feat: add caching to GapicCallable

[daniel-sanche]

_GapicCallable currently does a lot of work on each call, re-building each wrapped function, using a lot of calls to helper functions. This cost is added to every single rpc, so it can really add up

This PR does the following optimizations

• removes the _apply_decorators and _is_not_none_or_false to build the wrapped call more directly.
  • This seems to make it ~10% faster
• add a new helper that builds a wrapped call using a timeout and retry object, and then cache the result with @lru_cache
  • This seems to make it 50% faster
  • In practice, I think it's safe to assume most calls will be re-using timeout and retry values
  • I currently have the cache size of 4, but this can be changed

---

Benchmark:

from google.api_core.gapic_v1.method import _GapicCallable
from google.api_core.retry import Retry

callable =  _GapicCallable(lambda *a, **k: 1, retry=Retry(), timeout=1010, compression=False)

from timeit import timeit
timeit(lambda: callable())

Before: 20.43s
After: 9.48s

BEGIN_COMMIT_OVERRIDE
chore: add caching to GapicCallable
END_COMMIT_OVERRIDE

[parthea]

This seems to make it 50% faster

Thanks for fixing this!

Please could you add a simple benchmarking presubmit, similar to the test that you ran manually, to avoid future regressions in performance?

[parthea]

Assigning back to @daniel-sanche to resolve the presubmit failure

[daniel-sanche]

Hmm good point, the benchmark result will be machine-specific, and I was doing my tests locally instead of with the CI workers.

I guess I'll have to find an assertion value that works well for the CI nodes, and I'll add a comment explaining that it may be flake on slower hardware. Or let me know if you have other suggestions for how to approach this

[parthea]

Can you set it high enough that we don't get flaky results, but low enough that we can detect performance regressions.

Perhaps set the threshold to 0.4 for now and create an issue in https://github.com/googleapis/python-api-core/issues to add a proper benchmarking test ? I believe @ohmayr started looking into a benchmarking presubmit so please tag him on the issue.

[daniel-sanche]

Sure, I opened an issue to track this here: #616

I adjusted the value to 0.4. Feel free to merge it with that number, but I suspect we can find a lower value that still avoids flakiness. Let me know if you want me to do some investigation

[parthea]

@vchudnov-g Please could you review?

[vchudnov-g]

Minor code comment, and an idea about tightening benchmarks.

[vchudnov-g]

Do you mean to self-reference this PR?

[daniel-sanche]

It was intentional, to give the context on this test. But on second thought, git blame should be enough. Removed

[vchudnov-g]

Idea: If the assertion fails, print both the actual time it took and enough platform information so that in the future we can add the right threshold for the platform. The latter would be something like this

platform_threshold = { "foo": 0.2, "bar": 0.6 }
current_platform = ...
...
assert avg_time < platform_threshold.get(current_platform, 0.4)

In fact, you could implement platform_threshold now, and start with whatever your current machine is.

[daniel-sanche]

That's an interesting idea, but it's not completely clear to me what we'd need to capture for the platform. Number of CPUs? Architecture? OS? Let me know if you have thoughts

We already have #616 to track improving this though, so if it's alright with you, I'll merge this as-is and we can discuss follow-up there