Re: [COMMITTERS] pgsql: Fix an O(N^2) problem in foreign key references.

I wrote:
> Alvaro Herrera <alvherre(at)2ndquadrant(dot)com> writes:
>> Would it make sense to remove the only the few oldest entries, instead
>> of all of them? As is, I think this causes a storm of reloads every
>> once in a while, if the number of FKs in the system is large enough.
>> Maybe on a cache hit we could push the entry to the head of the list,
>> and then remove N entries from the back of the list when the threshold
>> is reached.

> (FWIW, it might take less code to put the recently-used entries at the
> back of the list. Then the loop in InvalidateConstraintCacheCallBack
> could just invalidate/delete entries if either they're targets, or
> the current list length exceeds the threshold.)

Specifically, we could change it as per attached. But this adds some
cycles to the mainline usage of fetching a valid cache entry, so I'd
want to see some evidence that there are use-cases it helps before
applying it.

regards, tom lane