Four additions to replications internals README in the MongoDB upstream are relevant reading.

• SERVER-43780 Describe lifetime of unprepared transactions

• SERVER-43781 Describe the prepare state and lifetime of a prepare transaction

• SERVER-43782 describe transaction oplog application

• SERVER-43783 Describe how recovery works with transactions

@Former user The test we'd like to do here is:

1. Prepare 2+ shard cluster

2. Create collection that is spread over 2+ shards. E.g.: use a shard key of the "_id" field; Insert doc with "_id": 99; and another with "_id": 101; sh.splitAt({"_id": 100}); sh.moveChunk("test.foo", {"_id": 101}, "shard2") (whatever the other shard is).

3. Start a transaction, make it a multi-shard transaction by deleting both docs "_id": 99 and 101. (It could be inserts or updates, but I think deletes is simpler to check). But don't commit the transaction

4. Without committing the transaction, stop and immediately make the backup. Important to finish the backup before the 60-second transaction timeout finishes.

5. Let the 60 second expire.

6. Run pbm restore.

7. Allow 60 seconds to pass.

8. Confirm the transaction writes didn't happen. E.g. if there were deletes confirm that docs with id 99 and 101 exist.

N.b. We're not 100% sure that this doesn't already work. That is, we're not sure if (or if not) a v4.2 primary node that is having these "prepare": true oplog docs applied via the applyOps command will buffer rather than immediately apply them.

Two ways forward: read the code, or do the test.