PYTHON-4780 Implement fast path for server selection with Primary

[blink1073]

Tested with the following script:

from pymongo import MongoClient
from pymongo.read_preferences import ReadPreference
from concurrent.futures import ThreadPoolExecutor, wait

client = MongoClient()
session = client.start_session()
read_preference = ReadPreference.PRIMARY


def test():
    for i in range(100):
        client._select_server(read_preference, session, "testOperation")



def test2_target():
    client._select_server(read_preference, session, "testOperation")


def test2():
    futures = []
    with ThreadPoolExecutor(max_workers=100) as pool:
        for i in range(100):
            futures.append(pool.submit(test2_target))
        wait(futures)
    

if __name__ == '__main__':
    import timeit
    print(timeit.timeit("test()", setup="from __main__ import test, client, read_preference, session", number=1000))
    print(timeit.timeit("test2()", setup="from __main__ import test2, client, read_preference, session", number=1000))

On master:
0.4806627919897437
1.5110677920019953

With this change:
0.17988054199668113
1.233099208009662

[blink1073]

I didn't schedule a perf build because our perf tests only run on standalone.

[ShaneHarvey]

Minor note for the benchmark, thread dispatching could be dominating test2. Could you change it to pool.submit(test) and just remove test2_target? That would probably show a larger perf increase for threaded.

[blink1073]

With your suggested benchmark changes:

On master:
0.4840990829980001
0.5537169170274865

With this PR:
0.1950377920002211
0.260104083979968

[ShaneHarvey]

Nice work! Could you add a changelog saying "Improved performance of selecting a server with the Primary selector."?

[ShaneHarvey]

My one concern is that technically this could be a breaking change if a user is subclassing Primary and overriding __call__ to do something special during server selection. One simple way to avoid that risk is doing type(selector) is Primary rather than isinstance. Or we could document the potential breaking change.

[blink1073]

I went with the code modification option

[ShaneHarvey]

Oh one more thing, don't we have to call the custom_selector here if one was provided?

[ShaneHarvey]

And add a test for it?

[blink1073]

I'll have to defer for now, working on a HELP ticket

[blink1073]

Done

[ShaneHarvey]

Is the custom_selector is None check supposed to be here?

[blink1073]

No, I forgot to back that change out

[ShaneHarvey]

Neat trick.

[ShaneHarvey]

LGTM!