> On Fri, Apr 23, 2021 at 01:56:57AM -0700, Mitar wrote:
> Hi!
>
> On Thu, Apr 15, 2021 at 12:11 PM Dmitry Dolgov <9erthalion6@gmail.com> wrote:
> > > My point was that for JSON, after validating that the input is
> > > syntactically correct, we just store it as-received. So in particular
> > > the amount of whitespace in the value would depend on how the client
> > > had chosen to format the JSON. This'd affect the stored size of
> > > course, and I think it would have an effect on compression time too.
> >
> > Yes, I got it and just wanted to confirm you were right - this was the
> > reason I've observed slowdown trying to reproduce the report.
>
> Thank you for trying to reproduce the report. I did a bit more digging
> myself and I am still confused.
>
> ...
>
> So I do not know what is happening and why you cannot reproduce it.
Could you maybe get a profile with perf for both cases? Since they're
executed within a single backend, you can profile only a single pid.
Having a reasonable profiling frequency, --call-graph dwarf and probably
limit events to only user space with precise tagging (cycles:uppp)
should give an impression what's going on.