parca-dev
symbolizer: Wrong order of segments from pprof heap profile flamegraph
Just for awareness--there might be bug in stack trace ordering in flame graph. For example, see the bytes.Split at the bottom. It comes before sum.Sum which is wrong:
It's the sum.Sum that calls bytes.Split. pprof profile obtained from HTTP looks fine. I downloaded using Parca download button and it shows correct order:
Thank you very much for reporting! Are both of these collected from the HTTP endpoints or is the Parca one from the Parca Agent?
Since this is not incorrect in the downloaded version, I suspect this is just a mistake in rendering inlined functions in the icicle graph in the wrong order.
hmm ... it seems like that already does the right thing https://github.com/parca-dev/parca/blob/790294ac3518c6edacc68645b3e678b626a6caff/pkg/query/flamegraph.go#L287-L301
Which makes it seem like either this is about asynchronous symbolization or the profile that was written was already symbolized and had an unexpected order.
Could you clarify where this data came from?
I might have reproduced this. I'll look into it in my first freecycle. https://pprof.me/f5c1a5d
Sorry, missed the discussion. It's either HTTP or benchmark --memprofile flag.