Merge remote-tracking branch 'origin/main' into yang/udf-secret-api

BREAKING CHANGE: list_tables now provides tables in arbitrary order and
the page token is now completely opaque. `table_names` retains the old
behavior of lexical ordering and `start-after` semantics.

Listing the tables in a directory database cost what the database held
rather than what the page held. `ListingDatabase::list_tables`
enumerated every child directory of the base path, sorted the names,
then discarded all but the requested page — on every request, for every
page. On object storage that is one full listing per page.

This PR pages the store instead. `list_tables` asks for one page at a
time through `ObjectStore::read_dir_page`, carrying the store's own
continuation token, so a page is one request. Non-table children can
leave a page short of its limit, so the walk continues until the page is
full or the store runs out.

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Cargo.lock

@@ -3455,8 +3455,8 @@ checksum = "42703706b716c37f96a77aea830392ad231f44c9e9a67872fa5548707e11b11c"
 
 [[package]]
 name = "fsst"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-array",
  "rand 0.9.5",
@@ -4815,8 +4815,8 @@ checksum = "e037a2e1d8d5fdbd49b16a4ea09d5d6401c1f29eca5ff29d03d3824dba16256a"
 
 [[package]]
 name = "lance"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arc-swap",
  "arrow",
@@ -4888,8 +4888,8 @@ dependencies = [
 
 [[package]]
 name = "lance-arrow"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-array",
  "arrow-buffer",
@@ -4911,7 +4911,7 @@ dependencies = [
 [[package]]
 name = "lance-arrow-scalar"
 version = "58.0.0"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-array",
  "arrow-buffer",
@@ -4925,7 +4925,7 @@ dependencies = [
 [[package]]
 name = "lance-arrow-stats"
 version = "58.0.0"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-array",
  "arrow-schema",
@@ -4934,8 +4934,8 @@ dependencies = [
 
 [[package]]
 name = "lance-bitpacking"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrayref",
  "crunchy",
@@ -4945,8 +4945,8 @@ dependencies = [
 
 [[package]]
 name = "lance-core"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-array",
  "arrow-buffer",
@@ -4983,8 +4983,8 @@ dependencies = [
 
 [[package]]
 name = "lance-datafusion"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow",
  "arrow-array",
@@ -5013,8 +5013,8 @@ dependencies = [
 
 [[package]]
 name = "lance-datagen"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow",
  "arrow-array",
@@ -5031,8 +5031,8 @@ dependencies = [
 
 [[package]]
 name = "lance-derive"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -5041,8 +5041,8 @@ dependencies = [
 
 [[package]]
 name = "lance-encoding"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-arith",
  "arrow-array",
@@ -5075,8 +5075,8 @@ dependencies = [
 
 [[package]]
 name = "lance-file"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-arith",
  "arrow-array",
@@ -5107,8 +5107,8 @@ dependencies = [
 
 [[package]]
 name = "lance-index"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arc-swap",
  "arrow",
@@ -5172,8 +5172,8 @@ dependencies = [
 
 [[package]]
 name = "lance-index-core"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-array",
  "arrow-schema",
@@ -5195,8 +5195,8 @@ dependencies = [
 
 [[package]]
 name = "lance-io"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow",
  "arrow-array",
@@ -5236,8 +5236,8 @@ dependencies = [
 
 [[package]]
 name = "lance-linalg"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-array",
  "arrow-schema",
@@ -5251,8 +5251,8 @@ dependencies = [
 
 [[package]]
 name = "lance-namespace"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow",
  "async-trait",
@@ -5264,8 +5264,8 @@ dependencies = [
 
 [[package]]
 name = "lance-namespace-impls"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow",
  "arrow-ipc",
@@ -5318,8 +5318,8 @@ dependencies = [
 
 [[package]]
 name = "lance-select"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-array",
  "arrow-buffer",
@@ -5333,8 +5333,8 @@ dependencies = [
 
 [[package]]
 name = "lance-table"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow",
  "arrow-array",
@@ -5374,8 +5374,8 @@ dependencies = [
 
 [[package]]
 name = "lance-testing"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "arrow-array",
  "arrow-schema",
@@ -5388,8 +5388,8 @@ dependencies = [
 
 [[package]]
 name = "lance-tokenizer"
-version = "12.0.0-beta.4"
-source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.4#474ae88dd75c437d5d506859b627510978e502e9"
+version = "12.0.0-beta.2"
+source = "git+https://github.com/lance-format/lance.git?tag=v12.0.0-beta.2#dafa4642658d996b3e31dde91e02f72db7860d7e"
 dependencies = [
  "frostem",
  "icu_segmenter",

Cargo.toml

@@ -13,20 +13,20 @@ categories = ["database-implementations"]
 rust-version = "1.91.0"
 
 [workspace.dependencies]
-lance = { "version" = "=12.0.0-beta.4", default-features = false, "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-core = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-datagen = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-file = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-io = { "version" = "=12.0.0-beta.4", default-features = false, "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-index = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-linalg = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-namespace = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-namespace-impls = { "version" = "=12.0.0-beta.4", default-features = false, "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-table = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-testing = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-datafusion = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-encoding = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
-lance-arrow = { "version" = "=12.0.0-beta.4", "tag" = "v12.0.0-beta.4", "git" = "https://github.com/lance-format/lance.git" }
+lance = { "version" = "=12.0.0-beta.2", default-features = false, "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-core = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-datagen = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-file = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-io = { "version" = "=12.0.0-beta.2", default-features = false, "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-index = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-linalg = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace-impls = { "version" = "=12.0.0-beta.2", default-features = false, "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-table = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-testing = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-datafusion = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-encoding = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "git" = "https://github.com/lance-format/lance.git" }
+lance-arrow = { "version" = "=12.0.0-beta.2", "tag" = "v12.0.0-beta.2", "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

docs/src/python/python.md

@@ -223,9 +223,13 @@ tokens = list(
 Blob columns store large binary values out of line so they can be read lazily
 instead of being materialized with the rest of the row.
 
-::: lancedb.blob
+`lancedb.BlobType` is `lance.blob.BlobType` when pylance is installed. Without
+pylance, LanceDB uses a matching `lance.blob.v2` extension type so blob columns
+still work. Queries return descriptors. Call
+[`fetch_blob_files`][lancedb.table.Table.fetch_blob_files] for lazy reads or
+[`fetch_blobs`][lancedb.table.Table.fetch_blobs] for eager bytes.
 
-::: lancedb.BlobType
+::: lancedb.blob
 
 ::: lancedb._blob.BlobFile
     options:

java/pom.xml

@@ -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.4</lance-core.version>
+        <lance-core.version>12.0.0-beta.2</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>

python/python/lancedb/__init__.py

@@ -6,7 +6,7 @@ import importlib.metadata
 import os
 from concurrent.futures import ThreadPoolExecutor
 from datetime import timedelta
-from typing import Dict, Optional, Union, Any, List, Iterable
+from typing import Dict, Optional, Union, Any, List, Iterable, TYPE_CHECKING
 
 __version__ = importlib.metadata.version("lancedb")
 
@@ -20,7 +20,7 @@ from .db import AsyncConnection, DBConnection, LanceDBConnection
 from .remote import ClientConfig
 from .remote.db import RemoteDBConnection
 from .expr import Expr, col, lit, func
-from .schema import blob, vector, BlobType
+from .schema import blob, vector
 from .job import AsyncJob, Job
 from .functions import (
     FunctionArtifactRequest as FunctionArtifactRequest,
@@ -49,6 +49,19 @@ from .namespace import (
 )
 
 
+if TYPE_CHECKING:
+    from lance.blob import BlobType as BlobType
+
+
+def __getattr__(name: str):
+    if name == "BlobType":
+        from .schema import BlobType
+
+        globals()["BlobType"] = BlobType
+        return BlobType
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+
+
 def _check_s3_bucket_with_dots(
     uri: str, storage_options: Optional[Dict[str, str]]
 ) -> None:

python/python/lancedb/_blob.py

@@ -12,7 +12,7 @@ from typing import TYPE_CHECKING, Optional, Union
 import pyarrow as pa
 
 from .expr import Expr
-from .schema import blob_v2_column_paths
+from .schema import row_addressable_blob_v2_paths
 from .types import BlobMode, QueryProjection, QueryProjectionSpec
 
 if TYPE_CHECKING:
@@ -119,7 +119,7 @@ def blob_v2_projection_sources(
     schema: pa.Schema,
     projection: QueryProjection,
 ) -> dict[str, str]:
-    blob_columns = blob_v2_column_paths(schema)
+    blob_columns = row_addressable_blob_v2_paths(schema)
     if not blob_columns:
         return {}
     columns = set(blob_columns)
@@ -140,7 +140,9 @@ def v2_projection_needs_row_id(
 ) -> bool:
     if with_row_id:
         return False
-    return projection_includes_blob_column(projection, blob_v2_column_paths(schema))
+    return projection_includes_blob_column(
+        projection, row_addressable_blob_v2_paths(schema)
+    )
 
 
 def blob_auto_row_id_for_scan(

python/python/lancedb/db.py

@@ -16,6 +16,7 @@ from typing import (
     Iterable,
     List,
     Literal,
+    Mapping,
     Optional,
     Union,
 )
@@ -687,17 +688,35 @@ class DBConnection(EnforceOverrides):
         """
         raise NotImplementedError("serialize is not supported for this connection type")
 
-    def create_function(self, definition: UdfDefinition) -> FunctionVersion:
+    def create_function(
+        self,
+        definition: UdfDefinition,
+        *,
+        secrets: Optional[Mapping[str, str]] = None,
+    ) -> FunctionVersion:
         """Register a scalar Python UDF and wait for its immutable version.
 
+        ``secrets`` must contain exactly the names declared by
+        ``@udf(secrets=[...])``. Values are sent in the create request and
+        stored server-side in the private execution artifact; returned
+        Function and Job metadata contain only the declared names.
         This is the blocking counterpart of :meth:`create_function_async`.
         Local connections raise ``NotImplementedError``.
         """
-        return self.create_function_async(definition).wait()
+        return self.create_function_async(definition, secrets=secrets).wait()
 
-    def create_function_async(self, definition: UdfDefinition) -> Job[FunctionVersion]:
+    def create_function_async(
+        self,
+        definition: UdfDefinition,
+        *,
+        secrets: Optional[Mapping[str, str]] = None,
+    ) -> Job[FunctionVersion]:
         """Register a scalar Python UDF through the remote Function catalog.
 
+        ``secrets`` must contain exactly the names declared by
+        ``@udf(secrets=[...])``. Values are sent in the create request and
+        stored server-side in the private execution artifact; returned
+        Function and Job metadata contain only the declared names.
         Submission returns a typed job. The immutable Function version becomes
         available only when :meth:`Job.wait` succeeds. Local connections raise
         ``NotImplementedError``.
@@ -1405,8 +1424,13 @@ class LanceDBConnection(DBConnection):
         return Job(self._conn.job(job_id))
 
     @override
-    def create_function_async(self, definition: UdfDefinition) -> Job[FunctionVersion]:
-        job = LOOP.run(self._conn.create_function_async(definition))
+    def create_function_async(
+        self,
+        definition: UdfDefinition,
+        *,
+        secrets: Optional[Mapping[str, str]] = None,
+    ) -> Job[FunctionVersion]:
+        job = LOOP.run(self._conn.create_function_async(definition, secrets=secrets))
         return Job(job)
 
     @override
@@ -2225,17 +2249,24 @@ class AsyncConnection(object):
         return AsyncJob(self._inner.job(job_id))
 
     async def create_function_async(
-        self, definition: UdfDefinition
+        self,
+        definition: UdfDefinition,
+        *,
+        secrets: Optional[Mapping[str, str]] = None,
     ) -> AsyncJob[FunctionVersion]:
         """Register a scalar Python UDF through the remote Function catalog.
 
+        ``secrets`` must contain exactly the names declared by
+        ``@udf(secrets=[...])``. Values are sent in the create request and
+        stored server-side in the private execution artifact; returned
+        Function and Job metadata contain only the declared names.
         The returned typed job resolves to the immutable Function version.
         Local connections raise ``NotImplementedError``.
         """
         if not isinstance(definition, UdfDefinition):
             raise TypeError("create_function_async requires a @udf definition")
         inner = await self._inner.create_function_async(
-            definition.registration_request.to_canonical_json()
+            definition._submission_json(secrets)
         )
         return _typed_job(inner, FunctionVersion.from_json)
 

python/python/lancedb/functions.py

@@ -4,7 +4,7 @@
 """Canonical Function values exchanged with LanceDB Enterprise services.
 
 These immutable models contain client/wire state only. Catalog persistence,
-environment bake, and execution are owned by Sophon.
+environment bake, secret resolution, and execution are owned by Sophon.
 ``RefreshColumnResult`` is also the backend-neutral result of a local
 expression-backed refresh job.
 """
@@ -229,7 +229,7 @@ class PythonEnvironmentSpec(_RemoteValue):
 
 
 class PythonRuntimeSpec(_RemoteValue):
-    """Remote runtime definition with environment values.
+    """Remote runtime definition with non-secret environment values.
 
     V1 supports ``kind="python"``. Newer runtime kinds remain readable, while
     their unknown payload fields are intentionally not retained by the client.
@@ -268,6 +268,7 @@ class FunctionVersion(_RemoteValue):
     runtime: PythonRuntimeSpec
     runtime_digest: str
     environment_digest: str
+    required_secrets: tuple[str, ...] = ()
     created_at: str
 
     def __call__(self, **inputs: Any) -> FunctionApplication:
@@ -329,12 +330,17 @@ class FunctionVersion(_RemoteValue):
 
 
 class FunctionRegistrationRequest(_RemoteValue):
-    """Stable remote registration envelope produced by :func:`udf`."""
+    """Stable remote registration envelope produced by :func:`udf`.
+
+    Only secret names are represented. Secret values are supplied separately
+    when the definition is submitted and are not part of this durable value.
+    """
 
     name: str
     artifact: FunctionArtifactRequest
     signature: FunctionSignature
     runtime: PythonRuntimeSpec
+    required_secrets: tuple[str, ...] = ()
 
 
 class FunctionVersionRef(_OpenRemoteValue):
@@ -479,6 +485,27 @@ class RefreshColumnResult(_RemoteValue):
 
 
 _FUNCTION_NAME = re.compile(r"^[A-Za-z_][A-Za-z0-9_.-]*$")
+_SECRET_NAME = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*$")
+# Keep this byte limit aligned with Sophon's MAX_FUNCTION_SECRET_VALUE_BYTES.
+_MAX_FUNCTION_SECRET_VALUE_BYTES = 64 * 1024
+_MAX_FUNCTION_SECRET_VALUES_BYTES = 512 * 1024
+
+
+def _validate_secret_value(name: str, value: Any) -> str:
+    """Validate one secret value before building the create request."""
+    if not isinstance(value, str):
+        raise TypeError(f"Function secret {name!r} value must be a string")
+    if not value:
+        raise ValueError(f"Function secret {name!r} value must be non-empty")
+    if "\0" in value:
+        raise ValueError(f"Function secret {name!r} value must not contain NUL")
+    value_bytes = len(value.encode("utf-8"))
+    if value_bytes > _MAX_FUNCTION_SECRET_VALUE_BYTES:
+        raise ValueError(
+            f"Function secret {name!r} value exceeds the "
+            f"{_MAX_FUNCTION_SECRET_VALUE_BYTES}-byte limit"
+        )
+    return value
 
 
 _GRAMMAR_PRIMITIVES = (
@@ -909,6 +936,7 @@ class UdfDefinition:
         output_schema: Optional[pa.DataType | pa.Field | pa.Schema],
         pip: tuple[str, ...],
         env: Mapping[str, str],
+        secrets: tuple[str, ...],
         python_version: Optional[str],
         conda: tuple[str, ...] = (),
         conda_channels: tuple[str, ...] = (),
@@ -935,6 +963,17 @@ class UdfDefinition:
             for key, value in environment.items()
         ):
             raise TypeError("Function env keys and values must be strings")
+        required_secrets = tuple(sorted(set(secrets)))
+        invalid_secrets = [
+            secret for secret in required_secrets if not _SECRET_NAME.fullmatch(secret)
+        ]
+        if invalid_secrets:
+            raise ValueError(f"invalid Function secret names: {invalid_secrets!r}")
+        overlap = set(environment) & set(required_secrets)
+        if overlap:
+            raise ValueError(
+                f"Function env and secret names must be disjoint: {sorted(overlap)!r}"
+            )
         signature = _infer_signature(function, input_schema, output_schema)
         source = _package_source(function)
         digest = f"sha256:{hashlib.sha256(source).hexdigest()}"
@@ -963,14 +1002,65 @@ class UdfDefinition:
             ),
             signature=signature,
             runtime=runtime,
+            required_secrets=required_secrets,
         )
         functools.update_wrapper(self, function)
 
     @property
     def registration_request(self) -> FunctionRegistrationRequest:
-        """The immutable request sent by ``create_function_async``."""
+        """The immutable, value-free client model for a Function submission."""
         return self._request
 
