `None` for an in-process job.
+    pub fn job_type(&self) -> Option<String> {
+        self.with_description(|description| description.job_type.clone())
+    }
+
+    /// When the job was created, in milliseconds since the epoch. `None` for
+    /// an in-process job.
+    pub fn creation_ms(&self) -> Option<i64> {
+        self.with_description(|description| description.creation_ms)
+    }
+
+    /// The job-type-specific specification it was submitted with.
+    pub fn spec(&self) -> Option<Value> {
+        self.with_description(|description| description.spec.clone())
+            .filter(|spec| !spec.is_null())
+    }
+
+    /// The job-type-specific terminal result, as reported data rather than the
+    /// typed model [`Job::wait`] returns. `None` until the job succeeds.
+    pub fn result(&self) -> Option<Value> {
+        self.with_description(|description| description.result.clone())
+            .flatten()
+    }
+
+    /// Why the job failed, when it failed and the server reports a reason.
+    pub fn failure(&self) -> Option<JobFailure> {
+        self.with_description(|description| description.failure.clone())
+            .flatten()
+    }
+
+    fn with_description<R>(&self, read: impl FnOnce(&JobDescription) -> R) -> Option<R> {
+        self.cache_read().description.as_ref().map(read)
+    }
+
+    /// This job's recorded lifecycle events.
+    ///
+    /// Unlike the detail accessors, which report a terminal result only once
+    /// the job reaches one, events are written as the job runs and outlive the
+    /// workers that produced them. A distributed job records a
+    /// `claim`/`claim_complete` pair per unit of work, each carrying
+    /// `rows_processed`, so a job that never finishes still accounts for what
+    /// it did. In-process operations keep no event history.
+    pub async fn events(&self, request: JobEventsRequest) -> Result<Vec<RecordBatch>> {
        match &self.inner {
-            JobInner::Handle { handle, .. } => handle.status().await,
-            JobInner::Completed(_) => Ok("finished".to_string()),
+            JobInner::Handle { handle, .. } => handle.events(request).await,
+            // The operation finished before the handle existed, so there is no
+            // id to query with even when a server ran it.
+            JobInner::Completed(_) => Err(Error::NotSupported {
+                message: "this operation completed before its handle was created, so it \
+                          carries no job id to query events with"
+                    .to_string(),
+            }),
        }
    }

@@ -190,8 +398,19 @@ where
    /// [`crate::Error::JobCancelled`] if it was cancelled.
    pub async fn wait(&self) -> Result<T> {
        match &self.inner {
-            JobInner::Handle { handle, decode } => (decode)(handle.wait().await?),
-            JobInner::Completed(result) => Ok(result.clone()),
+            JobInner::Handle { handle, decode } => {
+                let settled = handle.wait().await;
+                // Waiting already established a terminal state; record it so
+                // the detail accessors do not need another round trip for it.
+                if let Some(state) = terminal_state(&settled) {
+                    self.cache_write().state = Some(state.to_string());
+                }
+                (decode)(settled?)
+            }
+            JobInner::Completed(result) => {
+                self.cache_write().state = Some("finished".to_string());
+                Ok(result.clone())
+            }
        }
    }

@@ -224,20 +443,36 @@ where
        U: Clone + Send + Sync + 'static,
        F: Fn(T) -> U + Send + Sync + 'static,
    {
-        match self.inner {
+        // The mapped handle tracks the same job, so it inherits what this one
+        // has already learned about it.
+        let Self { inner, cache } = self;
+        match inner {
            JobInner::Handle { handle, decode } => Job {
                inner: JobInner::Handle {
                    handle,
                    decode: Arc::new(move |result| Ok(map((decode)(result)?))),
                },
+                cache,
            },
            JobInner::Completed(result) => Job {
                inner: JobInner::Completed(map(result)),
+                cache,
            },
        }
    }
 }

+/// The lifecycle state a settled [`JobHandle::wait`] implies.
+fn terminal_state(settled: &Result<TerminalResult>) -> Option<&'static str> {
+    match settled {
+        Ok(_) => Some("finished"),
+        Err(Error::JobFailed { .. }) => Some("failed"),
+        Err(Error::JobCancelled { .. }) => Some("cancelled"),
+        // Anything else is a transport failure, not a verdict on the job.
+        Err(_) => None,
+    }
+}
+
 /// How an in-process operation ended. Cloneable so every waiter can be given
 /// the outcome; [`Error`] is not, so failures share one behind an [`Arc`].
 #[derive(Clone)]
@@ -195,6 +195,7 @@ pub mod query;
 #[cfg(feature = "remote")]
 pub mod remote;
 pub mod rerankers;
+pub mod sql;
 pub mod table;
 #[cfg(test)]
 pub mod test_utils;
@@ -24,8 +24,8 @@ use std::time::{SystemTime, UNIX_EPOCH};

 use arrow_array::cast::AsArray;
 use arrow_array::types::UInt64Type;
-use arrow_array::{RecordBatch, UInt64Array};
-use arrow_schema::{Schema as ArrowSchema, SchemaRef};
+use arrow_array::{RecordBatch, UInt64Array, new_null_array};
+use arrow_schema::{FieldRef, Schema as ArrowSchema, SchemaRef};
 use datafusion::common::ScalarValue;
 use datafusion::error::DataFusionError;
 use datafusion::physical_plan::SendableRecordBatchStream;
@@ -34,7 +34,7 @@ use datafusion::prelude::{col, lit};
 use futures::{StreamExt, TryStreamExt};
 use lance::Dataset;
 use lance::dataset::mem_wal::DatasetMemWalExt;
-use lance::dataset::transaction::{Operation, Transaction};
+use lance::dataset::transaction::{Operation, Transaction, UpdateMode};
 use lance::dataset::write::delete::DeleteBuilder;
 use lance::dataset::write::merge_insert::inserted_rows::{
    KeyExistenceFilter, KeyExistenceFilterBuilder, KeyValue,
@@ -51,6 +51,9 @@ use super::{
    definition_to_metadata,
 };
 use crate::database::OpenTableRequest;
+use crate::table::computed_columns::{
+    computed_column_from_field, computed_columns, ensure_declarations_are_planned,
+};
 use crate::table::{NativeTable, NativeTableExt, Table};
 use crate::{Error, Result};

@@ -167,29 +170,52 @@ pub(crate) async fn execute_refresh(
        .map(|p| (p.output.clone(), p.expression.clone()))
        .collect();
    validate_inputs(&source_ds, definition)?;
-    let (replanned, mut planned_fields, _renames) = super::plan(
+    let (replanned, planned_fields, _renames) = super::plan(
        source_schema,
        &definition.source_table,
-        &projections,
+        &definition.source_namespace,
+        Some(&projections),
        definition.filter.as_deref(),
        definition.limit,
    )?;
+    let mut planned_fields = planned_fields;
    planned_fields.push(arrow_schema::Field::new(
        SOURCE_ROW_ID_COLUMN,
        arrow_schema::DataType::UInt64,
        false,
    ));
+    // A computed column is not planned from the source: refresh writes it
+    // NULL and its declaration's owner fills it. Its declaration must still
+    // be complete, and it must be able to hold NULL.
    let physical = ArrowSchema::from(view_ds.schema());
-    let planned_shape: Vec<_> = planned_fields
-        .iter()
-        .map(|f| (f.name().clone(), f.data_type().clone(), f.is_nullable()))
-        .collect();
-    let physical_shape: Vec<_> = physical
+    let mut computed = computed_columns(&physical).into_iter().map(|c| c.name);
+    if let Some(name) = computed.by_ref().find(|name| {
+        physical
+            .field_with_name(name)
+            .is_ok_and(|f| !f.is_nullable())
+    }) {
+        return Err(Error::Schema {
+            message: format!(
+                "computed column '{name}' of view '{}' cannot hold NULL; recreate the view",
+                view.name()
+            ),
+        });
+    }
+    ensure_declarations_are_planned(&physical)?;
+    let physical_planned: Vec<&FieldRef> = physical
        .fields()
        .iter()
-        .map(|f| (f.name().clone(), f.data_type().clone(), f.is_nullable()))
+        .filter(|f| computed_column_from_field(f).is_none())
        .collect();
-    if planned_shape != physical_shape {
+    // A projected column that became nullable at the source still fits the
+    // view's nullable field; the reverse would not.
+    let matches = planned_fields.len() == physical_planned.len()
+        && planned_fields.iter().zip(&physical_planned).all(|(e, p)| {
+            e.name() == p.name()
+                && e.data_type() == p.data_type()
+                && (p.is_nullable() || !e.is_nullable())
+        });
+    if !matches {
        return Err(Error::Schema {
            message: format!(
                "the stored definition of view '{}' does not produce this \
@@ -228,11 +254,18 @@ pub(crate) async fn execute_refresh(
        .get(SOURCE_VERSION_TS_META_KEY)
        .and_then(|raw| raw.parse().ok());
    // The watermark speaks only for the view state its refresh left behind;
-    // any other commit on the view since then is drift.
-    let view_intact = metadata
+    // any other commit on the view since then is drift, except a fill of its
+    // computed columns, which rewrites nothing refresh certifies.
+    let recorded_view_version = metadata
        .get(VIEW_VERSION_META_KEY)
-        .and_then(|raw| raw.parse::<u64>().ok())
-        == Some(view_ds.version().version);
+        .and_then(|raw| raw.parse::<u64>().ok());
+    let view_intact = match recorded_view_version {
+        Some(recorded) if recorded == view_ds.version().version => true,
+        Some(recorded) if recorded < view_ds.version().version => {
+            only_computed_rewrites_since(&view_ds, recorded).await?
+        }
+        _ => false,
+    };

    if !full && watermark == Some(source_version) && view_intact && recorded_ts == Some(source_ts) {
        return Ok(RefreshMaterializedViewResult {
@@ -590,7 +623,7 @@ async fn open_source(view: &Table, definition: &MaterializedViewDefinition) -> R
    let source = database
        .open_table(OpenTableRequest {
            name: definition.source_table.clone(),
-            namespace_path: Vec::new(),
+            namespace_path: definition.source_namespace.clone(),
            index_cache_size: None,
            lance_read_params: None,
            location: None,
@@ -1089,6 +1122,69 @@ struct RowScope {
    limit: Option<u64>,
 }

+/// Whether every commit on the view after `recorded` is a fill of its
+/// computed columns: a column rewrite or data replacement touching only
+/// those fields and neither adding nor removing rows. A version whose
+/// transaction cannot be read is not proven, so it counts as drift.
+async fn only_computed_rewrites_since(view_ds: &Dataset, recorded: u64) -> Result<bool> {
+    // A fill may write any field under a computed column, so the whole
+    // subtree counts, not only the root.
+    let physical = ArrowSchema::from(view_ds.schema());
+    fn subtree(field: &lance_core::datatypes::Field, ids: &mut Vec<u32>) {
+        ids.push(field.id as u32);
+        for child in &field.children {
+            subtree(child, ids);
+        }
+    }
+    let mut computed_fields = Vec::new();
+    for column in computed_columns(&physical) {
+        if let Some(field) = view_ds.schema().field(&column.name) {
+            subtree(field, &mut computed_fields);
+        }
+    }
+    if computed_fields.is_empty() {
+        return Ok(false);
+    }
+    for version in recorded + 1..=view_ds.version().version {
+        let Some(transaction) = view_ds.read_transaction_by_version(version).await? else {
+            return Ok(false);
+        };
+        let fill = match &transaction.operation {
+            Operation::Update {
+                removed_fragment_ids,
+                new_fragments,
+                fields_modified,
+                update_mode: Some(UpdateMode::RewriteColumns),
+                ..
+            } => {
+                removed_fragment_ids.is_empty()
+                    && new_fragments.is_empty()
+                    && !fields_modified.is_empty()
+                    && fields_modified
+                        .iter()
+                        .all(|field| computed_fields.contains(field))
+            }
+            // What `refresh_column` commits for a SQL declaration.
+            Operation::DataReplacement { replacements } => {
+                !replacements.is_empty()
+                    && replacements.iter().all(|group| {
+                        !group.1.fields.is_empty()
+                            && group
+                                .1
+                                .fields
+                                .iter()
+                                .all(|field| computed_fields.contains(&(*field as u32)))
+                    })
+            }
+            _ => false,
+        };
+        if !fill {
+            return Ok(false);
+        }
+    }
+    Ok(true)
+}
+
 async fn compute_stream(
    source: &Dataset,
    definition: &MaterializedViewDefinition,
@@ -1157,6 +1253,10 @@ async fn compute_stream(
        let batch = batch.map_err(|e| DataFusionError::External(Box::new(e)))?;
        let mut columns = Vec::with_capacity(out_schema.fields().len());
        for field in out_schema.fields() {
+            if computed_column_from_field(field).is_some() {
+                columns.push(new_null_array(field.data_type(), batch.num_rows()));
+                continue;
+            }
            let name = if field.name() == SOURCE_ROW_ID_COLUMN {
                ROW_ID
            } else {
@@ -2767,7 +2867,7 @@ mod tests {
        let (conn, source) = db_with_source(vec![1]).await;
        let prepared = crate::materialized_view::prepare_declaration(
            &source,
-            &[("x".into(), "x".into()), ("twice".into(), "x * 2".into())],
+            Some(&[("x".into(), "x".into()), ("twice".into(), "x * 2".into())]),
            None,
            None,
        )
@@ -2919,6 +3019,7 @@ mod tests {

        let replacement = crate::materialized_view::MaterializedViewDefinition {
            source_table: "src".into(),
+            source_namespace: Vec::new(),
            projections: vec![
                crate::materialized_view::ViewProjection {
                    output: "x".into(),
@@ -2958,6 +3059,7 @@ mod tests {

