Repository navigation
fix(python): return the session supplied to LanceDBConnection - #4433
Conversation
LanceDBConnection.session read AsyncConnection.session, but AsyncConnection has no session attribute, so access raised AttributeError. Store the session passed to the constructor and return that object. Assisted by Cursor (Grok).
There was a problem hiding this comment.
Returning the caller-supplied session fixes the attribute failure and keeps property inspection synchronous for direct and reconstructed connections. Callers that retain closed connection wrappers should account for the supplied session’s cache lifetime.
| # The native connection does not expose the Python Session wrapper. | ||
| # Keep the object supplied by the caller so this property stays | ||
| # synchronous for debugger inspection. | ||
| self._session = session |
There was a problem hiding this comment.
A closed connection wrapper now retains its supplied session’s caches. The new strong reference survives close(), which only closes _conn. Passing a temporary Session and retaining the closed connection can therefore keep populated caches alive up to their configured limits.
This is non-blocking because dropping the closed wrapper releases its session reference. Clearing _session during close() is an alternative if explicit connection cleanup should release that reference.
Reproducer
import gc
import tempfile
import lancedb
with tempfile.TemporaryDirectory() as uri:
db = lancedb.connect(
uri,
session=lancedb.Session(
index_cache_size_bytes=1024 * 1024,
metadata_cache_size_bytes=512 * 1024,
),
)
table = db.create_table("test", [{"id": i} for i in range(100)])
table.to_arrow()
del table
gc.collect()
before = db.session.size_bytes
db.close()
gc.collect()
assert not db.is_open()
print(before, db.session.size_bytes)On this revision, the connection is closed but this prints 5133 5133: the populated session cache is still retained.
Fixes #4400. Sync remote connections fail on `uri`, `len(db)`, and `name in db`, and remote tables lack `namespace`, `id`, and `replace_field_metadata`. Share the paginated connection helpers and table identifier contract through the base classes, and delegate remote metadata replacement to the existing async implementation. Unsupported remote session and read-consistency properties raise explicit `NotImplementedError`s. Reconciled with main at `3dd57978`, including the synchronous method signatures from #4498. The local session fix has already landed in #4433, so this PR preserves its supplied-session-or-None contract and adds no Rust/PyO3 session access. Namespace connections retain their writable properties. Existing custom `Table` subclasses remain instantiable: `namespace` defaults to a fresh empty list, and deprecated `replace_field_metadata` delegates to `update_field_metadata` with `replace=True` while retaining its `None` return. Native and remote overrides are preserved. Regression tests instantiate an old-style subclass through normal ABC class creation and verify identifiers, independent namespace lists, deprecation, and exact metadata delegation. Historical validation at `03c43f7b` before the latest main integration and subclass compatibility repair, Ubuntu x86-64, Python 3.12.13: - Rebuilt the native extension with the repository `ci` profile and locked Python dependencies. - 273 tests passed across `test_sync_api_parity.py`, `test_db.py`, `test_remote_db.py`, and `test_session.py`. - 84 namespace and catalog compatibility tests passed (357 tests total). - Python SDK lint and formatting pass (158 files); changed-file checks and `git diff --check` pass. - Repository-wide lint/format checks report existing issues in unchanged `docs/web-tests` files (49 lint findings and 11 formatting files). These are outside this patch. Remote tests use the existing HTTP fixtures through the rebuilt native client; the original aarch64 server deployment was not recreated. Validation after merging main `3dd57978` and the compatibility repair (macOS ARM64, uv-managed Python 3.12.12, rebuilt editable native extension): - 60 focused parity, signature, and error-contract tests passed. - 381 database, remote, session, namespace, catalog, and sync parity tests passed; 2 platform-dependent tests skipped. - The two new subclass regressions fail against the pre-repair implementation with the expected abstract-member instantiation error, then pass with the repair. - Python SDK Ruff lint/format passes (167 files); repository-wide findings remain confined to unchanged docs/web-tests (49 lint findings, 11 formatting files). --------- Co-authored-by: Xuanwo <[email protected]>
LanceDBConnection.session read AsyncConnection.session, but AsyncConnection has no session attribute, so access raised AttributeError. Store the session passed to the constructor and return that object.
Assisted by Cursor (Grok).