view roundup/rate_limit.py @ 6565:2c2dbfc332ba

Try to handle multiple connections better. The session database is a hot spot. When multiple requests (e.g. 20) come in at the same time session database contention can get great. The original code didn't retry session database access when the open failed. This resulted in errors at the client. The second pass delayed 0.01 seconds and retried. It was better but we still had multiple second stalls. I think the first request got in, everybody else backed up and then retried at the same time. Again they stepped on each other. With logging I would see many counters go all the way to low single digits or to -1 indicating falure. This pass uses randomint to generate delays from 0-.125 seconds in 5ms increments. This performs better in testing. I rarely saw a counter less than 13 (2 failed retries). Current logging starts after 6 failures and counts down until success or failure.
author John Rouillard <rouilj@ieee.org>
date Thu, 16 Dec 2021 20:02:00 -0500
parents 69a35d164a69
children 8f29e4ea05ce
line wrap: on
line source

# Originaly from
# https://smarketshq.com/implementing-gcra-in-python-5df1f11aaa96?gi=4b9725f99bfa
# with imports, modifications for python 2, implementation of
# set/get_tat and marshaling as string, support for testonly
# and status method.

from datetime import timedelta, datetime


class RateLimit:  # pylint: disable=too-few-public-methods
    def __init__(self, count, period):
        self.count = count
        self.period = period

    @property
    def inverse(self):
        return self.period.total_seconds() / self.count


class Gcra:
    def __init__(self):
        self.memory = {}

    def get_tat(self, key):
        # This should return a previous tat for the key or the current time.
        if key in self.memory:
            return self.memory[key]
        else:
            return datetime.min

    def set_tat(self, key, tat):
        self.memory[key] = tat

    def get_tat_as_string(self, key):
        # get value as string:
        #  YYYY-MM-DDTHH:MM:SS.mmmmmm
        # to allow it to be marshalled/unmarshaled
        if key in self.memory:
            return self.memory[key].isoformat()
        else:
            return datetime.min.isoformat()

    def set_tat_as_string(self, key, tat):
        # Take value as string and unmarshall:
        #  YYYY-MM-DDTHH:MM:SS.mmmmmm
        # to datetime
        self.memory[key] = datetime.strptime(tat, "%Y-%m-%dT%H:%M:%S.%f")

    def update(self, key, limit, testonly=False):
        '''Determine if the item associated with the key should be
           rejected given the RateLimit limit.
        '''
        now = datetime.utcnow()
        tat = max(self.get_tat(key), now)
        separation = (tat - now).total_seconds()
        max_interval = limit.period.total_seconds() - limit.inverse
        if separation > max_interval:
            reject = True
        else:
            reject = False
            if not testonly:
                new_tat = max(tat, now) + timedelta(seconds=limit.inverse)
                self.set_tat(key, new_tat)
        return reject

    def status(self, key, limit):
        '''Return status suitable for displaying as headers:
            X-RateLimit-Limit: calls allowed per period. Period/window
                is not specified in any api I found.
            X-RateLimit-Limit-Period: Non standard. Defines period in
                seconds for RateLimit-Limit.
            X-RateLimit-Remaining: How many calls are left in this window.
            X-RateLimit-Reset: window ends in this many seconds (not an
                 epoch timestamp) and all RateLimit-Limit calls are
                 available again.
            Retry-After: if user's request fails, this is the next time there
                 will be at least 1 available call to be consumed.
        '''

        ret = {}
        tat = self.get_tat(key)
        # static defined headers according to limit
        # all values are strings as that is required when used as headers
        ret['X-RateLimit-Limit'] = str(limit.count)
        ret['X-RateLimit-Limit-Period'] = str(
                                           int(
                                            limit.period.total_seconds())
                                          )

        # status of current limit as of now
        now = datetime.utcnow()

        current_count = int((limit.period - (tat - now)).total_seconds() /
                            limit.inverse)
        ret['X-RateLimit-Remaining'] = str(min(current_count, limit.count))

        # tat_in_epochsec = (tat - datetime(1970, 1, 1)).total_seconds()
        seconds_to_tat = (tat - now).total_seconds()
        ret['X-RateLimit-Reset'] = str(max(seconds_to_tat, 0))
        ret['X-RateLimit-Reset-date'] = "%s" % tat
        ret['Now'] = str((now - datetime(1970, 1, 1)).total_seconds())
        ret['Now-date'] = "%s" % now

        if self.update(key, limit, testonly=True):
            # A new request would be rejected if it was processes.
            # The user has to wait until an item is dequeued.
            # One item is dequeued every limit.inverse seconds.
            ret['Retry-After'] = str(int(limit.inverse))
            ret['Retry-After-Timestamp'] = "%s" % \
                    (now + timedelta(seconds=limit.inverse))  # noqa: E127
        else:
            # if we are not rejected, the user can post another
            # attempt immediately.
            # Do we even need this header if not rejected?
            # RFC implies this is used with a 503 (or presumably
            # 429 which may postdate the rfc). So if no error, no header?
            # ret['Retry-After'] = '0'
            # ret['Retry-After-Timestamp'] = str(ret['Now-date'])
            pass

        return ret

Roundup Issue Tracker: http://roundup-tracker.org/