        let narrower = crate::materialized_view::MaterializedViewDefinition {
            source_table: "src".into(),
+            source_namespace: Vec::new(),
            projections: vec![crate::materialized_view::ViewProjection {
                output: "x".into(),
                expression: "x".into(),
@@ -3129,4 +3231,424 @@ mod tests {
        let err = view.refresh().execute().await.unwrap_err();
        assert!(err.to_string().contains("source table 'src'"), "{err}");
    }
+
+    /// A view with a computed column, declared over `people` and refreshed.
+    async fn refreshed_computed_view(conn: &Connection) -> MaterializedView {
+        use crate::materialized_view::tests::{computed_field, people, test_binding};
+        let source = people(conn).await;
+        let view = crate::materialized_view::prepare_declaration(
+            &source,
+            Some(&[
+                ("id".to_string(), "id".to_string()),
+                ("name".to_string(), "name".to_string()),
+            ]),
+            None,
+            None,
+        )
+        .await
+        .unwrap()
+        .with_computed_columns(
+            vec![(2, computed_field("emb", "fb_1", "name"))],
+            &[test_binding("fb_1", "name", "emb")],
+        )
+        .unwrap()
+        .create("v")
+        .await
+        .unwrap();
+        let result = view.refresh().execute().await.unwrap();
+        assert_eq!(result.mode, RefreshMode::Rebuild);
+        view
+    }
+
+    async fn unfilled(view: &MaterializedView) -> usize {
+        view.table()
+            .count_rows(Some("emb IS NULL".to_string()))
+            .await
+            .unwrap()
+    }
+
+    async fn append_people(conn: &Connection, ids: Vec<i32>, names: Vec<&str>) {
+        let batch = record_batch!(("id", Int32, ids), ("name", Utf8, names)).unwrap();
+        conn.open_table("people")
+            .execute()
+            .await
+            .unwrap()
+            .add(batch)
+            .execute()
+            .await
+            .unwrap();
+    }
+
+    /// Commit the fill job's shape on the view: a column rewrite of
+    /// `fields`, touching no rows. The data is left as it is; what matters
+    /// here is how the next refresh classifies the commit.
+    async fn commit_column_rewrite(view: &MaterializedView, fields: &[&str]) {
+        let native = view.table().as_native().unwrap();
+        native.dataset.reload().await.unwrap();
+        let dataset = native.dataset.get().await.unwrap().as_ref().clone();
+        let fields_modified = fields
+            .iter()
+            .map(|name| dataset.schema().field(name).unwrap().id as u32)
+            .collect();
+        let updated_fragments = dataset
+            .get_fragments()
+            .iter()
+            .map(|fragment| fragment.metadata().clone())
+            .collect();
+        let operation = Operation::Update {
+            removed_fragment_ids: Vec::new(),
+            updated_fragments,
+            new_fragments: Vec::new(),
+            fields_modified,
+            compacted_sstables: Vec::new(),
+            fields_for_preserving_frag_bitmap: Vec::new(),
+            update_mode: Some(UpdateMode::RewriteColumns),
+            inserted_rows_filter: None,
+            updated_fragment_offsets: None,
+        };
+        let read_version = dataset.version().version;
+        CommitBuilder::new(WriteDestination::Dataset(Arc::new(dataset)))
+            .execute(Transaction::new(read_version, operation, None))
+            .await
+            .unwrap();
+    }
+
+    /// Refresh never computes a computed column: every row it writes, on a
+    /// rebuild, an append and a rewrite, carries NULL there, and the
+    /// declaration survives all three.
+    #[tokio::test]
+    async fn test_computed_columns_are_written_null_and_kept() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+        assert_eq!(unfilled(&view).await, 3);
+
+        append_people(&conn, vec![4, 5], vec!["d", "e"]).await;
+        let result = view.refresh().execute().await.unwrap();
+        assert_eq!(result.mode, RefreshMode::Incremental);
+        assert_eq!(unfilled(&view).await, 5);
+
+        conn.open_table("people")
+            .execute()
+            .await
+            .unwrap()
+            .update()
+            .column("name", "'z'")
+            .only_if("id = 1")
+            .execute()
+            .await
+            .unwrap();
+        view.refresh().execute().await.unwrap();
+        assert_eq!(unfilled(&view).await, 5);
+        assert_eq!(read(view.table(), "id").await, vec![1, 2, 3, 4, 5]);
+
+        let schema = view.table().schema().await.unwrap();
+        assert!(
+            crate::table::computed_columns::function_bindings(&schema)
+                .unwrap()
+                .iter()
+                .any(|b| b.binding_id() == "fb_1"),
+            "the binding envelope was lost"
+        );
+        assert!(
+            computed_column_from_field(schema.field_with_name("emb").unwrap()).is_some(),
+            "the declaration was lost"
+        );
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+    }
+
+    /// The fill job's commit rewrites only computed columns. It is the one
+    /// commit on a view that is not drift: the next refresh carries on from
+    /// its watermark instead of rebuilding, which would null what the fill
+    /// just wrote.
+    #[tokio::test]
+    async fn test_a_computed_column_fill_is_not_drift() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+
+        commit_column_rewrite(&view, &["emb"]).await;
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+
+        commit_column_rewrite(&view, &["emb"]).await;
+        append_people(&conn, vec![4], vec!["d"]).await;
+        let result = view.refresh().execute().await.unwrap();
+        assert_eq!(result.mode, RefreshMode::Incremental);
+        assert_eq!(result.rows_written, 1);
+        assert_eq!(read(view.table(), "id").await, vec![1, 2, 3, 4]);
+    }
+
+    /// A column rewrite that reaches a projected column is drift like any
+    /// other write: refresh certifies those columns and must recompute them.
+    #[tokio::test]
+    async fn test_a_rewrite_of_a_projected_column_is_drift() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+
+        commit_column_rewrite(&view, &["emb", "name"]).await;
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Rebuild
+        );
+    }
+
+    /// The declaration contract is checked before any refresh mutation: a
+    /// missing binding envelope and a column that lost its declaration both
+    /// fail closed.
+    #[tokio::test]
+    async fn test_a_broken_declaration_is_refused_before_refresh() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+        let native = view.table().as_native().unwrap();
+        let mut dataset = native.dataset.get().await.unwrap().as_ref().clone();
+        dataset
+            .update_schema_metadata(vec![(
+                crate::table::computed_columns::FUNCTION_BINDINGS_META_KEY.to_string(),
+                None,
+            )])
+            .await
+            .unwrap();
+        let err = view.refresh().execute().await.unwrap_err().to_string();
+        assert!(err.contains("references missing binding 'fb_1'"), "{err}");
+
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+        let native = view.table().as_native().unwrap();
+        let mut dataset = native.dataset.get().await.unwrap().as_ref().clone();
+        dataset
+            .replace_field_metadata(vec![(
+                dataset.schema().field("emb").unwrap().id as u32,
+                HashMap::new(),
+            )])
+            .await
+            .unwrap();
+        let err = view.refresh().execute().await.unwrap_err().to_string();
+        assert!(err.contains("does not match binding 'fb_1'"), "{err}");
+    }
+
+    /// An input the view does not project is materialized on every refresh
+    /// path, before the provenance column, with the source's values.
+    #[tokio::test]
+    async fn test_internal_inputs_are_materialized_and_refreshed() {
+        use crate::materialized_view::tests::{computed_field, strict_people, test_binding};
+        let conn = connect("memory://").execute().await.unwrap();
+        let source = strict_people(&conn).await;
+        let mut prepared = crate::materialized_view::prepare_declaration(
+            &source,
+            Some(&[("id".to_string(), "id".to_string())]),
+            None,
+            None,
+        )
+        .await
+        .unwrap();
+        let input = prepared.input_column("name").unwrap();
+        let view = prepared
+            .with_computed_columns(
+                vec![(1, computed_field("emb", "fb_1", &input))],
+                &[test_binding("fb_1", &input, "emb")],
+            )
+            .unwrap()
+            .create("v")
+            .await
+            .unwrap();
+        let names: Vec<String> = view
+            .table()
+            .schema()
+            .await
+            .unwrap()
+            .fields()
+            .iter()
+            .map(|f| f.name().clone())
+            .collect();
+        assert_eq!(names, ["id", "emb", "__input_name", SOURCE_ROW_ID_COLUMN]);
+
+        let unfilled_inputs = || async {
+            view.table()
+                .count_rows(Some("__input_name IS NULL".to_string()))
+                .await
+                .unwrap()
+        };
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Rebuild
+        );
+        assert_eq!(view.table().count_rows(None).await.unwrap(), 3);
+        assert_eq!(unfilled_inputs().await, 0);
+
+        let more = arrow_array::RecordBatch::try_new(
+            source.schema().await.unwrap(),
+            vec![
+                Arc::new(Int32Array::from(vec![4])),
+                Arc::new(arrow_array::StringArray::from(vec!["d"])),
+            ],
+        )
+        .unwrap();
+        source.add(more).execute().await.unwrap();
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Incremental
+        );
+        assert_eq!(unfilled_inputs().await, 0);
+        assert_eq!(
+            view.table()
+                .count_rows(Some("__input_name = 'd'".to_string()))
+                .await
+                .unwrap(),
+            1
+        );
+
+        source
+            .update()
+            .column("name", "'z'")
+            .only_if("id = 1")
+            .execute()
+            .await
+            .unwrap();
+        view.refresh().execute().await.unwrap();
+        assert_eq!(
+            view.table()
+                .count_rows(Some("__input_name = 'z'".to_string()))
+                .await
+                .unwrap(),
+            1
+        );
+        assert_eq!(
+            unfilled(&view).await,
+            4,
+            "rewritten and new rows are unfilled"
+        );
+    }
+
+    /// A SQL declaration is filled by `refresh_column` on the view, which
+    /// commits a data replacement; the next refresh continues from its
+    /// watermark and keeps what the fill wrote, and only rows the view added
+    /// since come back unfilled.
+    #[tokio::test]
+    async fn test_a_sql_fill_is_not_drift() {
+        use crate::materialized_view::tests::{people, sql_field};
+        let conn = connect("memory://").execute().await.unwrap();
+        let source = people(&conn).await;
+        let view = crate::materialized_view::prepare_declaration(
+            &source,
+            Some(&[("id".to_string(), "id".to_string())]),
+            None,
+            None,
+        )
+        .await
+        .unwrap()
+        .with_computed_columns(
+            vec![(
+                1,
+                sql_field("next", arrow_schema::DataType::Int32, "id + 1", r#"["id"]"#),
+            )],
+            &[],
+        )
+        .unwrap()
+        .create("v")
+        .await
+        .unwrap();
+        let filled = || async {
+            view.table()
+                .count_rows(Some("next = id + 1".to_string()))
+                .await
+                .unwrap()
+        };
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Rebuild
+        );
+        assert_eq!(
+            view.table()
+                .refresh_column("next")
+                .await
+                .unwrap()
+                .rows_filled,
+            3
+        );
+        assert_eq!(filled().await, 3);
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+        assert_eq!(filled().await, 3);
+
+        append_people(&conn, vec![4], vec!["d"]).await;
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Incremental
+        );
+        assert_eq!(filled().await, 3);
+        assert_eq!(
+            view.table()
+                .refresh_column("next")
+                .await
+                .unwrap()
+                .rows_filled,
+            1
+        );
+        assert_eq!(filled().await, 4);
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+    }
+
+    /// A fill of a nested computed column writes its child fields; that is
+    /// still a fill, not drift.
+    #[tokio::test]
+    async fn test_a_nested_sql_fill_is_not_drift() {
+        use crate::materialized_view::tests::{people, sql_field};
+        let conn = connect("memory://").execute().await.unwrap();
+        let source = people(&conn).await;
+        let payload = sql_field(
+            "payload",
+            arrow_schema::DataType::Struct(
+                vec![arrow_schema::Field::new(
+                    "value",
+                    arrow_schema::DataType::Utf8,
+                    true,
+                )]
+                .into(),
+            ),
+            "named_struct('value', name)",
+            r#"["name"]"#,
+        );
+        let view = crate::materialized_view::prepare_declaration(
+            &source,
+            Some(&[("name".to_string(), "name".to_string())]),
+            None,
+            None,
+        )
+        .await
+        .unwrap()
+        .with_computed_columns(vec![(1, payload)], &[])
+        .unwrap()
+        .create("v")
+        .await
+        .unwrap();
+        view.refresh().execute().await.unwrap();
+        assert_eq!(
+            view.table()
+                .refresh_column("payload")
+                .await
+                .unwrap()
+                .rows_filled,
+            3
+        );
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+        assert_eq!(
+            view.table()
+                .count_rows(Some("payload.value = name".to_string()))
+                .await
+                .unwrap(),
+            3
+        );
+    }
 }
@@ -11,6 +11,7 @@ pub(crate) mod db;
 pub(crate) mod job;
 pub mod oauth;
 mod retry;
+pub(crate) mod sql;
 pub(crate) mod table;
 pub(crate) mod util;

@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright The LanceDB Authors

-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::sync::Arc;

 use async_trait::async_trait;
@@ -20,19 +20,20 @@ use lance_namespace::models::{

 use crate::Error;
 use crate::database::{
-    CloneTableRequest, CreateTableMode, CreateTableRequest, Database, DatabaseOptions,
-    JobDescription, JobInfo, OpenTableRequest, ReadConsistency, TableNamesRequest,
+    CloneTableRequest, CreateTableMode, CreateTableRequest, Database, DatabaseOptions, JobInfo,
+    OpenTableRequest, ReadConsistency, TableNamesRequest,
 };
 use crate::error::Result;
 use crate::function::{FunctionRegistrationRequest, FunctionVersion};
 use crate::job::Job;
-use crate::remote::job::{DescribeJobResponse, RemoteJob, job_state_to_client};
+use crate::remote::job::{RemoteJob, job_state_to_client};
 use crate::remote::util::stream_as_body;
 use crate::table::BaseTable;

 use super::client::{
    ClientConfig, HeaderProvider, HttpSend, RequestResultExt, RestfulLanceDbClient, Sender,
 };
+use super::sql::SqlClient;
 use super::table::RemoteTable;
 use super::util::parse_server_version;
 use super::{ARROW_STREAM_CONTENT_TYPE, extract_job_id};
@@ -97,6 +98,7 @@ pub const OPT_REMOTE_PREFIX: &str = "remote_database_";
 pub const OPT_REMOTE_API_KEY: &str = "remote_database_api_key";
 pub const OPT_REMOTE_REGION: &str = "remote_database_region";
 pub const OPT_REMOTE_HOST_OVERRIDE: &str = "remote_database_host_override";
+pub const OPT_REMOTE_SQL_HOST_OVERRIDE: &str = "remote_database_sql_host_override";
 // TODO: add support for configuring client config via key/value options

 #[derive(Clone, Debug, Default)]
@@ -212,6 +214,7 @@ pub struct RemoteDatabase<S: HttpSend = Sender> {
    namespace_context_provider: Option<Arc<dyn DynamicContextProvider>>,
    /// TLS configuration for mTLS support
    tls_config: Option<super::client::TlsConfig>,
+    sql_client: Option<SqlClient>,
 }

 #[derive(Clone)]
@@ -269,22 +272,35 @@ impl DynamicContextProvider for NamespaceHeaderProviderContext {
    }
 }

+pub struct RemoteHostOverrides {
+    pub rest: Option<String>,
+    pub sql: Option<String>,
+}
+
 impl RemoteDatabase {
-    pub fn try_new(
+    pub(crate) fn try_new(
        uri: &str,
        api_key: &str,
        region: &str,
-        host_override: Option<String>,
+        host_overrides: RemoteHostOverrides,
        client_config: ClientConfig,
        options: RemoteOptions,
        read_consistency_interval: Option<std::time::Duration>,
    ) -> Result<Self> {
        let parsed = super::client::parse_db_url(uri)?;
+        let sql_client = SqlClient::new(
+            parsed.db_name.clone(),
+            parsed.db_prefix.clone(),
+            api_key.to_string(),
+            host_overrides.rest.clone(),
+            host_overrides.sql,
+            client_config.clone(),
+        );
        let header_map = RestfulLanceDbClient::<Sender>::default_headers(
            api_key,
            region,
            &parsed.db_name,
-            host_override.is_some(),
+            host_overrides.rest.is_some(),
            &options,
            parsed.db_prefix.as_deref(),
            &client_config,
@@ -312,7 +328,7 @@ impl RemoteDatabase {
        let client = RestfulLanceDbClient::try_new(
            &parsed,
            region,
-            host_override,
+            host_overrides.rest,
            header_map,
            client_config.clone(),
            read_consistency_interval,
@@ -330,6 +346,7 @@ impl RemoteDatabase {
            namespace_headers,
            namespace_context_provider,
            tls_config: client_config.tls_config,
+            sql_client: Some(sql_client),
        })
    }
 }
@@ -427,6 +444,7 @@ mod test_utils {
                namespace_headers: HashMap::new(),
                namespace_context_provider: None,
                tls_config: None,
+                sql_client: None,
            }
        }

@@ -449,6 +467,7 @@ mod test_utils {
                namespace_headers: config.extra_headers.clone(),
                namespace_context_provider,
                tls_config: config.tls_config.clone(),
+                sql_client: None,
            }
        }
    }
@@ -533,6 +552,19 @@ struct RemoteListJobsResponse {
    page_token: Option<String>,
 }

+#[derive(serde::Deserialize)]
+struct RemoteListedFunctionVersion {
+    definition: FunctionVersion,
+}
+
+#[derive(serde::Deserialize)]
+struct RemoteListFunctionsResponse {
+    #[serde(default)]
+    functions: Vec<RemoteListedFunctionVersion>,
+    #[serde(default)]
+    page_token: Option<String>,
+}
+
 #[derive(serde::Deserialize)]
 struct RemoteDropFunctionResponse {
    dropped: bool,
@@ -588,6 +620,43 @@ impl<S: HttpSend> Database for RemoteDatabase<S> {
        response.json().await.err_to_http(request_id)
    }

+    async fn list_functions(&self) -> Result<Vec<FunctionVersion>> {
+        let mut functions = Vec::new();
+        let mut page_token: Option<String> = None;
+        let mut seen_page_tokens = HashSet::new();
+        loop {
+            let mut body = serde_json::json!({ "include_definition": true });
+            if let Some(token) = &page_token {
+                body["page_token"] = serde_json::Value::String(token.clone());
+            }
+            let req = self.client.post("/v1/functions/list").json(&body);
+            let (request_id, response) = self.client.send(req).await?;
+            let response = self.client.check_response(&request_id, response).await?;
+            let status = response.status();
+            let response: RemoteListFunctionsResponse =
+                response.json().await.err_to_http(request_id.clone())?;
+            functions.extend(
+                response
+                    .functions
+                    .into_iter()
+                    .map(|listed| listed.definition),
+            );
+            let Some(next_page_token) = response.page_token.filter(|token| !token.is_empty())
+            else {
+                break;
+            };
+            if !seen_page_tokens.insert(next_page_token.clone()) {
+                return Err(Error::Http {
+                    source: "Function listing response repeated a page_token".into(),
+                    request_id,
+                    status_code: Some(status),
+                });
+            }
+            page_token = Some(next_page_token);
+        }
+        Ok(functions)
+    }
+
    async fn drop_function(&self, name: &str, version: &str) -> Result<bool> {
        let req = self
            .client
@@ -602,11 +671,18 @@ impl<S: HttpSend> Database for RemoteDatabase<S> {
        Ok(response.dropped)
    }

