[python] Occasional mis-shaping issue for to_h5ad/to_anndata on obsm/varm

[johnkerl]

Using the following dataset:
https://cellxgene.cziscience.com/collections/2b02dff7-e427-4cdc-96fb-c0f354c099aa

Ingest this as a SOMA using tiledbsoma.io.from_h5ad. Then outgest using tiledbsoma.io.to_h5ad. Resulting error:

tiledbsoma._exception.SOMAError: internal error: encountered non-rectangular varm[PCs]: 22068 does not divide 46200

For this dataset:

• nobs is 32458
• nvar is 22068
• varm/PCs is 22038 tall (not 22068) x 30 wide

The to_h5ad is being unit-tested in the repo, but, for data wherein varm height matches nvar. We need to deep-dive what it means for this dataset for varm height to differ from nvar, and, if this is something that's not downright pathological about this particular dataset, find out how to handle it more elegantly.

A good solution here might benefit from the bounding-box-as-metadata topic discussed on #1445.

[johnkerl]

Long since resolved