Tornado profiling shows calls to multiple endpoints for async endpoints

[tuukkamustonen]

How do you use Sentry?

Sentry Saas (sentry.io)

Version

1.22.2

Steps to Reproduce

Do some profiling with Tornado app, nothing special in setup.

Expected Result

Correct profiling data.

Actual Result

Traces are showing mixed stuff from multiple endpoints. For example:

The issue seems to only apply to async endpoints. Sync endpoints show correct traces.

[Zylphrex]

What you see here is the underlying execution model of most async io based frameworks. The async endpoints are run in the event loop which all happens on 1 thread, while sync endpoints are dispatched and run on a separate thread to avoid blocking the event loop.

Are you expecting to be able to view the profile per request? If so, may I ask what are you expecting to see in the profile when the event loop is handling another request?

[tuukkamustonen]

Are you expecting to be able to view the profile per request?

Yes. (Per task, I believe.)

If so, may I ask what are you expecting to see in the profile when the event loop is handling another request?

Doesn't matter too much - either blank or some pause/wait/pending/whatever. Anything that indicates that this specific request/task is not being processed (but that time is being passed).

Could be cool, perhaps, to toggle between showing those gaps and not showing them.

[Zylphrex]

Thanks for sharing :) This has come up previously but isn't something that we're actively working on at the moment as the profiler shows the data per thread.

We'll share an update once we have one

[tuukkamustonen]

Ok, just note that it's quite "broken" right now.

You click on certain transaction, e.g. "MyHandler.get" and what you get is a mixed gibberish from processing of multiple requests. Certainly should mention this in docs before you plan to fix/improve it.

[szokeasaurusrex]

@Zylphrex this issue has been open for awhile. Has this perhaps been addressed yet, or should we keep the issue open?

[Zylphrex]

This has not been addressed as there is no easy way to separate the stack samples into the originating request today.