Add support for FlameGraph output

[andreasf]

• New, optional command line parameter: -f
• If enabled, writes folded call stacks to output file, instead
  of statistics
• Use with the FlameGraph tool: https://github.com/brendangregg/FlameGraph

[bdarnell]

Cool.

• Instead of -f to mean flamegraph, it's better to add --format=flamegraph or something like that, to allow room for other formats in the future.
• The command-line mode of plop.collector is not the only one (I rarely use it myself), so the part that actually writes to a file should probably be moved to a method of Collector (instead of callers like main() reaching into stack_counts).

[xmo-odoo]

The command-line mode of plop.collector is not the only one (I rarely use it myself), so the part that actually writes to a file should probably be moved to a method of Collector (instead of callers like main() reaching into stack_counts).

Maybe have the collector strictly just collect stack samples, and a different object would filter and format samples, with —format only configuring the latter object?

That means the collector would always collect a list of frames though (in case the formatter needs that), the default collector would be the one folding those into a counter.

And assuming the formatters don't alter the collector, this would allow multiple outputs from a single collection pass.

[andreasf]

Good idea. What do you think of the changes?

[andreasf]

I just fixed the last commit, in case you're wondering. After a semi-failed git reset the diff was referring to a hidden commit and thus rather confusing.

[xmo-odoo]

@andreasf the ArgumentParser conversion should probably live in a separate PR (well played on -m being a switch, I would have missed that completely and tried to set up an exclusion group of some sort).

Could use a list literal for Collector.stacks, and maybe in a future PR look into preallocating it (after profiling) to avoid a possible list resizing in the middle of stack collection.

BTW if you initialise ArgumentParser with formatter_class=ArgumentDefaultsHelpFormatter it'll automatically insert the default values after the help text: https://docs.python.org/3/library/argparse.html#formatter-class

Maybe rejiggle things around in 3 different PRs, one for the ArgumentParser conversion, one for the Formatter extraction (if @bdarnell thinks those are good ideas) and the 3rd one with just the flamegraph output?

[andreasf]

You're right on all points. I thought about doing a separate PR for the argparse changes, but since the last commit is from 8 months ago, I was too lazy to split it up. People over processes ;-)

[bdarnell]

Wow, this looks great. For future reference I'd have a slight preference for separate PRs for the PEP8 changes and the move to argparse, but it's not worth splitting it up.

The one thing I might change is actually the -m flag: it's intended to be like python's -m flag and take an argument. But in any case this is a big improvement over what was there before and going to go ahead and merge it; don't worry about changing -m unless you feel like it.

[andreasf]

Thanks!