Feature or enhancement

Proposal:

I'm thinking how to make#145464 more managable, and I realized that sole--jsonl is actually a good first step.

Adding--jsonl does not depend actually on streaming and will be immediately useful anyway.

The format described in#145464 for:

% cat /tmp/hello_world.pyimport timeprint("Hello, World!")time.sleep(0.1)

could be as simple as:

{"type":"meta","v":1,"run_id":"ab2e464d9d814af4b653302f03256520","sample_interval_usec":1000,"mode":"wall"}{"type":"str_def","v":1,"run_id":"ab2e464d9d814af4b653302f03256520","defs":[{"str_id":1,"value":"<module>"},{"str_id":2,"value":"/tmp/hello_world.py"},{"str_id":3,"value":"_run_code"},{"str_id":4,"value":"<frozen runpy>"},{"str_id":5,"value":"_run_module_as_main"}]}{"type":"frame_def","v":1,"run_id":"ab2e464d9d814af4b653302f03256520","defs":[{"frame_id":1,"path_str_id":2,"func_str_id":1,"line":3,"end_line":3,"col":0,"end_col":15},{"frame_id":2,"path_str_id":4,"func_str_id":3,"line":87,"end_line":87,"col":4,"end_col":27},{"frame_id":3,"path_str_id":4,"func_str_id":5,"line":196,"end_line":197,"col":11,"end_col":42}]}{"type":"agg","v":1,"run_id":"ab2e464d9d814af4b653302f03256520","kind":"frame","scope":"final","samples_total":105,"entries":[{"frame_id":1,"self":105,"cumulative":105},{"frame_id":2,"self":0,"cumulative":105},{"frame_id":3,"self":0,"cumulative":105}]}{"type":"end","v":1,"run_id":"ab2e464d9d814af4b653302f03256520","samples_total":105}

Please usejq . to format. I omitted "\n"s for brevity.

That's without fullmeta, withoutraw,heartbeat,stats,error, as well as withoutseq and timestamps, northread_id.

What's the minimum viable format here? Perhaps it could be marked even as"v":0 while it's polished?

Open Question

The format builds on the existingprofiling.sampling internals.

There's a very interesting question what should be the first class objects / semantics of the format?

My honest feeling is thatframe_def does not really describe frames but places in the code. pprof itself seems to operate more onlocation_def level.

I don't know the previous conversations to make an informed decision here.frame_def is much more simpler.

Perhaps there's a more complex discussion to be held? I'm open to any suggestions.

Running a much deeper comparison between various formats might be warranted.

Has this already been discussed elsewhere?

#145464

Links to previous discussion of this feature:

No response

Linked PRs

• gh-146256: Add--jsonl flag to theprofiling.sampling #146257