Re: Re: Heaps of read() syscalls by the postmaster

"Matthias Urlichs" <smurf@noris.net> writes:
> NB: The same benchmark revealed that CREATE TABLE (or maybe it's CREATE
> INDEX) leaks about 2k of memory.

Following up on this other point: this could simply be the new table's
relcache entry (2K seems high though).  Currently the relcache doesn't
have any procedure for discarding uninteresting entries, so once a
table is referenced by a backend that relcache entry will be there until
the backend quits or has some specific reason for flushing the entry.

I wouldn't expect a CREATE TABLE / DELETE TABLE cycle to show any memory
leak, since the DELETE would flush the relcache entry.  But creating a
few thousand tables in a row would start to eat up memory a little bit.
What is the benchmark doing exactly?

We could add a mechanism for aging relcache entries out of the cache
when they haven't been touched recently, but so far it hasn't seemed
worth the trouble...
        regards, tom lane