MemPalace/mempalace · error · BackendError

pgvector backend requires a local palace path to anchor mism

Error message

pgvector backend requires a local palace path to anchor mismatch protection; pure-remote palaces (local_path=None) are not supported yet

What it means

The pgvector backend's only mismatch protection is the local marker file inside the palace directory. With palace.local_path=None (pure-remote/hosted mode) it can neither write nor validate the marker, so opening would silently drop protection against DSN/namespace drift — therefore it refuses loudly with BackendError. Remote marker storage for pure-remote palaces is tracked as follow-up work.

Source

Thrown at mempalace/backends/pgvector.py:1515

    def get_collection(self, *args, **kwargs) -> PgVectorCollection:
        palace, collection_name, create, options = self._normalize_args(args, kwargs)
        config = _PgVectorConfig.from_options(options)
        if palace.namespace and palace.namespace != config.namespace:
            config = _PgVectorConfig(dsn=config.dsn, namespace=palace.namespace)
        client = self._client(config)
        if palace.local_path:
            marker_path = self._marker_path(palace.local_path)
            if os.path.isfile(marker_path):
                self._validate_marker_target(palace, config)
            elif not create:
                raise PalaceNotFoundError(marker_path)
        else:
            # The local marker is this backend's only mismatch-protection
            # anchor. With no local_path (pure-remote / hosted mode) we can
            # neither write nor validate it, so opening would silently drop
            # protection against DSN/namespace drift. Refuse loudly. A remote
            # marker store for pure-remote palaces is tracked as a follow-up.
            raise BackendError(
                "pgvector backend requires a local palace path to anchor mismatch "
                "protection; pure-remote palaces (local_path=None) are not "
                "supported yet"
            )
        table = self._table_name(palace=palace, collection_name=collection_name, config=config)
        if not create and not client.table_exists(table):
            raise CollectionNotInitializedError(collection_name)
        collection = PgVectorCollection(
            backend=self,
            client=client,
            config=config,
            palace=palace,
            collection_name=collection_name,
            table=table,
        )
        with self._lock:
            self._collections_by_palace.setdefault(palace.id, []).append(collection)
        return collection

View on GitHub (pinned to 06cb6987f0)

Solutions

1. Provide a local palace directory (local_path set to a writable path) so the marker can anchor the palace.
2. Until remote markers ship, keep a small local state dir per palace even in server deployments.
3. Track the upstream remote-marker-store feature if pure-remote mode is required.

Example fix

# before
palace = PalaceRef(name="notes", local_path=None)
col = pgvector_backend.get_collection(palace, "notes")

# after
palace = PalaceRef(name="notes", local_path="/var/lib/mempalace/notes")
col = pgvector_backend.get_collection(palace, "notes", create=True)

if palace.local_path is None:
    palace = PalaceRef(name=palace.name, local_path=default_state_dir(palace.name))
col = backend.get_collection(palace, "notes", create=True)

def supports_pgvector(palace) -> bool:
    return palace.local_path is not None

try:
    col = backend.get_collection(palace, "notes")
except BackendError as e:
    if "local palace path" in str(e):
        raise RuntimeError("pgvector backend needs a local palace dir; set local_path")

When it happens

Trigger: Constructing a PalaceRef with local_path=None and requesting the pgvector backend: backend.get_collection(PalaceRef(name="x", local_path=None), "notes").

Common situations: Server/hosted deployments that keep palace state only in Postgres; refactoring from a local path to a remote-only PalaceRef; early experiments with headless palaces.

Related errors

• facet_counts does not support local-only filters
• pgvector does not support maintenance kind {kind!r}
• update requires at least one of documents, metadatas, embedd
• query requires exactly one of query_texts or query_embedding
• ChromaBackend has been closed

AI-assisted analysis of MemPalace/mempalace@06cb6987f0 (2026-08-15). Data as JSON: /api/errors/0715701c096a6b8f. Report an issue: GitHub.