This chapter's title implies that MongDB has a performance stability issue. It does not.
MongoDB performance stability is primarily about application load stability. And this is true of all database's I've worked with.
DBAs: you must not take responsibility EVER for application load instability. Unless it is your responsibility to be the development team's manager as well, in which case make sure you get that in writing and a pay rise to match.
When there are a lot of writes (think many GB/s per minute as a rough guide) there is variation in the MongoDB performance during checkpoints. Checkpoints are when the changes that have been made in memory are flushed to disk. If you get slow command latency appearing once per minute for a second or several, and you see disk utilization rising to 100% at the same time, this would be what it is.
One internal thing that can add real load to a given pair of shards is the movement of data (and deletion clean-up after) between shards to balance the number of chunks for sharded collections between them. ("chunks" are ranges of docs up to a certain size, default 64MB.)
One chunk being moved is barely noticeable. But when you shard a collection for the first time, or add a new shard, they will happen continuously for a long time. This means there will be batch inserts on the recipient shard(s) and deletes on the donating shard(s). A signficant impact if there is already continually heavy load from the users on the database.
Don't overlook the cost of deletes, and be aware they get asynchronously scheduled to run after. ASAP if not too much load, but it could be delayed for a long time.
Two things to be aware of regarding time-to-live indexes
If you would like a much more constant cost and can accept a policy based on size rather than a timestamp-value limit, please use capped collections.
This is not really a DB issue of course. But if your DB server is a shared physical server you may have other big applications competing for the same resources, affecting the mongod process's top processing rate.
Another form of OS instability is when the kernel or library performance is not even. This is rare given how well-written the Linux kernel is, but I can think of at least two known issues.