Close tdbMatrix and tdbMatrixWithIds Array's when we have nothing left to to read

[jparismorgan]

What

When I tested with the new TileDB-Py 0.31.0 wheels (downloaded manually from this run: https://github.com/TileDB-Inc/TileDB-Py/actions/runs/10096606074), I saw a new error when running Flat and IVF Flat:

>           Index.clear_history(uri=index_uri, timestamp=19)

apis/python/test/test_ingestion.py:720: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
/opt/homebrew/anaconda3/envs/TileDB-Vector-Search-2/lib/python3.9/site-packages/tiledb/vector_search/index.py:641: in clear_history
    tiledb.Array.delete_fragments(db_uri, 0, timestamp)
libtiledb.pyx:1337: in tiledb.libtiledb.Array.delete_fragments
    ???
libtiledb.pyx:337: in tiledb.libtiledb._raise_ctx_err
    ???
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

>   ???
E   tiledb.cc.TileDBError: TileDB internal: [Array::set_array_closed] May not perform simultaneous open or close operations.

This happens from code like this:

def test_ingestion_timetravel(tmp_path):
    for index_type, index_class in zip(INDEXES, INDEX_CLASSES):
        index_uri = os.path.join(tmp_path, f"array_{index_type}")

        data = np.array([[1.0, 1.1, 1.2, 1.3], [2.0, 2.1, 2.2, 2.3]], dtype=np.float32)
        default_result_d = [[np.finfo(np.float32).max], [np.finfo(np.float32).max]]
        default_result_i = [[np.iinfo(np.uint64).max], [np.iinfo(np.uint64).max]]

        # We ingest at timestamp 10.
        index = ingest(
            index_type=index_type,
            index_uri=index_uri,
            input_vectors=data,
            index_timestamp=10,
            num_subspaces=2,
        )
        
        Index.clear_history(uri=index_uri, timestamp=19)

and is coming from load_as_matrix():

flat_index.py
           self._db = load_as_matrix(
                self.db_uri,
                ctx=self.ctx,
                config=config,
                size=self.size,
                timestamp=self.base_array_timestamp,
            )

which does:

module.py
def load_as_matrix(
    ...    
    if dtype == np.float32:
        m = tdbColMajorMatrix_f32(ctx, path, 0, None, 0, size, 0, timestamp)
    ...
    m.load()
    return m

I fixed a similar bug for type-erased matrixes by having tdbPartitionedMatrix automatically close Array's when done reading (#448). So here we do the same thing for tdbMatrix and tdbMatrixWithIds.

Testing

• Existing unit tests pass.
• Added new unit tests that we can call load() many times and nothing bad happens.
• With these changes test_ingestion.py now passes with TileDB-Py 0.31.0 wheels.

[NikolaosPapailiou]

Thanks for debugging this!