fix(python): use one blobv2 type and coerce blob writes by metadata (#4065)

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/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_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()})