Re: Thousands of schemas and ANALYZE goes out of memory

Sorry, I thought you had already committed this change.

It definitely fixes the worst CPU bottleneck.  To ANALYZE (manually, no table name specified) a database with 210,000 tables goes from 1280 seconds to 120 seconds.  It does not meaningfully change the memory usage, as tabstat is not the bottleneck on memory usage, CacheMemoryContext is.  It takes about 1.5GB either with this patch or without it.

I'm having trouble coming up with an adverse scenario to test for performance regressions.  Perhaps a platform where IPC is pathologically slow?  I don't have one of those at hand, nor even know what such a one might be.  And presumably such a platform would be unsuited to the task of running a very large database anyway.

By the way, this issue can now be partially worked around using "vacuumdb -Z -j8".  The quadratic CPU issue is fixed because with the parallel option it issues a series of "ANALYZE table" commands rather than one unqualified command for the whole database.  And the memory usage originally reported is ameliorated because each backend gets own address space. The total amount of memory used remains the same, but the 32 bit limit doesn't apply to the aggregate, only to each process.

Cheers,

Jeff