runtime: deprecate SetCPUProfileRate and replace body with panic

[rsc]

Long ago we had runtime.SetCPUProfileRate and runtime.CPUProfile as the interface to profiling. And runtime/pprof was layered on top of them. At some point we realized that this was not a good enough interface, and we made the runtime export hooks directly to runtime/pprof.

We marked CPUProfile deprecated and replaced the body with a panic.

But we forgot to do the same to SetCPUProfileRate. The result is that today this function is documented as if it is useful, when in fact it is not. We should mark it deprecated and replace the body with a panic, same as CPUProfile, leaving a back-door hook for runtime/pprof.

[cherrymui]

I think SetCPUProfileRate is, albeit confusing, still functioning. It needs to be called in the right order with StartCPUProfile. Last time I looked (~a month ago), it probably prints some warnings, but it does change the rate.

The pprof package doesn't seem to have a way to set the profiling rate. Not sure if this is really a useful feature.

[rhysh]

The pprof package doesn't seem to have a way to set the profiling rate. Not sure if this is really a useful feature.

I have an unfortunate example of it being useful: Go's CPU profiling gives inaccurate results if the profiling rate is "too high". On the Linux machines I use, that's "more than 250 samples per second across the entire process". I've used SetCPUProfileRate to work around #35057, though only for the purpose of filing that bug report. It's the only practical workaround I know for that issue, and the warning it prints ("runtime: cannot set cpu profile rate until previous profile has finished." as of go1.13.3) is the main indication I see that it's not suitable for production use.

[rsc]

@cherrymui I don't see how SetCPUProfileRate can be useful. It enables the profiler but there's no way to read the profiles back out. If you later call pprof.StartCPUProfile, that errors out if the profiler is on, and otherwise it calls SetCPUProfileRate(100). What am I missing?

[cherrymui]

If you later call pprof.StartCPUProfile, that errors out if the profiler is on, and otherwise it calls SetCPUProfileRate(100).

Yeah. SetCPUProfileRate(100) then, seeing the profiler is already started, print a message and return, without changing the rate ( https://tip.golang.org/src/runtime/cpuprof.go#L65 ). It will collect profile at the rate set by the previous SetCPUProfileRate call.

[rhysh]

If you later call pprof.StartCPUProfile, that errors out if the profiler is on

That checks runtime/pprof.cpu.profiling. It does not check runtime.cpuprof.on.

An early call to runtime.SetCPUProfileRate does not cause a subsequent call to pprof.StartCPUProfile to return an error (though it does cause it to print a log line). CPU profiling data at the rate requested by the first call will be written to the io.Writer provided in the second call.

[rsc]

OK, now I see how that could possibly work. Do we actually want to support that? It seems like an accident.

[cherrymui]

The printed message suggests it is probably an accident (such that for a long time I thought it just doesn't work, until I found out the message is not actually an error).

Do we actually want to support that?

Personally, I only used it once, and I don't mind taking it out. It seems @rhysh used it only once as well. Not sure about other people's uses, though. (Maybe there are not many uses, as it appears not working.)

[tpaschalis]

In my humble opinion, the way that profiling rates might work differently between OSes/kernel settings when you start nearing millisecond-scale resolutions is an issue as well.

~I'd like to open a CL if there's no objection; I'll maintain the 100Hz hardcoded value and keep changes at a minimum.~
Edit: I'm really sorry for opening a CL before consensus was reached in the discussion.

[gopherbot]

Change https://golang.org/cl/253398 mentions this issue: runtime: deprecate SetCPUProfileRate

[rogerlucena]

I have just opened a separate Issue (#42502) defending runtime.SetCPUProfileRate, or at least its functionality in the context of pprof.StartCPUProfile, explaining one use case on which it was important along with some supporting links as well.

[aclements]

Given that this still needs a decision and there's a proposal for improvements working through the proposal process, I'm kicking this to 1.17. (Let me know if I misunderstood.)

[mknyszek]

Since this issue is very much related to #42502 and although that issue did not make a decision on what to do here (AFAICT), that issue is in the backlog. So I'm going to move this to the backlog to align the two, under the assumption that nothing is going to happen here before the freeze (May 1st).

If that's incorrect, please let me know or feel free move it back. Just trying to clean out the milestone a bit.

[mknyszek]

Moving this to backlog, though the discussion here seems related to @rhysh's recent work.

CC @prattmic

[BrennaEpp]

I'll add a comment here in case someone else gets bitten by this before (if) this gets deprecated:

Calling runtime.SetCPUProfileRate() also turns the profiling on, before the call to pprof.StartCPUProfile. I was getting extra functions that should not have been in the profile and it turns out runtime.SetCPUProfileRate() was the culprit.