+    def _submission_json(self, secrets: Optional[Mapping[str, str]]) -> str:
+        """Build one registration submission without retaining values on self."""
+        if secrets is None:
+            secret_values: Mapping[str, str] = {}
+        elif not isinstance(secrets, Mapping):
+            raise TypeError("Function secrets must be a mapping of names to strings")
+        else:
+            secret_values = secrets
+
+        if any(not isinstance(name, str) for name in secret_values):
+            raise TypeError("Function secret names must be strings")
+        expected = set(self._request.required_secrets)
+        provided = set(secret_values)
+        if provided != expected:
+            missing = sorted(expected - provided)
+            unexpected = sorted(provided - expected)
+            details = []
+            if missing:
+                details.append(f"missing: {missing!r}")
+            if unexpected:
+                details.append(f"unexpected: {unexpected!r}")
+            raise ValueError(
+                "Function secret values must exactly match the declared secrets ("
+                + "; ".join(details)
+                + ")"
+            )
+
+        canonical_values = {}
+        total_bytes = 0
+        for name in sorted(secret_values):
+            value = _validate_secret_value(name, secret_values[name])
+            total_bytes += len(value.encode("utf-8"))
+            if total_bytes > _MAX_FUNCTION_SECRET_VALUES_BYTES:
+                raise ValueError(
+                    "Function secret values exceed the "
+                    f"{_MAX_FUNCTION_SECRET_VALUES_BYTES}-byte request limit"
+                )
+            canonical_values[name] = value
+
+        submission = self._request._known_dict()
+        if canonical_values:
+            submission["secret_values"] = canonical_values
+        return json.dumps(
+            submission,
+            ensure_ascii=False,
+            allow_nan=False,
+            sort_keys=True,
+            separators=(",", ":"),
+        )
+
     def __call__(self, *args, **kwargs):
         return self._function(*args, **kwargs)
 
@@ -988,6 +1078,7 @@ def udf(
     output_schema: Optional[pa.DataType | pa.Field | pa.Schema] = None,
     pip: tuple[str, ...] | list[str] = (),
     env: Optional[Mapping[str, str]] = None,
+    secrets: tuple[str, ...] | list[str] = (),
     python_version: Optional[str] = None,
     conda: tuple[str, ...] | list[str] = (),
     conda_channels: tuple[str, ...] | list[str] = (),
@@ -1002,6 +1093,7 @@ def udf(
     output_schema: Optional[pa.DataType | pa.Field | pa.Schema] = None,
     pip: tuple[str, ...] | list[str] = (),
     env: Optional[Mapping[str, str]] = None,
+    secrets: tuple[str, ...] | list[str] = (),
     python_version: Optional[str] = None,
     conda: tuple[str, ...] | list[str] = (),
     conda_channels: tuple[str, ...] | list[str] = (),
@@ -1032,7 +1124,10 @@ def udf(
     conda_channels : sequence of str, optional
         Conda channels in priority order; requires ``conda``.
     env : mapping of str to str, optional
-        Environment variables included in the Function definition.
+        Non-secret environment variables. Use ``secrets`` for credentials.
+    secrets : sequence of str, optional
+        Names of secrets required by the callable. Supply their values separately
+        to ``create_function`` or ``create_function_async``.
     python_version : str, optional
         Remote Python major/minor version. Defaults to the client version.
 
@@ -1054,11 +1149,15 @@ def udf(
     Examples
     --------
     >>> from lancedb import udf
-    >>> @udf(pip=["numpy==2.2.0"])
+    >>> @udf(pip=["numpy==2.2.0"], secrets=["MODEL_TOKEN"])
     ... def score(value: float) -> float:
     ...     return value * 2
     >>> score(1.5)
     3.0
+    >>> db.create_function(  # doctest: +SKIP
+    ...     score, secrets={"MODEL_TOKEN": "user-secret-value"}
+    ... )
+
     """
 
     def decorate(target: Callable[..., Any]) -> UdfDefinition:
@@ -1069,6 +1168,7 @@ def udf(
             output_schema=output_schema,
             pip=tuple(pip),
             env={} if env is None else env,
+            secrets=tuple(secrets),
             python_version=python_version,
             conda=tuple(conda),
             conda_channels=tuple(conda_channels),

python/python/lancedb/remote/db.py

@@ -7,7 +7,7 @@ import json
 import logging
 from concurrent.futures import ThreadPoolExecutor
 import sys
-from typing import TYPE_CHECKING, Any, Dict, Iterable, List, Optional, Union
+from typing import TYPE_CHECKING, Any, Dict, Iterable, List, Mapping, Optional, Union
 from urllib.parse import urlparse
 import warnings
 
@@ -742,8 +742,15 @@ class RemoteDBConnection(DBConnection):
         return Job(self._conn.job(job_id))
 
     @override
-    def create_function_async(self, definition: UdfDefinition) -> Job[FunctionVersion]:
-        return Job(LOOP.run(self._conn.create_function_async(definition)))
+    def create_function_async(
+        self,
+        definition: UdfDefinition,
+        *,
+        secrets: Optional[Mapping[str, str]] = None,
+    ) -> Job[FunctionVersion]:
+        return Job(
+            LOOP.run(self._conn.create_function_async(definition, secrets=secrets))
+        )
 
     @override
     def get_function(self, name: str, *, version: str) -> FunctionVersion:

python/python/lancedb/schema.py

@@ -4,30 +4,34 @@
 
 """Schema helpers for Lance blob columns."""
 
+import importlib
+from typing import TYPE_CHECKING
+
 import pyarrow as pa
+import pyarrow.ipc
+
+if TYPE_CHECKING:
+    from lance.blob import BlobType as BlobType
 
 _BLOB_EXTENSION_NAME = "lance.blob.v2"
 _BLOB_V1_KEY = "lance-encoding:blob"
 _ARROW_EXT_NAME_KEY = "ARROW:extension:name"
+_BLOB_V2_STORAGE_TYPE = pa.struct(
+    [
+        pa.field("data", pa.large_binary(), nullable=True),
+        pa.field("uri", pa.utf8(), nullable=True),
+        pa.field("position", pa.uint64(), nullable=True),
+        pa.field("size", pa.uint64(), nullable=True),
+    ]
+)
+_resolved_blob_type = None
 
 
-class BlobType(pa.ExtensionType):
-    """PyArrow extension type for a Lance blob v2 column.
-
-    Queries return descriptors; call :meth:`~lancedb.table.Table.fetch_blob_files`
-    for lazy reads or :meth:`~lancedb.table.Table.fetch_blobs` for eager bytes.
-    """
+class _FallbackBlobType(pa.ExtensionType):
+    """lance.blob.v2 extension type used when pylance is not installed."""
 
     def __init__(self) -> None:
-        storage_type = pa.struct(
-            [
-                pa.field("data", pa.large_binary(), nullable=True),
-                pa.field("uri", pa.utf8(), nullable=True),
-                pa.field("position", pa.uint64(), nullable=True),
-                pa.field("size", pa.uint64(), nullable=True),
-            ]
-        )
-        super().__init__(storage_type, _BLOB_EXTENSION_NAME)
+        pa.ExtensionType.__init__(self, _BLOB_V2_STORAGE_TYPE, _BLOB_EXTENSION_NAME)
 
     def __arrow_ext_serialize__(self) -> bytes:
         return b""
@@ -35,23 +39,16 @@ class BlobType(pa.ExtensionType):
     @classmethod
     def __arrow_ext_deserialize__(
         cls, storage_type: pa.DataType, serialized: bytes
-    ) -> "BlobType":
+    ) -> "_FallbackBlobType":
         return cls()
 
     def __reduce__(self):
-        # Ensure pickle round-trips on older pyarrow (apache/arrow#35599).
         return type(self).__arrow_ext_deserialize__, (
             self.storage_type,
             self.__arrow_ext_serialize__(),
         )
 
 
-try:
-    pa.register_extension_type(BlobType())  # type: ignore[arg-type]
-except pa.ArrowKeyError:
-    pass
-
-
 def _metadata_value(metadata: dict, key: str):
     return metadata.get(key.encode()) or metadata.get(key)
 
@@ -92,43 +89,105 @@ def is_blob_like_field(field: pa.Field) -> bool:
     return is_blob_v2_field(field) or _metadata_marks_legacy_blob(field.metadata or {})
 
 
-def _collect_blob_paths(schema: pa.Schema, is_blob) -> list[str]:
-    paths: list[str] = []
+def _collect_blob_paths(schema: pa.Schema, is_blob) -> list[tuple[str, bool]]:
+    """Walk the schema and return (path, has_list_ancestor) for each blob field."""
+    paths: list[tuple[str, bool]] = []
 
-    def walk(fields, prefix: str) -> None:
+    def walk(fields, prefix: str, has_list_ancestor: bool) -> None:
         for field in fields:
             path = f"{prefix}.{field.name}" if prefix else field.name
             if is_blob(field):
-                paths.append(path)
+                paths.append((path, has_list_ancestor))
             elif pa.types.is_struct(field.type):
-                walk(field.type, path)
+                walk(field.type, path, has_list_ancestor)
             elif (
                 pa.types.is_list(field.type)
                 or pa.types.is_large_list(field.type)
                 or pa.types.is_fixed_size_list(field.type)
             ):
-                walk([field.type.value_field], path)
+                walk([field.type.value_field], path, True)
 
-    walk(schema, "")
+    walk(schema, "", False)
     return paths
 
 
 def blob_column_paths(schema: pa.Schema) -> list[str]:
     """Dotted paths of blob-like columns (v2 extension or legacy metadata)."""
-    return _collect_blob_paths(schema, is_blob_like_field)
+    return [path for path, _ in _collect_blob_paths(schema, is_blob_like_field)]
 
 
 def blob_v2_column_paths(schema: pa.Schema) -> list[str]:
-    return _collect_blob_paths(schema, is_blob_v2_field)
+    return [path for path, _ in _collect_blob_paths(schema, is_blob_v2_field)]
+
+
+def row_addressable_blob_v2_paths(schema: pa.Schema) -> list[str]:
+    """Blob v2 paths with one blob addressable by table row id.
+
+    ``fetch_blobs`` and the descriptor row-id ride-along address one blob per
+    row, so a blob inside a list container has no row-id slot and no fetch
+    path. Those columns still store and query as raw descriptors.
+    """
+    return [
+        path
+        for path, has_list_ancestor in _collect_blob_paths(schema, is_blob_v2_field)
+        if not has_list_ancestor
+    ]
 
 
 def schema_has_blob_field(schema: pa.Schema) -> bool:
     return bool(blob_column_paths(schema))
 
 
+def _deserialize_registered_type(extension_type: pa.ExtensionType) -> pa.DataType:
+    """Return the type Arrow reconstructs for this extension name."""
+    schema = pa.schema([pa.field("value", extension_type)])
+    restored = pa.ipc.read_schema(schema.serialize())
+    return restored.field("value").type
+
+
+def _resolve_blob_type():
+    """Return the BlobType class this process should use.
+
+    pylance's class when it owns the lance.blob.v2 registry entry,
+    otherwise LanceDB's fallback. A different registered class is an error.
+    """
+    global _resolved_blob_type
+    if _resolved_blob_type is not None:
+        return _resolved_blob_type
+    try:
+        blob_module = importlib.import_module("lance.blob")
+    except ModuleNotFoundError as err:
+        if err.name not in ("lance", "lance.blob"):
+            raise
+    else:
+        blob_type = getattr(blob_module, "BlobType", None)
+        if blob_type is not None:
+            registered_type = _deserialize_registered_type(blob_type())
+            if type(registered_type) is not blob_type:
+                registered_cls = type(registered_type)
+                raise ValueError(
+                    "lance.blob.v2 is already registered by "
+                    f"{registered_cls.__module__}.{registered_cls.__qualname__}"
+                )
+            _resolved_blob_type = blob_type
+            return blob_type
+    try:
+        pa.register_extension_type(_FallbackBlobType())  # type: ignore[arg-type]
+    except pa.ArrowKeyError as err:
+        raise ValueError(
+            "lance.blob.v2 is already registered by another extension class"
+        ) from err
+    _resolved_blob_type = _FallbackBlobType
+    return _resolved_blob_type
+
+
 def blob(name: str, nullable: bool = True) -> pa.Field:
-    """Create a Lance blob v2 column field."""
-    return pa.field(name, BlobType(), nullable=nullable)
+    """Create a Lance blob v2 column field.
+
+    When pylance is installed this is ``lance.blob.BlobType``.
+    """
+    blob_type = _resolve_blob_type()
+    return pa.field(name, blob_type(), nullable=nullable)
 
 
 def vector(dimension: int, value_type: pa.DataType = pa.float32()) -> pa.DataType:
@@ -155,3 +214,11 @@ def vector(dimension: int, value_type: pa.DataType = pa.float32()) -> pa.DataTyp
     ... ])
     """
     return pa.list_(value_type, dimension)
+
+
+def __getattr__(name: str):
+    if name == "BlobType":
+        blob_type = _resolve_blob_type()
+        globals()["BlobType"] = blob_type
+        return blob_type
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

python/python/lancedb/table.py

@@ -104,7 +104,12 @@ from .util import (
     value_to_sql,
 )
 from .index import lang_mapping
-from .schema import blob_v2_column_paths, schema_has_blob_field
+from .schema import (
+    blob_v2_column_paths,
+    is_blob_v2_field,
+    row_addressable_blob_v2_paths,
+    schema_has_blob_field,
+)
 
 
 def _should_push_down_query_table(
@@ -426,6 +431,7 @@ def _cast_to_target_schema(
 
     def gen():
         for batch in reader:
+            batch = _coerce_blob_write_columns(batch, reordered_schema)
             # Table but not RecordBatch has cast.
             cast_batches = (
                 pa.Table.from_batches([batch]).cast(reordered_schema).to_batches()
@@ -438,6 +444,166 @@ def _cast_to_target_schema(
     return pa.RecordBatchReader.from_batches(reordered_schema, gen())
 
