Files
edx-platform/lms/djangoapps/experiments/flags.py

272 lines
11 KiB
Python

"""
Feature flag support for experiments
"""
import datetime
import logging
from contextlib import contextmanager
import dateutil
import pytz
from crum import get_current_request
from edx_django_utils.cache import RequestCache
from experiments.stable_bucketing import stable_bucketing_hash_group
from openedx.core.djangoapps.waffle_utils import CourseWaffleFlag
from track import segment
log = logging.getLogger(__name__)
class ExperimentWaffleFlag(CourseWaffleFlag):
"""
ExperimentWaffleFlag handles logic around experimental bucketing and whitelisting.
You'll have one main flag that gates the experiment. This allows you to control the scope
of your experiment and always provides a quick kill switch.
But you'll also have smaller related flags that can force bucketing certain users into
specific buckets of your experiment. Those can be set using a waffle named like
"main_flag.BUCKET_NUM" (e.g. "course_experience.animated_exy.0") to force
users that pass the first main waffle check into a specific bucket experience.
If a user is not forced into a specific bucket by one of the aforementioned smaller flags,
then they will be randomly assigned a default bucket based on a consistent hash of:
* (flag_name, course_key, username) if use_course_aware_bucketing=True, or
* (flag_name, username) if use_course_aware_bucketing=False.
Note that you may call `.get_bucket` and `.is_enabled` without a course_key, in which case:
* the smaller flags will be evaluated without course context, and
* the default bucket will be evaluated as if use_course_aware_bucketing=False.
You can also control whether the experiment only affects future enrollments by setting
an ExperimentKeyValue model object with a key of 'enrollment_start' to the date of the
first enrollments that should be bucketed.
Bucket 0 is assumed to be the control bucket.
See a HOWTO here: https://openedx.atlassian.net/wiki/spaces/AC/pages/1250623700/Bucketing+users+for+an+experiment
When writing tests involving an ExperimentWaffleFlag you must not use the
override_waffle_flag utility. That will only turn the experiment on or off and won't
override bucketing. Instead use ExperimentWaffleFlag's override method which
will do both. Example:
with MY_EXPERIMENT_WAFFLE_FLAG.override(active=True, bucket=1):
...
or as a decorator:
@MY_EXPERIMENT_WAFFLE_FLAG.override(active=True, bucket=1)
def test_my_experiment(self):
...
"""
def __init__(
self,
waffle_namespace,
flag_name,
num_buckets=2,
experiment_id=None,
use_course_aware_bucketing=True,
**kwargs
):
super().__init__(waffle_namespace, flag_name, **kwargs)
self.num_buckets = num_buckets
self.experiment_id = experiment_id
self.bucket_flags = [
CourseWaffleFlag(waffle_namespace, '{}.{}'.format(flag_name, bucket))
for bucket in range(num_buckets)
]
self.use_course_aware_bucketing = use_course_aware_bucketing
def _cache_bucket(self, key, value):
request_cache = RequestCache('experiments')
request_cache.set(key, value)
return value
def _is_enrollment_inside_date_bounds(self, experiment_values, user, course_key):
""" Returns True if the user's enrollment (if any) is valid for the configured experiment date range """
from student.models import CourseEnrollment
enrollment_start = experiment_values.get('enrollment_start')
enrollment_end = experiment_values.get('enrollment_end')
if not enrollment_start and not enrollment_end:
return True # early exit just to avoid any further lookups
now = datetime.datetime.now(pytz.utc)
enrollment = CourseEnrollment.get_enrollment(user, course_key)
# If the user isn't enrolled, act like they would enroll right now (this keeps the pre-enroll and post-enroll
# experiences the same, if they decide to enroll right now)
enrollment_creation_date = enrollment.created if enrollment else now
# Enrollment must be after any enrollment_start date, if specified
if enrollment_start:
try:
start_date = dateutil.parser.parse(enrollment_start).replace(tzinfo=pytz.UTC)
except ValueError:
log.exception('Could not parse enrollment start date for experiment %d', self.experiment_id)
return False
if enrollment_creation_date < start_date:
return False
# Enrollment must be before any enrollment_end date, if specified
if enrollment_end:
try:
end_date = dateutil.parser.parse(enrollment_end).replace(tzinfo=pytz.UTC)
except ValueError:
log.exception('Could not parse enrollment end date for experiment %d', self.experiment_id)
return False
if enrollment_creation_date >= end_date:
return False
# All good! Either because the key was not set or because the enrollment was valid
return True
def get_bucket(self, course_key=None, track=True):
"""
Return which bucket number the specified user is in.