-    fn job(&self, job_id: &str) -> Result<crate::job::Job> {
-        Ok(crate::job::Job::new(Box::new(super::job::RemoteJob::new(
-            self.client.clone(),
-            job_id.to_string(),
-        ))))
+    async fn open_job(&self, job_id: &str) -> Result<Job> {
+        let handle = super::job::RemoteJob::new(self.client.clone(), job_id.to_string());
+        match crate::job::JobHandle::describe(&handle).await {
+            Ok(description) => Ok(Job::opened(Box::new(handle), description)),
+            Err(Error::Http {
+                status_code: Some(StatusCode::NOT_FOUND),
+                ..
+            }) => Err(Error::JobNotFound {
+                job_id: job_id.to_string(),
+            }),
+            Err(err) => Err(err),
+        }
    }

    async fn list_jobs(&self) -> Result<Vec<JobInfo>> {
@@ -643,31 +719,6 @@ impl<S: HttpSend> Database for RemoteDatabase<S> {
        Ok(out)
    }

-    async fn get_job(&self, job_id: &str) -> Result<Option<JobDescription>> {
-        let req = self
-            .client
-            .post("/v1/jobs/describe")
-            .json(&serde_json::json!({ "job_id": job_id }));
-        let (request_id, rsp) = self.client.send(req).await?;
-        let rsp = match self.client.check_response(&request_id, rsp).await {
-            Ok(rsp) => rsp,
-            Err(Error::Http {
-                status_code: Some(StatusCode::NOT_FOUND),
-                ..
-            }) => return Ok(None),
-            Err(err) => return Err(err),
-        };
-        let body: DescribeJobResponse = rsp.json().await.err_to_http(request_id)?;
-        Ok(Some(JobDescription {
-            job_id: body.job_id,
-            job_type: body.job_type,
-            state: job_state_to_client(&body.job_state),
-            creation_ms: body.creation_ms,
-            spec: body.spec,
-            failure: body.failure.map(|reported| reported.into_job_failure()),
-        }))
-    }
-
    async fn cancel_job(&self, job_id: &str) -> Result<bool> {
        let req = self
            .client
@@ -684,19 +735,28 @@ impl<S: HttpSend> Database for RemoteDatabase<S> {
        }
    }

-    async fn job_history(&self, job_id: Option<&str>) -> Result<Vec<arrow_array::RecordBatch>> {
-        let mut body = serde_json::json!({});
-        if let Some(job_id) = job_id {
-            body["job_id"] = serde_json::Value::String(job_id.to_string());
-        }
-        let req = self.client.post("/v1/jobs/query_events").json(&body);
-        let (request_id, rsp) = self.client.send(req).await?;
-        let rsp = self.client.check_response(&request_id, rsp).await?;
-        let bytes = rsp.bytes().await.err_to_http(request_id)?;
-        let reader = arrow_ipc::reader::StreamReader::try_new(std::io::Cursor::new(bytes), None)?;
-        reader
-            .collect::<std::result::Result<Vec<_>, _>>()
-            .map_err(Into::into)
+    async fn execute_query_async(
+        &self,
+        query: &str,
+        default_namespace_path: &[String],
+    ) -> Result<crate::sql::Query> {
+        let client = self
+            .sql_client
+            .as_ref()
+            .ok_or_else(|| Error::NotSupported {
+                message: "SQL is unavailable for this remote database client".to_string(),
+            })?;
+        client.submit(query, default_namespace_path).await
+    }
+
+    async fn describe_query(&self, query_id: uuid::Uuid) -> Result<crate::sql::QueryDescription> {
+        let client = self
+            .sql_client
+            .as_ref()
+            .ok_or_else(|| Error::NotSupported {
+                message: "SQL is unavailable for this remote database client".to_string(),
+            })?;
+        client.describe(query_id).await
    }

    async fn table_names(&self, request: TableNamesRequest) -> Result<Vec<String>> {
@@ -1214,6 +1274,7 @@ mod tests {
    use crate::{
        Connection, Error,
        database::CreateTableMode,
+        job::JobEventsRequest,
        remote::{ARROW_STREAM_CONTENT_TYPE, ClientConfig, HeaderProvider, JSON_CONTENT_TYPE},
    };

@@ -2562,7 +2623,7 @@ mod tests {
    }

    #[tokio::test]
-    async fn test_get_job() {
+    async fn test_open_job() {
        let conn = Connection::new_with_handler(|request| {
            assert_eq!(request.method(), &reqwest::Method::POST);
            assert_eq!(request.url().path(), "/v1/jobs/describe");
@@ -2576,51 +2637,55 @@ mod tests {
                )
                .unwrap()
        });
-        let job = conn.get_job("job-1").await.unwrap().unwrap();
-        assert_eq!(job.job_id, "job-1");
-        assert_eq!(job.job_type, "create_index");
-        assert_eq!(job.state, "failed");
-        assert_eq!(job.creation_ms, 1000);
-        assert_eq!(job.spec["column"], "vec");
-        let failure = job.failure.unwrap();
+        // Opening populates the handle, so the accessors answer without a
+        // second round trip.
+        let job = conn.open_job("job-1").await.unwrap();
+        assert_eq!(job.id(), Some("job-1"));
+        assert_eq!(job.job_type().as_deref(), Some("create_index"));
+        assert_eq!(job.state().as_deref(), Some("failed"));
+        assert_eq!(job.creation_ms(), Some(1000));
+        assert_eq!(job.spec().unwrap()["column"], "vec");
+        assert!(job.result().is_none());
+        let failure = job.failure().unwrap();
        assert_eq!(failure.phase.as_deref(), Some("execute"));
        assert_eq!(failure.message.as_deref(), Some("worker died"));
        assert_eq!(failure.retryable, Some(true));
    }

    #[tokio::test]
-    async fn test_get_job_missing_is_none() {
+    async fn test_open_job_reports_the_terminal_result() {
        let conn = Connection::new_with_handler(|_| {
-            http::Response::builder()
-                .status(404)
-                .body("no such job")
-                .unwrap()
-        });
-        assert!(conn.get_job("nope").await.unwrap().is_none());
-    }
-
-    #[tokio::test]
-    async fn test_cancel_job() {
-        let conn = Connection::new_with_handler(|request| {
-            assert_eq!(request.url().path(), "/v1/jobs/cancel");
            http::Response::builder()
                .status(200)
-                .body(r#"{"job_id": "job-1"}"#)
+                .body(
+                    r#"{"job_id": "job-1", "job_type": "refresh_column", "job_state": "DONE", "creation_ms": 1000, "result": {"rows_assigned": 1000000, "rows_failed": 0}}"#,
+                )
                .unwrap()
        });
-        assert!(conn.cancel_job("job-1").await.unwrap());
+        let job = conn.open_job("job-1").await.unwrap();
+        assert_eq!(job.state().as_deref(), Some("finished"));
+        let result = job.result().unwrap();
+        assert_eq!(result["rows_assigned"], 1_000_000);
+        assert_eq!(result["rows_failed"], 0);
+    }

+    #[tokio::test]
+    async fn test_open_job_missing_fails() {
        let conn = Connection::new_with_handler(|_| {
            http::Response::builder()
                .status(404)
                .body("no such job")
                .unwrap()
        });
-        assert!(!conn.cancel_job("nope").await.unwrap());
+        let err = conn.open_job("nope").await.unwrap_err();
+        assert!(
+            matches!(&err, Error::JobNotFound { job_id } if job_id == "nope"),
+            "{err:?}"
+        );
    }

    #[tokio::test]
-    async fn test_job_history_parses_arrow_stream() {
+    async fn test_job_events_scope_to_that_job() {
        let schema = Arc::new(Schema::new(vec![Field::new(
            "state",
            DataType::Utf8,
@@ -2629,29 +2694,91 @@ mod tests {
        let batch = RecordBatch::try_new(
            schema.clone(),
            vec![Arc::new(arrow_array::StringArray::from(vec![
-                "created", "done",
+                "claim_complete",
            ]))],
        )
        .unwrap();
-        let mut body = Vec::new();
+        let mut events = Vec::new();
        {
-            let mut writer = arrow_ipc::writer::StreamWriter::try_new(&mut body, &schema).unwrap();
+            let mut writer =
+                arrow_ipc::writer::StreamWriter::try_new(&mut events, &schema).unwrap();
            writer.write(&batch).unwrap();
            writer.finish().unwrap();
        }
        let conn = Connection::new_with_handler(move |request| {
-            assert_eq!(request.url().path(), "/v1/jobs/query_events");
-            let req_body: serde_json::Value =
+            let body: serde_json::Value =
                serde_json::from_slice(request.body().unwrap().as_bytes().unwrap()).unwrap();
-            assert_eq!(req_body["job_id"], "job-1");
+            if request.url().path() == "/v1/jobs/describe" {
+                return http::Response::builder()
+                    .status(200)
+                    .body(
+                        r#"{"job_id": "job-1", "job_type": "refresh_column", "job_state": "IN_PROGRESS", "creation_ms": 1}"#
+                            .as_bytes()
+                            .to_vec(),
+                    )
+                    .unwrap();
+            }
+            assert_eq!(request.url().path(), "/v1/jobs/query_events");
+            // The handle supplies job_id; the caller only narrows the query.
+            assert_eq!(body["job_id"], "job-1");
+            assert_eq!(body["limit"], 500);
+            assert_eq!(body["filter"], "state = 'claim_complete'");
            http::Response::builder()
                .status(200)
-                .body(body.clone())
+                .body(events.clone())
                .unwrap()
        });
-        let batches = conn.job_history(Some("job-1")).await.unwrap();
+        let job = conn.open_job("job-1").await.unwrap();
+        let batches = job
+            .events(
+                JobEventsRequest::default()
+                    .limit(500)
+                    .filter("state = 'claim_complete'"),
+            )
+            .await
+            .unwrap();
        assert_eq!(batches.len(), 1);
-        assert_eq!(batches[0].num_rows(), 2);
+        assert_eq!(batches[0].num_rows(), 1);
+    }
+
+    #[tokio::test]
+    async fn test_job_events_keep_the_schema_when_nothing_matches() {
+        let schema = Arc::new(Schema::new(vec![Field::new(
+            "state",
+            DataType::Utf8,
+            false,
+        )]));
+        let mut events = Vec::new();
+        {
+            let mut writer =
+                arrow_ipc::writer::StreamWriter::try_new(&mut events, &schema).unwrap();
+            writer.finish().unwrap();
+        }
+        let conn = Connection::new_with_handler(move |request| {
+            if request.url().path() == "/v1/jobs/describe" {
+                return http::Response::builder()
+                    .status(200)
+                    .body(
+                        r#"{"job_id": "job-1", "job_type": "refresh_column", "job_state": "IN_PROGRESS", "creation_ms": 1}"#
+                            .as_bytes()
+                            .to_vec(),
+                    )
+                    .unwrap();
+            }
+            let body: serde_json::Value =
+                serde_json::from_slice(request.body().unwrap().as_bytes().unwrap()).unwrap();
+            // Only the job id when the caller narrows nothing.
+            assert_eq!(body, serde_json::json!({ "job_id": "job-1" }));
+            http::Response::builder()
+                .status(200)
+                .body(events.clone())
+                .unwrap()
+        });
+        let job = conn.open_job("job-1").await.unwrap();
+        let batches = job.events(JobEventsRequest::default()).await.unwrap();
+        assert_eq!(batches.len(), 1);
+        assert_eq!(batches[0].num_rows(), 0);
+        assert_eq!(batches[0].schema(), schema);
    }

    #[tokio::test]
@@ -2708,6 +2835,119 @@ mod tests {
        assert_eq!(version.version(), "fv_01K3EXACT");
    }

+    #[tokio::test]
+    async fn test_list_functions_requests_definitions_and_paginates() {
+        const VERSION: &str = include_str!(
+            "../../tests/fixtures/first_class_functions/v1/remote_function_version.canonical.json"
+        );
+        let version: serde_json::Value = serde_json::from_str(VERSION).unwrap();
+        let page = Arc::new(AtomicUsize::new(0));
+        let conn = Connection::new_with_handler(move |request| {
+            assert_eq!(request.method(), &reqwest::Method::POST);
+            assert_eq!(request.url().path(), "/v1/functions/list");
+            let body: serde_json::Value =
+                serde_json::from_slice(request.body().unwrap().as_bytes().unwrap()).unwrap();
+            assert_eq!(body["include_definition"], true);
+            match page.fetch_add(1, Ordering::SeqCst) {
+                0 => {
+                    assert!(body.get("page_token").is_none());
+                    http::Response::builder()
+                        .status(200)
+                        .body(r#"{"functions": [], "page_token": "next"}"#.to_string())
+                        .unwrap()
+                }
+                _ => {
+                    assert_eq!(body["page_token"], "next");
+                    http::Response::builder()
+                        .status(200)
+                        .body(
+                            serde_json::json!({
+                                "functions": [{
+                                    "name": "embed",
+                                    "version": "fv_01K3EXACT",
+                                    "definition": version.clone(),
+                                }],
+                            })
+                            .to_string(),
+                        )
+                        .unwrap()
+                }
+            }
+        });
+        let functions = conn.list_functions().await.unwrap();
+        assert_eq!(functions.len(), 1);
+        assert_eq!(functions[0].name(), "embed");
+        assert_eq!(functions[0].version(), "fv_01K3EXACT");
+    }
+
+    #[tokio::test]
+    async fn test_list_functions_stops_on_an_empty_page_token() {
+        let requests = Arc::new(AtomicUsize::new(0));
+        let seen = requests.clone();
+        let conn = Connection::new_with_handler(move |request| {
+            seen.fetch_add(1, Ordering::SeqCst);
+            let body: serde_json::Value =
+                serde_json::from_slice(request.body().unwrap().as_bytes().unwrap()).unwrap();
+            assert!(body.get("page_token").is_none());
+            http::Response::builder()
+                .status(200)
+                .body(r#"{"functions": [], "page_token": ""}"#)
+                .unwrap()
+        });
+
+        let functions = conn.list_functions().await.unwrap();
+        assert!(functions.is_empty());
+        assert_eq!(requests.load(Ordering::SeqCst), 1);
+    }
+
+    #[tokio::test]
+    async fn test_list_functions_rejects_a_page_token_cycle() {
+        let page = Arc::new(AtomicUsize::new(0));
+        let requests = page.clone();
+        let conn = Connection::new_with_handler(move |request| {
+            let body: serde_json::Value =
+                serde_json::from_slice(request.body().unwrap().as_bytes().unwrap()).unwrap();
+            let next_page_token = match page.fetch_add(1, Ordering::SeqCst) {
+                0 => {
+                    assert!(body.get("page_token").is_none());
+                    "one"
+                }
+                1 => {
+                    assert_eq!(body["page_token"], "one");
+                    "two"
+                }
+                2 => {
+                    assert_eq!(body["page_token"], "two");
+                    "one"
+                }
+                page => panic!("unexpected page: {page}"),
+            };
+            http::Response::builder()
+                .status(200)
+                .body(
+                    serde_json::json!({
+                        "functions": [],
+                        "page_token": next_page_token,
+                    })
+                    .to_string(),
+                )
+                .unwrap()
+        });
+
+        let error = conn.list_functions().await.unwrap_err();
+        assert!(
+            matches!(
+                &error,
+                Error::Http {
+                    status_code: Some(http::StatusCode::OK),
+                    ..
+                }
+            ),
+            "got {error:?}"
+        );
+        assert_eq!(requests.load(Ordering::SeqCst), 3);
+    }
+
    #[tokio::test]
    async fn test_drop_function_sends_exact_version_and_decodes_replay() {
        let conn = Connection::new_with_handler(|request| {
@@ -2733,7 +2973,9 @@ mod tests {
        let polls_ref = polls.clone();
        let conn = Connection::new_with_handler(move |request| {
            assert_eq!(request.url().path(), "/v1/jobs/describe");
-            let state = if polls_ref.fetch_add(1, Ordering::SeqCst) == 0 {
+            // Two in-progress answers: one for the load, one for the first
+            // status poll.
+            let state = if polls_ref.fetch_add(1, Ordering::SeqCst) < 2 {
                "IN_PROGRESS"
            } else {
                "DONE"
@@ -2746,11 +2988,13 @@ mod tests {
                ))
                .unwrap()
        });
-        let job = conn.job("job-1").unwrap();
+        let job = conn.open_job("job-1").await.unwrap();
        assert_eq!(job.id(), Some("job-1"));
+        // Opening already answered the state; no extra call needed for it.
+        assert_eq!(job.state().as_deref(), Some("running"));
        assert_eq!(job.status().await.unwrap(), "running");
        job.wait().await.unwrap();
        assert_eq!(job.status().await.unwrap(), "finished");
-        assert!(polls.load(Ordering::SeqCst) >= 3);
+        assert!(polls.load(Ordering::SeqCst) >= 4);
    }
 }
@@ -5,13 +5,15 @@

 use std::time::Duration;

