With the changes committed under
SERVER-47699 (auto-yielding), it should be safe to get rid of batched deletes and - instead - simply delete documents belonging to a whole range in one shot. Some experiments showed that deleting hundred of thousands of documents takes less than a minute and the yielding should theoretically prevent any kind of starvation.
Keeping the current behavior of the range-deleter can easily result in the balancer starving waiting for a specific range deletion to drain (the more range deletions, the more likely this happens - because the task gets re-enqueued behind all the other ones).
Objective of this ticket is to decide whether we need to:
- Get rid of the batching policy
- Keep it, but drastically increasing rangeDeleterBatchSize (e.g. set it to 100K). This may be useful in case a user wants to throttle orphan deletions in favor of CRUD operations.