lower level api

[imposeren]

I'm currently using your module to benchmark some external api performance. Currently this is implemented using severala coroutines running concurrent parallel taks where one coroutine calls some "processor" and this one processor should be benchmarked. using weave here is possible but problematic bacause of complex "setup" and concurrency issues. I have a workaround to solve my problem:

def some_test(benchmark, event_loop):
    benchmark._mode = 'benchmark(...)'
    stats = benchmark._make_stats(1)
    primary_task = event_loop.create_task(primary_task_factory(data_source, duration_callback=stats.update))

@asyncio.coroutine
def process_docs(
        docs_iterator, max_frequency, max_amount, processor,
        doc_index_in_args=1,
        duration_callback=None,
        *args, **kwargs):
    # ...
    p_t0 = default_timer()
    processor(*args, **kwargs)
    processor_time = default_timer()-p_t0
    total_processor_time += processor_time
    if duration_callback:
        duration_callback(processor_time)

It would be great to have some "legal" and documented way to do similar things. I had to read pytest-benchmark source code to do this. I think that public api for similar tasks should be exposed:

def test_something(benchmark):
    duration_handler = benchmark.manual()
    for x in something():
        foo()
        baz()

        duration_handler.start()
        some_processing_to_benchmark()
        duration_handler.end()

[ionelmc]

How come you don't wrap the processor function instead? Does the processor have side-effects?

[imposeren]

I'm sorry my example was incorrect it's more like:

@asyncio.coroutine
def process_docs(
        docs_iterator, max_frequency, max_amount, processor,
        doc_index_in_args=1,
        duration_callback=None,
        *args, **kwargs):
    for doc in docs_iterator:
        args[doc_index_in_args] = doc

        p_t0 = default_timer()
        processor(*args, **kwargs)
        processor_time = default_timer()-p_t0
        total_processor_time += processor_time
        if duration_callback:
            duration_callback(processor_time)
        for x in asyncio.sleep(wait_t):
            yield x

def test_something(benchmark, event_loop):
        benchmark._mode = 'benchmark(...)'
        stats = benchmark._make_stats(1)
        def done_callback(future):
            for task in unlimited_tasks:
                if not task.done():
                    task.cancel()
            event_loop.stop()

        limited_primary_task = event_loop.create_task(primary_task_factory(es_data, duration_callback=stats.update))
        # primary_task_factory will call `process_docs` internally
        limited_task.add_done_callback(done_callback)

        unlimited_tasks = []

        for i in range(SECONDARY_TASKS_CONCURRENCY):
            unlimited_tasks += [
                event_loop.create_task(secondary_task_factory(es_data))
                for secondary_task_factory in secondary_tasks_factories
            ]

        try:
            event_loop.run_forever()
        finally:
            event_loop.close()
            for task in [limited_task] + unlimited_tasks:
                if (not task.cancelled()) and task.exception():
                    raise task.exception()

Processor is invoked multiple times (see coroutine in my initial message): this will cause exception in benchmark when self._mode is checked.

Wrapping function with for loop is possible but that way i will get timing for whole run instead of single run and this for-loop is a generator with yield from asyncio.sleep() which also causes problems

Wrapping only a "processor" with benchmark seems to be hard because I also need to have "concurrent" secondary tasks to be run too but benchmark(processor) will run it outside of event loop

[ionelmc]

I still think something is wrong there - you shouldn't have runs with different data aggregated in the same benchmark. What the point with this? Get avg time for the "docs building thing"? In that case min and deviation values are not useful (you would have no way to know how stable your benchmark was).

How fast is this process function? Is it less than 100ms?

My understanding is that you want to isolate the asyncio overhead (or whatever) and only get metrics for the "hot" part in your coroutines (the process func). However, that hot part is called multiple times during a test.

I think a better way would be to just isolate the process function so you only call it once in a test (and let pytest-benchmark repeat it as necessary).

[imposeren]

I\m using pytest-benchmark for a kind of unrelated purpose: I'm testing API performance that is a part of our infrastructure. This is a ElasticSearch API and i'm benchmarking search using current index configuration and queries structure. ES performance have high dependency on data diversity and also depends on concurrent queries. So I need to

1. benchmark 500 runs of search
2. during this benchmark other queries should also be performed to simulate situation closer to reality.

That's why I can't simply let pytest-benchmark to rerun call 500 times: I can't run parallel concurrent tasks with controlled frequency using current api. I have a workaround:

        benchmark._mode = 'benchmark(...)'
        stats = benchmark._make_stats(1)

        # ...
        prev_t = default_timer()
        benchmarked_function_call()
        processor_time = default_timer()-prev_t
        stats.update(main_call_duration)

        # check concurrent queries frequency
        # run concurrent tasks zero or several times to reach target frequency

But this is not documented anywhere and I think that there may also be other possible usecases of pytest-benchmark where such low level api will be useful

[ionelmc]

I need to read your comment a bit more but have you looked at the manual mode? http://pytest-benchmark.readthedocs.io/en/latest/pedantic.html

[imposeren]

Maybe my situation can be solved by using pedantic(..., setup=foo, ... ) It still have some limitations that require workarounds but still it's better because I don't need to call or to know any internals of pytest-benchmark. Thanks