 
+def _coerce_blob_write_columns(
+    batch: pa.RecordBatch, target_schema: pa.Schema
+) -> pa.RecordBatch:
+    """Materialize blob storage structs before the stream leaves Python.
+
+    merge_insert requires its source reader to already match the table's
+    physical schema. Unlike add and insert, it does not pass through
+    LanceDB's Rust blob coercion, so preserving binary input here would
+    reach Lance as binary and fail the schema check.
+    """
+    columns = []
+    fields = []
+    changed = False
+    for field, column in zip(batch.schema, batch.columns):
+        target_field = target_schema.field(field.name)
+        coerced = _coerce_blob_value(column, target_field)
+        if coerced is not column:
+            column = coerced
+            field = pa.field(
+                field.name,
+                coerced.type,
+                field.nullable,
+                target_field.metadata,
+            )
+            changed = True
+        columns.append(column)
+        fields.append(field)
+    if not changed:
+        return batch
+    return pa.RecordBatch.from_arrays(
+        columns, schema=pa.schema(fields, metadata=batch.schema.metadata)
+    )
+
+
+def _coerce_blob_value(column: pa.Array, target_field: pa.Field) -> pa.Array:
+    if is_blob_v2_field(target_field) and _can_coerce_to_blob(column.type):
+        return _coerce_value_to_blob(column, target_field)
+
+    target_type = target_field.type
+    if pa.types.is_struct(target_type) and pa.types.is_struct(column.type):
+        children = []
+        fields = []
+        changed = False
+        for source_field in column.type:
+            source_column = column.field(source_field.name)
+            nested_target = next(
+                (field for field in target_type if field.name == source_field.name),
+                None,
+            )
+            if nested_target is None:
+                children.append(source_column)
+                fields.append(source_field)
+                continue
+            coerced = _coerce_blob_value(source_column, nested_target)
+            if coerced is not source_column:
+                changed = True
+            child_array, child_type = _physical_array_and_type(coerced)
+            children.append(child_array)
+            fields.append(
+                pa.field(
+                    source_field.name,
+                    child_type,
+                    source_field.nullable,
+                    nested_target.metadata,
+                )
+            )
+        if not changed:
+            return column
+        return pa.StructArray.from_arrays(
+            children,
+            fields=fields,
+            mask=column.is_null() if column.null_count else None,
+        )
+
+    if _is_list_like(target_type) and _is_list_like(column.type):
+        return _coerce_blob_list_values(column, target_type.value_field)
+
+    return column
+
+
+def _coerce_blob_list_values(
+    column: pa.Array, target_value_field: pa.Field
+) -> pa.Array:
+    """Coerce blob values inside a list column, preserving offsets and nulls.
+
+    Works on the raw child values window instead of ``pc.list_flatten`` because
+    flatten drops values spanned by null slots, which would misalign offsets.
+    """
+    mask = column.is_null() if column.null_count else None
+    if pa.types.is_fixed_size_list(column.type):
+        list_size = column.type.list_size
+        values = column.values.slice(column.offset * list_size, len(column) * list_size)
+        coerced = _coerce_blob_value(values, target_value_field)
+        if coerced is values:
+            return column
+        physical_values, _ = _physical_array_and_type(coerced)
+        return pa.FixedSizeListArray.from_arrays(physical_values, list_size, mask=mask)
+    offsets = column.offsets
+    first_offset = offsets[0].as_py()
+    values = column.values.slice(
+        first_offset,
+        offsets[-1].as_py() - first_offset,
+    )
+    coerced = _coerce_blob_value(values, target_value_field)
+    if coerced is values:
+        return column
+    physical_values, _ = _physical_array_and_type(coerced)
+    if first_offset:
+        offsets = pc.subtract(offsets, pa.scalar(first_offset, offsets.type))
+    if pa.types.is_large_list(column.type):
+        return pa.LargeListArray.from_arrays(offsets, physical_values, mask=mask)
+    return pa.ListArray.from_arrays(offsets, physical_values, mask=mask)
+
+
+def _coerce_value_to_blob(values: pa.Array, target_field: pa.Field) -> pa.Array:
+    if pa.types.is_null(values.type):
+        data = pa.nulls(len(values), type=pa.large_binary())
+    elif pa.types.is_large_binary(values.type):
+        data = values
+    else:
+        data = values.cast(pa.large_binary())
+    length = len(values)
+    storage_type = target_field.type
+    if isinstance(storage_type, pa.ExtensionType):
+        storage_type = storage_type.storage_type
+    storage_fields = list(storage_type)
+    children = []
+    for storage_field in storage_fields:
+        if storage_field.name == "data":
+            children.append(data)
+        else:
+            children.append(pa.nulls(length, type=storage_field.type))
+    storage = pa.StructArray.from_arrays(
+        children,
+        fields=storage_fields,
+        mask=values.is_null() if values.null_count else None,
+    )
+    if isinstance(target_field.type, pa.ExtensionType):
+        return pa.ExtensionArray.from_storage(target_field.type, storage)
+    return storage
+
+
+def _physical_array_and_type(array: pa.Array) -> tuple[pa.Array, pa.DataType]:
+    if isinstance(array.type, pa.ExtensionType):
+        return array.storage, array.type.storage_type
+    return array, array.type
+
+
+def _can_coerce_to_blob(data_type: pa.DataType) -> bool:
+    return _is_binary_like(data_type) or pa.types.is_null(data_type)
+
+
+def _is_binary_like(data_type: pa.DataType) -> bool:
+    return (
+        pa.types.is_binary(data_type)
+        or pa.types.is_large_binary(data_type)
+        or pa.types.is_binary_view(data_type)
+    )
+
+
 def _field_extension_name(field: pa.Field) -> Optional[str]:
     extension_name = getattr(field.type, "extension_name", None)
     if extension_name is not None:
@@ -464,63 +630,71 @@ def _align_field_types(
         target_field = next((f for f in target_fields if f.name == field.name), None)
         if target_field is None:
             raise ValueError(f"Field '{field.name}' not found in target schema")
-        # Preserve arrow.json input until it reaches Lance. LanceDB exposes stored
-        # JSON columns as lance.json (JSONB-backed LargeBinary), but casting the
-        # input to that storage type here merely relabels the raw JSON bytes as
-        # JSONB. Lance must see arrow.json so it can perform the JSONB encoding.
-        if (
-            _field_extension_name(field) == "arrow.json"
-            and _field_extension_name(target_field) == "lance.json"
-        ):
-            new_fields.append(field)
-            continue
-        if pa.types.is_struct(target_field.type):
-            if pa.types.is_struct(field.type):
-                new_type = pa.struct(
-                    _align_field_types(
-                        field.type.fields,
-                        target_field.type.fields,
-                    )
+        new_fields.append(_align_field(field, target_field))
+    return new_fields
+
+
+def _align_list_value_field(
+    value_field: pa.Field, target_value_field: pa.Field
+) -> pa.Field:
+    # A list has exactly one child, so the inferred child name ("item") aligns
+    # positionally and adopts the table's child name; pa.Table.cast renames it.
+    return _align_field(value_field, target_value_field).with_name(
+        target_value_field.name
+    )
+
+
+def _align_field(field: pa.Field, target_field: pa.Field) -> pa.Field:
+    # Preserve arrow.json input until it reaches Lance. LanceDB exposes stored
+    # JSON columns as lance.json (JSONB-backed LargeBinary), but casting the
+    # input to that storage type here merely relabels the raw JSON bytes as
+    # JSONB. Lance must see arrow.json so it can perform the JSONB encoding.
+    if (
+        _field_extension_name(field) == "arrow.json"
+        and _field_extension_name(target_field) == "lance.json"
+    ):
+        return field
+    if pa.types.is_struct(target_field.type):
+        if pa.types.is_struct(field.type):
+            new_type = pa.struct(
+                _align_field_types(
+                    field.type.fields,
+                    target_field.type.fields,
                 )
-            else:
-                new_type = target_field.type
-        elif pa.types.is_list(target_field.type):
-            if _is_list_like(field.type):
-                new_type = pa.list_(
-                    _align_field_types(
-                        [field.type.value_field],
-                        [target_field.type.value_field],
-                    )[0]
-                )
-            else:
-                new_type = target_field.type
-        elif pa.types.is_large_list(target_field.type):
-            if _is_list_like(field.type):
-                new_type = pa.large_list(
-                    _align_field_types(
-                        [field.type.value_field],
-                        [target_field.type.value_field],
-                    )[0]
-                )
-            else:
-                new_type = target_field.type
-        elif pa.types.is_fixed_size_list(target_field.type):
-            if _is_list_like(field.type):
-                new_type = pa.list_(
-                    _align_field_types(
-                        [field.type.value_field],
-                        [target_field.type.value_field],
-                    )[0],
-                    target_field.type.list_size,
-                )
-            else:
-                new_type = target_field.type
+            )
         else:
             new_type = target_field.type
-        new_fields.append(
-            pa.field(field.name, new_type, field.nullable, target_field.metadata)
-        )
-    return new_fields
+    elif pa.types.is_list(target_field.type):
+        if _is_list_like(field.type):
+            new_type = pa.list_(
+                _align_list_value_field(
+                    field.type.value_field, target_field.type.value_field
+                )
+            )
+        else:
+            new_type = target_field.type
+    elif pa.types.is_large_list(target_field.type):
+        if _is_list_like(field.type):
+            new_type = pa.large_list(
+                _align_list_value_field(
+                    field.type.value_field, target_field.type.value_field
+                )
+            )
+        else:
+            new_type = target_field.type
+    elif pa.types.is_fixed_size_list(target_field.type):
+        if _is_list_like(field.type):
+            new_type = pa.list_(
+                _align_list_value_field(
+                    field.type.value_field, target_field.type.value_field
+                ),
+                target_field.type.list_size,
+            )
+        else:
+            new_type = target_field.type
+    else:
+        new_type = target_field.type
+    return pa.field(field.name, new_type, field.nullable, target_field.metadata)
 
 
 def _infer_subschema(
@@ -589,7 +763,7 @@ def sanitize_create_table(
         schema = data.schema
     else:
         if schema is not None:
-            data = pa.Table.from_pylist([], schema)
+            data = pa.Table.from_batches([], schema=schema)
     if schema is None:
         if data is None:
             raise ValueError("Either data or schema must be provided")
@@ -2698,7 +2872,7 @@ class LanceTable(Table):
             arrow_tbl = self.to_arrow()
             if blob_mode == "descriptions":
                 arrow_tbl = strip_auto_row_ids(
-                    arrow_tbl, blob_v2_column_paths(self.schema)
+                    arrow_tbl, row_addressable_blob_v2_paths(self.schema)
                 )
             return arrow_tbl.to_pandas(**kwargs)
 
@@ -5102,7 +5276,9 @@ class AsyncTable:
         if blob_mode == "descriptions" or not schema_has_blob_field(schema):
             arrow_tbl = await self.to_arrow()
             if blob_mode == "descriptions":
-                arrow_tbl = strip_auto_row_ids(arrow_tbl, blob_v2_column_paths(schema))
+                arrow_tbl = strip_auto_row_ids(
+                    arrow_tbl, row_addressable_blob_v2_paths(schema)
+                )
             return arrow_tbl.to_pandas(**kwargs)
 
         if blob_mode == "lazy" and get_uri_scheme(await self.uri()) == "memory":

python/python/tests/test_blob.py

@@ -2,10 +2,15 @@
 # SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
 import io
+import subprocess
+import sys
+import textwrap
 
+import lance
 import pyarrow as pa
 import pyarrow.compute as pc
 import pytest
+from lance.blob import BlobType as LanceBlobType
 
 import lancedb
 from lancedb._blob import (
@@ -18,6 +23,20 @@ from lancedb.index import FTS
 from lancedb.schema import blob_column_paths, blob_v2_column_paths
 
 
+_HIDE_LANCE_BLOB = """\
+import importlib.abc
+import sys
+
+class _MissingLanceBlob(importlib.abc.MetaPathFinder):
+    def find_spec(self, fullname, path, target=None):
+        if fullname == "lance.blob" or fullname.startswith("lance.blob."):
+            raise ModuleNotFoundError(fullname, name="lance.blob")
+
+sys.modules.pop("lance.blob", None)
+sys.meta_path.insert(0, _MissingLanceBlob())
+"""
+
+
 def _blob_table(name, rows):
     db = lancedb.connect("memory:///")
     schema = pa.schema([pa.field("id", pa.int64()), lancedb.blob("image")])
@@ -51,6 +70,181 @@ def test_blob_factory_declares_v2_field():
     field = lancedb.blob("image")
     assert isinstance(field.type, pa.ExtensionType)
     assert field.type.extension_name == "lance.blob.v2"
+    assert lancedb.BlobType is LanceBlobType
+    assert type(field.type) is LanceBlobType
+
+
+def test_blob_type_works_without_pylance():
+    script = _HIDE_LANCE_BLOB + textwrap.dedent(
+        """\
+        import lancedb
+        import pyarrow as pa
+
+        field = lancedb.blob("image")
+        if not isinstance(field.type, pa.ExtensionType):
+            raise SystemExit("expected an extension type")
+        if field.type.extension_name != "lance.blob.v2":
+            raise SystemExit(field.type.extension_name)
+        if lancedb.BlobType is not type(field.type):
+            raise SystemExit("BlobType is not the field type class")
+        if lancedb.BlobType.__module__ != "lancedb.schema":
+            raise SystemExit(lancedb.BlobType.__module__)
+
+        db = lancedb.connect("memory:///")
+        table = db.create_table(
+            "images",
+            schema=pa.schema([pa.field("id", pa.int64()), field]),
+        )
+        table.add([{"id": 1, "image": b"hello"}])
+        result = (
+            table.merge_insert("id")
+            .when_matched_update_all()
+            .when_not_matched_insert_all()
+            .execute([{"id": 1, "image": b"updated"}, {"id": 2, "image": b"inserted"}])
+        )
+        if result.num_updated_rows != 1 or result.num_inserted_rows != 1:
+            raise SystemExit(
+                f"merge_insert rows updated={result.num_updated_rows} "
+                f"inserted={result.num_inserted_rows}"
+            )
+        """
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode == 0, result.stderr
+
+
+def test_blob_resolves_pylance_type_without_eager_import():
+    script = textwrap.dedent(
+        """\
+        import sys
+        import lancedb
+
+        if "lance.blob" in sys.modules:
+            raise SystemExit("import lancedb imported lance.blob")
+        field = lancedb.blob("image")
+        from lance.blob import BlobType
+
+        if type(field.type) is not BlobType:
+            raise SystemExit(f"{type(field.type)} is not {BlobType}")
+        import lance
+
+        image = lance.blob_array([b"x"])
+        if type(image.type) is not BlobType:
+            raise SystemExit("blob_array used a different class")
+        if type(image.type) is not type(field.type):
+            raise SystemExit("field and array classes differ")
+        """
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode == 0, result.stderr
+
+
+def test_blob_fallback_fails_if_name_already_registered():
+    script = _HIDE_LANCE_BLOB + textwrap.dedent(
+        """\
+        import pyarrow as pa
+
+        class OtherBlobType(pa.ExtensionType):
+            def __init__(self):
+                super().__init__(
+                    pa.struct([pa.field("data", pa.large_binary())]),
+                    "lance.blob.v2",
+                )
+
+            def __arrow_ext_serialize__(self):
+                return b""
+
+            @classmethod
+            def __arrow_ext_deserialize__(cls, storage_type, serialized):
+                return cls()
+
+        pa.register_extension_type(OtherBlobType())
+        import lancedb
+
+        try:
+            lancedb.blob("image")
+        except ValueError as err:
+            if "already registered" not in str(err):
+                raise SystemExit(err)
+        else:
+            raise SystemExit("expected ValueError")
+        """
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode == 0, result.stderr
+
+
+def test_blob_type_rejects_competing_registration_with_pylance():
+    script = textwrap.dedent(
+        """\
+        import pyarrow as pa
+        import pyarrow.ipc
+
+        class OtherBlobType(pa.ExtensionType):
+            def __init__(self):
+                super().__init__(
+                    pa.struct(
+                        [
+                            pa.field("data", pa.large_binary()),
+                            pa.field("uri", pa.utf8()),
+                            pa.field("position", pa.uint64()),
+                            pa.field("size", pa.uint64()),
+                        ]
+                    ),
+                    "lance.blob.v2",
+                )
+
+            def __arrow_ext_serialize__(self):
+                return b""
+
+            @classmethod
+            def __arrow_ext_deserialize__(cls, storage_type, serialized):
+                return cls()
+
+        pa.register_extension_type(OtherBlobType())
+
+        from lance.blob import BlobType
+
+        if BlobType is OtherBlobType:
+            raise SystemExit("pylance BlobType was replaced")
+        schema = pa.schema([pa.field("value", BlobType())])
+        restored = pa.ipc.read_schema(schema.serialize())
+        if type(restored.field("value").type) is not OtherBlobType:
+            raise SystemExit(type(restored.field("value").type))
+
+        import lancedb
+
+        try:
+            lancedb.blob("image")
+        except ValueError as err:
+            if "__main__.OtherBlobType" not in str(err):
+                raise SystemExit(err)
+        else:
+            raise SystemExit("expected ValueError")
+        """
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode == 0, result.stderr
 
 
 def test_blob_v2_column_paths_include_list_children():
@@ -203,6 +397,292 @@ def test_fetch_blobs_round_trip():
     assert [blobs[0].as_py(), blobs[1].as_py()] == [b"alpha", b"beta"]
 
