I'm sure it's been done before, but right now I feel smart about the recognition.
I want to make postgres' buffer cache take into account whether previous accesses where synchronous or asynchronous:
Conversation
I.e. tag pages that needed to be read in synchronously with that fact (or reversely ones that were prefetchable).
That's somewhat similar to the old idea of (IIRC) to just always keep inner index pages around more aggressively.
2
But I hope it'd be more generic (i.e. also keep randomly accessed table pages around) and resilient (doesn't trigger due to vacuum scanning the index).
1
Replying to
I agree with this. I just really like the framing from the LRU-K paper, which is very simple but totally convincing -- see example 1.1: cs.cmu.edu/~christos/cour