Skip to content

Codebase intelligence for AI and humans: code health scores, auto-generated docs, git analytics, dead code detection, and architectural decisions via MCP.

perf(search): cap LanceDB's metadata cache so generation memory stops climbing (#2868)

Raghav Chamadiya7077cf0aef82
diff --git a/packages/core/src/repowise/core/persistence/vector_store/lancedb_store.py b/packages/core/src/repowise/core/persistence/vector_store/lancedb_store.py
index 98bc56112..8dc8e3ea4 100644
--- a/packages/core/src/repowise/core/persistence/vector_store/lancedb_store.py
+++ b/packages/core/src/repowise/core/persistence/vector_store/lancedb_store.py
@@ -23,6 +23,12 @@ __all__ = ["STORED_SNIPPET_CHARS", "LanceDBVectorStore", "read_recorded_vector_d
# gigabytes before returning even though the selected result is small.
_SUMMARY_PATH_BATCH_SIZE = 100
+# LanceDB caches each table version's manifest (default cap 1 GiB). Generation
+# upserts in embedder-sized chunks, one version each, and every manifest lists
+# all fragments so far, so the cache grew quadratically: about 0.55 GiB over
+# 22k pages, held until exit. Reads only need the latest version.
+_METADATA_CACHE_BYTES = 64 * 1024 * 1024
+
# ``STORED_SNIPPET_CHARS`` — how much of a page's content each row keeps — is
# defined with the embed recipe and re-exported here so the historical import
# path keeps working.
@@ -136,7 +142,15 @@ class LanceDBVectorStore(VectorStore):
"LanceDB is not installed. Install it with: pip install repowise-core[search]"
) from exc
- self._db = await lancedb.connect_async(self._db_path)
+ # ``Session`` arrived after the oldest supported LanceDB; without it
+ # the connection keeps the library default.
+ session_cls = getattr(lancedb, "Session", None)
+ kwargs = (
+ {"session": session_cls(metadata_cache_size_bytes=_METADATA_CACHE_BYTES)}
+ if session_cls is not None
+ else {}
+ )
+ self._db = await lancedb.connect_async(self._db_path, **kwargs)
table_names = await self._db.table_names()
if self._table_name in table_names:
self._table = await self._db.open_table(self._table_name)
diff --git a/tests/unit/persistence/test_vector_store.py b/tests/unit/persistence/test_vector_store.py
index 2987dca67..eafc84284 100644
--- a/tests/unit/persistence/test_vector_store.py
+++ b/tests/unit/persistence/test_vector_store.py
@@ -356,6 +356,32 @@ async def test_lancedb_embed_batch(tmp_path, mock_embedder):
await store.close()
+async def test_lancedb_connects_with_bounded_metadata_cache(tmp_path, mock_embedder, monkeypatch):
+ # Every upsert adds a table version whose manifest the session caches, so
+ # the library's 1 GiB default grew with the page count until exit.
+ lancedb = pytest.importorskip("lancedb")
+ if not hasattr(lancedb, "Session"):
+ pytest.skip("LanceDB without Session keeps its default cache")
+ from repowise.core.persistence.vector_store import LanceDBVectorStore
+
+ seen: dict = {}
+ original = lancedb.connect_async
+
+ async def _spy(*args, **kwargs):
+ seen.update(kwargs)
+ return await original(*args, **kwargs)
+
+ monkeypatch.setattr(lancedb, "connect_async", _spy)
+ store = LanceDBVectorStore(str(tmp_path / "lance"), mock_embedder)
+ try:
+ await store.embed_batch([("p1", "Python async generators", {"target_path": "a.py"})])
+ assert isinstance(seen.get("session"), lancedb.Session)
+ assert await store.list_page_ids() == {"p1"}
+ finally:
+ await store.close()
+
+
class _FixedDimEmbedder:
"""Deterministic embedder with a configurable output dimension."""