 
+def test_merge_insert_writes_python_bytes():
+    table = _blob_table("merge_bytes", [{"id": 1, "image": b"before"}])
+    result = (
+        table.merge_insert("id")
+        .when_matched_update_all()
+        .when_not_matched_insert_all()
+        .execute([{"id": 1, "image": b"updated"}, {"id": 2, "image": b"inserted"}])
+    )
+    assert result.num_updated_rows == 1
+    assert result.num_inserted_rows == 1
+    by_id = _row_ids_by_id(table)
+    blobs = table.fetch_blobs("image", [by_id[1], by_id[2]])
+    assert blobs.to_pylist() == [b"updated", b"inserted"]
+
+
+def test_merge_insert_bytes_after_reopen_without_touching_blob_type(tmp_path):
+    db = lancedb.connect(tmp_path)
+    schema = pa.schema([pa.field("id", pa.int64()), lancedb.blob("image")])
+    table = db.create_table("images", schema=schema)
+    table.add([{"id": 1, "image": b"hello"}])
+
+    script = textwrap.dedent(
+        f"""\
+        import lancedb
+
+        db = lancedb.connect({str(tmp_path)!r})
+        table = db.open_table("images")
+        image_type = table.schema.field("image").type
+        if type(image_type).__name__ != "StructType":
+            raise SystemExit(f"expected StructType, got {{type(image_type)}}")
+        result = (
+            table.merge_insert("id")
+            .when_matched_update_all()
+            .when_not_matched_insert_all()
+            .execute(
+                [{{"id": 1, "image": b"updated"}}, {{"id": 2, "image": b"inserted"}}]
+            )
+        )
+        if result.num_updated_rows != 1 or result.num_inserted_rows != 1:
+            raise SystemExit(
+                f"rows updated={{result.num_updated_rows}} "
+                f"inserted={{result.num_inserted_rows}}"
+            )
+        hits = table.search().with_row_id(True).limit(10).to_arrow()
+        by_id = dict(zip(hits["id"].to_pylist(), hits["_rowid"].to_pylist()))
+        blobs = table.fetch_blobs("image", [by_id[1], by_id[2]])
+        if blobs.to_pylist() != [b"updated", b"inserted"]:
+            raise SystemExit(blobs.to_pylist())
+        """
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode == 0, result.stderr
+
+
+def test_merge_insert_bytes_after_reopen_without_pylance(tmp_path):
+    db = lancedb.connect(tmp_path)
+    schema = pa.schema([pa.field("id", pa.int64()), lancedb.blob("image")])
+    table = db.create_table("images", schema=schema)
+    table.add([{"id": 1, "image": b"hello"}])
+
+    script = _HIDE_LANCE_BLOB + textwrap.dedent(
+        f"""\
+        import lancedb
+
+        db = lancedb.connect({str(tmp_path)!r})
+        table = db.open_table("images")
+        image_type = table.schema.field("image").type
+        if type(image_type).__name__ != "StructType":
+            raise SystemExit(f"expected StructType, got {{type(image_type)}}")
+        result = (
+            table.merge_insert("id")
+            .when_matched_update_all()
+            .when_not_matched_insert_all()
+            .execute(
+                [{{"id": 1, "image": b"updated"}}, {{"id": 2, "image": b"inserted"}}]
+            )
+        )
+        if result.num_updated_rows != 1 or result.num_inserted_rows != 1:
+            raise SystemExit(
+                f"rows updated={{result.num_updated_rows}} "
+                f"inserted={{result.num_inserted_rows}}"
+            )
+        hits = table.search().with_row_id(True).limit(10).to_arrow()
+        by_id = dict(zip(hits["id"].to_pylist(), hits["_rowid"].to_pylist()))
+        blobs = table.fetch_blobs("image", [by_id[1], by_id[2]])
+        if blobs.to_pylist() != [b"updated", b"inserted"]:
+            raise SystemExit(blobs.to_pylist())
+        """
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode == 0, result.stderr
+
+
+def test_merge_insert_blob_array_into_reopened_unregistered_table(tmp_path):
+    db = lancedb.connect(tmp_path)
+    schema = pa.schema([pa.field("id", pa.int64()), lancedb.blob("image")])
+    table = db.create_table("images", schema=schema)
+    table.add([{"id": 1, "image": b"before"}])
+
+    script = textwrap.dedent(
+        f"""\
+        import pyarrow as pa
+        import lancedb
+
+        db = lancedb.connect({str(tmp_path)!r})
+        table = db.open_table("images")
+        image_type = table.schema.field("image").type
+        if type(image_type).__name__ != "StructType":
+            raise SystemExit(
+                f"expected StructType before lance import, got {{type(image_type)}}"
+            )
+
+        import lance
+
+        updates = pa.Table.from_arrays(
+            [
+                pa.array([1, 2], type=pa.int64()),
+                lance.blob_array([b"updated", b"inserted"]),
+            ],
+            names=["id", "image"],
+        )
+        result = (
+            table.merge_insert("id")
+            .when_matched_update_all()
+            .when_not_matched_insert_all()
+            .execute(updates)
+        )
+        if result.num_updated_rows != 1 or result.num_inserted_rows != 1:
+            raise SystemExit(
+                f"rows updated={{result.num_updated_rows}} "
+                f"inserted={{result.num_inserted_rows}}"
+            )
+        hits = table.search().with_row_id(True).limit(10).to_arrow()
+        by_id = dict(zip(hits["id"].to_pylist(), hits["_rowid"].to_pylist()))
+        blobs = table.fetch_blobs("image", [by_id[1], by_id[2]])
+        if blobs.to_pylist() != [b"updated", b"inserted"]:
+            raise SystemExit(blobs.to_pylist())
+        """
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode == 0, result.stderr
+
+
+def test_add_all_null_blob_column():
+    db = lancedb.connect("memory:///")
+    schema = pa.schema([pa.field("id", pa.int64()), lancedb.blob("image")])
+    table = db.create_table("all_null", schema=schema)
+    table.add([{"id": 1, "image": None}, {"id": 2, "image": None}])
+    by_id = _row_ids_by_id(table)
+    blobs = table.fetch_blobs("image", [by_id[1], by_id[2]])
+    assert blobs.to_pylist() == [None, None]
+
+
+def test_create_table_nested_blob_schema_without_rows():
+    db = lancedb.connect("memory:///")
+    schema = pa.schema(
+        [
+            pa.field("id", pa.int64()),
+            pa.field("info", pa.struct([lancedb.blob("blob")])),
+            pa.field("images", pa.list_(lancedb.blob("image"))),
+        ]
+    )
+    table = db.create_table("nested_empty", schema=schema)
+    assert table.count_rows() == 0
+
+
+def test_merge_insert_nested_blob_dicts():
+    db = lancedb.connect("memory:///")
+    info = pa.StructArray.from_arrays(
+        [
+            pa.array(["first"], type=pa.string()),
+            _blob_array("blob", [b"before"]),
+        ],
+        names=["name", "blob"],
+    )
+    data = pa.Table.from_arrays(
+        [pa.array([1], type=pa.int64()), info],
+        names=["id", "info"],
+    )
+    table = db.create_table("nested_merge", data=data)
+    result = (
+        table.merge_insert("id")
+        .when_matched_update_all()
+        .execute([{"id": 1, "info": {"name": "first", "blob": b"after"}}])
+    )
+    assert result.num_updated_rows == 1
+    by_id = _row_ids_by_id(table)
+    blobs = table.fetch_blobs("info.blob", [by_id[1]])
+    assert blobs.to_pylist() == [b"after"]
+
+
+def _list_blob_table(name):
+    db = lancedb.connect("memory:///")
+    blob_field = lancedb.blob("image")
+    images = pa.ListArray.from_arrays(
+        pa.array([0, 1], type=pa.int32()), _blob_array("image", [b"before"])
+    )
+    data = pa.Table.from_arrays(
+        [pa.array([1], type=pa.int64()), images],
+        schema=pa.schema(
+            [pa.field("id", pa.int64()), pa.field("images", pa.list_(blob_field))]
+        ),
+    )
+    return db.create_table(name, data=data)
+
+
+def test_merge_insert_list_blob_dicts():
+    table = _list_blob_table("list_merge")
+    result = (
+        table.merge_insert("id")
+        .when_matched_update_all()
+        .when_not_matched_insert_all()
+        .execute([{"id": 1, "images": [b"one", b"two"]}, {"id": 2, "images": None}])
+    )
+    assert result.num_updated_rows == 1
+    assert result.num_inserted_rows == 1
+    hits = table.search().limit(10).to_arrow()
+    sizes = {
+        row["id"]: None if row["images"] is None else [d["size"] for d in row["images"]]
+        for row in hits.to_pylist()
+    }
+    assert sizes == {1: [3, 3], 2: None}
+
+
+def test_list_blob_column_queries_as_raw_descriptors():
+    table = _list_blob_table("list_query")
+    hits = table.search().limit(10).to_arrow()
+    element = hits.schema.field("images").type.value_type
+    assert pa.types.is_struct(element)
+    assert "_lance_row_id" not in element.names
+    with pytest.raises(ValueError, match="expected struct before segment"):
+        table.fetch_blobs("images.image", [0])
+
+
+def test_row_addressable_paths_exclude_list_children():
+    from lancedb.schema import row_addressable_blob_v2_paths
+
+    schema = pa.schema(
+        [
+            pa.field("id", pa.int64()),
+            pa.field("info", pa.struct([lancedb.blob("blob")])),
+            pa.field("images", pa.list_(lancedb.blob("image"))),
+        ]
+    )
+    assert blob_v2_column_paths(schema) == ["info.blob", "images.image"]
+    assert row_addressable_blob_v2_paths(schema) == ["info.blob"]
+
+
+def test_merge_insert_writes_pylance_blob_array():
+    table = _blob_table("merge_pylance", [{"id": 1, "image": b"before"}])
+    image = lance.blob_array([b"updated", b"inserted"])
+    assert type(image.type) is LanceBlobType
+    assert type(image.type) is type(lancedb.BlobType())
+    updates = pa.Table.from_arrays(
+        [pa.array([1, 2], type=pa.int64()), image], names=["id", "image"]
+    )
+
+    result = (
+        table.merge_insert("id")
+        .when_matched_update_all()
+        .when_not_matched_insert_all()
+        .execute(updates)
+    )
+
+    assert result.num_updated_rows == 1
+    assert result.num_inserted_rows == 1
+    by_id = _row_ids_by_id(table)
+    blobs = table.fetch_blobs("image", [by_id[1], by_id[2]])
+    assert blobs.to_pylist() == [b"updated", b"inserted"]
+
+
 def test_fetch_blobs_accepts_query_result():
     table = _blob_table("from_result", [{"id": 1, "image": b"gamma"}])
     hits = table.search().limit(10).to_arrow()

python/python/tests/test_first_class_function_slice1.py

@@ -37,6 +37,21 @@ def job_result(name: str) -> dict:
     return json.loads(fixture(name))["result"]
 
 
+def assert_no_secret_values(value):
+    if isinstance(value, dict):
+        for key, child in value.items():
+            assert key not in {
+                "secret_value",
+                "secret_values",
+                "resolved_secret",
+                "resolved_secrets",
+            }
+            assert_no_secret_values(child)
+    elif isinstance(value, list):
+        for child in value:
+            assert_no_secret_values(child)
+
+
 def test_public_function_values_are_in_api_reference():
     docs = Path(__file__).parents[3] / "docs" / "src" / "python" / "python.md"
     rendered = docs.read_text()
@@ -94,6 +109,7 @@ def test_function_version_identity_is_immutable_and_exact():
     version = FunctionVersion.from_json(json.dumps(value))
     assert version.name == "embed"
     assert version.version == "fv_01K3EXACT"
+    assert version.required_secrets == ("HF_TOKEN",)
 
     with pytest.raises((TypeError, ValueError)):
         version.version = "fv_changed"
@@ -276,6 +292,15 @@ def test_refresh_result_rejects_non_u64_values(field):
         RefreshColumnResult.from_json(json.dumps(value))
 
 
+def test_canonical_client_values_contain_secret_names_only():
+    version = FunctionVersion.from_json(
+        json.dumps(job_result("remote_function_job.json"))
+    )
+    canonical = json.loads(version.to_canonical_json())
+    assert canonical["required_secrets"] == ["HF_TOKEN"]
+    assert_no_secret_values(canonical)
+
+
 class _FunctionDeclarationInner:
     def __init__(self):
         self.calls = []

python/python/tests/test_first_class_function_slice2.py

@@ -19,7 +19,13 @@ import pyarrow as pa
 import pytest
 
 import lancedb
-from lancedb.functions import UdfDefinition, udf
+from lancedb.functions import (
+    _MAX_FUNCTION_SECRET_VALUE_BYTES,
+    _MAX_FUNCTION_SECRET_VALUES_BYTES,
+    FunctionRegistrationRequest,
+    UdfDefinition,
+    udf,
+)
 
 THRESHOLD = 20
 _CACHE = None
@@ -39,12 +45,28 @@ FIXTURES = (
 @udf(
     pip=["numpy>=2"],
     env={"MODE": "test"},
+    secrets=["API_TOKEN"],
     python_version="3.12",
 )
 def normalize_score(value: float) -> float:
     return value / 100.0
 
 
+def _assert_no_secret_values(value):
+    if isinstance(value, dict):
+        for key, child in value.items():
+            assert key not in {
+                "secret_value",
+                "secret_values",
+                "resolved_secret",
+                "resolved_secrets",
+            }
+            _assert_no_secret_values(child)
+    elif isinstance(value, list):
+        for child in value:
+            _assert_no_secret_values(child)
+
+
 def test_scalar_udf_matches_shared_registration_golden_and_remains_callable():
     assert isinstance(normalize_score, UdfDefinition)
     assert normalize_score(25.0) == 0.25
@@ -59,6 +81,8 @@ def test_scalar_udf_matches_shared_registration_golden_and_remains_callable():
         "kind": "scalar_to_arrow_batch",
         "version": 1,
     }
+    assert request["required_secrets"] == ["API_TOKEN"]
+    _assert_no_secret_values(request)
 
 
 def _run_packaged(definition, *args):
@@ -372,6 +396,7 @@ def test_udf_recursion_versus_a_rebound_module_name(tmp_path):
         output_schema=None,
         pip=(),
         env={},
+        secrets=(),
         python_version=None,
     )
     with pytest.raises(ValueError, match="binds that name to another value"):
@@ -526,13 +551,54 @@ def test_annotation_and_explicit_schema_validation_fail_closed():
             return value
 
 
+def test_secret_names_are_canonical_and_disjoint_from_environment():
+    @udf(secrets=["Z_TOKEN", "A_TOKEN", "Z_TOKEN"])
+    def canonical_secrets(value: int) -> int:
+        return value
+
+    assert canonical_secrets.registration_request.required_secrets == (
+        "A_TOKEN",
+        "Z_TOKEN",
+    )
+
+    with pytest.raises(ValueError, match="must be disjoint"):
+
+        @udf(env={"TOKEN": "plaintext"}, secrets=["TOKEN"])
+        def overlapping(value: int) -> int:
+            return value
+
+
+def test_declared_secret_api_still_requires_explicit_create_values():
+    @udf(secrets=["API_TOKEN"])
+    def declared_secret(value: int) -> int:
+        return value
+
+    with pytest.raises(ValueError, match="missing"):
+        declared_secret._submission_json(None)
+    submission = json.loads(
+        declared_secret._submission_json({"API_TOKEN": "explicit-secret"})
+    )
+    assert submission["required_secrets"] == ["API_TOKEN"]
+    assert submission["secret_values"] == {"API_TOKEN": "explicit-secret"}
+
+
+def test_no_secrets_preserve_canonical_registration_shape():
+    @udf
+    def no_secrets(value: int) -> int:
+        return value
+
+    canonical = json.loads(no_secrets.registration_request.to_canonical_json())
+    assert "required_secrets" not in canonical
+    assert json.loads(no_secrets._submission_json(None)) == canonical
+
+
 def test_local_function_catalog_operations_are_not_supported(tmp_path):
     db = lancedb.connect(tmp_path)
     message = "Function catalog operations are not supported by this database"
     with pytest.raises(NotImplementedError, match=message):
-        db.create_function(normalize_score)
+        db.create_function(normalize_score, secrets={"API_TOKEN": "value"})
     with pytest.raises(NotImplementedError, match=message):
-        db.create_function_async(normalize_score)
+        db.create_function_async(normalize_score, secrets={"API_TOKEN": "value"})
     with pytest.raises(NotImplementedError, match=message):
         db.get_function("normalize_score", version="fv_exact")
 
@@ -562,6 +628,7 @@ def _mock_remote_function_catalog():
                     "runtime": body["runtime"],
                     "runtime_digest": "sha256:runtime",
                     "environment_digest": "sha256:environment",
+                    "required_secrets": body.get("required_secrets", []),
                     "created_at": "2026-08-21T00:00:00Z",
                 }
                 response = {"job_id": "job-register"}
@@ -608,7 +675,9 @@ def test_remote_registration_job_and_exact_version_reopen_round_trip():
             host_override=host,
             client_config={"retry_config": {"retries": 0}},
         )
-        registration = db.create_function_async(normalize_score)
+        registration = db.create_function_async(
+            normalize_score, secrets={"API_TOKEN": "secret-value"}
+        )
         assert registration.id == "job-register"
         created = registration.wait()
         reopened = db.get_function("normalize_score", version=created.version)
@@ -617,9 +686,18 @@ def test_remote_registration_job_and_exact_version_reopen_round_trip():
     assert reopened.name == "normalize_score"
     assert reopened.version == "fv_exact"
     create_request = state["requests"][0][1]
-    assert create_request == json.loads(
+    expected = json.loads(normalize_score.registration_request.to_canonical_json())
+    expected["secret_values"] = {"API_TOKEN": "secret-value"}
+    assert create_request == expected
+    durable_request = FunctionRegistrationRequest.from_json(json.dumps(create_request))
+    assert not hasattr(durable_request, "secret_values")
+    assert "secret_values" not in json.loads(durable_request.to_canonical_json())
+    assert "secret_values" not in json.loads(
         normalize_score.registration_request.to_canonical_json()
     )
+    assert "secret-value" not in repr(normalize_score)
+    assert "secret-value" not in repr(normalize_score.registration_request)
+    assert not hasattr(created, "secret_values")
 
 
 def test_blocking_remote_registration_returns_function_version():
@@ -630,7 +708,9 @@ def test_blocking_remote_registration_returns_function_version():
             host_override=host,
             client_config={"retry_config": {"retries": 0}},
         )
-        created = db.create_function(normalize_score)
+        created = db.create_function(
+            normalize_score, secrets={"API_TOKEN": "blocking-secret"}
+        )
 
