Even faster stack traces for _connection.py

[neoncube2]

This PR optimizes _connection.py even further. This further fixes#2744.

In the current code, we:

1. Retrieve the entire stack, by calling getattr(task, "__pw_stack__", inspect.stack(0))
2. Parse the entire stack
3. Throw away all of the parsed stack except for the first relevant frame, unless we're tracing (e.g. unless self._tracing_count > 0)

This PR makes it so that we only retrieve and parse as much of the stack as necessary.

Test code:

asyncdeftest_speed_test(page: Page) ->None: start_time=time.time() foriinrange(0, 1000): awaitpage.evaluate('() => 1 + 1') print(time.time() -start_time)

Before this PR: 4s
With this PR: 1.75s

Speed improvement: 2.3x

I'd like add tests for when self._is_internal_type = True, but I'm not sure the best way to go about constructing an internal type and then triggering an exception in it. (Any suggestions?)

Similarly, I'd like to add tests for when getattr(task, "__pw_stack__") returns a value, but again, I'm not sure when this occurs. If anyone has suggestions, please let me know! :)

[neoncube2]

Ready to rerun tests

[neoncube2]

Probably best to wait until #2840 is merged, and then I can run checks locally.

[mxschmitt]

Just thinking out loud - why does it need to be an Iterator in order to get the faster stacktraces? I assume that we call it only when we actually need it in line 546 and 566?

Also should we maybe use this similar function in playwright/_impl/_sync_base.pyplaywright/_impl/_sync_base.py:108` ? Then sync users - what the majority is - would benefit as well?

[neoncube2]

As to your question: In the original code, when we were calling inspect.stack(0), it would immediately walk the entire stack (essentially, it would call frame.f_back in a loop until there were no more frames). This is pretty expensive. By using an iterator, we delay walking the stack until we need it, and then we can only walk part of the stack (because frame.f_back is only called when we need another frame)

Basically: Calling frame.f_back is more expensive than one might think.

Thanks for the heads up about playwright/_impl/_sync_base.pyplaywright/_impl/_sync_base.py:108! :) No wonder I couldn't find any information about __pw_stack__... I thought it was being set by asyncio, but it's actually being set by this package! XD

Let me take a look.

[neoncube2]

@mxschmitt Now that I'm working on this again, I see what you mean about the iterator. I think that once I merge the logic in _sync_base.py and _connection.py, this PR should be a lot cleaner (and faster!) :)

[neoncube2]

@mxschmitt I'm still working on this, but I've split off a few simple and easy to review optimizations into #2844

[neoncube2]

@mxschmitt While continuing to work on this, I've discovered something interesting.

1. In _sync_base.py, we set __pw_stack__ and __pw_stack_trace__
2. In _connection.py, we check to see if the current task has __pw_stack__ or __pw_stack_trace__ set, and if not, we generate a new stack trace to use.

Interestingly, for all of the tests in the tests\ folder, __pw_stack__ and __pw_stack_trace__ are always None whenever they're retrieved in _connection.py. This makes me wonder if we should stop setting __pw_stack__ and __pw_stack_trace__ in _sync_base.py (because these values are never being used) or if we just need to better tests to cover this case.

Any thoughts?

[mxschmitt]

I wonder when it was introduced if it was any different back then: c05cad2#diff-291261f4050ff453e2c661edf96c77af65a779b75815708d924809ee49d5dc64 - would be great to find out if it was used back in the days and not anymore or if our tests don't have the coverage.

[neoncube2]

@mxschmitt Similarly, it might be nice to know if sending apiName and internal as part of the metadata is necessary or was just added because it was nice to have:

playwright-python/playwright/_impl/_connection.py

Line 351 in e87e340

metadata={

[neoncube2]

I think that probably the best way to move this PR forward would be to modify the monkeypatch by @SichangHe (#2744 (comment)), as it looks like it preserved the best of all worlds and has the least chance of breakage :)

I'm thinking it would probably make sense to open a new PR for that, but let's keep this PR open for now.