The user may be force-bucketed if matching subordinate flags of the form
"main_flag.BUCKET_NUM" exist. Otherwise, they will be hashed into a default
bucket based on their username, the experiment name, and the course-run key.
If `self.use_course_aware_bucketing` is False, the course-run key will
be omitted from the hashing formula, thus making it so a given user
has the same default bucket across all course runs; however, subordinate
flags that match the course-run key will still apply.
If `course_key` argument is omitted altogether, then subordinate flags
will be evaluated outside of the course-run context, and the default bucket
will be calculated as if `self.use_course_aware_bucketing` is False.
Finally, Bucket 0 is assumed to be the control bucket and will be returned if the
experiment is not enabled for this user and course.
Arguments:
course_key (Optional[CourseKey])
track (bool):
Whether an analytics event should be generated if the user is
bucketed for the first time.
Returns: int
"""
# Keep some imports in here, because this class is commonly used at a module level, and we want to avoid
# circular imports for any models.
from experiments.models import ExperimentKeyValue
from lms.djangoapps.courseware.masquerade import get_specific_masquerading_user
request = get_current_request()
if not request:
return 0
if not hasattr(request, 'user') or not request.user.id:
# We need username for stable bucketing and id for tracking, so just skip anonymous (not-logged-in) users
return 0
user = get_specific_masquerading_user(request.user, course_key)
if user is None:
user = request.user
masquerading_as_specific_student = False
else:
masquerading_as_specific_student = True
# If a course key is passed in, include it in the experiment name
# in order to separate caches and analytics calls per course-run.
# If we are using course-aware bucketing, then also append that course key
# to `bucketing_group_name`, such that users can be hashed into different
# buckets for different course-runs.
experiment_name = bucketing_group_name = self.namespaced_flag_name
if course_key:
experiment_name += ".{}".format(course_key)
if course_key and self.use_course_aware_bucketing:
bucketing_group_name += ".{}".format(course_key)
# Check if we have a cache for this request already
request_cache = RequestCache('experiments')
cache_response = request_cache.get_cached_response(experiment_name)
if cache_response.is_found:
return cache_response.value
# Check if the main flag is even enabled for this user and course.
if not self.is_experiment_on(course_key): # grabs user from the current request, if any
return self._cache_bucket(experiment_name, 0)
# Check if the enrollment should even be considered (if it started before the experiment wants, we ignore)
if course_key and self.experiment_id is not None:
values = ExperimentKeyValue.objects.filter(experiment_id=self.experiment_id).values('key', 'value')
values = {pair['key']: pair['value'] for pair in values}
if not self._is_enrollment_inside_date_bounds(values, user, course_key):
return self._cache_bucket(experiment_name, 0)
# Determine the user's bucket.
# First check if forced into a particular bucket, using our subordinate bucket flags.
# If not, calculate their default bucket using a consistent hash function.
for i, bucket_flag in enumerate(self.bucket_flags):
if bucket_flag.is_enabled(course_key):
bucket = i
break
else:
bucket = stable_bucketing_hash_group(
bucketing_group_name, self.num_buckets, user.username
)
session_key = 'tracked.{}'.format(experiment_name)
if (
track and hasattr(request, 'session') and
session_key not in request.session and
not masquerading_as_specific_student
):
segment.track(
user_id=user.id,
event_name='edx.bi.experiment.user.bucketed',
properties={
'site': request.site.domain,
'app_label': self.waffle_namespace.name,
'experiment': self.flag_name,
'course_id': str(course_key) if course_key else None,
'bucket': bucket,
'is_staff': user.is_staff,
'nonInteraction': 1,
}
)
# Mark that we've recorded this bucketing, so that we don't do it again this session
request.session[session_key] = True
return self._cache_bucket(experiment_name, bucket)
def is_enabled(self, course_key=None):
"""
Return whether the requesting user is in a nonzero bucket for the given course.
See the docstring of `.get_bucket` for more details.
Arguments:
course_key (Optional[CourseKey])
Returns: bool
"""
return self.get_bucket(course_key) != 0
def is_experiment_on(self, course_key=None):
"""
Return whether the overall experiment flag is enabled for this user.
This disregards `.bucket_flags`.
"""
return super().is_enabled(course_key)
@contextmanager
def override(self, active=True, bucket=1): # pylint: disable=arguments-differ
# Let CourseWaffleFlag override the base waffle flag value
with super().override(active=active):
# Now override the experiment bucket value
from mock import patch
if not active:
bucket = 0
with patch.object(self, 'get_bucket', return_value=bucket):
yield