After SERVER-55033 there is still one situation where OpObserverShardingImpl::shardObserveDeleteOp will not throw upon attempting to delete a document in a txn on a chunk that has moved since the atClusterTime. This situation happens when the collection has a nested shard key.

For instance:

OpObserverImpl::getDocumentKey() calls dotted_path_support::extractElementsBasedOnTemplate() and would return an object {"x.y": 10}.

OpObserverShardingImpl::shardObserveDeleteOp() calls ShardKeyPattern::extractShardKeyFromDocThrows() which would fail to find the shard key because there is no field call "x" in the object {"x.y": 10}. There is only a field called "x.y".

Possible solutions could be:
a) Make shardObserveDeleteOp() receive the full document, so it can call extractShardKeyFromDoc() itself. This would likely cause a performance regression since we would need to save a copy of the full document in aboutToDelete().
b) Make shardObserveDeleteOp() receive the shard key (without the _id, unless _id was part of the shard key pattern) so that it can skip calling extractShardKeyFromDoc*() itself. shardObserveDeleteOp() also needs to receive the _id separately so it can be given to MigrationChunkClonerSourceLegacy::onDeleteOp()
c) Make OpObserverImpl::getDocumentKey set the shard key BSONObj as nested object (if the shard key was nested), instead of setting it as dot-notation.