     assert created.name == "normalize_score"
     assert created.version == "fv_exact"
@@ -638,3 +718,121 @@ def test_blocking_remote_registration_returns_function_version():
         "/v1/functions/create",
         "/v1/jobs/describe",
     ]
+    assert state["requests"][0][1]["secret_values"] == {"API_TOKEN": "blocking-secret"}
+
+
+@pytest.mark.parametrize(
+    ("secret_values", "error_type", "message"),
+    [
+        (None, ValueError, "missing"),
+        ({}, ValueError, "missing"),
+        ({"OTHER": "value"}, ValueError, "missing.*unexpected"),
+        ({"API_TOKEN": ""}, ValueError, "non-empty"),
+        ({"API_TOKEN": "bad\0value"}, ValueError, "NUL"),
+        ({"API_TOKEN": 123}, TypeError, "must be a string"),
+        ([("API_TOKEN", "value")], TypeError, "must be a mapping"),
+    ],
+)
+def test_secret_values_are_validated_before_remote_request(
+    secret_values, error_type, message
+):
+    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}},
+        )
+        with pytest.raises(error_type, match=message):
+            db.create_function_async(normalize_score, secrets=secret_values)
+        assert state["requests"] == []
+
+
+@pytest.mark.parametrize(
+    "value",
+    [
+        "x" * _MAX_FUNCTION_SECRET_VALUE_BYTES,
+        "é" * (_MAX_FUNCTION_SECRET_VALUE_BYTES // len("é".encode("utf-8"))),
+    ],
+    ids=["ascii", "multibyte"],
+)
+def test_secret_value_accepts_exact_utf8_byte_limit(value):
+    submission = json.loads(normalize_score._submission_json({"API_TOKEN": value}))
+    assert submission["secret_values"]["API_TOKEN"] == value
+    assert len(value.encode("utf-8")) == _MAX_FUNCTION_SECRET_VALUE_BYTES
+
+
+@pytest.mark.parametrize(
+    "value",
+    [
+        "x" * (_MAX_FUNCTION_SECRET_VALUE_BYTES + 1),
+        "é" * (_MAX_FUNCTION_SECRET_VALUE_BYTES // len("é".encode("utf-8")) + 1),
+    ],
+    ids=["ascii", "multibyte"],
+)
+def test_secret_value_rejects_over_utf8_byte_limit_before_json_construction(
+    monkeypatch, value
+):
+    def fail_if_json_construction_starts(self):
+        pytest.fail("oversized secret reached JSON construction")
+
+    monkeypatch.setattr(
+        FunctionRegistrationRequest, "_known_dict", fail_if_json_construction_starts
+    )
+    with pytest.raises(ValueError, match=r"exceeds the 65536-byte limit"):
+        normalize_score._submission_json({"API_TOKEN": value})
+
+
+def test_secret_values_accept_exact_aggregate_utf8_byte_limit(monkeypatch):
+    names = tuple(f"SECRET_{index}" for index in range(8))
+    value = "é" * (_MAX_FUNCTION_SECRET_VALUE_BYTES // len("é".encode("utf-8")))
+    values = {name: value for name in names}
+    monkeypatch.setattr(
+        normalize_score,
+        "_request",
+        normalize_score._request._copy(update={"required_secrets": names}),
+    )
+
+    submission = json.loads(normalize_score._submission_json(values))
+
+    assert submission["secret_values"] == values
+    assert sum(len(item.encode("utf-8")) for item in values.values()) == (
+        _MAX_FUNCTION_SECRET_VALUES_BYTES
+    )
+
+
+def test_secret_values_reject_aggregate_over_limit_before_construction(monkeypatch):
+    names = tuple(f"SECRET_{index}" for index in range(9))
+    values = {name: "x" * _MAX_FUNCTION_SECRET_VALUE_BYTES for name in names}
+    monkeypatch.setattr(
+        normalize_score,
+        "_request",
+        normalize_score._request._copy(update={"required_secrets": names}),
+    )
+
+    def fail_if_json_construction_starts(self):
+        pytest.fail("oversized aggregate reached JSON construction")
+
+    monkeypatch.setattr(
+        FunctionRegistrationRequest, "_known_dict", fail_if_json_construction_starts
+    )
+    with pytest.raises(ValueError, match=r"exceed.*524288-byte request limit"):
+        normalize_score._submission_json(values)
+
+
+@pytest.mark.asyncio
+async def test_async_remote_registration_submits_secret_values_only_once():
+    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, secrets={"API_TOKEN": "async-secret"}
+        )
+        created = await registration.wait()
+
+    assert state["requests"][0][1]["secret_values"] == {"API_TOKEN": "async-secret"}
+    assert not hasattr(created, "secret_values")

python/python/tests/test_util.py

@@ -7,6 +7,7 @@ import pathlib
 from typing import Optional
 
 import lance
+from lance.blob import BlobType as LanceBlobType
 from lancedb.conftest import MockTextEmbeddingFunction
 from lancedb.embeddings.base import EmbeddingFunctionConfig
 from lancedb.embeddings.registry import EmbeddingFunctionRegistry
@@ -907,6 +908,165 @@ def test_cast_to_target_schema():
     assert output == expected
 
 
+def test_cast_to_target_schema_coerces_binary_to_blob_v2():
+    data = pa.table({"image": pa.array([b"hello", None], type=pa.binary())})
+    target = pa.schema([lancedb.blob("image")])
+
+    output = _cast_to_target_schema(data.to_reader(), target).read_all()
+
+    image = output["image"].chunk(0)
+    assert type(image.type) is lancedb.BlobType
+    assert image.storage.to_pylist() == [
+        {"data": b"hello", "uri": None, "position": None, "size": None},
+        None,
+    ]
+
+
+def test_cast_to_target_schema_coerces_binary_to_metadata_blob_struct():
+    storage = lancedb.blob("image").type.storage_type
+    target = pa.schema(
+        [
+            pa.field(
+                "image",
+                storage,
+                metadata={
+                    b"ARROW:extension:name": b"lance.blob.v2",
+                    b"ARROW:extension:metadata": b"",
+                },
+            )
+        ]
+    )
+    data = pa.table({"image": pa.array([b"hello", None], type=pa.binary())})
+
+    output = _cast_to_target_schema(data.to_reader(), target).read_all()
+
+    image = output["image"].chunk(0)
+    assert not isinstance(image.type, pa.ExtensionType)
+    assert image.to_pylist() == [
+        {"data": b"hello", "uri": None, "position": None, "size": None},
+        None,
+    ]
+
+
+def test_cast_to_target_schema_coerces_nested_binary_blob():
+    data = pa.table(
+        {
+            "info": pa.array(
+                [{"blob": b"hello"}, {"blob": None}],
+                type=pa.struct([pa.field("blob", pa.binary())]),
+            )
+        }
+    )
+    target = pa.schema([pa.field("info", pa.struct([lancedb.blob("blob")]))])
+
+    output = _cast_to_target_schema(data.to_reader(), target).read_all()
+
+    blob = output["info"].chunk(0).field("blob")
+    assert type(blob.type) is lancedb.BlobType
+    assert blob.storage.to_pylist() == [
+        {"data": b"hello", "uri": None, "position": None, "size": None},
+        None,
+    ]
+
+
+def test_cast_to_target_schema_coerces_list_binary_blob_with_inferred_child_name():
+    data = pa.table(
+        {"images": pa.array([[b"a", b"b"], None], type=pa.list_(pa.binary()))}
+    )
+    target = pa.schema([pa.field("images", pa.list_(lancedb.blob("image")))])
+
+    output = _cast_to_target_schema(data.to_reader(), target).read_all()
+
+    images = output["images"].chunk(0)
+    assert images.type.value_field.name == "image"
+    assert type(images.type.value_type) is lancedb.BlobType
+    assert images.to_pylist()[1] is None
+    assert images.values.storage.to_pylist() == [
+        {"data": b"a", "uri": None, "position": None, "size": None},
+        {"data": b"b", "uri": None, "position": None, "size": None},
+    ]
+
+
+def test_list_blob_coercion_preserves_null_slots_with_nonzero_extent():
+    child = pa.field("image", pa.binary())
+    source = pa.ListArray.from_arrays(
+        pa.array([0, 2, 4], type=pa.int32()),
+        pa.array([b"a", b"b", b"dead", b"beef"], type=pa.binary()),
+        mask=pa.array([False, True]),
+    ).cast(pa.list_(child))
+    target = pa.schema([pa.field("images", pa.list_(lancedb.blob("image")))])
+
+    output = _cast_to_target_schema(
+        pa.table({"images": source}).to_reader(), target
+    ).read_all()
+
+    images = output["images"].chunk(0)
+    assert images.to_pylist()[1] is None
+    assert [b["data"] for b in images.to_pylist()[0]] == [b"a", b"b"]
+
+
+def test_fixed_size_list_blob_coercion_keeps_null_rows():
+    child = pa.field("frame", pa.binary())
+    source = (
+        pa.FixedSizeListArray.from_arrays(
+            pa.array([b"a", b"b", b"c", b"d"], type=pa.binary()), 2
+        )
+        .take(pa.array([0, None], type=pa.int32()))
+        .cast(pa.list_(child, 2))
+    )
+    target = pa.schema([pa.field("frames", pa.list_(lancedb.blob("frame"), 2))])
+
+    output = _cast_to_target_schema(
+        pa.table({"frames": source}).to_reader(), target
+    ).read_all()
+
+    frames = output["frames"].chunk(0)
+    assert frames.to_pylist()[1] is None
+    assert [b["data"] for b in frames.to_pylist()[0]] == [b"a", b"b"]
+
+
+def test_cast_to_target_schema_accepts_pylance_blob_v2():
+    target_type = lancedb.BlobType()
+    source = lance.blob_array([b"hello", None])
+    assert type(source.type) is LanceBlobType
+    assert type(source.type) is type(target_type)
+    data = pa.table({"image": source})
+    target = pa.schema([pa.field("image", target_type)])
+
+    output = _cast_to_target_schema(data.to_reader(), target).read_all()
+
+    image = output["image"].chunk(0)
+    assert type(image.type) is LanceBlobType
+    assert image.type == target_type
+    assert image.storage.to_pylist() == [
+        {"data": b"hello", "uri": None, "position": None, "size": None},
+        None,
+    ]
+
+
+def test_cast_to_target_schema_rejects_different_blob_v2_class():
+    class OtherBlobType(pa.ExtensionType):
+        def __init__(self):
+            super().__init__(lancedb.BlobType().storage_type, "lance.blob.v2")
+
+        def __arrow_ext_serialize__(self) -> bytes:
+            return b""
+
+        @classmethod
+        def __arrow_ext_deserialize__(
+            cls, storage_type: pa.DataType, serialized: bytes
+        ) -> "OtherBlobType":
+            return cls()
+
+    storage = lance.blob_array([b"hello"]).storage
+    source = pa.ExtensionArray.from_storage(OtherBlobType(), storage)
+    data = pa.table({"image": source})
+    target = pa.schema([lancedb.blob("image")])
+
+    with pytest.raises(pa.ArrowTypeError, match="different extension type"):
+        _cast_to_target_schema(data.to_reader(), target).read_all()
+
+
 def test_sanitize_data_stream():
     # Make sure we don't collect the whole stream when running sanitize_data
     schema = pa.schema({"a": pa.int32()})

rust/lancedb/src/function.rs

@@ -5,9 +5,9 @@
 //! backend-neutral terminal result of a computed-column refresh.
 //!
 //! This module contains client/wire values only. Catalog persistence,
-//! environment bake, and execution are owned by Sophon.
+//! environment bake, secret resolution, and execution are owned by Sophon.
 
-use std::collections::BTreeMap;
+use std::collections::{BTreeMap, BTreeSet};
 
 use serde::de::{self, DeserializeOwned};
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
@@ -15,6 +15,16 @@ use serde_json::Value;
 
 use crate::{Error, Result};
 
+// Keep these byte limits aligned with Sophon's Function submission validation.
+pub(crate) const MAX_FUNCTION_SECRET_VALUE_BYTES: usize = 64 * 1024;
+const MAX_FUNCTION_SECRET_VALUES_BYTES: usize = 512 * 1024;
+
+fn is_portable_environment_name(name: &str) -> bool {
+    let mut bytes = name.bytes();
+    matches!(bytes.next(), Some(b'A'..=b'Z' | b'a'..=b'z' | b'_'))
+        && bytes.all(|byte| matches!(byte, b'A'..=b'Z' | b'a'..=b'z' | b'0'..=b'9' | b'_'))
+}
+
 fn invalid_json(error: impl std::fmt::Display) -> Error {
     Error::InvalidInput {
         message: format!("invalid remote Function JSON: {error}"),
@@ -198,6 +208,11 @@ pub struct PythonEnvironmentSpec {
 }
 
 /// Reproducible Python runtime definition understood by Sophon.
+///
+/// `env` contains non-secret values. Secret values are submission-only in the
+/// client model and do not become part of this public runtime identity;
+/// [`FunctionVersion::required_secrets`] contains names only. Sophon persists
+/// submitted values separately in the private execution artifact.
 #[derive(Debug, Clone, PartialEq, Eq)]
 #[non_exhaustive]
 pub enum PythonRuntimeSpec {
@@ -239,7 +254,7 @@ impl PythonRuntimeSpec {
         }
     }
 
-    /// Environment variables, or `None` for an unknown kind.
+    /// Non-secret environment variables, or `None` for an unknown kind.
     pub fn env(&self) -> Option<&BTreeMap<String, String>> {
         match self {
             Self::Python { env, .. } => Some(env),
@@ -324,6 +339,8 @@ pub struct FunctionVersion {
     runtime: PythonRuntimeSpec,
     runtime_digest: String,
     environment_digest: String,
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    required_secrets: Vec<String>,
     created_at: String,
 }
 
@@ -356,6 +373,12 @@ impl FunctionVersion {
         &self.environment_digest
     }
 
+    /// Required secret names. Resolved values exist only in Sophon's private
+    /// execution artifact and worker launch path.
+    pub fn required_secrets(&self) -> &[String] {
+        &self.required_secrets
+    }
+
     pub fn created_at(&self) -> &str {
         &self.created_at
     }
@@ -397,12 +420,115 @@ pub struct FunctionArtifactRequest {
 }
 
 /// Stable request envelope for remote immutable Function registration.
-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+///
+/// Secret values are submission-only in the client model. Sophon persists them
+/// in the database-scoped private execution artifact; returned
+/// [`FunctionVersion`] and Job metadata contain only
+/// [`Self::required_secrets`] names. Debug formatting always redacts values.
+#[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct FunctionRegistrationRequest {
     pub name: String,
     pub artifact: FunctionArtifactRequest,
     pub signature: FunctionSignature,
     pub runtime: PythonRuntimeSpec,
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub required_secrets: Vec<String>,
+    #[serde(default, skip_serializing_if = "BTreeMap::is_empty")]
+    pub secret_values: BTreeMap<String, String>,
+}
+
+impl FunctionRegistrationRequest {
+    pub(crate) fn validate_secret_values(&self) -> Result<()> {
+        let mut required = BTreeSet::new();
+        for name in &self.required_secrets {
+            if !is_portable_environment_name(name) {
+                return Err(Error::InvalidInput {
+                    message: format!(
+                        "Function secret name {name:?} must be a portable environment variable name"
+                    ),
+                });
+            }
+            if !required.insert(name) {
+                return Err(Error::InvalidInput {
+                    message: format!("Function required_secrets contains duplicate name {name:?}"),
+                });
+            }
+        }
+
+        if let PythonRuntimeSpec::Python { env, .. } = &self.runtime
+            && let Some(name) = required.iter().find(|name| env.contains_key(**name))
+        {
+            return Err(Error::InvalidInput {
+                message: format!(
+                    "Function runtime env and secret names must be disjoint: {name:?}"
+                ),
+            });
+        }
+
+        let provided = self.secret_values.keys().collect::<BTreeSet<_>>();
+        if required != provided {
+            return Err(Error::InvalidInput {
+                message: "Function secret_values keys must exactly match required_secrets"
+                    .to_string(),
+            });
+        }
+
+        let mut total_bytes = 0usize;
+        for (name, value) in &self.secret_values {
+            if value.is_empty() {
+                return Err(Error::InvalidInput {
+                    message: format!("Function secret {name:?} value must be non-empty"),
+                });
+            }
+            if value.contains('\0') {
+                return Err(Error::InvalidInput {
+                    message: format!("Function secret {name:?} value must not contain NUL"),
+                });
+            }
+            if value.len() > MAX_FUNCTION_SECRET_VALUE_BYTES {
+                return Err(Error::InvalidInput {
+                    message: format!(
+                        "Function secret {name:?} value exceeds the \
+                         {MAX_FUNCTION_SECRET_VALUE_BYTES}-byte limit"
+                    ),
+                });
+            }
+            total_bytes =
+                total_bytes
+                    .checked_add(value.len())
+                    .ok_or_else(|| Error::InvalidInput {
+                        message: "Function secret values exceed the request byte limit".to_string(),
+                    })?;
+        }
+        if total_bytes > MAX_FUNCTION_SECRET_VALUES_BYTES {
+            return Err(Error::InvalidInput {
+                message: format!(
+                    "Function secret values exceed the \
+                     {MAX_FUNCTION_SECRET_VALUES_BYTES}-byte request limit"
+                ),
+            });
+        }
+        Ok(())
+    }
+}
+
+impl std::fmt::Debug for FunctionRegistrationRequest {
+    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let secret_values = self
+            .secret_values
+            .keys()
+            .map(|name| (name, "[REDACTED]"))
+            .collect::<BTreeMap<_, _>>();
+        formatter
+            .debug_struct("FunctionRegistrationRequest")
+            .field("name", &self.name)
+            .field("artifact", &self.artifact)
+            .field("signature", &self.signature)
+            .field("runtime", &self.runtime)
+            .field("required_secrets", &self.required_secrets)
+            .field("secret_values", &secret_values)
+            .finish()
+    }
 }
 
 impl_json!(FunctionRegistrationRequest);
@@ -587,6 +713,185 @@ impl RefreshColumnResult {
 
 impl_json!(RefreshColumnResult);
 