+use arrow_array::RecordBatch;
 use async_trait::async_trait;
 use tokio::time::sleep;

 use serde::Deserialize;

+use crate::database::JobDescription;
 use crate::error::{Error, JobFailure, Result};
-use crate::job::{JobHandle, TerminalResult};
+use crate::job::{JobEventsRequest, JobHandle, TerminalResult};
 use crate::remote::client::{HttpSend, RequestResultExt, RestfulLanceDbClient};

 /// Delay before the second job-state poll; doubles up to [`MAX_POLL_INTERVAL`].
@@ -86,7 +88,7 @@ pub(super) struct DescribeJobResponse {
    #[serde(default)]
    pub(super) spec: serde_json::Value,
    #[serde(default)]
-    result: Option<serde_json::Value>,
+    pub(super) result: Option<serde_json::Value>,
    #[serde(default)]
    pub(super) failure: Option<ReportedFailure>,
 }
@@ -110,6 +112,39 @@ impl DescribeJobResponse {
    fn into_terminal_result(self, request_id: String) -> TerminalResult {
        TerminalResult::remote(self.result, request_id)
    }
+
+    /// The public description this wire envelope stands for.
+    pub(super) fn into_description(self) -> JobDescription {
+        JobDescription {
+            job_id: self.job_id,
+            job_type: self.job_type,
+            state: JobState::from(self.job_state.as_str()).client_label(),
+            creation_ms: self.creation_ms,
+            spec: self.spec,
+            result: self.result,
+            failure: self.failure.map(ReportedFailure::into_job_failure),
+        }
+    }
+}
+
+/// One `/v1/jobs/query_events` round trip.
+pub(super) async fn fetch_job_events<S: HttpSend>(
+    client: &RestfulLanceDbClient<S>,
+    body: serde_json::Value,
+) -> Result<Vec<RecordBatch>> {
+    let request = client.post("/v1/jobs/query_events").json(&body);
+    let (request_id, response) = client.send(request).await?;
+    let response = client.check_response(&request_id, response).await?;
+    let bytes = response.bytes().await.err_to_http(request_id)?;
+    let reader = arrow_ipc::reader::StreamReader::try_new(std::io::Cursor::new(bytes), None)?;
+    let schema = reader.schema();
+    let mut batches = reader.collect::<std::result::Result<Vec<_>, _>>()?;
+    // A query that matched nothing still describes the event columns.
+    // Keep that schema so callers can build a typed empty result.
+    if batches.is_empty() {
+        batches.push(RecordBatch::new_empty(schema));
+    }
+    Ok(batches)
 }

 pub struct RemoteJob<S: HttpSend> {
@@ -123,7 +158,7 @@ impl<S: HttpSend> RemoteJob<S> {
    }

    /// One `/v1/jobs/describe` round trip.
-    async fn describe(&self) -> Result<(String, DescribeJobResponse)> {
+    async fn fetch_description(&self) -> Result<(String, DescribeJobResponse)> {
        let request = self
            .client
            .post("/v1/jobs/describe")
@@ -148,13 +183,28 @@ impl<S: HttpSend> JobHandle for RemoteJob<S> {
    }

    async fn status(&self) -> Result<String> {
-        Ok(self.describe().await?.1.state().client_label())
+        Ok(self.fetch_description().await?.1.state().client_label())
+    }
+
+    async fn describe(&self) -> Result<JobDescription> {
+        Ok(self.fetch_description().await?.1.into_description())
+    }
+
+    async fn events(&self, request: JobEventsRequest) -> Result<Vec<RecordBatch>> {
+        let mut body = serde_json::json!({ "job_id": self.job_id });
+        if let Some(limit) = request.limit {
+            body["limit"] = serde_json::Value::from(limit);
+        }
+        if let Some(filter) = request.filter {
+            body["filter"] = serde_json::Value::String(filter);
+        }
+        fetch_job_events(&self.client, body).await
    }

    async fn wait(&self) -> Result<TerminalResult> {
        let mut interval = INITIAL_POLL_INTERVAL;
        loop {
-            let (request_id, description) = self.describe().await?;
+            let (request_id, description) = self.fetch_description().await?;
            match description.state() {
                JobState::Done => return Ok(description.into_terminal_result(request_id)),
                JobState::Failed => {
@@ -466,7 +466,9 @@ impl TokenSource for AzureImdsSource {
 /// OAuth header provider that manages the full token lifecycle.
 ///
 /// Implements [`HeaderProvider`] to inject `Authorization: Bearer <token>`
-/// headers into every LanceDB request, with automatic token refresh.
+/// headers into every LanceDB request, with automatic token refresh. It also
+/// identifies the bearer credential as OIDC so LanceDB's SQL service selects
+/// OIDC validation instead of API-key validation.
 pub struct OAuthHeaderProvider {
    token_source: Box<dyn TokenSource>,
    token_state: Arc<RwLock<TokenState>>,
@@ -554,10 +556,10 @@ impl OAuthHeaderProvider {
 impl HeaderProvider for OAuthHeaderProvider {
    async fn get_headers(&self) -> Result<HashMap<String, String>> {
        let token = self.get_valid_token().await?;
-        Ok(HashMap::from([(
-            "authorization".to_string(),
-            format!("Bearer {token}"),
-        )]))
+        Ok(HashMap::from([
+            ("authorization".to_string(), format!("Bearer {token}")),
+            ("x-lancedb-credential-type".to_string(), "oidc".to_string()),
+        ]))
    }
 }

@@ -72,7 +72,7 @@ use lance_datafusion::exec::{OneShotExec, execute_plan};
 use reqwest::{RequestBuilder, Response};
 use serde::{Deserialize, Serialize};
 use serde_json::Number;
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::io::Cursor;
 use std::pin::Pin;
 use std::sync::{Arc, Mutex};
@@ -264,6 +264,17 @@ impl<S: HttpSend> crate::job::JobHandle for FreshnessJob<S> {
        crate::job::JobHandle::status(&self.inner).await
    }

+    async fn describe(&self) -> Result<crate::database::JobDescription> {
+        crate::job::JobHandle::describe(&self.inner).await
+    }
+
+    async fn events(
+        &self,
+        request: crate::job::JobEventsRequest,
+    ) -> Result<Vec<arrow_array::RecordBatch>> {
+        crate::job::JobHandle::events(&self.inner, request).await
+    }
+
    async fn wait(&self) -> Result<crate::job::TerminalResult> {
        let result = crate::job::JobHandle::wait(&self.inner).await?;
        let version = self.version.read().await;
@@ -527,6 +538,10 @@ impl<S: HttpSend> RemoteTable<S> {
            "column": canonical_column
        });

+        if !index.replace {
+            body["replace"] = false.into();
+        }
+
        // Add name parameter if provided (for backwards compatibility, only include if Some)
        if let Some(ref name) = index.name {
            body["name"] = serde_json::Value::String(name.clone());
@@ -3647,7 +3662,12 @@ impl<S: HttpSend> BaseTable for RemoteTable<S> {

 #[derive(Serialize, Clone, Debug)]
 pub struct MergeInsertRequest {
-    on: String,
+    // Sent as one repeated `on` query parameter per column, which is how the
+    // namespace spec encodes an array-valued `on`. serde_urlencoded (which
+    // reqwest's `query()` uses) cannot serialize a sequence nested in a struct,
+    // so this field is emitted separately by [`Self::on_query_params`].
+    #[serde(skip_serializing)]
+    on: Vec<String>,
    when_matched_update_all: bool,
    when_matched_update_all_filt: Option<String>,
    when_not_matched_insert_all: bool,
@@ -3663,6 +3683,17 @@ pub struct MergeInsertRequest {
    use_lsm: Option<bool>,
 }

+impl MergeInsertRequest {
+    /// The `on` columns as repeated query parameters: `?on=a&on=b`.
+    ///
+    /// A single column serializes to `?on=a`, exactly what clients sent before
+    /// `on` became a list, so a server that predates composite keys sees no
+    /// change from a single-column caller.
+    pub(crate) fn on_query_params(&self) -> Vec<(&str, &str)> {
+        self.on.iter().map(|col| ("on", col.as_str())).collect()
+    }
+}
+
 fn is_true(b: &bool) -> bool {
    *b
 }
@@ -3675,12 +3706,15 @@ impl TryFrom<MergeInsertBuilder> for MergeInsertRequest {
            return Err(Error::InvalidInput {
                message: "MergeInsertBuilder missing required 'on' field".into(),
            });
-        } else if value.on.len() > 1 {
-            return Err(Error::NotSupported {
-                message: "MergeInsertBuilder only supports a single 'on' column".into(),
+        }
+        // The server rejects a repeated column with a 400; catching it here
+        // names the offending column and costs no round trip.
+        let mut seen = HashSet::with_capacity(value.on.len());
+        if let Some(dup) = value.on.iter().find(|col| !seen.insert(*col)) {
+            return Err(Error::InvalidInput {
+                message: format!("MergeInsertBuilder 'on' column '{dup}' is repeated"),
            });
        }
-        let on = value.on[0].clone();

        let when_matched_update_all_filt = match value.when_matched_update_all_filt {
            Some(MergeFilter::Sql(sql)) => Some(sql),
@@ -3704,7 +3738,7 @@ impl TryFrom<MergeInsertBuilder> for MergeInsertRequest {
            };

        Ok(Self {
-            on,
+            on: value.on,
            when_matched_update_all: value.when_matched_update_all,
            when_matched_update_all_filt,
            when_not_matched_insert_all: value.when_not_matched_insert_all,
@@ -4548,6 +4582,76 @@ mod tests {
        }
    }

+    #[tokio::test]
+    async fn test_merge_insert_composite_key() {
+        let batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![Field::new("a", DataType::Int32, false)])),
+            vec![Arc::new(Int32Array::from(vec![1, 2, 3]))],
+        )
+        .unwrap();
+        let data: Box<dyn RecordBatchReader + Send> = Box::new(RecordBatchIterator::new(
+            [Ok(batch.clone())],
+            batch.schema(),
+        ));
+
+        let table = Table::new_with_handler("my_table", move |request| {
+            assert_eq!(request.url().path(), "/v1/table/my_table/merge_insert/");
+
+            // One repeated `on` per column, in the order the caller gave them.
+            let on = request
+                .url()
+                .query_pairs()
+                .filter(|(key, _)| key == "on")
+                .map(|(_, value)| value.into_owned())
+                .collect::<Vec<_>>();
+            assert_eq!(on, vec!["shard_key".to_string(), "id".to_string()]);
+
+            let params = request.url().query_pairs().collect::<HashMap<_, _>>();
+            assert_eq!(params["when_matched_update_all"], "true");
+            assert_eq!(params["when_not_matched_insert_all"], "true");
+
+            http::Response::builder()
+                .status(200)
+                .body(r#"{"version": 43, "num_deleted_rows": 0, "num_inserted_rows": 3, "num_updated_rows": 0}"#)
+                .unwrap()
+        });
+
+        let mut merge = table.merge_insert(&["shard_key", "id"]);
+        merge.when_matched_update_all(None);
+        merge.when_not_matched_insert_all();
+        let result = table.base_table().merge_insert(merge, data).await.unwrap();
+
+        assert_eq!(result.num_inserted_rows, 3);
+    }
+
+    #[tokio::test]
+    async fn test_merge_insert_rejects_repeated_on_column() {
+        let batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![Field::new("a", DataType::Int32, false)])),
+            vec![Arc::new(Int32Array::from(vec![1]))],
+        )
+        .unwrap();
+        let data: Box<dyn RecordBatchReader + Send> = Box::new(RecordBatchIterator::new(
+            [Ok(batch.clone())],
+            batch.schema(),
+        ));
+
+        let table = Table::new_with_handler::<&str>("my_table", |request| {
+            panic!("Unexpected request: {}", request.url());
+        });
+
+        let merge = table.merge_insert(&["id", "id"]);
+        let err = table
+            .base_table()
+            .merge_insert(merge, data)
+            .await
+            .unwrap_err();
+        assert!(
+            matches!(&err, Error::InvalidInput { message } if message.contains("'id' is repeated")),
+            "unexpected error: {err}"
+        );
+    }
+
    #[tokio::test]
    async fn test_merge_insert_retries_on_409() {
        let batch = RecordBatch::try_new(
@@ -6233,6 +6337,40 @@ mod tests {
        }
    }

+    #[tokio::test]
+    async fn test_create_index_forwards_replace_false_on_existing_route() {
+        let table = Table::new_with_handler("my_table", move |request| {
+            assert_eq!(request.method(), "POST");
+            match request.url().path() {
+                "/v1/table/my_table/describe/" => {
+                    let schema = Schema::new(vec![Field::new("a", DataType::Int32, false)]);
+                    http::Response::builder()
+                        .status(200)
+                        .body(describe_response(&schema))
+                        .unwrap()
+                }
+                "/v1/table/my_table/create_index/" => {
+                    let body = request.body().unwrap().as_bytes().unwrap();
+                    let body: serde_json::Value = serde_json::from_slice(body).unwrap();
+                    assert_eq!(body["replace"], json!(false));
+
+                    http::Response::builder()
+                        .status(200)
+                        .body("{}".to_string())
+                        .unwrap()
+                }
+                path => panic!("Unexpected path: {}", path),
+            }
+        });
+
+        table
+            .create_index(&["a"], Index::BTree(Default::default()))
+            .replace(false)
+            .execute()
+            .await
+            .unwrap();
+    }
+
    #[tokio::test]
    async fn test_create_index_returns_job() {
        let describe_calls = std::sync::Arc::new(std::sync::atomic::AtomicUsize::new(0));
@@ -7855,6 +7993,72 @@ mod tests {
        );
    }

+    /// The refresh handle is wrapped for read-freshness tracking, so it has to
+    /// forward the detail APIs too -- this is the job an operator is holding
+    /// when a backfill goes quiet.
+    #[tokio::test]
+    async fn test_refresh_job_handle_reports_detail_and_events() {
+        let schema = Arc::new(Schema::new(vec![Field::new(
+            "state",
+            DataType::Utf8,
+            false,
+        )]));
+        let batch = RecordBatch::try_new(
+            schema.clone(),
+            vec![Arc::new(arrow_array::StringArray::from(vec![
+                "claim_complete",
+            ]))],
+        )
+        .unwrap();
+        let mut events = Vec::new();
+        {
+            let mut writer =
+                arrow_ipc::writer::StreamWriter::try_new(&mut events, &schema).unwrap();
+            writer.write(&batch).unwrap();
+            writer.finish().unwrap();
+        }
+        let table = Table::new_with_handler("my_table", move |request| {
+            match request.url().path() {
+                "/v1/table/my_table/backfill_column" => http::Response::builder()
+                    .status(202)
+                    .body(br#"{"job_id": "j-42"}"#.to_vec())
+                    .unwrap(),
+                "/v1/jobs/describe" => http::Response::builder()
+                    .status(200)
+                    .body(
+                        r#"{"job_id": "j-42", "job_type": "refresh_column", "job_state": "IN_PROGRESS", "creation_ms": 7, "spec": {"column": "doubled"}}"#
+                            .as_bytes()
+                            .to_vec(),
+                    )
+                    .unwrap(),
+                "/v1/jobs/query_events" => {
+                    let body: serde_json::Value =
+                        serde_json::from_slice(request.body().unwrap().as_bytes().unwrap())
+                            .unwrap();
+                    assert_eq!(body["job_id"], "j-42");
+                    http::Response::builder()
+                        .status(200)
+                        .body(events.clone())
+                        .unwrap()
+                }
+                other => panic!("unexpected path {other}"),
+            }
+        });
+
+        let job = table.refresh_column_async("doubled").await.unwrap();
+        job.refresh().await.unwrap();
+        assert_eq!(job.state().as_deref(), Some("running"));
+        assert_eq!(job.job_type().as_deref(), Some("refresh_column"));
+        assert_eq!(job.creation_ms(), Some(7));
+        assert_eq!(job.spec().unwrap()["column"], "doubled");
+
+        let batches = job
+            .events(crate::job::JobEventsRequest::default())
+            .await
+            .unwrap();
+        assert_eq!(batches.iter().map(|b| b.num_rows()).sum::<usize>(), 1);
+    }
+
    #[tokio::test]
    async fn test_refresh_submission_uses_add_columns_version_fence() {
        let table = Table::new_with_handler("my_table", |request| match request.url().path() {
@@ -734,6 +734,7 @@ impl<S: HttpSend + 'static> ExecutionPlan for RemoteWriteExec<S> {
                WriteOp::MergeInsert { query, timeout } => {
                    let mut request = client
                        .post(&format!("/v1/table/{}/merge_insert/", identifier))
+                        .query(&query.on_query_params())
                        .query(query)
                        .header(CONTENT_TYPE, ARROW_STREAM_CONTENT_TYPE);
                    if let Some(timeout) = timeout {
@@ -1489,7 +1490,7 @@ mod tests {
        });

