runtime/trace: execution trace doesn't include pprof labels

[nsrip-dd]

Labels set through runtime/pprof.SetGoroutineLabels don't show up in execution traces. In the CPU and goroutine profiles, samples include the labels from the sampled goroutine. Labels are quite useful for associating CPU usage with trace span IDs, HTTP endpoints, user IDs, etc.

It would be useful to have these labels in execution traces as well. For example, since execution traces now include timestamped CPU profile samples, trace events could be associated with an application-level trace span ID using labels to get a breakdown of on-CPU and off-CPU time for an individual request. We can further bridge runtime/pprof and runtime/trace by making labels, which are already available in some profiles, available in traces as well.

Arguably there's already support for something similar to this through trace.WithRegion and trace.NewTask. However, these values are one-dimensional and seem to be intended for breaking down a process into steps (read an event stream, query a database, etc...). To record additional contextual information, like a specific distributed trace ID or user ID, you'd need to do something ad-hoc like pack that information into the task name, or show it in a designated trace.Log category.

Perhaps profile labels could be added to the trace as a new event that is emitted whenever the labels for a goroutine are updated? (See this comment)

See also #41554

[dr2chase]

@mknyszek @prattmic

[gopherbot]

Change https://go.dev/cl/446975 mentions this issue: runtime/trace: record goroutine labels in execution trace

[mknyszek]

Thanks. I think this is addition makes sense. How should these labels be surfaced in the current UI?

[nsrip-dd]

I was picturing that the label event could be connected to every other event for the same G while the G had those labels. Ideally you'd then be able to search for events based on the labels. Possibly you'd also have an instantaneous event (like a CPU sample) for when the labels for a goroutine are changed.

This seems possible in principle through references in the Chrome trace format. I've experimented a bit and have yet to find a good way to do that, though. It's not clear to me whether resolving references, as described in that doc, is supported in the trace UI. I've tried it, but it didn't work like I expected (id_ref values didn't get replaced with the arguments for the label event I added). I can't tell if Perfetto can use that information, either.

I also tried creating an async event for each time a set of labels is active for a goroutine. I can get something to show up in the UI as a span for the duration the labels existed. However, the span shows up above the row of events for the P where the event originated, even though the goroutine with those labels can move between Ps.

I'm not too familiar with the trace UI, though, so there may be some obvious way to do what I'm picturing that I just don't know about :)

We could also have a command-line option to filter the trace event stream by specific label sets before its read into the UI.

Also, if we have labels in the event stream, the labels could also be added to the pprof profiles that are generated from the trace.

[felixge]

We could also have a command-line option to filter the trace event stream by specific label sets before its read into the UI.

Also, if we have labels in the event stream, the labels could also be added to the pprof profiles that are generated from the trace.

+1 to both of these options for now.

In the future Perfetto should be able to display and filter events based on labels in the UI if we go in that direction. But I haven't tested it.

[nsrip-dd]

I've mailed https://go-review.googlesource.com/c/go/+/456239 to demonstrate adding pprof labels in a trace event stream to pprof profiles derived from the stream. You'll get profiles with the labels attached to samples, which you can see in the pprof graph UI like in the screenshot below from a trace-derived network latency profile. You can also filter by labels (a.k.a "tags") with options like go tool pprof -tagfocus=....

I'm still working on a demonstration of filtering/focusing on events within the UI based on labels, but I'll mail something once I get a good solution.

[mknyszek]

We could also have a command-line option to filter the trace event stream by specific label sets before its read into the UI.

Also, if we have labels in the event stream, the labels could also be added to the pprof profiles that are generated from the trace.

Both of these sound good to me as well.

You'll get profiles with the labels attached to samples, which you can see in the pprof graph UI like in the screenshot below from a trace-derived network latency profile.

Very cool! Slightly tangential, but it reminds me that I think eventually we're going to want to document these features at like, go.dev/doc/diagnostics (you don't need to do that now). I'll go scribble that down somewhere else...