Change buffer
InnoDB's mechanism for deferring maintenance of secondary indexes whose pages happen not to be in memory. An optimisation built for spinning disks, and it shows.
Write a row and every secondary index has to be updated, and those entries land in random places within their respective indexes — which on a mechanical disk means a single insert can turn into a dozen head movements. Rather than reading the page in just to modify it, the change is stored in a dedicated tree structure in the system tablespace and merged with the real index page later, when the page is read anyway or when the background threads catch up. The condition is that the index must not be unique, since a uniqueness check requires actually looking at the page. Two things made the technique less attractive: random reads cost almost nothing on SSD, and a large accumulated buffer makes restarts slow and crash recovery unpredictable. MySQL has accordingly marked the setting as going away.