        let query = MergeInsertRequest {
-            on: "id".to_string(),
+            on: vec!["id".to_string()],
            when_matched_update_all: false,
            when_matched_update_all_filt: None,
            when_not_matched_insert_all: false,
@@ -0,0 +1,124 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+//! Handles to SQL queries running on a remote database.
+
+use std::{fmt, sync::Arc};
+
+use async_trait::async_trait;
+use chrono::{DateTime, Utc};
+use uuid::Uuid;
+
+use crate::{Result, arrow::SendableRecordBatchStream};
+
+/// The externally visible lifecycle state of a submitted SQL query.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum QueryStatus {
+    /// The server is still executing the query.
+    Running,
+    /// The server has made the complete result available.
+    Finished,
+    /// The server accepted cancellation but has not confirmed it yet.
+    Cancelling,
+    /// The server confirmed cancellation.
+    Cancelled,
+}
+
+impl fmt::Display for QueryStatus {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        formatter.write_str(match self {
+            Self::Running => "running",
+            Self::Finished => "finished",
+            Self::Cancelling => "cancelling",
+            Self::Cancelled => "cancelled",
+        })
+    }
+}
+
+/// A point-in-time description of a submitted SQL query.
+#[derive(Clone, Debug, PartialEq)]
+pub struct QueryDescription {
+    /// The stable, connection-scoped identifier assigned when the query was submitted.
+    pub id: Uuid,
+    /// The server-visible lifecycle state.
+    pub status: QueryStatus,
+    /// Server-reported completion progress, when known. Values are in `[0.0, 1.0]`,
+    /// with `1.0` meaning complete.
+    pub progress: Option<f64>,
+    /// When the server may stop accepting this query's continuation token.
+    pub expires_at: Option<DateTime<Utc>>,
+}
+
+#[async_trait]
+pub(crate) trait QueryHandle: Send + Sync {
+    fn id(&self) -> Uuid;
+    async fn describe(&self) -> Result<QueryDescription>;
+    async fn reader(&self) -> Result<SendableRecordBatchStream>;
+    async fn cancel(&self) -> Result<()>;
+}
+
+/// A handle to a submitted SQL query.
+///
+/// The handle can be inspected, opened as an Arrow reader, or cancelled.
+/// Dropping it does not cancel the server-side query.
+/// Identifier lookup is scoped to the connection that submitted the query and
+/// is not a durable resume mechanism.
+pub struct Query {
+    handle: Arc<dyn QueryHandle>,
+}
+
+impl std::fmt::Debug for Query {
+    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        formatter
+            .debug_struct("Query")
+            .field("id", &self.id())
+            .finish()
+    }
+}
+
+impl Query {
+    #[cfg(feature = "remote")]
+    pub(crate) fn new(handle: Arc<dyn QueryHandle>) -> Self {
+        Self { handle }
+    }
+
+    /// Return the stable, connection-scoped identifier for this query.
+    pub fn id(&self) -> Uuid {
+        self.handle.id()
+    }
+
+    /// Get a point-in-time description of the query.
+    pub async fn describe(&self) -> Result<QueryDescription> {
+        self.handle.describe().await
+    }
+
+    /// Wait for the initial result stream and return its Arrow record batches.
+    ///
+    /// The stream can begin yielding partial results before query execution is
+    /// complete. It continues polling for newly available result endpoints
+    /// until the query finishes and all endpoints have been consumed.
+    ///
+    /// Results are single-consumer. Calling this method more than once on the
+    /// same handle returns an error.
+    pub async fn reader(&self) -> Result<SendableRecordBatchStream> {
+        self.handle.reader().await
+    }
+
+    /// Request cancellation of the query.
+    pub async fn cancel(&self) -> Result<()> {
+        self.handle.cancel().await
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::QueryStatus;
+
+    #[test]
+    fn query_status_display_is_stable() {
+        assert_eq!(QueryStatus::Running.to_string(), "running");
+        assert_eq!(QueryStatus::Finished.to_string(), "finished");
+        assert_eq!(QueryStatus::Cancelling.to_string(), "cancelling");
+        assert_eq!(QueryStatus::Cancelled.to_string(), "cancelled");
+    }
+}
@@ -1192,6 +1192,9 @@ impl Table {
    /// valid empty blobs contain empty byte strings. Prefer
    /// [`Self::fetch_blob_files`] for large selections.
    ///
+    /// `_rowid` values stay valid after compaction when the table has stable
+    /// row ids.
+    ///
    /// ```
    /// use arrow_array::UInt64Array;
    /// use futures::TryStreamExt;
@@ -1233,6 +1236,9 @@ impl Table {
    /// the requests. Null blobs produce null output slots; empty ranges on
    /// non-null blobs produce empty byte strings.
    ///
+    /// `_rowid` values stay valid after compaction when the table has stable
+    /// row ids.
+    ///
    /// ```
    /// use lancedb::blob::BlobRangeRequest;
    ///
@@ -1271,6 +1277,9 @@ impl Table {
    /// Same length and order as `row_ids`. Null rows are `None`. Bytes are not
    /// read from disk until a call to [`BlobFile::read`].
    ///
+    /// `_rowid` values stay valid after compaction when the table has stable
+    /// row ids.
+    ///
    /// ```
    /// # use lancedb::Table;
    /// # async fn lazy_read(table: &Table, row_ids: &[u64]) -> Result<(), Box<dyn std::error::Error>> {
@@ -1506,7 +1515,9 @@ impl Table {
    ///
    /// * `on` One or more columns to join on.  This is how records from the
    ///   source table and target table are matched.  Typically this is some
-    ///   kind of key or id column.
+    ///   kind of key or id column.  Several columns match on the composite
+    ///   key: a source row updates a target row only when it agrees on every
+    ///   one of them.
    ///
    /// # Examples
    ///
@@ -2793,7 +2804,7 @@ impl NativeTable {
        namespace_client: Option<Arc<dyn LanceNamespace>>,
        pushdown_operations: HashSet<NamespaceClientPushdownOperation>,
    ) -> Result<Self> {
-        computed_columns::ensure_no_foreign_declarations(batches.arrow_schema().fields())?;
+        let batches = computed_columns::admit_create_source(batches)?;
        // Default params uses format v1.
        let params = params.unwrap_or(WriteParams {
            ..Default::default()
@@ -2893,6 +2904,7 @@ impl NativeTable {
        pushdown_operations: HashSet<NamespaceClientPushdownOperation>,
        session: Option<Arc<lance::session::Session>>,
    ) -> Result<Self> {
+        let batches = computed_columns::admit_create_source(batches)?;
        // Build table_id from namespace + name for the storage options provider
        let mut table_id = namespace.clone();
        table_id.push(name.to_string());
@@ -5666,7 +5678,7 @@ mod tests {
            TableStatistics {
                num_rows: 250,
                num_indices: 0,
-                total_bytes: 8925,
+                total_bytes: 8969,
                fragment_stats: FragmentStatistics {
                    num_fragments: 11,
                    num_small_fragments: 11,
@@ -21,6 +21,7 @@
 //! [`computed_columns`] and [`computed_column_from_field`] read declarations
 //! back off a schema.

+use futures::StreamExt;
 use std::collections::{BTreeSet, HashMap, HashSet};
 use std::sync::Arc;

@@ -60,6 +61,10 @@ pub const FUNCTION_BINDING_ID_META_KEY: &str = "computed_column.function.binding
 /// Field metadata key holding this sibling's ordered Function output ordinal.
 pub const FUNCTION_OUTPUT_ORDINAL_META_KEY: &str = "computed_column.function.output_ordinal";

+/// Reserved Function output ordinal for an internal flattened-result
+/// assignment column.
+pub const FUNCTION_ASSIGNMENT_OUTPUT_ORDINAL: u32 = u32::MAX;
+
 /// Schema metadata key holding all immutable Function bindings.
 pub const FUNCTION_BINDINGS_META_KEY: &str = "lancedb::function_bindings";

@@ -312,22 +317,29 @@ pub(crate) fn ensure_supported_function_metadata(schema: &ArrowSchema) -> Result
                                binding_id
                            ),
                        })?;
-                let output = binding
-                    .outputs()
-                    .get(output_ordinal as usize)
-                    .ok_or_else(|| Error::InvalidInput {
-                        message: format!(
-                            "Function output '{}' has invalid ordinal {}",
-                            field.name(),
-                            output_ordinal
-                        ),
-                    })?;
-                if output.output_name != field.name().as_str() {
+                let destination = if output_ordinal == FUNCTION_ASSIGNMENT_OUTPUT_ORDINAL {
+                    binding
+                        .assignment()
+                        .map(|assignment| assignment.output_name.as_str())
+                } else {
+                    binding
+                        .outputs()
+                        .get(output_ordinal as usize)
+                        .map(|output| output.output_name.as_str())
+                }
+                .ok_or_else(|| Error::InvalidInput {
+                    message: format!(
+                        "Function output '{}' has invalid ordinal {}",
+                        field.name(),
+                        output_ordinal
+                    ),
+                })?;
+                if destination != field.name().as_str() {
                    return Err(Error::InvalidInput {
                        message: format!(
                            "Function output '{}' does not match binding destination '{}'",
                            field.name(),
-                            output.output_name
+                            destination
                        ),
                    });
                }
@@ -498,6 +510,7 @@ fn ensure_known_binding_shape(value: &Value) -> Result<()> {
            "function",
            "inputs",
            "outputs",
+            "assignment",
            "input_schema",
            "output_schema",
        ],
@@ -546,6 +559,13 @@ fn ensure_known_binding_shape(value: &Value) -> Result<()> {
            "output mapping",
        )?;
    }
+    if let Some(assignment) = object.get("assignment") {
+        reject_unknown_object_fields(
+            assignment,
+            &["output_name", "output_field_id"],
+            "assignment mapping",
+        )?;
+    }
    Ok(())
 }

@@ -589,16 +609,13 @@ fn canonical_input_arrow_type(field: &JsonArrowField) -> Result<String> {
        .and_then(|metadata| metadata.get(ARROW_EXT_NAME_KEY))
        .map(String::as_str)
        == Some(BLOB_V2_EXT_NAME);
-    if is_blob_v2 {
+    if is_blob_v2 || field.r#type.fields.is_some() {
        let arrow_field = lance_namespace::schema::convert_json_arrow_field(field)
            .map_err(|e| invalid_function(format!("invalid Function input field: {e}")))?;
-        if !has_supported_blob_v2_layout(&arrow_field) {
-            return Err(invalid_function(format!(
-                "Function input '{}' has an invalid Blob v2 storage layout",
-                arrow_field.name()
-            )));
+        validate_function_blob_nesting(&arrow_field, false)?;
+        if is_blob_v2 {
+            return Ok(FUNCTION_BLOB_V2_TYPE.to_string());
        }
-        return Ok(FUNCTION_BLOB_V2_TYPE.to_string());
    }
    if field.r#type.fields.is_none() && field.r#type.length.is_none() {
        Ok(field.r#type.r#type.clone())
@@ -617,6 +634,34 @@ fn has_supported_blob_v2_layout(field: &ArrowField) -> bool {
        )
 }

+fn validate_function_blob_nesting(field: &ArrowField, inside_collection: bool) -> Result<()> {
+    if field.is_blob_v2() {
+        if inside_collection {
+            return Err(invalid_function(format!(
+                "Function field '{}' nests Blob v2 under a collection, which Function signatures do not support",
+                field.name()
+            )));
+        }
+        if !has_supported_blob_v2_layout(field) {
+            return Err(invalid_function(format!(
+                "Function field '{}' has an invalid Blob v2 storage layout",
+                field.name()
+            )));
+        }
+        return Ok(());
+    }
+    match field.data_type() {
+        DataType::Struct(fields) => fields
+            .iter()
+            .try_for_each(|field| validate_function_blob_nesting(field, inside_collection)),
+        DataType::List(field)
+        | DataType::LargeList(field)
+        | DataType::FixedSizeList(field, _)
+        | DataType::Map(field, _) => validate_function_blob_nesting(field, true),
+        _ => Ok(()),
+    }
+}
+
 /// `fixed_size_list<item, size>` -> (`item`, `size`); the comma must sit outside
 /// any nested `<...>`.
 fn split_fixed_size_list(raw: &str) -> Option<(&str, i32)> {
@@ -697,31 +742,51 @@ fn parse_output_arrow_type(raw: &str) -> Result<JsonArrowDataType> {
 }

 fn function_output_field(name: &str, nullable: bool, raw: &str) -> Result<JsonArrowField> {
-    if raw == FUNCTION_BLOB_V2_TYPE {
-        return lance_namespace::schema::arrow_schema_to_json(&ArrowSchema::new(vec![
-            crate::blob(name, nullable),
-        ]))
+    let field = if raw == FUNCTION_BLOB_V2_TYPE {
+        lance_namespace::schema::arrow_schema_to_json(&ArrowSchema::new(vec![crate::blob(
+            name, nullable,
+        )]))
        .map_err(|e| invalid_function(format!("could not encode Blob v2 output field: {e}")))?
        .fields
        .into_iter()
        .next()
-        .ok_or_else(|| invalid_function("Blob v2 output field is missing"));
-    }
-    Ok(JsonArrowField::new(
-        name.to_string(),
-        nullable,
-        parse_output_arrow_type(raw)?,
-    ))
+        .ok_or_else(|| invalid_function("Blob v2 output field is missing"))?
+    } else {
+        JsonArrowField::new(name.to_string(), nullable, parse_output_arrow_type(raw)?)
+    };
+    let arrow_field = lance_namespace::schema::convert_json_arrow_field(&field)
+        .map_err(|e| invalid_function(format!("invalid Function output field: {e}")))?;
+    validate_function_blob_nesting(&arrow_field, false)?;
+    Ok(field)
 }

-fn function_output_field_matches(expected: &ArrowField, actual: &ArrowField) -> bool {
-    expected.name() == actual.name()
-        && expected.is_nullable() == actual.is_nullable()
-        && if expected.is_blob_v2() {
+/// Whether two fields describe the same Function output.
+///
+/// `compare_identity` covers the field's own name and nullability. Struct
+/// children carry both as part of the declaration and compare with it on. List
+/// children do not: Lance rewrites a list item's name and nullability when it
+/// writes, so a stored `fixed_size_list<item: float not null>` comes back as
+/// `fixed_size_list<item: float>` and never matches the declaration again.
+/// Comparing those by type alone keeps this agreeing with the server, which
+/// draws the same distinction and is what accepted the column when it was
+/// declared.
+fn function_output_field_matches(
+    expected: &ArrowField,
+    actual: &ArrowField,
+    compare_identity: bool,
+) -> bool {
+    if compare_identity
+        && (expected.name() != actual.name() || expected.is_nullable() != actual.is_nullable())
+    {
+        return false;
+    }
+    match (expected.is_blob_v2(), actual.is_blob_v2()) {
+        (false, false) => function_output_type_matches(expected.data_type(), actual.data_type()),
+        (true, true) => {
            has_supported_blob_v2_layout(expected) && has_supported_blob_v2_layout(actual)
-        } else {
-            function_output_type_matches(expected.data_type(), actual.data_type())
        }
+        _ => false,
+    }
 }

