It appears that cache entries in memcached are not invalidated by replication packets containing changes that affect the entities they are concerned with. This is exacerbated by the fact that many cache entries are not set with a TTL.

Whilst I can understand that invalidating cache entries during the replication process is probably too complex and may not be feasible for infrastructure reasons, it should be made clear to users of a slave server that this is the case.

This could be mitigated by the provision of a 'default cache TTL' configuration setting, which, if present, sets a TTL on cache entries where no expiry is specifically passed.