+#[cfg(test)]
+mod secret_value_tests {
+    use super::{
+        FunctionRegistrationRequest, MAX_FUNCTION_SECRET_VALUE_BYTES,
+        MAX_FUNCTION_SECRET_VALUES_BYTES, PythonRuntimeSpec,
+    };
+    use crate::Error;
+
+    fn request() -> FunctionRegistrationRequest {
+        FunctionRegistrationRequest::from_json(include_str!(
+            "../tests/fixtures/first_class_functions/v1/remote_function_registration_request.json"
+        ))
+        .unwrap()
+    }
+
+    #[test]
+    fn validates_secret_name_and_value_invariants() {
+        let missing = request();
+        assert!(matches!(
+            missing.validate_secret_values(),
+            Err(Error::InvalidInput { message }) if message.contains("exactly match")
+        ));
+
+        let mut empty = request();
+        empty
+            .secret_values
+            .insert("API_TOKEN".to_string(), String::new());
+        assert!(matches!(
+            empty.validate_secret_values(),
+            Err(Error::InvalidInput { message }) if message.contains("non-empty")
+        ));
+
+        let mut nul = request();
+        nul.secret_values
+            .insert("API_TOKEN".to_string(), "before\0after".to_string());
+        assert!(matches!(
+            nul.validate_secret_values(),
+            Err(Error::InvalidInput { message }) if message.contains("NUL")
+        ));
+
+        let mut unexpected = request();
+        unexpected
+            .secret_values
+            .insert("OTHER".to_string(), "value".to_string());
+        assert!(matches!(
+            unexpected.validate_secret_values(),
+            Err(Error::InvalidInput { message }) if message.contains("exactly match")
+        ));
+    }
+
+    #[test]
+    fn rejects_invalid_duplicate_and_overlapping_secret_declarations() {
+        let mut invalid_name = request();
+        invalid_name.required_secrets = vec!["BAD=NAME".to_string()];
+        invalid_name
+            .secret_values
+            .insert("BAD=NAME".to_string(), "secret".to_string());
+
+        let mut duplicate = request();
+        duplicate.required_secrets = vec!["API_TOKEN".to_string(), "API_TOKEN".to_string()];
+        duplicate
+            .secret_values
+            .insert("API_TOKEN".to_string(), "secret".to_string());
+
+        let mut overlap = request();
+        overlap
+            .secret_values
+            .insert("API_TOKEN".to_string(), "secret".to_string());
+        if let PythonRuntimeSpec::Python { env, .. } = &mut overlap.runtime {
+            env.insert("API_TOKEN".to_string(), "public".to_string());
+        }
+
+        assert!(matches!(
+            invalid_name.validate_secret_values(),
+            Err(Error::InvalidInput { message }) if message.contains("portable environment variable")
+        ));
+        assert!(matches!(
+            duplicate.validate_secret_values(),
+            Err(Error::InvalidInput { message }) if message.contains("duplicate")
+        ));
+        assert!(matches!(
+            overlap.validate_secret_values(),
+            Err(Error::InvalidInput { message }) if message.contains("must be disjoint")
+        ));
+    }
+
+    #[test]
+    fn enforces_portable_secret_name_boundaries() {
+        for name in ["A", "_", "A0_"] {
+            let mut request = request();
+            request.required_secrets = vec![name.to_string()];
+            request
+                .secret_values
+                .insert(name.to_string(), "secret".to_string());
+            request.validate_secret_values().unwrap();
+        }
+
+        for name in ["", "0TOKEN", "BAD-NAME", "TÖKEN"] {
+            let mut request = request();
+            request.required_secrets = vec![name.to_string()];
+            request
+                .secret_values
+                .insert(name.to_string(), "secret".to_string());
+            assert!(matches!(
+                request.validate_secret_values(),
+                Err(Error::InvalidInput { message })
+                    if message.contains("portable environment variable")
+            ));
+        }
+    }
+
+    #[test]
+    fn accepts_exact_secret_value_utf8_byte_limit() {
+        for value in [
+            "x".repeat(MAX_FUNCTION_SECRET_VALUE_BYTES),
+            "é".repeat(MAX_FUNCTION_SECRET_VALUE_BYTES / "é".len()),
+        ] {
+            assert_eq!(value.len(), MAX_FUNCTION_SECRET_VALUE_BYTES);
+            let mut request = request();
+            request.secret_values.insert("API_TOKEN".to_string(), value);
+            request.validate_secret_values().unwrap();
+        }
+    }
+
+    #[test]
+    fn rejects_secret_value_over_utf8_byte_limit() {
+        for value in [
+            "x".repeat(MAX_FUNCTION_SECRET_VALUE_BYTES + 1),
+            "é".repeat(MAX_FUNCTION_SECRET_VALUE_BYTES / "é".len() + 1),
+        ] {
+            assert!(value.len() > MAX_FUNCTION_SECRET_VALUE_BYTES);
+            let mut request = request();
+            request.secret_values.insert("API_TOKEN".to_string(), value);
+            assert!(matches!(
+                request.validate_secret_values(),
+                Err(Error::InvalidInput { message }) if message.contains("65536-byte limit")
+            ));
+        }
+    }
+
+    #[test]
+    fn rejects_aggregate_secret_value_bytes_over_server_limit() {
+        let mut request = request();
+        request.required_secrets = (0..9).map(|index| format!("SECRET_{index}")).collect();
+        request.secret_values = request
+            .required_secrets
+            .iter()
+            .map(|name| (name.clone(), "x".repeat(MAX_FUNCTION_SECRET_VALUE_BYTES)))
+            .collect();
+
+        assert!(matches!(
+            request.validate_secret_values(),
+            Err(Error::InvalidInput { message })
+                if message.contains(&format!("{MAX_FUNCTION_SECRET_VALUES_BYTES}-byte request limit"))
+        ));
+    }
+
+    #[test]
+    fn accepts_exact_aggregate_secret_value_byte_limit() {
+        let mut request = request();
+        request.required_secrets = (0..8).map(|index| format!("SECRET_{index}")).collect();
+        request.secret_values = request
+            .required_secrets
+            .iter()
+            .map(|name| (name.clone(), "x".repeat(MAX_FUNCTION_SECRET_VALUE_BYTES)))
+            .collect();
+
+        assert_eq!(
+            request
+                .secret_values
+                .values()
+                .map(String::len)
+                .sum::<usize>(),
+            MAX_FUNCTION_SECRET_VALUES_BYTES
+        );
+        request.validate_secret_values().unwrap();
+    }
+}
+
 #[cfg(test)]
 mod conda_environment_tests {
     use super::PythonEnvironmentSpec;

rust/lancedb/src/remote/client.rs

@@ -7,6 +7,7 @@ use reqwest::{
     Body, Request, RequestBuilder, Response,
     header::{HeaderMap, HeaderValue},
 };
+use serde_json::Value;
 use std::{collections::HashMap, future::Future, str::FromStr, sync::Arc, time::Duration};
 
 use crate::error::{Error, Result};
@@ -14,6 +15,60 @@ use crate::remote::db::RemoteOptions;
 use crate::remote::retry::{ResolvedRetryConfig, RetryCounter};
 
 const REQUEST_ID_HEADER: HeaderName = HeaderName::from_static("x-request-id");
+const REDACTED_JSON_VALUE: &str = "[REDACTED]";
+const SUPPRESSED_JSON_BODY: &str = "[JSON BODY SUPPRESSED]";
+
+fn is_sensitive_json_field(name: &str) -> bool {
+    name.to_ascii_lowercase().contains("secret")
+}
+
+fn redact_sensitive_json_fields(value: &mut Value) {
+    match value {
+        Value::Object(fields) => {
+            for (name, child) in fields {
+                if is_sensitive_json_field(name) {
+                    *child = Value::String(REDACTED_JSON_VALUE.to_string());
+                } else {
+                    redact_sensitive_json_fields(child);
+                }
+            }
+        }
+        Value::Array(values) => values.iter_mut().for_each(redact_sensitive_json_fields),
+        _ => {}
+    }
+}
+
+fn redacted_json_body(request: &Request) -> Option<String> {
+    let body = request.body()?.as_bytes()?;
+    let mut value = serde_json::from_slice(body).ok()?;
+    redact_sensitive_json_fields(&mut value);
+    serde_json::to_string(&value).ok()
+}
+
+fn request_log_message(request: &Request, request_id: &str) -> String {
+    let prefix = format!(
+        "Sending request_id={}: {} {}",
+        request_id,
+        request.method(),
+        request.url()
+    );
+    let content_type = request
+        .headers()
+        .get("content-type")
+        .and_then(|value| value.to_str().ok())
+        .and_then(|value| value.split(';').next());
+    if content_type.is_some_and(|value| value.eq_ignore_ascii_case("application/json")) {
+        // Never format the raw Request here: its Debug representation is not a
+        // redaction boundary and may include the original body. If the JSON body
+        // cannot be structurally parsed, suppress it instead of logging raw bytes.
+        let body = redacted_json_body(request).unwrap_or_else(|| SUPPRESSED_JSON_BODY.to_string());
+        format!("{prefix} with body {body}")
+    } else {
+        // Method and URL are sufficient request context. Raw Request formatting
+        // may expose headers or a non-JSON body, so it is never a logging fallback.
+        prefix
+    }
+}
 
 /// Configuration for TLS/mTLS settings.
 #[derive(Clone, Debug)]
@@ -839,22 +894,9 @@ impl<S: HttpSend> RestfulLanceDbClient<S> {
         }
     }
 
-    pub(crate) fn log_request(&self, request: &Request, request_id: &String) {
+    pub(crate) fn log_request(&self, request: &Request, request_id: &str) {
         if log::log_enabled!(log::Level::Debug) {
-            let content_type = request
-                .headers()
-                .get("content-type")
-                .map(|v| v.to_str().unwrap());
-            if content_type == Some("application/json") {
-                let body = request.body().as_ref().unwrap().as_bytes().unwrap();
-                let body = String::from_utf8_lossy(body);
-                debug!(
-                    "Sending request_id={}: {:?} with body {}",
-                    request_id, request, body
-                );
-            } else {
-                debug!("Sending request_id={}: {:?}", request_id, request);
-            }
+            debug!("{}", request_log_message(request, request_id));
         }
     }
 
@@ -1077,6 +1119,49 @@ mod tests {
         ENV_MUTEX.lock().unwrap_or_else(|e| e.into_inner())
     }
 
+    #[test]
+    fn test_request_log_message_redacts_secrets_and_never_formats_raw_requests() {
+        const SECRET_SENTINEL: &str = "udf-secret-log-sentinel-7e4e";
+        const MALFORMED_SENTINEL: &str = "malformed-secret-log-sentinel-b652";
+        const NON_JSON_SENTINEL: &str = "non-json-secret-log-sentinel-7fd1";
+
+        let request = reqwest::Client::new()
+            .post("https://example.com/v1/functions/create")
+            .json(&serde_json::json!({
+                "name": "uses_secret",
+                "nested": {
+                    "secret_values": {"OPENAI_API_KEY": SECRET_SENTINEL},
+                    "safe": "visible-value"
+                }
+            }))
+            .build()
+            .unwrap();
+        let log_message = request_log_message(&request, "valid-json");
+
+        let malformed_request = reqwest::Client::new()
+            .post("https://example.com/v1/functions/create")
+            .header("content-type", "application/json; charset=utf-8")
+            .body(format!(r#"{{"secret_values":"{MALFORMED_SENTINEL}""#))
+            .build()
+            .unwrap();
+        let malformed_log_message = request_log_message(&malformed_request, "malformed-json");
+
+        let non_json_request = reqwest::Client::new()
+            .post("https://example.com/v1/functions/create")
+            .header("content-type", "text/plain")
+            .body(NON_JSON_SENTINEL)
+            .build()
+            .unwrap();
+        let non_json_log_message = request_log_message(&non_json_request, "non-json");
+
+        assert!(log_message.contains("visible-value"));
+        assert!(log_message.contains(REDACTED_JSON_VALUE));
+        assert!(!log_message.contains(SECRET_SENTINEL));
+        assert!(malformed_log_message.contains(SUPPRESSED_JSON_BODY));
+        assert!(!malformed_log_message.contains(MALFORMED_SENTINEL));
+        assert!(!non_json_log_message.contains(NON_JSON_SENTINEL));
+    }
+
     #[test]
     fn test_timeout_config_default() {
         let config = TimeoutConfig::default();

rust/lancedb/src/remote/db.rs

@@ -554,6 +554,7 @@ impl<S: HttpSend> Database for RemoteDatabase<S> {
         &self,
         request: FunctionRegistrationRequest,
     ) -> Result<Job<FunctionVersion>> {
+        request.validate_secret_values()?;
         let req = self.client.post("/v1/functions/create").json(&request);
         let (request_id, response) = self.client.send(req).await?;
         let response = self.client.check_response(&request_id, response).await?;
@@ -2642,7 +2643,8 @@ mod tests {
         );
         const FUNCTION_JOB: &str =
             include_str!("../../tests/fixtures/first_class_functions/v1/remote_function_job.json");
-        let expected: serde_json::Value = serde_json::from_str(REQUEST).unwrap();
+        let mut expected: serde_json::Value = serde_json::from_str(REQUEST).unwrap();
+        expected["secret_values"] = serde_json::json!({"API_TOKEN": "secret-value"});
         let conn = Connection::new_with_handler(move |request| match request.url().path() {
             "/v1/functions/create" => {
                 assert_eq!(request.method(), &reqwest::Method::POST);
@@ -2660,7 +2662,10 @@ mod tests {
                 .unwrap(),
             path => panic!("unexpected path: {path}"),
         });
-        let request = crate::function::FunctionRegistrationRequest::from_json(REQUEST).unwrap();
+        let mut request = crate::function::FunctionRegistrationRequest::from_json(REQUEST).unwrap();
+        request
+            .secret_values
+            .insert("API_TOKEN".to_string(), "secret-value".to_string());
         let job = conn.create_function_async(request).await.unwrap();
         assert_eq!(job.id(), Some("job-function-1"));
         let version = job.wait().await.unwrap();
@@ -2668,6 +2673,32 @@ mod tests {
         assert_eq!(version.version(), "fv_01K3EXACT");
     }
 
+    #[tokio::test]
+    async fn test_create_function_async_validates_secrets_before_serialization_and_send() {
+        const REQUEST: &str = include_str!(
+            "../../tests/fixtures/first_class_functions/v1/remote_function_registration_request.json"
+        );
+        let sends = Arc::new(AtomicUsize::new(0));
+        let sends_ref = sends.clone();
+        let conn = Connection::new_with_handler(move |_| {
+            sends_ref.fetch_add(1, Ordering::SeqCst);
+            http::Response::builder().status(500).body("").unwrap()
+        });
+        let mut request = crate::function::FunctionRegistrationRequest::from_json(REQUEST).unwrap();
+        request.secret_values.insert(
+            "API_TOKEN".to_string(),
+            "x".repeat(crate::function::MAX_FUNCTION_SECRET_VALUE_BYTES + 1),
+        );
+
+        let error = conn.create_function_async(request).await.unwrap_err();
+
+        assert!(matches!(
+            error,
+            Error::InvalidInput { message } if message.contains("65536-byte limit")
+        ));
+        assert_eq!(sends.load(Ordering::SeqCst), 0);
+    }
+
     #[tokio::test]
     async fn test_get_function_requires_and_sends_exact_version() {
         const VERSION: &str = include_str!(

rust/lancedb/src/remote/table.rs

@@ -5734,8 +5734,9 @@ mod tests {
             ))
             .execute()
             .await;
-        let Err(err) = result else {
-            panic!("legacy remote query unexpectedly succeeded")
+        let err = match result {
+            Ok(_) => panic!("legacy remote query unexpectedly succeeded"),
+            Err(err) => err,
         };
 
         assert!(

rust/lancedb/src/table/computed_columns.rs

@@ -22,7 +22,7 @@
 use std::collections::{BTreeSet, HashMap};
 use std::sync::Arc;
 
