AWS Lambda support

[basepi]

Remaining:

• All the context fields -- waiting on spec to be merged AWS Lambda spec apm#470

Related issues

Ref #428

[ghost]

💚 Build Succeeded

the below badges are clickable and redirect to their specific view in the CI or DOCS

Expand to view the summary

Build stats

• Start Time: 2021-10-18T14:57:40.848+0000

• Duration: 29 min 14 sec

• Commit: 744a39e

Test stats 🧪

Test	Results
Failed	0
Passed	10127
Skipped	8977
Total	19104

💚 Flaky test report

Tests succeeded.

🤖 GitHub comments

To re-run your PR in the CI, just comment with:

• /test : Re-trigger the build.

[beniwohli]

Regarding time_to_perf_counter, it generally LGTM. I worry however that especially in an environment like Lambda, where the runtime is put to sleep and woken up on a regular basis, clock skew could be a phenomenon that is more regular than usual. But that's just a hunch...

[basepi]

Regarding time_to_perf_counter, it generally LGTM. I worry however that especially in an environment like Lambda, where the runtime is put to sleep and woken up on a regular basis, clock skew could be a phenomenon that is more regular than usual. But that's just a hunch...

I don't think that's actually an issue in this case. perf_counter is designed to be monotonic and take sleep into account. time.time() is designed to be "real-world" (take sleep into account) the same as perf_counter, and will be monotonic absent a system clock change. I can't imagine we will have to deal with system clock changes in AWS but if we do, we might have a weird outlier but there's not much we can do about that.

Edit: I am going to do some testing around perf_counter and the freezing that lambda does. I think it should still work, but if it doesn't we definitely want to make sure we take these measurements each invocation.

[basepi]

Did some testing today and the conversion measurement we take at the beginning of the cold start appears to stay accurate throughout multiple invocations, with freezes in between.

I'm still going to implement some re-measurement for long-running processes but I don't think we need to measure every time with Lambda like I was fearing.

[beniwohli]

Great work, looks good!

I think we need a better solution for passing state along in the capture_serverless object. Especially when used as a module-level decorator, the chance of things going awry are pretty big IMO.