Re: [PATCH] Identify LWLocks in tracepoints

Hi,

On 2021-04-12 14:31:32 +0800, Craig Ringer wrote:
> * There is no easy way to look up the tranche name by ID from outside the
> backend

But it's near trivial to add that.

> It's annoying that we have to pay the cost of computing the tranche name
> though. It never used to matter, but now that T_NAME() expands to
> GetLWTrancheName() calls as of 29c3e2dd5a6 it's going to cost a little more
> on such a hot path. I might see if I can do a little comparison and see how
> much. I could add TRACE_POSTGRESQL_<<tracepointname>>_ENABLED() guards since we
> do in fact build with SDT semaphore support. That adds a branch for each
> tracepoint, but they're already marshalling arguments and making a function
> call that does lots more than a single branch, so that seems pretty
> sensible.

I am against adding any overhead for this feature. I honestly think the
probes we have right now in postgres do not provide a commensurate
benefit.

> (It'd be possible to instead generate probes for each GUC at compile-time
> using the preprocessor and the DTRACE_ macros. But as noted above, that
> doesn't currently work properly in the same compilation unit that a dtrace
> script-generated probes.h is included in. I think it's probably nicer to
> have specific probes for GUCs of high interest, then generic probes that
> capture all GUCs anyway.)
>
> There are a TON of probes I want to add, and I have a tree full of them
> waiting to submit progressively. Yes, ability to probe all GUCs is in
> there. So is detail on walsender, reorder buffer, and snapshot builder
> activity. Heavyweight lock SDTs. A probe that identifies the backend type
> at startup. SDT probe events emitted for every wait-event. Probes in elog.c
> to let probes observe error unwinding, capture error messages,
> etc. [...] A probe that fires whenever debug_query_string
> changes. Lots. But I can't submit them all at once, especially without
> some supporting use cases and scripts that other people can use so
> they can understand why these probes are useful.

-1. This is not scalable. Adding static probes all over has both a
runtime (L1I, branches, code optimization) and maintenance overhead.

> (Those can also be used with systemtap guru mode scripts to do things
> like turn a particular elog(DEBUG) into a PANIC at runtime for
> diagnostic purposes).

Yikes.

Greetings,

Andres Freund