-use arrow_schema::{DataType, Field as ArrowField, Fields, Schema as ArrowSchema, SchemaRef};
+use arrow_schema::{DataType, Field as ArrowField, Schema as ArrowSchema, SchemaRef};
 use datafusion_common::tree_node::TreeNode;
 use datafusion_physical_plan::PhysicalExpr;
 use lance::dataset::NewColumnTransform;
@@ -1273,11 +1273,6 @@ pub(crate) fn bind(schema: SchemaRef, column: &str, expression: &str) -> Result<
 /// refresh time: that the expression parses, that every column it reads
 /// exists, and that the target name is free. A declaration that survives this
 /// is one a refresh can always act on.
-///
-/// Each accepted column joins the schema the next one resolves against, so a
-/// batch may declare `a` and then `b = a + 1` in one commit. Refresh fills a
-/// column's computed inputs before the column, so the order of refresh calls
-/// does not matter.
 pub(crate) fn plan(schema: SchemaRef, columns: &[(String, String)]) -> Result<Vec<ArrowField>> {
     if columns.is_empty() {
         return Err(Error::InvalidInput {
@@ -1285,11 +1280,11 @@ pub(crate) fn plan(schema: SchemaRef, columns: &[(String, String)]) -> Result<Ve
         });
     }
 
-    let mut schema = schema;
     let mut fields = Vec::with_capacity(columns.len());
+    let mut declared: Vec<&str> = Vec::with_capacity(columns.len());
 
     for (name, expression) in columns {
-        if schema.field_with_name(name).is_ok() {
+        if schema.field_with_name(name).is_ok() || declared.contains(&name.as_str()) {
             return Err(Error::ColumnAlreadyExists { name: name.clone() });
         }
 
@@ -1297,50 +1292,16 @@ pub(crate) fn plan(schema: SchemaRef, columns: &[(String, String)]) -> Result<Ve
 
         // Declared columns start entirely null, so nullability is a property
         // of the declaration rather than of what the expression yields.
-        let field = ArrowField::new(name, bound.data_type, true)
-            .with_metadata(computed_column_metadata(expression, &bound.inputs));
-        schema = Arc::new(ArrowSchema::new_with_metadata(
-            schema
-                .fields()
-                .iter()
-                .cloned()
-                .chain(std::iter::once(Arc::new(field.clone())))
-                .collect::<Fields>(),
-            schema.metadata().clone(),
-        ));
-        fields.push(field);
+        fields.push(
+            ArrowField::new(name, bound.data_type, true)
+                .with_metadata(computed_column_metadata(expression, &bound.inputs)),
+        );
+        declared.push(name);
     }
 
     Ok(fields)
 }
 
-/// Run the schema-level checks of
-/// [`AddColumnsBuilder::computed`](super::AddColumnsBuilder::computed) against
-/// `schema` without committing: the Function-binding guard and the planning of
-/// every declaration. For callers that stage declarations behind other work
-/// and need those rejections before any of it lands.
-///
-/// Only the schema is consulted. Declaring also refuses a table with an LSM
-/// write spec or retained SSTables; that is table state, checked at commit.
-///
-/// ```
-/// # use std::sync::Arc;
-/// # use arrow_schema::{DataType, Field, Schema};
-/// use lancedb::table::computed_columns::validate_declarations;
-///
-/// let schema = Arc::new(Schema::new(vec![Field::new("x", DataType::Int32, false)]));
-/// let declarations = vec![
-///     ("a".to_string(), "x + 1".to_string()),
-///     ("b".to_string(), "a * 2".to_string()),
-/// ];
-/// assert!(validate_declarations(schema.clone(), &declarations).is_ok());
-/// assert!(validate_declarations(schema, &[("c".into(), "random()".into())]).is_err());
-/// ```
-pub fn validate_declarations(schema: SchemaRef, columns: &[(String, String)]) -> Result<()> {
-    ensure_no_function_bindings_for_mutation(schema.as_ref(), "schema evolution")?;
-    plan(schema, columns).map(drop)
-}
-
 /// Build the transform that declares `columns` against `schema`.
 ///
 /// An all-null column is how a binding with no values yet is carried into a
@@ -1379,22 +1340,6 @@ pub(super) async fn add_foreign_kind(table: &crate::Table, name: &str, kind: &st
 
 #[cfg(test)]
 mod tests {
-    /// The gate's reproducer: the validator applies the same schema-level
-    /// guard declaring does, so a staging caller is refused before it commits
-    /// anything else.
-    #[test]
-    fn test_validate_declarations_matches_schema_admission_barriers() {
-        let schema = Arc::new(ArrowSchema::new_with_metadata(
-            vec![ArrowField::new("x", DataType::Int32, true)],
-            HashMap::from([(
-                FUNCTION_BINDINGS_META_KEY.to_string(),
-                "not valid binding metadata".to_string(),
-            )]),
-        ));
-        let declarations = vec![("a".to_string(), "x + 1".to_string())];
-        assert!(super::validate_declarations(schema, &declarations).is_err());
-    }
-
     #[test]
     fn output_arrow_type_grammar_matches_the_shared_golden() {
         let golden: serde_json::Value = serde_json::from_str(include_str!(
@@ -1637,40 +1582,6 @@ mod tests {
         assert!(declared(&table).await.is_empty());
     }
 
-    /// A batch may build on itself: one commit, and the later entry's inputs
-    /// name the earlier one.
-    #[tokio::test]
-    async fn test_a_declaration_may_read_one_declared_before_it() {
-        let table = table_with_ints("chain").await;
-        let before = table.version().await.unwrap();
-        add_computed(
-            &table,
-            &[("a".into(), "x + 1".into()), ("b".into(), "a * 2".into())],
-        )
-        .await
-        .unwrap();
-        assert_eq!(table.version().await.unwrap(), before + 1);
-        let declared = declared(&table).await;
-        assert_eq!(declared[1].name, "b");
-        assert_eq!(declared[1].inputs, vec!["a".to_string()]);
-
-        // Order is the dependency order; reading ahead is still unknown.
-        let err = add_computed(
-            &table,
-            &[("c".into(), "d + 1".into()), ("d".into(), "x + 1".into())],
-        )
-        .await
-        .unwrap_err();
-        assert!(matches!(err, Error::InvalidExpression { column, .. } if column == "c"));
-        assert!(
-            validate_declarations(
-                table.schema().await.unwrap(),
-                &[("e".into(), "random()".into())]
-            )
-            .is_err()
-        );
-    }
-
     /// A column added by an ordinary transform is materialized, not bound, so
     /// it carries no declaration to report.
     #[tokio::test]

rust/lancedb/src/table/datafusion/blob_coerce.rs

@@ -36,6 +36,14 @@ pub(super) fn coerce_blob_expr(
     };
 
     let input_shape = match input_field.data_type() {
+        DataType::Null => {
+            let expr: Arc<dyn PhysicalExpr> = Arc::new(CastExpr::new(
+                input_expr,
+                table_field.data_type().clone(),
+                None,
+            ));
+            return Ok((expr, table_field.clone()));
+        }
         DataType::Binary | DataType::LargeBinary | DataType::BinaryView => BlobInputShape::Bytes,
         DataType::Utf8 | DataType::LargeUtf8 | DataType::Utf8View => BlobInputShape::String,
         DataType::Struct(children) => {
@@ -155,7 +163,7 @@ mod tests {
     use crate::blob::blob;
     use arrow_array::{
         Array, ArrayRef, BinaryArray, BinaryViewArray, Int32Array, Int64Array, LargeBinaryArray,
-        RecordBatch, StringArray, StringViewArray, StructArray, UInt8Array, UInt64Array,
+        NullArray, RecordBatch, StringArray, StringViewArray, StructArray, UInt8Array, UInt64Array,
     };
     use arrow_schema::Schema;
     use datafusion::prelude::SessionContext;
@@ -279,6 +287,18 @@ mod tests {
         assert_eq!(data.value(0), b"view");
     }
 
+    #[tokio::test]
+    async fn null_column_coerces_to_all_null_blob_struct() {
+        let batch = batch_with_image(
+            Field::new("image", DataType::Null, true),
+            Arc::new(NullArray::new(2)),
+        );
+        let coerced = coerce(batch, &blob_table_schema()).await;
+        let image = image_struct(&coerced);
+        assert!(image.is_null(0));
+        assert!(image.is_null(1));
+    }
+
     #[tokio::test]
     async fn binary_nulls_stay_null_after_coercion() {
         let batch = batch_with_image(

rust/lancedb/src/table/refresh.rs

@@ -7,16 +7,6 @@
 //! therefore idempotent and does not observe input mutation -- once a row is
 //! filled, changing what the expression reads leaves the stored result alone.
 //!
-//! A column's computed inputs are filled first -- the dependency graph is
-//! walked once, each reachable column filled once in dependency order, each
-//! fill its own commit. Every fill in the pass, the requested column's
-//! included, covers only the fragments of the snapshot the pass started
-//! from: a commit may rebase over a concurrent append, and the fragment that
-//! admits carries placeholder nulls no earlier fill covered, so it waits for
-//! a later refresh rather than being read as values. Two concurrent fills of
-//! one input collide on its field in lance's conflict check, so a dependent
-//! fill can only commit over inputs that were durable when it read them.
-//!
 //! Two passes per fragment. The first scans only the unfilled live rows and
 //! evaluates the expression over them, which yields the exact fill count and
 //! decides whether the fragment is staged at all -- a fragment where nothing
@@ -29,7 +19,6 @@
 //! inputs masked to null first, so a poison value in a row nobody is filling
 //! cannot fail the refresh.
 
-use std::collections::HashSet;
 use std::sync::Arc;
 
 use arrow_array::{ArrayRef, BooleanArray, RecordBatch, RecordBatchOptions};
@@ -52,8 +41,7 @@ use crate::{Error, Result};
 /// The result of refreshing a computed column.
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Default)]
 pub struct RefreshColumnResult {
-    /// Rows that had a value computed, in the requested column only; inputs
-    /// filled on its behalf are not counted.
+    /// Rows that had a value computed.
     #[serde(default)]
     pub rows_filled: u64,
     /// The commit version associated with the operation.
@@ -63,79 +51,7 @@ pub struct RefreshColumnResult {
 
 struct RefreshExecution {
     result: RefreshColumnResult,
-    /// The snapshot the requested column was evaluated against, after its
-    /// inputs were filled.
     source_version: u64,
-    /// The last version any fill in the pass committed, inputs included.
-    published_version: Option<u64>,
-    /// Unfilled live rows of the requested column in fragments the pass did
-    /// not cover, counted on the snapshot the result reports -- the published
-    /// one, or the source when nothing was published.
-    rows_deferred: u64,
-}
-
-/// One column's fill against one snapshot.
-struct Fill {
-    rows_filled: u64,
-    /// The dataset the commit produced, if anything was filled.
-    committed: Option<Arc<Dataset>>,
-}
-
-/// Test-only one-shot pauses before a named column's fill:
-/// `(column, reached, resume)`, one per column, consumed when hit.
-#[cfg(test)]
-type Pause = (String, Arc<tokio::sync::Notify>, Arc<tokio::sync::Notify>);
-#[cfg(test)]
-static PAUSES_BEFORE_FILL: std::sync::Mutex<Vec<Pause>> = std::sync::Mutex::new(Vec::new());
-
-/// The SQL-computed columns `column` reads, transitively, each once, in an
-/// order that fills every column after the columns it reads.
-///
-/// Declarations are acyclic by construction: a column can only read what
-/// existed when it was declared.
-fn dependency_order(schema: &ArrowSchema, column: &str) -> Result<Vec<String>> {
-    fn visit(
-        schema: &ArrowSchema,
-        column: &str,
-        visited: &mut HashSet<String>,
-        order: &mut Vec<String>,
-    ) -> Result<()> {
-        let Some(declaration) = schema
-            .field_with_name(column)
-            .ok()
-            .and_then(computed_column_from_field)
-        else {
-            return Ok(());
-        };
-        for input in &declaration.inputs {
-            let input = super::computed_columns::root(input);
-            if visited.contains(input) {
-                continue;
-            }
-            let Some(input_declaration) = schema
-                .field_with_name(input)
-                .ok()
-                .and_then(computed_column_from_field)
-            else {
-                continue;
-            };
-            if !matches!(input_declaration.kind, ComputedColumnKind::Sql { .. }) {
-                return Err(Error::NotSupported {
-                    message: format!(
-                        "computed column '{column}' reads '{input}', which this refresh \
-                         cannot fill first; refresh '{input}' before '{column}'"
-                    ),
-                });
-            }
-            visit(schema, input, visited, order)?;
-            visited.insert(input.to_string());
-            order.push(input.to_string());
-        }
-        Ok(())
-    }
-    let mut order = Vec::new();
-    visit(schema, column, &mut HashSet::new(), &mut order)?;
-    Ok(order)
 }
 
 /// Internal implementation of the refresh logic.
@@ -154,96 +70,9 @@ async fn execute_refresh_column_with_source(
 ) -> Result<RefreshExecution> {
     table.dataset.ensure_mutable()?;
     ensure_no_lsm_write_spec(table).await?;
-    let mut dataset = table.dataset.get().await?;
-    declared_expression(&dataset, column)?;
+    let dataset = table.dataset.get().await?;
 
-    // The pass covers exactly these fragments. A commit below may rebase over
-    // a concurrent append, and the fragment that admits was scanned by no
-    // earlier fill, so it is excluded from every later one.
-    let eligible: HashSet<u64> = dataset
-        .get_fragments()
-        .iter()
-        .map(|f| f.id() as u64)
-        .collect();
-
-    let mut published_version = None;
-    for input in dependency_order(&ArrowSchema::from(dataset.schema()), column)? {
-        pause_before_fill(&input).await;
-        if let Some(committed) = fill_column(table, &dataset, &input, &eligible)
-            .await?
-            .committed
-        {
-            published_version = Some(committed.version().version);
-            dataset = committed;
-        }
-    }
-
-    pause_before_fill(column).await;
-    let source_version = dataset.version().version;
-    let fill = fill_column(table, &dataset, column, &eligible).await?;
-    if let Some(committed) = &fill.committed {
-        published_version = Some(committed.version().version);
-    }
-    // Counted after the commit: it may have rebased over an append, and the
-    // fragment that admits is deferred but part of the published version.
-    let terminal = fill.committed.as_ref().unwrap_or(&dataset);
-    let rows_deferred = count_deferred(terminal, column, &eligible).await?;
-    Ok(RefreshExecution {
-        result: RefreshColumnResult {
-            rows_filled: fill.rows_filled,
-            version: published_version.unwrap_or(source_version),
-        },
-        source_version,
-        published_version,
-        rows_deferred,
-    })
-}
-
-/// Unfilled live rows of `column` in `dataset`'s fragments outside `eligible`.
-async fn count_deferred(dataset: &Dataset, column: &str, eligible: &HashSet<u64>) -> Result<u64> {
-    let deferred: Vec<_> = dataset
-        .get_fragments()
-        .into_iter()
-        .filter(|fragment| !eligible.contains(&(fragment.id() as u64)))
-        .map(|fragment| fragment.metadata().clone())
-        .collect();
-    if deferred.is_empty() {
-        return Ok(0);
-    }
-    let mut scanner = dataset.scan();
-    scanner
-        .with_fragments(deferred)
-        .filter(&format!("{} IS NULL", quote_identifier(column)))?;
-    Ok(scanner.count_rows().await?)
-}
-
-#[cfg(test)]
-async fn pause_before_fill(column: &str) {
-    let pause = {
-        let mut pauses = PAUSES_BEFORE_FILL.lock().unwrap();
-        pauses
-            .iter()
-            .position(|(paused, _, _)| paused == column)
-            .map(|index| pauses.remove(index))
-    };
-    if let Some((_, reached, resume)) = pause {
-        reached.notify_one();
-        resume.notified().await;
-    }
-}
-
-#[cfg(not(test))]
-async fn pause_before_fill(_column: &str) {}
-
-/// Fill `column`'s unfilled live rows in the `eligible` fragments as they
-/// stand in `dataset`, committing against that snapshot.
-async fn fill_column(
-    table: &NativeTable,
-    dataset: &Arc<Dataset>,
-    column: &str,
-    eligible: &HashSet<u64>,
-) -> Result<Fill> {
-    let expression = declared_expression(dataset, column)?;
+    let expression = declared_expression(&dataset, column)?;
     let schema = Arc::new(ArrowSchema::from(dataset.schema()));
     let bound = Arc::new(super::computed_columns::bind(schema, column, &expression)?);
     let field = dataset
@@ -262,22 +91,23 @@ async fn fill_column(
     let mut rows_filled = 0u64;
     let mut replacements = Vec::new();
     for fragment in dataset.get_fragments() {
-        if !eligible.contains(&(fragment.id() as u64)) {
-            continue;
-        }
-        let gained = count_fragment_gains(dataset, &fragment, &bound, column).await?;
+        let gained = count_fragment_gains(&dataset, &fragment, &bound, column).await?;
         if gained == 0 {
             continue;
         }
         rows_filled += gained;
-        let values = fill_stream(dataset, &fragment, bound.clone(), column).await?;
+        let values = fill_stream(&dataset, &fragment, bound.clone(), column).await?;
         replacements.push(fragment.write_columns(values, &column_schema).await?);
     }
 
     if replacements.is_empty() {
-        return Ok(Fill {
-            rows_filled: 0,
-            committed: None,
+        let source_version = dataset.version().version;
+        return Ok(RefreshExecution {
+            result: RefreshColumnResult {
+                rows_filled: 0,
+                version: source_version,
+            },
+            source_version,
         });
     }
 
