Simon Riggs <simon@2ndquadrant.com> writes:
> My view would be that the LockMgrLock is not relevant for all workloads,
> but I want even more to be able to discuss whether it is, or is not, on
> an accepted basis before discussions begin.
Certainly. I showed the evidence that it is currently a significant
problem for pgbench-like workloads, but pgbench is of course not
representative of everything.
My feeling about it is that different workloads are going to expose
different weak spots, and so as long as a given test case isn't
obviously artificial, whatever bottleneck it exposes is fair game
to work on. pgbench seems reasonably representative of a class of
applications with relatively short transactions, so I don't doubt that
if pgbench has a problem with LockMgrLock contention, there are real-
world cases out there that do too.
regards, tom lane