each request (even same) adds to memory footprint and is increasing time to process

[ecamaj]

@MathiasDesch Do you have any idea why each call to SyntaxNetWrapper adds to memory and adds to time needed to process?

Clean docker installation starts with 280MB memory and after maybe 200 hundred calls it is already at 500MB+.
Also, time needed to process a call starts from 0.5s and goes to over 1s for the same amount of requests.

Thank you,
Eduard

[MathiasDesch]

Could please provide a code snippet, recreating this behaviour? Is it with the subprocess version?

[ecamaj]

There not much to explain - I've put your SyntaxNetWrapper object (hence, not the subprocess version) on the UWSGI server and have put it in static variable so that each new request uses the same object:

from syntaxnet_wrapper import SyntaxNetWrapper

sn_wrapper = {}  # language is the key

def query(data):
    ... getting language and sentences from POST data

    if not sn_wrapper.get(language):
            sn_wrapper[language] = SyntaxNetWrapper(language=language)

    dependency_output = sn_wrapper[language].parse_sentences(sentences)

    return dependency_output 

And all works nice. But the thing with increasing memory and response time is making it unusable.

Any ideas?

Thanks!

[MathiasDesch]

Thanks. Could you check the process stack when your memory is increasing? There should be only one SyntaxNet process. If more exists, the problem comes from here. If not, I am not sure what changes the new SyntaxNet version has brought.

Maybe the way the lib handle instantiation is not efficient anymore and processes now launch at each call of parse_sentences. That would be for me the main lead for memory and response time increasing after a few calls.