view roundup/rate_limit.py @ 6433:c1d3fbcdbfbd

issue2551142 - Import of retired node ... unique constraint failure. Title: Import of retired node with username after active node fails with unique constraint failure. More fixes needed for mysql and postgresql. mysql: add unique constraint for (keyvalue, __retired__) when creating class in the database. On schema change if class is changed, remove the unique constraint too. upgrade version of rdbms database from 5 to 6 to add constraint to all version 5 databases that were created as version 5 and didn't get the unique constraint. Make no changes on version 5 databases upgraded from version 4, the upgrade process to 5 added the constraint. Make no changes to other databases (sqlite, postgres) during upgrade from version 5 to 6. postgres: Handle the exception raised on unique constraint violation. The exception invalidates the database connection so it can't be used to recover from the exception. Added two new database methods: checkpoint_data - performs a db.commit under postgres does nothing on other backends restore_connection_on_error - does a db.rollback on postgres, does nothing on other backends with the rollback() done on the connection I can use the database connection to fixup the import that failed on the unique constraint. This makes postgres slower but without the commit after every imported object, the rollback will delete all the entries done up to this point. Trying to figure out how to make the caller do_import batch and recover from this failure is beyond me. Also dismissed having to process the export csv file before importing. Pushing that onto a user just seems wrong. Also since import/export isn't frequently done the lack of surprise on having a failing import and reduced load/frustration for the user seems worth it. Also the import can be run in verbose mode where it prints out a row as it is processed, so it may take a while, ut the user can get feedback. db_test-base.py: add test for upgrade from 5 to 6.
author John Rouillard <rouilj@ieee.org>
date Thu, 10 Jun 2021 12:52:05 -0400
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/