Currently the write path (i.e any insert, delete, update) acquire collections either via the shard role API or via AutoGetCollection. 

Ideally, the "query" part - which requires sharding information (usually to skip orphans, but not necessarily) - should only be using shard role API.
For the "replica-set" part, the AutoGetCollection should be enough.

There are some places left that would benefit from having access to the acquisition (for example, the preWriteFilter) instead of accessing the CSS directly. 

A clear separation should be made so that we can determine how much work is left to correctly propagate/use acquisition wherever needed in the write path. 

Any direct access to the CSS within the write path should be replaced by acquisitions.