Split this from MDL-82511

>  I suspect the del used in purge caches may also be problematic. It's listed as O(m) for hashes and we have also seen instances of this exceeding our read timeouts when it contains a large number of items (a little over 3 seconds for ~3 million). Should we also be looking to switch this to unlink as part of this tracker? https://redis.io/docs/latest/commands/unlink/

We have seen evidence of a large cache purge which runs DEL blocking redis reads in other threads / processes. We want to gather more info on this but pretty sure the solution is swapping to UNLINK which is run async on the server and does not block.

The potential downside is the latency of the deletions not being immediate which could introduce weird timing issues eg in unit tests.