 fn function_output_type_matches(expected: &DataType, actual: &DataType) -> bool {
@@ -734,33 +799,19 @@ fn function_output_type_matches(expected: &DataType, actual: &DataType) -> bool
                && expected
                    .iter()
                    .zip(actual)
-                    .all(|(expected, actual)| function_output_field_matches(expected, actual))
+                    .all(|(expected, actual)| function_output_field_matches(expected, actual, true))
        }
        (DataType::List(expected), DataType::List(actual))
        | (DataType::LargeList(expected), DataType::LargeList(actual)) => {
-            function_output_field_matches(expected, actual)
+            function_output_field_matches(expected, actual, false)
        }
        (
            DataType::FixedSizeList(expected, expected_size),
            DataType::FixedSizeList(actual, actual_size),
-        ) => expected_size == actual_size && function_output_field_matches(expected, actual),
+        ) => expected_size == actual_size && function_output_field_matches(expected, actual, false),
        (DataType::Map(expected, expected_sorted), DataType::Map(actual, actual_sorted)) => {
-            expected_sorted == actual_sorted && function_output_field_matches(expected, actual)
-        }
-        _ => false,
-    }
-}
-
-fn function_output_type_has_blob(data_type: &DataType) -> bool {
-    match data_type {
-        DataType::Struct(fields) => fields
-            .iter()
-            .any(|field| field.is_blob_v2() || function_output_type_has_blob(field.data_type())),
-        DataType::List(field)
-        | DataType::LargeList(field)
-        | DataType::FixedSizeList(field, _)
-        | DataType::Map(field, _) => {
-            field.is_blob_v2() || function_output_type_has_blob(field.data_type())
+            expected_sorted == actual_sorted
+                && function_output_field_matches(expected, actual, true)
        }
        _ => false,
    }
@@ -839,23 +890,20 @@ fn ensure_binding_matches_schema(schema: &ArrowSchema, binding: &FunctionBinding
                output.output_name
            ))
        })?;
-        if field.name() != &output.output_name || !field.is_nullable() || output.nullable {
+        if field.name() != &output.output_name || !field.is_nullable() {
            return Err(invalid_function(format!(
                "Function output '{}' no longer matches binding '{}'",
                output.output_name,
                binding.binding_id()
            )));
        }
-        let (type_matches, has_semantic_blob) = if output.arrow_type == FUNCTION_BLOB_V2_TYPE {
-            (has_supported_blob_v2_layout(field), true)
+        let type_matches = if output.arrow_type == FUNCTION_BLOB_V2_TYPE {
+            has_supported_blob_v2_layout(field)
        } else {
            let expected_type = parse_output_arrow_type(&output.arrow_type)?;
            let expected_type = lance_namespace::schema::convert_json_arrow_type(&expected_type)
                .map_err(|e| invalid_function(format!("invalid Function output type: {e}")))?;
-            (
-                function_output_type_matches(&expected_type, field.data_type()),
-                function_output_type_has_blob(&expected_type),
-            )
+            function_output_type_matches(&expected_type, field.data_type())
        };
        if !type_matches {
            return Err(invalid_function(format!(
@@ -886,19 +934,71 @@ fn ensure_binding_matches_schema(schema: &ArrowSchema, binding: &FunctionBinding
                binding.binding_id()
            )));
        }
-        if has_semantic_blob {
-            output_fields.push(function_output_field(
-                field.name(),
-                true,
-                &output.arrow_type,
-            )?);
-        } else {
-            let json = lance_namespace::schema::arrow_schema_to_json(&ArrowSchema::new(vec![
-                ArrowField::new(field.name().clone(), field.data_type().clone(), true),
-            ]))
-            .map_err(|e| invalid_function(format!("invalid Function output schema: {e}")))?;
-            output_fields.push(json.fields.into_iter().next().unwrap());
+        // Rebuild from the declaration rather than from the stored field. The
+        // stored field carries Lance's write-time normalization, which would
+        // never round-trip back to the schema the binding recorded -- the same
+        // reason list children compare by type above. Whether the column on
+        // disk still matches is settled by that comparison, not here.
+        output_fields.push(function_output_field(
+            field.name(),
+            true,
+            &output.arrow_type,
+        )?);
+    }
+    if let Some(assignment) = binding.assignment() {
+        if binding
+            .outputs()
+            .iter()
+            .any(|output| output.result_field == WHOLE_RESULT_FIELD)
+        {
+            return Err(invalid_function(format!(
+                "Function binding '{}' cannot attach an assignment column to a whole result",
+                binding.binding_id()
+            )));
        }
+        let field = schema
+            .field_with_name(&assignment.output_name)
+            .map_err(|_| {
+                invalid_function(format!(
+                    "Function binding '{}' assignment column '{}' is missing",
+                    binding.binding_id(),
+                    assignment.output_name
+                ))
+            })?;
+        let metadata = field.metadata();
+        if field.data_type() != &DataType::Boolean
+            || !field.is_nullable()
+            || metadata.get(COMPUTED_COLUMN_META_KEY).map(String::as_str) != Some("true")
+            || metadata.get(KIND_META_KEY).map(String::as_str) != Some(FUNCTION_KIND)
+            || metadata
+                .get(FUNCTION_BINDING_ID_META_KEY)
+                .map(String::as_str)
+                != Some(binding.binding_id())
+            || metadata
+                .get(FUNCTION_OUTPUT_ORDINAL_META_KEY)
+                .and_then(|value| value.parse::<u32>().ok())
+                != Some(FUNCTION_ASSIGNMENT_OUTPUT_ORDINAL)
+        {
+            return Err(invalid_function(format!(
+                "Function binding '{}' assignment column no longer matches its declaration",
+                binding.binding_id()
+            )));
+        }
+        let json = lance_namespace::schema::arrow_schema_to_json(&ArrowSchema::new(vec![
+            ArrowField::new(assignment.output_name.clone(), DataType::Boolean, true),
+        ]))
+        .map_err(|e| invalid_function(format!("invalid Function assignment schema: {e}")))?;
+        output_fields.push(json.fields.into_iter().next().unwrap());
+    } else if binding.outputs().iter().all(|output| output.nullable)
+        && binding
+            .outputs()
+            .iter()
+            .all(|output| output.result_field != WHOLE_RESULT_FIELD)
+    {
+        return Err(invalid_function(format!(
+            "Function binding '{}' has no flattened-result assignment column",
+            binding.binding_id()
+        )));
    }
    let output_schema = JsonArrowSchema::new(output_fields);
    let output_schema = serde_json::to_value(output_schema).map_err(|e| {
@@ -1055,11 +1155,6 @@ pub(crate) fn plan_function_application(
                    "named-struct Function result field names must be unique",
                ));
            }
-            if output.fields.iter().any(|field| field.nullable) {
-                return Err(invalid_function(
-                    "Function logical outputs must be non-nullable during NULL assignment",
-                ));
-            }
            let unknown = application
                .columns()
                .keys()
@@ -1081,7 +1176,9 @@ pub(crate) fn plan_function_application(
                let fields = output
                    .fields
                    .iter()
-                    .map(|field| function_output_field(&field.name, false, &field.arrow_type))
+                    .map(|field| {
+                        function_output_field(&field.name, field.nullable, &field.arrow_type)
+                    })
                    .collect::<Result<Vec<_>>>()?;
                let mut data_type = JsonArrowDataType::new("struct".to_string());
                data_type.fields = Some(fields);
@@ -1242,6 +1339,106 @@ pub(crate) fn ensure_batch_writes_no_computed_values(
    Ok(())
 }

+/// Validate every computed-column declaration `schema` carries against the
+/// schema itself: every field with declaration metadata is a complete
+/// declaration, a SQL declaration re-plans to the field it declares, a
+/// Function declaration satisfies the binding contract, and no declaration
+/// reads another computed column. What passes here is what `refresh_column`
+/// can execute.
+pub(crate) fn ensure_declarations_are_planned(schema: &ArrowSchema) -> Result<()> {
+    let invalid = |message: String| Error::InvalidInput { message };
+    // A field with any declaration key is a declaration; a partial one is
+    // not "no declaration", it is a broken one.
+    for field in schema.fields() {
+        if field.metadata().keys().any(|k| is_declaration_key(k))
+            && computed_column_from_field(field).is_none()
+        {
+            return Err(invalid(format!(
+                "field '{}' carries an incomplete computed-column declaration",
+                field.name()
+            )));
+        }
+    }
+    let declared: HashSet<String> = computed_columns(schema)
+        .into_iter()
+        .map(|c| c.name)
+        .collect();
+    for column in computed_columns(schema) {
+        let field = schema.field_with_name(&column.name)?;
+        if !field.is_nullable() {
+            return Err(invalid(format!(
+                "computed column '{}' must be nullable until a refresh fills it",
+                column.name
+            )));
+        }
+        match &column.kind {
+            ComputedColumnKind::Sql { expression } => {
+                let others: Vec<ArrowField> = schema
+                    .fields()
+                    .iter()
+                    .filter(|f| f.name() != &column.name)
+                    .map(|f| f.as_ref().clone())
+                    .collect();
+                let bound = bind(Arc::new(ArrowSchema::new(others)), &column.name, expression)?;
+                if let Some(input) = bound.roots.iter().find(|r| declared.contains(*r)) {
+                    return Err(invalid(format!(
+                        "computed column '{}' reads computed column '{input}'",
+                        column.name
+                    )));
+                }
+                if &bound.data_type != field.data_type() {
+                    return Err(invalid(format!(
+                        "computed column '{}' is declared as {} but its expression yields {}",
+                        column.name,
+                        field.data_type(),
+                        bound.data_type
+                    )));
+                }
+                let mut declared_inputs = column.inputs.clone();
+                declared_inputs.sort();
+                if declared_inputs != bound.inputs {
+                    return Err(invalid(format!(
+                        "computed column '{}' declares inputs {:?} but its expression reads {:?}",
+                        column.name, declared_inputs, bound.inputs
+                    )));
+                }
+            }
+            ComputedColumnKind::Function { binding_id, .. } => {
+                // The binding validator resolves each input's leaf; the
+                // no-computed-input rule is about the root it hangs from.
+                let bindings = function_bindings(schema)?;
+                let Some(binding) = bindings.iter().find(|b| b.binding_id() == binding_id) else {
+                    continue; // reported by the binding validator below
+                };
+                // Roots come from the canonical path parser: a quoted
+                // top-level name may itself contain a dot.
+                if let Some(input) = binding
+                    .inputs()
+                    .iter()
+                    .filter_map(|input| resolve_field_path(schema, &input.field_path).ok())
+                    .map(|resolved| resolved.root.name().as_str())
+                    .find(|r| declared.contains(*r))
+                {
+                    return Err(invalid(format!(
+                        "computed column '{}' reads computed column '{input}'",
+                        column.name
+                    )));
+                }
+            }
+            ComputedColumnKind::Unrecognized { kind } => {
+                return Err(Error::NotSupported {
+                    message: format!(
+                        "computed column '{}' is defined by '{kind}', which this version \
+                         of lancedb cannot fill",
+                        column.name
+                    ),
+                });
+            }
+        }
+    }
+    ensure_supported_function_metadata(schema)
+}
+
 /// Reject fields carrying declaration metadata that did not come through
 /// [`plan`]. One authority for creation, overwrite and raw transforms.
 pub(crate) fn ensure_no_foreign_declarations<'a>(
@@ -1700,6 +1897,54 @@ pub(super) async fn add_foreign_kind(table: &crate::Table, name: &str, kind: &st
    .unwrap();
 }

+/// Admit a table's initial data: every declaration it carries is validated,
+/// and the stream refuses any batch with values in a computed column, whose
+/// values come from refresh alone. One boundary for every way a table is
+/// created.
+pub(crate) fn admit_create_source<S: lance_datafusion::utils::StreamingWriteSource>(
+    batches: S,
+) -> Result<UnfilledDeclarations<S>> {
+    let schema = batches.arrow_schema();
+    ensure_declarations_are_planned(&schema)?;
+    let declared = computed_columns(&schema)
+        .into_iter()
+        .map(|c| c.name)
+        .collect();
+    Ok(UnfilledDeclarations {
+        inner: batches,
+        declared,
+    })
+}
+
+/// A write source whose computed columns must arrive unfilled.
+pub(crate) struct UnfilledDeclarations<S> {
+    inner: S,
+    declared: Vec<String>,
+}
+
+impl<S: lance_datafusion::utils::StreamingWriteSource> lance_datafusion::utils::StreamingWriteSource
+    for UnfilledDeclarations<S>
+{
+    fn arrow_schema(&self) -> SchemaRef {
+        self.inner.arrow_schema()
+    }
+
+    fn into_stream(self) -> datafusion_physical_plan::SendableRecordBatchStream {
+        if self.declared.is_empty() {
+            return self.inner.into_stream();
+        }
+        let schema = self.inner.arrow_schema();
+        let declared = self.declared;
+        let stream = self.inner.into_stream().map(move |batch| {
+            let batch = batch?;
+            ensure_batch_writes_no_computed_values(&declared, &batch)
+                .map_err(|e| datafusion_common::DataFusionError::External(Box::new(e)))?;
+            Ok(batch)
+        });
+        Box::pin(datafusion_physical_plan::stream::RecordBatchStreamAdapter::new(schema, stream))
+    }
+}
+
 #[cfg(test)]
 mod tests {
    /// The gate's reproducer: the validator applies the same schema-level
@@ -1718,6 +1963,69 @@ mod tests {
        assert!(super::validate_declarations(schema, &declarations).is_err());
    }

+    #[test]
+    fn list_children_match_by_type_but_struct_children_by_identity() {
+        use arrow_schema::Field as F;
+
+        // Lance rewrites a list item's name and nullability on write, so the
+        // stored field is no longer identical to what was declared. Comparing
+        // those by type keeps a table with a vector output usable.
+        let declared =
+            DataType::FixedSizeList(Arc::new(F::new("item", DataType::Float32, false)), 4);
+        let stored = DataType::FixedSizeList(Arc::new(F::new("item", DataType::Float32, true)), 4);
+        assert!(super::function_output_type_matches(&declared, &stored));
+
+        let renamed =
+            DataType::FixedSizeList(Arc::new(F::new("element", DataType::Float32, true)), 4);
+        assert!(super::function_output_type_matches(&declared, &renamed));
+
+        // The dimension is still part of the declaration.
+        let resized = DataType::FixedSizeList(Arc::new(F::new("item", DataType::Float32, true)), 8);
+        assert!(!super::function_output_type_matches(&declared, &resized));
+
+        // Struct children keep comparing by name and nullability.
+        let struct_declared =
+            DataType::Struct(vec![F::new("changed", DataType::Boolean, false)].into());
+        let struct_nullable =
+            DataType::Struct(vec![F::new("changed", DataType::Boolean, true)].into());
+        let struct_renamed =
+            DataType::Struct(vec![F::new("altered", DataType::Boolean, false)].into());
+        assert!(super::function_output_type_matches(
+            &struct_declared,
+            &struct_declared
+        ));
+        assert!(!super::function_output_type_matches(
+            &struct_declared,
+            &struct_nullable
+        ));
+        assert!(!super::function_output_type_matches(
+            &struct_declared,
+            &struct_renamed
+        ));
+
+        // A list nested inside a struct gets the list rule.
+        let nested_declared = DataType::Struct(
+            vec![F::new(
+                "tokens",
+                DataType::List(Arc::new(F::new("item", DataType::Utf8, false))),
+                true,
+            )]
+            .into(),
+        );
+        let nested_stored = DataType::Struct(
+            vec![F::new(
+                "tokens",
+                DataType::List(Arc::new(F::new("item", DataType::Utf8, true))),
+                true,
+            )]
+            .into(),
+        );
+        assert!(super::function_output_type_matches(
+            &nested_declared,
+            &nested_stored
+        ));
+    }
+
    #[test]
    fn output_arrow_type_grammar_matches_the_shared_golden() {
        let golden: serde_json::Value = serde_json::from_str(include_str!(
@@ -2487,6 +2795,8 @@ mod tests {
        );
    }

+    /// A create carries a declaration only if it re-plans completely; this
+    /// one lacks its inputs and is refused before its forged value matters.
    #[tokio::test]
    async fn test_create_table_cannot_inject_a_declaration() {
        let conn = connect("memory://").execute().await.unwrap();
@@ -2514,7 +2824,7 @@ mod tests {
            .await
            .unwrap_err();
        assert!(
-            matches!(&err, Error::InvalidInput { message } if message.contains("computed()")),
+            matches!(&err, Error::InvalidInput { message } if message.contains("computed column 'doubled'")),
            "{err:?}"
        );
    }
@@ -2719,6 +3029,28 @@ mod tests {
        .unwrap()
    }

+    fn exact_arrow_type(field: ArrowField) -> String {
+        let json =
+            lance_namespace::schema::arrow_schema_to_json(&ArrowSchema::new(vec![field])).unwrap();
+        serde_json::to_string(json.fields[0].r#type.as_ref()).unwrap()
+    }
+
+    fn single_input_application(path: &str) -> FunctionApplication {
+        FunctionApplication::from_json(
+            &serde_json::json!({
+                "function": {"name": "inspect", "version": "fv_nested_blob"},
+                "inputs": [{
+                    "parameter": "value",
+                    "kind": "column",
+                    "value": {"path": path}
+                }],
+                "output": {"kind": "scalar", "arrow_type": "int64", "nullable": false}
+            })
+            .to_string(),
+        )
+        .unwrap()
+    }
+
    fn binding_from_plan(plan: &FunctionDeclarationPlan) -> FunctionBinding {
        let inputs = plan
            .input_bindings
@@ -2810,6 +3142,17 @@ mod tests {
                    &inputs,
                ));
        }
+        if let Some(assignment) = binding.assignment() {
+            fields.push(
+                ArrowField::new(&assignment.output_name, DataType::Boolean, true).with_metadata(
+                    function_computed_column_metadata(
+                        binding.binding_id(),
+                        FUNCTION_ASSIGNMENT_OUTPUT_ORDINAL,
+                        &inputs,
+                    ),
+                ),
+            );
+        }
        ArrowSchema::new(fields)
    }

