Heikki Linnakangas <heikki(dot)linnakangas(at)enterprisedb(dot)com> wrote:
> But the question Tom raised about doing this even for
> non-serializable transactions is more serious..
This particular call from heapam.c went in much later than most of
the code, since we had trouble proving that anything needed to be
done there in the first place. As such, it wasn't there during the
many benchmarks which were done along the way. None of those showed
any performance degradation for other isolation levels due to the
SSI code.
Dan and I are sorting out the best way to isolate this particular
call to assess the performance impact. We also think there should
probably be an additional test or two in the regression tests to
exercise this area. I think we should wait for the results of
performance tests and (if warranted by benchmark results) profiling
before we start talking about possible ways to address this. The
evils of premature optimization and all.
Dan came to the same conclusion a fix before seeing Heikki's post on
the issue, so it would be good get that committed.
More to follow, once we have benchmark results on this and a
suggested patch for regression test coverage.
-Kevin