MongoDB doesn’t handle locking too well. It’s getting better in new versions (3.2 looks to be a strong contender) with collection and object level locking, but it’s still not great.
In an environment I was working on some time back, there was a need to remove lots of data on a regular basis to keep the size of the database manageable. The developers had implemented a snazzy maintenance system that NULL-ed the large fields that contained payloads for the data in order to save space after their utility was over. This maintenance system was run once a day and locked the collections/databases for that time; the brief period of locking was considered acceptable in this environment.
However, this didn’t free up any space in the system, because of the way MongoDB, (at the time 2.6, but 3.0 seems to have the same issue without Wired Tiger) NULL-ing the field did free up the space, but did not actually allow for new data to be inserted into its place. The only way to free the space was to take the database offline and repair or compact it once a month. Not an option.
Because of the needs of the system, we were able to delete the whole object and reinsert it without the payload, but this caused excessive overhead and just turned our disk space problem into a CPU and memory one.
It was decided that since we did nightly backups, deleting the entire object was the way to go, but we’d need to do it manually until development could make the appropriate changes to the maintenance system. The problem then arose of deleting at least a week’s worth of data all in one go. This could take several hours and lots of locking if done all at once.
The solution was to have a process that deleted a set of the data from the database, pause for a moment to let other systems insert/delete/update, and it couldn’t require a large development cost (ie the DBA [spoiler: me] had to do it).