Reduce memory usage for long duration tests

[msaf1980]

Feature Description

Statistic require a lot of memory when run long duration tests. Any way for have statistic and thresholds and reduce memory usage ?

Suggested Solution (optional)

May be a flag (may be cycles), than allow to run N executions (by default is 1) in cycle.
Statistic are flushed on every cycle.

Draft in https://github.com/msaf1980/k6/tree/iterate_loop
But required of change some internals, so maybe not complete.

Already existing or connected issues / PRs (optional)

No response

[na--]

We have plans (e.g. #763, #1321) for reducing the memory usage that the built-in Trend metrics require, but I am not sure if something like what you propose is a viable option to include in the k6 core 🤔 At best, maybe we can add a new instance.flushMetrics() method to the k6/execution module that can be called from the script, but I imagine there will be quite a lot of unexpected issues even from that... 😕

For now, the official recommendation for long-running tests is to run k6 with --no-summary and --no-thresholds and stream the metrics to an external output (e.g. json, csv, k6 cloud, etc.) with the --out option. Then k6 won't keep all of the metrics data internally for the duration of the test, it will flush it to the external output at regular intervals. Can you explain why this is insufficient for your use case?

[msaf1980]

If no thresholds - I can't auto-stop on testing system degradation.
Long duration tests - usually regression tests. So auto-stop is really needed for this cases.

I agree that's my solution is ugly and not a viable option to include in the k6 core. It's example for oen way for reduce memory usage. Use histograms (like #763) is a better solution. And may require minimal refactoring ?

[na--]

Long duration tests - usually regression tests. So auto-stop is really needed for this cases.

You could monitor the response times in the script by comparing the Response.timings.duration, for example, and use the test.abort() method in k6/execution to stop the test (introduced in the recently released k6 v0.36.0)

And may require minimal refactoring ?

That's the hope, though the last time I looked at it, the Go HDR histogram library was unmaintained, so we might need to fork and polish it... 🤷‍♂️

[metaturso]

Any solutions or workarounds for long tests that run out of memory because of Trend on K6 cloud?

[mstoykov]

@metaturso k6 cloud already does the exact thing that is recommended in this issue - so there is no problem with Trend in cloud.

If you have problems please contact support