@@ -295,10 +125,15 @@ async fn fill_column(
         false,
     )
     .await?;
-    table.dataset.update(new_dataset.clone());
-    Ok(Fill {
-        rows_filled,
-        committed: Some(Arc::new(new_dataset)),
+
+    let version = new_dataset.version().version;
+    table.dataset.update(new_dataset);
+    Ok(RefreshExecution {
+        result: RefreshColumnResult {
+            rows_filled,
+            version,
+        },
+        source_version: read_version,
     })
 }
 
@@ -323,9 +158,10 @@ pub(crate) async fn execute_refresh_column_async(
         Ok(crate::function::RefreshColumnResult {
             rows_assigned: execution.result.rows_filled,
             rows_failed: 0,
-            rows_remaining: execution.rows_deferred,
+            rows_remaining: 0,
             source_version: execution.source_version,
-            published_version: execution.published_version,
+            published_version: (execution.result.rows_filled > 0)
+                .then_some(execution.result.version),
         })
     })))
 }
@@ -578,181 +414,6 @@ mod tests {
         table.add(batch).execute().await.unwrap();
     }
 
-    /// The gate's reproducer: `b = coalesce(a, 0)` refreshed before `a`
-    /// must not bake zeros from `a`'s placeholder null.
-    #[tokio::test]
-    async fn test_dependent_refresh_cannot_fill_from_placeholder_null() {
-        let table = table_with("dependent_refresh_order", vec![1, 2, 3]).await;
-        table
-            .add_columns()
-            .computed("a", "x + 1")
-            .computed("b", "coalesce(a, 0)")
-            .execute()
-            .await
-            .unwrap();
-
-        let result = table.refresh_column("b").await.unwrap();
-        assert_eq!(result.rows_filled, 3);
-        assert_eq!(read(&table, "a").await, vec![Some(2), Some(3), Some(4)]);
-        assert_eq!(
-            table.count_rows(Some("b = a".to_string())).await.unwrap(),
-            3
-        );
-        assert_eq!(table.refresh_column("a").await.unwrap().rows_filled, 0);
-
-        // Appended rows: the input is filled in the new fragment first too.
-        append(&table, vec![10]).await;
-        assert_eq!(table.refresh_column("b").await.unwrap().rows_filled, 1);
-        assert_eq!(
-            table.count_rows(Some("b = 0".to_string())).await.unwrap(),
-            0
-        );
-    }
-
-    /// The gate's reproducer: a fragment appended between the input fill and
-    /// the requested column's fill has inputs the pass never covered, so it
-    /// is left unfilled rather than read as null.
-    #[tokio::test]
-    async fn test_dependent_refresh_fences_an_append_after_input_fill() {
-        let table = table_with("dependent_refresh_append_gap", vec![1, 2, 3]).await;
-        table
-            .add_columns()
-            .computed("a_gap", "x + 1")
-            .computed("b_gap", "coalesce(a_gap, 0)")
-            .execute()
-            .await
-            .unwrap();
-        let reached = Arc::new(tokio::sync::Notify::new());
-        let resume = Arc::new(tokio::sync::Notify::new());
-        super::PAUSES_BEFORE_FILL.lock().unwrap().push((
-            "b_gap".to_string(),
-            reached.clone(),
-            resume.clone(),
-        ));
-
-        let refresh = {
-            let table = table.clone();
-            tokio::spawn(async move {
-                table
-                    .refresh_column_async("b_gap")
-                    .await
-                    .unwrap()
-                    .wait()
-                    .await
-            })
-        };
-        reached.notified().await;
-        append(&table, vec![10]).await;
-        resume.notify_one();
-        let result = refresh.await.unwrap().unwrap();
-        assert_eq!(result.rows_assigned, 3);
-        // The target commit rebased over the append: the published version
-        // holds the deferred row, and the count is taken there.
-        assert_eq!(result.rows_remaining, 1);
-
-        table.refresh_column("a_gap").await.unwrap();
-        assert_eq!(table.count_rows(Some("b_gap = 0".into())).await.unwrap(), 0);
-        assert_eq!(table.refresh_column("b_gap").await.unwrap().rows_filled, 1);
-    }
-
-    /// The gate's reproducer: an append landing between two input fills is
-    /// rebased into the second's snapshot, but the first never covered it.
-    #[tokio::test]
-    async fn test_dependent_refresh_fences_an_append_between_input_fills() {
-        let table = table_with("dependent_refresh_between_inputs", vec![1, 2, 3]).await;
-        table
-            .add_columns()
-            .computed("a_mid", "x + 1")
-            .computed("c_mid", "x + 2")
-            .computed("b_mid", "coalesce(a_mid, 0) + coalesce(c_mid, 0)")
-            .execute()
-            .await
-            .unwrap();
-        let reached = Arc::new(tokio::sync::Notify::new());
-        let resume = Arc::new(tokio::sync::Notify::new());
-        super::PAUSES_BEFORE_FILL.lock().unwrap().push((
-            "c_mid".to_string(),
-            reached.clone(),
-            resume.clone(),
-        ));
-
-        let refresh = {
-            let table = table.clone();
-            tokio::spawn(async move {
-                table
-                    .refresh_column_async("b_mid")
-                    .await
-                    .unwrap()
-                    .wait()
-                    .await
-            })
-        };
-        reached.notified().await;
-        append(&table, vec![10]).await;
-        resume.notify_one();
-        let result = refresh.await.unwrap().unwrap();
-        assert_eq!(result.rows_assigned, 3);
-        // The appended row is in the reported source version but was deferred.
-        assert_eq!(result.rows_remaining, 1);
-
-        table.refresh_column("a_mid").await.unwrap();
-        table.refresh_column("c_mid").await.unwrap();
-        assert_eq!(table.count_rows(Some("b_mid = 0".into())).await.unwrap(), 0);
-        assert_eq!(table.refresh_column("b_mid").await.unwrap().rows_filled, 1);
-        assert_eq!(
-            table.count_rows(Some("b_mid = 23".into())).await.unwrap(),
-            1
-        );
-    }
-
-    /// A dependency commit is a publication even when the requested column
-    /// itself fills nothing.
-    #[tokio::test]
-    async fn test_refresh_async_reports_a_dependency_publication() {
-        let table = table_with("refresh_async_dependency_publication", vec![1, 2, 3]).await;
-        table
-            .add_columns()
-            .computed("a", "x + 1")
-            .computed("b", "nullif(a, a)")
-            .execute()
-            .await
-            .unwrap();
-        let result = table
-            .refresh_column_async("b")
-            .await
-            .unwrap()
-            .wait()
-            .await
-            .unwrap();
-        assert_eq!(result.rows_assigned, 0);
-        assert_eq!(result.published_version, Some(result.source_version));
-    }
-
-    /// The gate's reproducer: a dense graph is walked once, not once per path.
-    #[tokio::test]
-    async fn test_dependency_refresh_visits_each_column_once() {
-        let table = table_with("dependency_refresh_deduplicates", vec![1]).await;
-        let mut declaration = table.add_columns();
-        for i in 0..=6 {
-            let expression = if i == 0 {
-                "x + 1".to_string()
-            } else {
-                (0..i)
-                    .map(|j| format!("a{j}"))
-                    .collect::<Vec<_>>()
-                    .join(" + ")
-            };
-            declaration = declaration.computed(format!("a{i}"), expression);
-        }
-        declaration.execute().await.unwrap();
-        let order = super::dependency_order(table.schema().await.unwrap().as_ref(), "a6").unwrap();
-        assert_eq!(order, (0..6).map(|i| format!("a{i}")).collect::<Vec<_>>());
-        // One commit per column, the requested one included.
-        let before = table.version().await.unwrap();
-        table.refresh_column("a6").await.unwrap();
-        assert_eq!(table.version().await.unwrap(), before + 7);
-    }
-
     #[tokio::test]
     async fn test_refresh_fills_a_declared_column() {
         let table = table_with("refresh_fills", vec![1, 2, 3]).await;

rust/lancedb/tests/first_class_function_slice1.rs

@@ -20,6 +20,25 @@ fn job_result(name: &str) -> Value {
     serde_json::from_str::<Value>(&fixture(name)).expect("remote Job fixture")["result"].clone()
 }
 
+fn assert_no_secret_values(value: &Value) {
+    match value {
+        Value::Object(values) => {
+            for (key, value) in values {
+                assert!(
+                    !matches!(
+                        key.as_str(),
+                        "secret_value" | "secret_values" | "resolved_secret" | "resolved_secrets"
+                    ),
+                    "client canonical value must not model resolved secret material"
+                );
+                assert_no_secret_values(value);
+            }
+        }
+        Value::Array(values) => values.iter().for_each(assert_no_secret_values),
+        _ => {}
+    }
+}
+
 #[test]
 fn function_version_job_result_matches_shared_canonical_golden() {
     let result = job_result("remote_function_job.json");
@@ -28,6 +47,7 @@ fn function_version_job_result_matches_shared_canonical_golden() {
     assert_eq!(version.name(), "embed");
     assert_eq!(version.version(), "fv_01K3EXACT");
     assert_eq!(version.runtime_digest(), "sha256:runtime");
+    assert_eq!(version.required_secrets(), &["HF_TOKEN"]);
     assert_eq!(
         version.to_canonical_json().expect("canonical JSON"),
         fixture("remote_function_version.canonical.json").trim()
@@ -142,3 +162,21 @@ fn floating_point_application_literals_are_rejected_consistently() {
             .contains("floating-point Function literals")
     );
 }
+
+#[test]
+fn canonical_client_values_contain_secret_names_only() {
+    let result = job_result("remote_function_job.json");
+    let version = FunctionVersion::from_json(&result.to_string()).expect("FunctionVersion result");
+    let canonical: Value = serde_json::from_str(
+        &version
+            .to_canonical_json()
+            .expect("canonical FunctionVersion"),
+    )
+    .expect("canonical JSON");
+
+    assert_eq!(
+        canonical["required_secrets"],
+        serde_json::json!(["HF_TOKEN"])
+    );
+    assert_no_secret_values(&canonical);
+}

rust/lancedb/tests/first_class_function_slice2.rs

@@ -6,6 +6,7 @@ use std::path::PathBuf;
 
 use lancedb::Error;
 use lancedb::function::FunctionRegistrationRequest;
+use serde_json::Value;
 
 fn fixture(name: &str) -> String {
     let path = PathBuf::from(env!("CARGO_MANIFEST_DIR"))
@@ -14,6 +15,25 @@ fn fixture(name: &str) -> String {
     fs::read_to_string(path).expect("fixture must be readable")
 }
 
+fn assert_no_secret_values(value: &Value) {
+    match value {
+        Value::Object(values) => {
+            for (key, value) in values {
+                assert!(
+                    !matches!(
+                        key.as_str(),
+                        "secret_value" | "secret_values" | "resolved_secret" | "resolved_secrets"
+                    ),
+                    "registration requests must not model resolved secret material"
+                );
+                assert_no_secret_values(value);
+            }
+        }
+        Value::Array(values) => values.iter().for_each(assert_no_secret_values),
+        _ => {}
+    }
+}
+
 #[test]
 fn registration_request_matches_shared_canonical_golden() {
     let request = FunctionRegistrationRequest::from_json(&fixture(
@@ -22,10 +42,33 @@ fn registration_request_matches_shared_canonical_golden() {
     .expect("registration request");
     assert_eq!(request.name, "normalize_score");
     assert_eq!(request.artifact.adapter.kind, "scalar_to_arrow_batch");
+    assert_eq!(request.required_secrets, ["API_TOKEN"]);
+    assert!(request.secret_values.is_empty());
     assert_eq!(
         request.to_canonical_json().expect("canonical request"),
         fixture("remote_function_registration_request.canonical.json").trim()
     );
+
+    let value: Value =
+        serde_json::from_str(&request.to_canonical_json().expect("canonical request"))
+            .expect("request JSON");
+    assert_no_secret_values(&value);
+}
+
+#[test]
+fn registration_request_serializes_secret_values_but_redacts_debug_output() {
+    let mut value: Value =
+        serde_json::from_str(&fixture("remote_function_registration_request.json")).unwrap();
+    value["secret_values"] = serde_json::json!({"API_TOKEN": "secret-plaintext"});
+    let request = FunctionRegistrationRequest::from_json(&value.to_string()).unwrap();
+
+    assert_eq!(request.secret_values["API_TOKEN"], "secret-plaintext");
+    let canonical = request.to_canonical_json().unwrap();
+    assert!(canonical.contains("secret-plaintext"));
+    let debug = format!("{request:?}");
+    assert!(debug.contains("API_TOKEN"));
+    assert!(debug.contains("[REDACTED]"));
+    assert!(!debug.contains("secret-plaintext"));
 }
 
 #[tokio::test]

rust/lancedb/tests/fixtures/first_class_functions/v1/remote_function_job.json

@@ -24,6 +24,7 @@
     },
     "runtime_digest": "sha256:runtime",
     "environment_digest": "sha256:environment",
+    "required_secrets": ["HF_TOKEN"],
     "created_at": "2026-08-21T00:00:00Z"
   },
   "future_job": {"trace_id": "trace-1"}

rust/lancedb/tests/fixtures/first_class_functions/v1/remote_function_registration_request.canonical.json

@@ -1 +1 @@
-{"artifact":{"adapter":{"kind":"scalar_to_arrow_batch","version":1},"content":{"data":"ZnJvbSBfX2Z1dHVyZV9fIGltcG9ydCBhbm5vdGF0aW9ucwoKZGVmIG5vcm1hbGl6ZV9zY29yZSh2YWx1ZTogZmxvYXQpIC0+IGZsb2F0OgogICAgcmV0dXJuIHZhbHVlIC8gMTAwLjAK","encoding":"base64"},"digest":"sha256:760784bdcef57b802f389b97804cc0b618aae39e86044733451bcd0b13089a7f","entrypoint":"normalize_score","kind":"python_callable"},"name":"normalize_score","runtime":{"env":{"MODE":"test"},"environment":{"kind":"pip","packages":["numpy>=2"]},"kind":"python","python_version":"3.12"},"signature":{"inputs":[{"arrow_type":"float64","name":"value","nullable":false}],"output":{"arrow_type":"float64","kind":"scalar","nullable":false}}}
+{"artifact":{"adapter":{"kind":"scalar_to_arrow_batch","version":1},"content":{"data":"ZnJvbSBfX2Z1dHVyZV9fIGltcG9ydCBhbm5vdGF0aW9ucwoKZGVmIG5vcm1hbGl6ZV9zY29yZSh2YWx1ZTogZmxvYXQpIC0+IGZsb2F0OgogICAgcmV0dXJuIHZhbHVlIC8gMTAwLjAK","encoding":"base64"},"digest":"sha256:760784bdcef57b802f389b97804cc0b618aae39e86044733451bcd0b13089a7f","entrypoint":"normalize_score","kind":"python_callable"},"name":"normalize_score","required_secrets":["API_TOKEN"],"runtime":{"env":{"MODE":"test"},"environment":{"kind":"pip","packages":["numpy>=2"]},"kind":"python","python_version":"3.12"},"signature":{"inputs":[{"arrow_type":"float64","name":"value","nullable":false}],"output":{"arrow_type":"float64","kind":"scalar","nullable":false}}}

rust/lancedb/tests/fixtures/first_class_functions/v1/remote_function_registration_request.json

@@ -39,5 +39,8 @@
     "env": {
       "MODE": "test"
     }
-  }
+  },
+  "required_secrets": [
+    "API_TOKEN"
+  ]
 }

rust/lancedb/tests/fixtures/first_class_functions/v1/remote_function_version.canonical.json

@@ -1 +1 @@
-{"artifact":{"digest":"sha256:code","entrypoint":"embed","kind":"python_callable"},"created_at":"2026-08-21T00:00:00Z","environment_digest":"sha256:environment","name":"embed","runtime":{"env":{"TOKENIZERS_PARALLELISM":"false"},"environment":{"kind":"pip","packages":["sentence-transformers>=3"]},"kind":"python","python_version":"3.12"},"runtime_digest":"sha256:runtime","signature":{"inputs":[{"arrow_type":"utf8","name":"text","nullable":true}],"output":{"arrow_type":"list<float32>","kind":"scalar","nullable":false}},"version":"fv_01K3EXACT"}
+{"artifact":{"digest":"sha256:code","entrypoint":"embed","kind":"python_callable"},"created_at":"2026-08-21T00:00:00Z","environment_digest":"sha256:environment","name":"embed","required_secrets":["HF_TOKEN"],"runtime":{"env":{"TOKENIZERS_PARALLELISM":"false"},"environment":{"kind":"pip","packages":["sentence-transformers>=3"]},"kind":"python","python_version":"3.12"},"runtime_digest":"sha256:runtime","signature":{"inputs":[{"arrow_type":"utf8","name":"text","nullable":true}],"output":{"arrow_type":"list<float32>","kind":"scalar","nullable":false}},"version":"fv_01K3EXACT"}