@@ -2827,6 +3170,74 @@ mod tests {
        .unwrap();
    }

+    #[test]
+    fn test_binding_preserves_all_nullable_outputs_with_an_assignment_column() {
+        let mut raw_binding: Value = serde_json::from_str(include_str!(
+            "../../tests/fixtures/first_class_functions/v1/remote_function_binding.json"
+        ))
+        .unwrap();
+        raw_binding["outputs"][0]["nullable"] = Value::Bool(true);
+        raw_binding["outputs"][1]["nullable"] = Value::Bool(true);
+        let without_assignment: FunctionBinding =
+            serde_json::from_value(raw_binding.clone()).unwrap();
+        let error = ensure_binding_matches_schema(
+            &valid_function_binding_schema(true, true, &without_assignment),
+            &without_assignment,
+        )
+        .unwrap_err();
+        assert!(
+            error
+                .to_string()
+                .contains("flattened-result assignment column")
+        );
+
+        raw_binding["assignment"] = serde_json::json!({
+            "output_name": "__function_assignment_fb_01K3TEXT",
+            "output_field_id": -1,
+        });
+        raw_binding["output_schema"]["fields"]
+            .as_array_mut()
+            .unwrap()
+            .push(serde_json::json!({
+                "name": "__function_assignment_fb_01K3TEXT",
+                "nullable": true,
+                "type": {"type": "bool"},
+            }));
+        let binding: FunctionBinding = serde_json::from_value(raw_binding).unwrap();
+        ensure_binding_matches_schema(
+            &valid_function_binding_schema(true, true, &binding),
+            &binding,
+        )
+        .unwrap();
+
+        let schema = ArrowSchema::new_with_metadata(
+            valid_function_binding_schema(true, true, &binding)
+                .fields()
+                .to_vec(),
+            HashMap::from([(
+                FUNCTION_BINDINGS_META_KEY.to_string(),
+                function_bindings_metadata(std::slice::from_ref(&binding)).unwrap(),
+            )]),
+        );
+        ensure_supported_function_metadata(&schema).unwrap();
+
+        let mut metadata: Value =
+            serde_json::from_str(schema.metadata().get(FUNCTION_BINDINGS_META_KEY).unwrap())
+                .unwrap();
+        metadata["bindings"][0]["assignment"]["future"] = Value::Bool(true);
+        let future_schema = ArrowSchema::new_with_metadata(
+            schema.fields().to_vec(),
+            HashMap::from([(
+                FUNCTION_BINDINGS_META_KEY.to_string(),
+                serde_json::to_string(&metadata).unwrap(),
+            )]),
+        );
+        assert!(matches!(
+            ensure_supported_function_metadata(&future_schema),
+            Err(Error::NotSupported { .. })
+        ));
+    }
+
    #[test]
    fn test_nullable_function_input_cannot_bind_to_non_nullable_parameter() {
        let mut raw_binding: Value = serde_json::from_str(include_str!(
@@ -3046,6 +3457,35 @@ mod tests {
        );
    }

+    #[test]
+    fn test_named_struct_plan_preserves_nullable_result_fields() {
+        let mut value = serde_json::to_value(named_struct_application("{}")).unwrap();
+        value["output"]["fields"][0]["nullable"] = Value::Bool(true);
+        value["output"]["fields"][1]["nullable"] = Value::Bool(true);
+        let application = FunctionApplication::from_json(&value.to_string()).unwrap();
+
+        let expanded =
+            plan_function_application(&function_input_schema(), &application, None).unwrap();
+        assert!(
+            expanded
+                .output_schema
+                .fields
+                .iter()
+                .all(|field| field.nullable)
+        );
+
+        let whole =
+            plan_function_application(&function_input_schema(), &application, Some("features"))
+                .unwrap();
+        let fields = whole.output_schema.fields[0]
+            .r#type
+            .fields
+            .as_ref()
+            .unwrap();
+        assert!(fields[0].nullable);
+        assert!(fields[1].nullable);
+    }
+
    #[test]
    fn test_blob_function_plans_semantic_input_and_scalar_output() {
        let schema = ArrowSchema::new(vec![crate::blob("image", false)]);
@@ -3062,6 +3502,70 @@ mod tests {
        assert!(output_schema.field(0).is_blob_v2());
    }

+    #[test]
+    fn binding_accepts_a_lance_normalized_list_child() {
+        // The whole guard, not just the type helper: this also reaches the
+        // output-schema comparison at the end of ensure_binding_matches_schema,
+        // which used to rebuild the schema from the stored field and so failed
+        // on exactly the same normalization.
+        let input = ArrowField::new("value", DataType::Int64, false);
+        let application = FunctionApplication::from_json(
+            &serde_json::json!({
+                "function": {"name": "embed", "version": "fv_embed"},
+                "inputs": [{
+                    "parameter": "value",
+                    "kind": "column",
+                    "value": {"path": "value"}
+                }],
+                "output": {
+                    "kind": "scalar",
+                    "arrow_type": "fixed_size_list<float32, 4>",
+                    "nullable": false
+                }
+            })
+            .to_string(),
+        )
+        .unwrap();
+        let plan = plan_function_application(
+            &ArrowSchema::new(vec![input.clone()]),
+            &application,
+            Some("embedding"),
+        )
+        .unwrap();
+        let binding = binding_from_plan(&plan);
+
+        // The declaration says the item is non-nullable; Lance rewrites it to
+        // nullable on write, so this is what the column looks like on disk.
+        let stored = DataType::FixedSizeList(
+            Arc::new(ArrowField::new("item", DataType::Float32, true)),
+            4,
+        );
+        let output = ArrowField::new("embedding", stored, true).with_metadata(
+            function_computed_column_metadata(binding.binding_id(), 0, &["value".into()]),
+        );
+
+        ensure_binding_matches_schema(&ArrowSchema::new(vec![input.clone(), output]), &binding)
+            .unwrap();
+
+        // A different element type is still a mismatch.
+        let wrong = ArrowField::new(
+            "embedding",
+            DataType::FixedSizeList(
+                Arc::new(ArrowField::new("item", DataType::Float64, true)),
+                4,
+            ),
+            true,
+        )
+        .with_metadata(function_computed_column_metadata(
+            binding.binding_id(),
+            0,
+            &["value".into()],
+        ));
+        assert!(
+            ensure_binding_matches_schema(&ArrowSchema::new(vec![input, wrong]), &binding).is_err()
+        );
+    }
+
    #[test]
    fn test_blob_scalar_binding_accepts_full_logical_layout() {
        let input = crate::blob("image", false);
@@ -3158,6 +3662,128 @@ mod tests {
        assert_eq!(fields[1].data_type(), &DataType::Int32);
    }

+    #[test]
+    fn test_struct_blob_input_preserves_exact_schema_and_nullability() {
+        let payload = ArrowField::new(
+            "payload",
+            DataType::Struct(Fields::from(vec![
+                ArrowField::new("mime_type", DataType::Utf8, false),
+                ArrowField::new(
+                    "nested",
+                    DataType::Struct(Fields::from(vec![crate::blob("image", true)])),
+                    true,
+                ),
+            ])),
+            true,
+        );
+        let plan = plan_function_application(
+            &ArrowSchema::new(vec![payload]),
+            &single_input_application("payload"),
+            Some("size"),
+        )
+        .unwrap();
+
+        let declared: JsonArrowDataType =
+            serde_json::from_str(&plan.input_bindings[0].arrow_type).unwrap();
+        let DataType::Struct(fields) =
+            lance_namespace::schema::convert_json_arrow_type(&declared).unwrap()
+        else {
+            panic!("expected a struct Function input")
+        };
+        assert!(fields[1].is_nullable());
+        let DataType::Struct(nested) = fields[1].data_type() else {
+            panic!("expected a recursive struct Function input")
+        };
+        assert!(nested[0].is_blob_v2());
+        assert!(nested[0].is_nullable());
+
+        let exact = lance_namespace::schema::convert_json_arrow_schema(&plan.input_schema).unwrap();
+        let DataType::Struct(fields) = exact.field(0).data_type() else {
+            panic!("expected exact input schema to retain the struct")
+        };
+        let DataType::Struct(nested) = fields[1].data_type() else {
+            panic!("expected exact input schema to retain the nested struct")
+        };
+        assert!(nested[0].is_blob_v2());
+    }
+
+    #[test]
+    fn test_recursive_blob_result_plans_one_whole_named_struct_column() {
+        let details_type = exact_arrow_type(ArrowField::new(
+            "details",
+            DataType::Struct(Fields::from(vec![crate::blob("image", true)])),
+            false,
+        ));
+        let application = FunctionApplication::from_json(
+            &serde_json::json!({
+                "function": {"name": "inspect", "version": "fv_nested_blob"},
+                "inputs": [],
+                "output": {
+                    "kind": "named_struct",
+                    "fields": [
+                        {"name": "mime_type", "arrow_type": "utf8", "nullable": false},
+                        {"name": "details", "arrow_type": details_type, "nullable": false}
+                    ]
+                }
+            })
+            .to_string(),
+        )
+        .unwrap();
+        let plan = plan_function_application(&ArrowSchema::empty(), &application, Some("payload"))
+            .unwrap();
+
+        assert_eq!(plan.outputs.len(), 1);
+        assert_eq!(plan.outputs[0].result_field, WHOLE_RESULT_FIELD);
+        let schema =
+            lance_namespace::schema::convert_json_arrow_schema(&plan.output_schema).unwrap();
+        assert_eq!(schema.field(0).name(), "payload");
+        let DataType::Struct(fields) = schema.field(0).data_type() else {
+            panic!("whole named result must be one struct column")
+        };
+        assert_eq!(
+            fields.iter().map(|field| field.name()).collect::<Vec<_>>(),
+            ["mime_type", "details"]
+        );
+        let DataType::Struct(details) = fields[1].data_type() else {
+            panic!("expected recursive result struct")
+        };
+        assert!(details[0].is_blob_v2());
+        assert!(!fields.iter().any(|field| field.name() == "payload"));
+    }
+
+    #[test]
+    fn test_blob_children_under_collections_are_rejected() {
+        let collections = vec![
+            DataType::List(Arc::new(crate::blob("item", false))),
+            DataType::LargeList(Arc::new(crate::blob("item", false))),
+            DataType::FixedSizeList(Arc::new(crate::blob("item", false)), 2),
+            DataType::Map(
+                Arc::new(ArrowField::new(
+                    "entries",
+                    DataType::Struct(Fields::from(vec![
+                        ArrowField::new("key", DataType::Utf8, false),
+                        crate::blob("value", false),
+                    ])),
+                    false,
+                )),
+                false,
+            ),
+        ];
+        for data_type in collections {
+            let schema = ArrowSchema::new(vec![ArrowField::new("value", data_type, false)]);
+            let error = plan_function_application(
+                &schema,
+                &single_input_application("value"),
+                Some("size"),
+            )
+            .unwrap_err();
+            assert!(
+                error.to_string().contains("under a collection"),
+                "got: {error}"
+            );
+        }
+    }
+
    #[test]
    fn test_blob_whole_struct_binding_accepts_full_logical_layout() {
        let input = crate::blob("image", false);
@@ -133,7 +133,7 @@ impl NativeTable {
                    ),
                });
            }
-            (resolved.canonical_path, resolved.field)
+            (resolved.canonical_path, resolved.terminal_field)
        } else {
            Self::resolve_index_field(dataset.schema(), &opts.columns[0])?
        };
@@ -439,7 +439,8 @@ mod tests {
    use arrow_array::record_batch;
    use arrow_array::{
        Array, ArrayRef, BinaryArray, BooleanArray, FixedSizeListArray, Float32Array, Int32Array,
-        LargeBinaryArray, LargeStringArray, RecordBatch, StringArray, StructArray,
+        LargeBinaryArray, LargeStringArray, ListArray, RecordBatch, StringArray, StructArray,
+        UInt32Array,
    };
    use arrow_data::ArrayDataBuilder;
    use arrow_schema::{DataType, Field, Schema};
@@ -458,6 +459,7 @@ mod tests {
    use crate::query::{ExecutableQuery, QueryBase};
    use crate::table::optimize::{CompactionOptions, OptimizeAction};
    use lance_index::scalar::FullTextSearchQuery;
+    use lance_index::scalar::inverted::query::{FtsQuery, MatchQuery};

    fn create_fixed_size_list<T: Array>(
        values: T,
@@ -599,6 +601,80 @@ mod tests {
        assert!(invalid_granularity.is_err());
    }

+    #[tokio::test]
+    async fn test_nested_list_fts_uses_deepest_document_coordinates() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let mut docs = ListBuilder::new(ListBuilder::new(StringBuilder::new()));
+
+        docs.values().values().append_value("alpha");
+        docs.values().values().append_value("beta");
+        docs.values().append(true);
+        docs.values().values().append_value("gamma");
+        docs.values().values().append_value("alpha delta");
+        docs.values().append(true);
+        docs.append(true);
+
+        docs.values().append(true);
+        docs.values().values().append_value("alpha");
+        docs.values().append(true);
+        docs.append(true);
+
+        let batch = RecordBatch::try_from_iter(vec![
+            ("id", Arc::new(Int32Array::from(vec![0, 1])) as ArrayRef),
+            ("docs", Arc::new(docs.finish()) as ArrayRef),
+        ])
+        .unwrap();
+        let table = conn.create_table("nested", batch).execute().await.unwrap();
+
+        let job = table
+            .create_index(
+                &["docs"],
+                Index::FTS(
+                    FtsIndexBuilder::default()
+                        .document_granularity(DocumentGranularity::ListElement),
+                ),
+            )
+            .execute_async()
+            .await
+            .unwrap();
+        job.wait().await.unwrap();
+
+        let query = FullTextSearchQuery::new_query(FtsQuery::Match(
+            MatchQuery::new("alpha".to_string())
+                .with_column(Some("docs".to_string()))
+                .with_document_granularity(DocumentGranularity::ListElement),
+        ));
+        let batches = table
+            .query()
+            .full_text_search(query)
+            .limit(10)
+            .execute()
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+
+        let mut hits = Vec::new();
+        for batch in batches {
+            let ids = batch["id"].as_any().downcast_ref::<Int32Array>().unwrap();
+            let coordinates = batch["_doc_index"]
+                .as_any()
+                .downcast_ref::<ListArray>()
+                .unwrap();
+            for row in 0..batch.num_rows() {
+                let coordinate = coordinates.value(row);
+                let coordinate = coordinate.as_any().downcast_ref::<UInt32Array>().unwrap();
+                hits.push((ids.value(row), coordinate.values().to_vec()));
+            }
+        }
+        hits.sort_unstable();
+        assert_eq!(
+            hits,
+            vec![(0, vec![0, 0]), (0, vec![1, 1]), (1, vec![1, 0])]
+        );
+    }
+
    /// Concurrent waiters, and a wait issued after the job settled, all
    /// succeed once the build does.
    #[tokio::test]
@@ -5,12 +5,17 @@
 //!
 //! [`super::cast::cast_to_table_schema`] calls [`coerce_blob_expr`].

