Prototype: Rudimentary stack profiler

[igorwwwwwwwwwwwwwwwwwwww]

We have a relatively large jsonnet code base (source), and I was trying to profile slow jsonnet invocations.

The golang-level profile via JSONNET_CPU_PROFILE wasn't very helpful in this regard. And tracing individual calls by adding std.trace was quite combersome.

I prototyped a little sampling stack profiler. It is not very sophisticated. It samples stacks for 1% of calls to EvalInCleanEnv. Those stacks can then be visualised as a flamegraph.

Usage:

JSONNET_STACK_PROFILE=stacks.out jsonnet main.jsonnet
cat stacks.out | flamegraph.pl --hash > flamegraph.svg

There's plenty of room for improvement:

• Making the code nicer
• Including a time component (since we currently assume each eval takes a fixed amount of time)
• Buffering stacks in memory with their counts and writing out aggregated counts in the end

That said, the results I got from it so far look quite usable.

It's unlikely that I'll have much time to work on this and develop it into a fully-fledged solution.

I wanted to share what I have in case it's useful for others, or in case someone feels inspired to take the idea and develop it further.

[sbarzowski]

This is very interesting. I was thinking about doing something like that myself, but didn't have much time recently.

I'd like to refactor it slightly (eliminate globals, extract all profiling logic to a separate function, so that in the normal flow it's just one check and one function call). Otherwise it looks good.