[benchmarker] Only inspect current step's operations when checking criteria

[barroco]

This PR improves the criteria evaluation process for large number of operations by only considering operations of interest for the current evaluated step.

[mickmis]

I don't understand why the operations are ordered within a 60 seconds tolerance because they are recorded as they complete?
Also: I think making this assumption here of logic defined elsewhere is not good. If that assumption changes (and I don't think it is guaranteed by anything?), this will break this logic here.
Maybe replace operations with a data structure that has some guarantees on the sorting?

[mickmis]

Have you considered instead doing something like summarize_and_report_step?

    step_ops = [
        op
        for op in operations
        if op.completed_at.datetime >= step_start_time
        and op.completed_at.datetime <= step_end_time
    ]

[BenjaminPelletier]

Is the purpose of this PR to reduce the processing time required by check_stability_criteria by reducing the number of operations to iterate over when checking various criteria? If so, I would probably expect a search for the cutoff index to use a more performant algorithm (like binary search) and then just pass the cutoff/start index as an additional parameter to check_stability_criteria rather than creating an entirely new list to pass to check_stability_criteria. Though I am surprised that the check_stability_criteria processing time would be appreciable, even with a pretty large volume of operations.

[barroco]

Thank you @BenjaminPelletier and @mickmis, let me re-open this issue with more evidence so we can better address root cause.