+use std::fmt;
+use std::hash::{Hash, Hasher};
 use std::sync::Arc;

-use arrow_schema::{DataType, Field, FieldRef, Fields};
+use arrow_array::{Array, BooleanArray, RecordBatch};
+use arrow_schema::{DataType, Field, FieldRef, Fields, Schema};
+use arrow_select::nullif::nullif;
 use datafusion::functions::core::{get_field, named_struct};
 use datafusion_common::ScalarValue;
 use datafusion_common::config::ConfigOptions;
+use datafusion_expr::ColumnarValue;
 use datafusion_physical_expr::ScalarFunctionExpr;
 use datafusion_physical_expr::expressions::{CastExpr, Literal};
 use datafusion_physical_plan::PhysicalExpr;
@@ -133,16 +138,102 @@ pub(super) fn coerce_blob_expr(
        ns_args.push(value);
    }

-    let expr: Arc<dyn PhysicalExpr> = Arc::new(ScalarFunctionExpr::new(
+    let built: Arc<dyn PhysicalExpr> = Arc::new(ScalarFunctionExpr::new(
        &format!("named_struct({})", table_field.name()),
        named_struct(),
        ns_args,
        table_field.clone(),
        config.clone(),
    ));
+
+    // `named_struct` always yields a valid struct, so a null input would land
+    // as a row that set neither `data` nor `uri` -- not an absent blob but a
+    // malformed one, which Lance rejects on write.
+    let expr: Arc<dyn PhysicalExpr> = Arc::new(AbsentBlobIsNull {
+        source: input_expr,
+        built,
+        field: table_field.clone(),
+    });
    Ok((expr, table_field.clone()))
 }

+/// Carries the source column's nullity onto the struct built for it.
+///
+/// This is its own expression rather than a `CASE` because the projection
+/// takes its output field from `return_field`, and the generic implementation
+/// rebuilds a bare field -- which would drop the `lance.blob.v2` extension
+/// metadata and stop the column being recognised as a blob at all.
+#[derive(Debug, Clone)]
+struct AbsentBlobIsNull {
+    source: Arc<dyn PhysicalExpr>,
+    built: Arc<dyn PhysicalExpr>,
+    field: FieldRef,
+}
+
+impl fmt::Display for AbsentBlobIsNull {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "absent_blob_is_null({}, {})", self.source, self.built)
+    }
+}
+
+impl PartialEq for AbsentBlobIsNull {
+    fn eq(&self, other: &Self) -> bool {
+        self.source.eq(&other.source) && self.built.eq(&other.built) && self.field == other.field
+    }
+}
+
+impl Eq for AbsentBlobIsNull {}
+
+impl Hash for AbsentBlobIsNull {
+    fn hash<H: Hasher>(&self, state: &mut H) {
+        self.source.hash(state);
+        self.built.hash(state);
+        self.field.hash(state);
+    }
+}
+
+impl PhysicalExpr for AbsentBlobIsNull {
+    fn return_field(&self, _input_schema: &Schema) -> datafusion_common::Result<FieldRef> {
+        Ok(self.field.clone())
+    }
+
+    fn nullable(&self, _input_schema: &Schema) -> datafusion_common::Result<bool> {
+        Ok(true)
+    }
+
+    fn evaluate(&self, batch: &RecordBatch) -> datafusion_common::Result<ColumnarValue> {
+        let rows = batch.num_rows();
+        let built = self.built.evaluate(batch)?.into_array(rows)?;
+        let source = self.source.evaluate(batch)?.into_array(rows)?;
+        let Some(nulls) = source.logical_nulls() else {
+            return Ok(ColumnarValue::Array(built));
+        };
+        // `nullif` nulls the rows the mask marks true, which is where the
+        // source had no value.
+        let absent = BooleanArray::new(!nulls.inner(), None);
+        Ok(ColumnarValue::Array(nullif(built.as_ref(), &absent)?))
+    }
+
+    fn children(&self) -> Vec<&Arc<dyn PhysicalExpr>> {
+        vec![&self.source, &self.built]
+    }
+
+    fn with_new_children(
+        self: Arc<Self>,
+        children: Vec<Arc<dyn PhysicalExpr>>,
+    ) -> datafusion_common::Result<Arc<dyn PhysicalExpr>> {
+        Ok(Arc::new(Self {
+            source: children[0].clone(),
+            built: children[1].clone(),
+            field: self.field.clone(),
+        }))
+    }
+
+    fn fmt_sql(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "{self}")
+    }
+}
+
 enum BlobInputShape<'a> {
    Bytes,
    String,
@@ -313,6 +404,11 @@ mod tests {
        let data = image.column_by_name("data").unwrap();
        assert!(!data.is_null(0));
        assert!(data.is_null(1));
+        // The row itself has to be null, not merely a struct whose children
+        // are. A present-but-empty struct set neither `data` nor `uri`, which
+        // Lance rejects as malformed rather than reading as an absent blob.
+        assert!(!image.is_null(0));
+        assert!(image.is_null(1));
    }

    #[tokio::test]
@@ -227,7 +227,7 @@ pub(crate) fn resolve_arrow_field_path(schema: &Schema, column: &str) -> Result<

 pub(crate) struct ResolvedFtsField {
    pub canonical_path: String,
-    pub field: Field,
+    pub terminal_field: Field,
    pub list_depth: usize,
 }

@@ -309,7 +309,7 @@ pub(crate) fn resolve_lance_fts_field_path(
    );
    Ok(ResolvedFtsField {
        canonical_path,
-        field: Field::from(field),
+        terminal_field: Field::from(terminal),
        list_depth,
    })
 }
@@ -375,7 +375,7 @@ pub(crate) fn resolve_arrow_fts_field_path(
            message: format!("Invalid schema: {}", e),
        })?;
    let resolved = resolve_lance_fts_field_path(&lance_schema, column)?;
-    Ok((resolved.canonical_path, resolved.field))
+    Ok((resolved.canonical_path, resolved.terminal_field))
 }

 pub fn supported_btree_data_type(dtype: &DataType) -> bool {
@@ -647,8 +647,9 @@ mod tests {
            Field::new("docs", text_list(), true),
        ]);

-        let (path, _) = resolve_arrow_fts_field_path(&schema, "docs.content").unwrap();
+        let (path, field) = resolve_arrow_fts_field_path(&schema, "docs.content").unwrap();
        assert_eq!(path, "docs.content");
+        assert_eq!(field.data_type(), &DataType::Utf8);

        let lance_schema = lance_core::datatypes::Schema::try_from(&schema).unwrap();
        let field_id = lance_schema
@@ -19,6 +19,7 @@ use lancedb::{
    connect, connect_namespace,
    database::listing::{
        ListingDatabaseOptions, NewTableConfig, OPT_NEW_TABLE_ENABLE_STABLE_ROW_IDS,
+        OPT_NEW_TABLE_STORAGE_VERSION,
    },
    query::{ExecutableQuery, QueryBase},
    table::{AddDataMode, CompactionOptions, OptimizeAction, OptimizeStats, WriteOptions},
@@ -111,7 +112,7 @@ async fn query_image_struct(table: &Table) -> StructArray {
 }

 #[tokio::test]
-async fn declaring_blob_column_bumps_format_and_enables_stable_row_ids() -> Result<()> {
+async fn declaring_blob_column_uses_v2_2_and_default_row_ids() -> Result<()> {
    let tmp = tempdir().unwrap();
    let db = connect(tmp.path().to_str().unwrap()).execute().await?;
    let table = db
@@ -120,12 +121,12 @@ async fn declaring_blob_column_bumps_format_and_enables_stable_row_ids() -> Resu
        .await?;

    assert!(supports_blob_v2(storage_format_version(&table).await));
-    assert!(uses_stable_row_ids(&table).await);
+    assert!(!uses_stable_row_ids(&table).await);
    Ok(())
 }

 #[tokio::test]
-async fn explicit_stable_row_id_setting_wins_over_blob_default() -> Result<()> {
+async fn blob_create_honors_disabled_stable_row_ids() -> Result<()> {
    let tmp = tempdir().unwrap();
    let db = connect(tmp.path().to_str().unwrap()).execute().await?;
    let table = db
@@ -146,7 +147,10 @@ async fn non_blob_table_keeps_default_format_and_row_id_setting() -> Result<()>
    let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int64, false)]));
    let table = db.create_empty_table("t", schema).execute().await?;

-    assert!(!supports_blob_v2(storage_format_version(&table).await));
+    assert_eq!(
+        storage_format_version(&table).await,
+        LanceFileVersion::Stable.resolve()
+    );
    assert!(!uses_stable_row_ids(&table).await);
    Ok(())
 }
@@ -179,7 +183,7 @@ async fn creating_with_blob_data_bumps_format() -> Result<()> {
    let table = db.create_table("t", batch).execute().await?;

    assert!(supports_blob_v2(storage_format_version(&table).await));
-    assert!(uses_stable_row_ids(&table).await);
+    assert!(!uses_stable_row_ids(&table).await);
    assert_eq!(table.count_rows(None).await?, 1);
    Ok(())
 }
@@ -277,7 +281,7 @@ async fn add_rejects_uncoercible_blob_input() -> Result<()> {
 }

 #[tokio::test]
-async fn connection_level_stable_row_id_setting_wins_over_blob_default() -> Result<()> {
+async fn connection_disables_stable_row_ids_on_blob_create() -> Result<()> {
    let tmp = tempdir().unwrap();
    let db = connect(tmp.path().to_str().unwrap())
        .storage_option(OPT_NEW_TABLE_ENABLE_STABLE_ROW_IDS, "false")
@@ -294,7 +298,7 @@ async fn connection_level_stable_row_id_setting_wins_over_blob_default() -> Resu
 }

 #[tokio::test]
-async fn namespace_create_applies_blob_defaults() -> Result<()> {
+async fn namespace_blob_create_uses_v2_2_and_default_row_ids() -> Result<()> {
    let tmp = tempdir().unwrap();
    let mut properties = std::collections::HashMap::new();
    properties.insert("root".to_string(), tmp.path().to_str().unwrap().to_string());
@@ -304,6 +308,23 @@ async fn namespace_create_applies_blob_defaults() -> Result<()> {
        .execute()
        .await?;

+    assert!(supports_blob_v2(storage_format_version(&table).await));
+    assert!(!uses_stable_row_ids(&table).await);
+    Ok(())
+}
+
+#[tokio::test]
+async fn namespace_create_honors_enabled_stable_row_ids() -> Result<()> {
+    let tmp = tempdir().unwrap();
+    let mut properties = std::collections::HashMap::new();
+    properties.insert("root".to_string(), tmp.path().to_str().unwrap().to_string());
+    let db = connect_namespace("dir", properties).execute().await?;
+    let table = db
+        .create_empty_table("t", blob_table_schema())
+        .storage_option(OPT_NEW_TABLE_ENABLE_STABLE_ROW_IDS, "true")
+        .execute()
+        .await?;
+
    assert!(supports_blob_v2(storage_format_version(&table).await));
    assert!(uses_stable_row_ids(&table).await);
    Ok(())
@@ -430,6 +451,35 @@ async fn collect_id_rowid(table: &Table) -> Result<Vec<(i64, u64)>> {
        .collect())
 }

+fn assert_missing_blob_row_ids(err: &Error) {
+    assert!(matches!(err, Error::InvalidInput { .. }), "got {err:?}");
+    let message = err.to_string();
+    assert!(message.contains("row ids"), "{message}");
+    assert!(!message.contains("rowaddr"), "{message}");
+    assert!(!message.contains("fragment"), "{message}");
+}
+
+async fn assert_fetch_apis_reject_missing_row_ids(table: &Table, row_ids: &[u64]) -> Result<()> {
+    let err = table.fetch_blobs("image", row_ids).await.unwrap_err();
+    assert_missing_blob_row_ids(&err);
+
+    let err = table.fetch_blob_files("image", row_ids).await.unwrap_err();
+    assert_missing_blob_row_ids(&err);
+
+    let err = table
+        .fetch_blob_ranges(
+            "image",
+            row_ids
+                .iter()
+                .copied()
+                .map(|row_id| BlobRangeRequest::new(row_id, 0, 1)),
+        )
+        .await
+        .unwrap_err();
+    assert_missing_blob_row_ids(&err);
+    Ok(())
+}
+
 #[tokio::test]
 async fn fetch_blobs_round_trips_bytes() -> Result<()> {
    let tmp = tempdir().unwrap();
@@ -482,7 +532,7 @@ async fn fetch_blobs_round_trips_nested_blob_column() -> Result<()> {
    let table = db.create_table("t", batch).execute().await?;

    assert!(supports_blob_v2(storage_format_version(&table).await));
-    assert!(uses_stable_row_ids(&table).await);
+    assert!(!uses_stable_row_ids(&table).await);

    let ids = collect_row_ids(&table).await?;
    let bytes = table.fetch_blobs("info.blob", &ids).await?;
@@ -656,8 +706,7 @@ async fn fetch_blob_ranges_validates_requests() -> Result<()> {
        .fetch_blob_ranges("image", [BlobRangeRequest::new(u64::MAX, 0, 1)])
        .await
        .unwrap_err();
-    assert!(matches!(&err, Error::InvalidInput { .. }), "got {err:?}");
-    assert!(err.to_string().contains("row IDs"));
+    assert_missing_blob_row_ids(&err);
    Ok(())
 }

@@ -690,7 +739,21 @@ async fn fetch_blobs_out_of_range_id_errors_without_panic() -> Result<()> {
    let table = create_inline_blob_table(&db, "t", &[1], &[Some(b"x".as_slice())]).await?;

    let err = table.fetch_blobs("image", &[u64::MAX]).await.unwrap_err();
-    assert!(err.to_string().contains("row IDs"));
+    assert_missing_blob_row_ids(&err);
+    Ok(())
+}
+
+#[tokio::test]
+async fn fetch_blob_files_rejects_missing_fragment_row_addr() -> Result<()> {
+    let tmp = tempdir().unwrap();
+    let db = connect(tmp.path().to_str().unwrap()).execute().await?;
+    let table = create_inline_blob_table(&db, "t", &[1], &[Some(b"x".as_slice())]).await?;
+
+    let err = table
+        .fetch_blob_files("image", &[1u64 << 32])
+        .await
+        .unwrap_err();
+    assert_missing_blob_row_ids(&err);
    Ok(())
 }

@@ -700,24 +763,25 @@ async fn fetch_blob_apis_reject_mixed_valid_and_missing_row_ids() -> Result<()>
    let db = connect(tmp.path().to_str().unwrap()).execute().await?;
    let table = create_inline_blob_table(&db, "t", &[1], &[Some(b"x".as_slice())]).await?;
    let row_id = collect_row_ids(&table).await?[0];
-    let row_ids = [u64::MAX, row_id];
+    let missing_row_addr = 1u64 << 32;
+    let row_ids = [missing_row_addr, row_id];
+    assert_fetch_apis_reject_missing_row_ids(&table, &row_ids).await
+}

-    let err = table.fetch_blobs("image", &row_ids).await.unwrap_err();
-    assert!(matches!(&err, Error::InvalidInput { .. }), "got {err:?}");
-    assert!(err.to_string().contains("row IDs"));
+#[tokio::test]
+async fn fetch_blob_apis_reject_deleted_row_ids() -> Result<()> {
+    let tmp = tempdir().unwrap();
+    let db = connect(tmp.path().to_str().unwrap()).execute().await?;
+    let table =
+        create_inline_blob_table(&db, "t", &[1, 2], &[Some(b"one".as_slice()), Some(b"two")])
+            .await?;
+    let pairs = collect_id_rowid(&table).await?;
+    let deleted_row_addr = pairs.iter().find(|(id, _)| *id == 2).unwrap().1;
+    let live_row_addr = pairs.iter().find(|(id, _)| *id == 1).unwrap().1;

-    let err = table.fetch_blob_files("image", &row_ids).await.unwrap_err();
-    assert!(matches!(&err, Error::InvalidInput { .. }), "got {err:?}");
-    assert!(err.to_string().contains("row IDs"));
+    table.delete("id = 2").await?;

-    let requests = row_ids.map(|row_id| BlobRangeRequest::new(row_id, 0, 1));
-    let err = table
-        .fetch_blob_ranges("image", requests)
-        .await
-        .unwrap_err();
-    assert!(matches!(&err, Error::InvalidInput { .. }), "got {err:?}");
-    assert!(err.to_string().contains("row IDs"));
-    Ok(())
+    assert_fetch_apis_reject_missing_row_ids(&table, &[deleted_row_addr, live_row_addr]).await
 }

 #[tokio::test]
@@ -749,7 +813,11 @@ async fn fetch_blobs_rejects_unknown_column() -> Result<()> {
 #[tokio::test]
 async fn fetch_blobs_rejects_legacy_v1_blob_column() -> Result<()> {
    let tmp = tempdir().unwrap();
-    let db = connect(tmp.path().to_str().unwrap()).execute().await?;
+    // Legacy v1 blob columns are only writable at file version <= 2.1.
+    let db = connect(tmp.path().to_str().unwrap())
+        .storage_options([(OPT_NEW_TABLE_STORAGE_VERSION, "2.1")])
+        .execute()
+        .await?;
    let legacy = Field::new("image", DataType::LargeBinary, true).with_metadata(
        std::collections::HashMap::from([("lance-encoding:blob".to_string(), "true".to_string())]),
    );
@@ -920,7 +988,10 @@ async fn fetch_blobs_after_delete() -> Result<()> {
 #[tokio::test]
 async fn fetch_blobs_with_precompaction_row_ids_survives_compaction() -> Result<()> {
    let tmp = tempdir().unwrap();
-    let db = connect(tmp.path().to_str().unwrap()).execute().await?;
+    let db = connect(tmp.path().to_str().unwrap())
+        .storage_option(OPT_NEW_TABLE_ENABLE_STABLE_ROW_IDS, "true")
+        .execute()
+        .await?;
    let table = create_inline_blob_table(&db, "t", &[1], &[Some(b"frag-one".as_slice())]).await?;
    table
        .add(binary_input_batch(&[2], &[Some(b"frag-two".as_slice())]))