At the moment a snapshot isolation transaction always keeps the same snapshot, and a read-committed transaction updates it's snapshot each time a new operations starts while there are no positioned cursors.

For some MongoDB operations they would like a semantic between those two. Where a transaction generally keeps the same snapshot, but can be moved forward periodically. Specifically for transactions on a node that are running at local or majority, which will be reading with a snapshot instead of a timestamp. They could be long running transactions, but are OK with read-committed semantics, where the snapshot gets updated when MongoDB chooses to yield the transaction.

This will probably involve adding a new API transaction_update_snapshot or similar, that can be used for either snapshot or read-committed transactions to trigger switching to a new snapshot.

It's possible this won't be necessary if MongoDB can use existing read-committed semantics.