Only preload interactive algorithms for reader studies with budget

[amickan]

Only preload interactive algorithms of reader studies with budget.

Part of https://github.com/DIAGNijmegen/rse-roadmap/issues/469

[amickan]

This query has an N+1 problem (due to expensive queries in credits_consumed). It's an asynchronous task, but still. Currently looking at annotating credits_consumed in the DB to make this more efficient.

[amickan]

Struggling with turning credits_consumed into an annotation. It involves nested aggregates, and gets quite complex. Maybe prefetching the related objects is good enough here? In general, I would like to know how bad the performance is before I continue trying to optimize this.

@jmsmkn could you compare timings for the following 2 queries:

import time
from grand_challenge.reader_studies.models import ReaderStudy


def query_without_prefetch():
    return [
        rs.pk
        for rs in ReaderStudy.objects.exclude(max_credits__isnull=True).only(
            "pk", "max_credits"
        )
        if not rs.has_budget
    ]


def query_with_prefetch():
    return [
        rs.pk
        for rs in ReaderStudy.objects.exclude(max_credits__isnull=True)
        .prefetch_related(
            "session_utilizations__reader_studies",
            "endpoint_utilizations__reader_studies",
        )
        .only("pk", "max_credits")
        if not rs.has_budget
    ]


def benchmark(label, func):
    start = time.perf_counter()
    result = func()
    end = time.perf_counter()

    return {
        "label": label,
        "time_seconds": end - start,
    }


def main():
    results = [
        benchmark("without_prefetch", query_without_prefetch),
        benchmark("with_prefetch", query_with_prefetch),
    ]

    for r in results:
        print(f"{r['label']}:")
        print(f"  time: {r['time_seconds']:.6f}s")


if __name__ == "__main__":
    main()

[jmsmkn]

without_prefetch:
  time: 0.294184s
with_prefetch:
  time: 0.203041s

[amickan]

Thanks. Those timings seem fine to me for an asynchronous task and not worth the added complexity of moving the consumed credits into a DB annotation.

[jmsmkn]

Would still be worth including the prefetching if those objects are going to be used.

[amickan]

Very true, I had that as a local commit, but forgot to push!

[jmsmkn]

Is there a reason to use the negative here? Filtering is going to be more efficient than excluding.

[amickan]

No. This seemed easier to read to me and I didn't know that filtering is more efficient than excluding. I'll change it to filter().