def _cohort_stream_transform(source, stream, start, end,
transform, grouping_key, unit):
start_stream = KronosSource(source, stream, start, end)
if transform:
transformed = transform(start_stream)
else:
transformed = start_stream
projected = Project(transformed,
[Property(TIMESTAMP_FIELD,
alias=TIMESTAMP_FIELD),
Property(grouping_key,
alias=grouping_key),
Floor([Property(TIMESTAMP_FIELD),
Constant(DateUnit.unit_to_kronos_time(unit)),
Constant(start)],
alias='date')])
# This leaves us with a single event per (user, unit time) pair.
aggregated = Aggregate(
projected,
GroupBy([Property(grouping_key,
alias=grouping_key),
Property('date',
alias='date')]),
# The first time the user performed the event in that bucket.
[Min([Property(TIMESTAMP_FIELD)], alias=TIMESTAMP_FIELD)])
return aggregated
def test_aggregate(self):
sums = defaultdict(int)
for i in xrange(200):
a = random.randint(0, 2)
self.kronos_client.put(
{self.stream: [{
constants.TIMESTAMP_FIELD: i,
'a': a
}]})
sums[50 * (i / 50)] += a
events = self.query(
Aggregate(
Project(KronosSource('kronos', self.stream, 0, 1000), [
Floor([Property(constants.TIMESTAMP_FIELD),
Constant(50)],
alias=constants.TIMESTAMP_FIELD)
],
merge=True),
GroupBy(
Property(constants.TIMESTAMP_FIELD,
alias=constants.TIMESTAMP_FIELD)), [
Count([], alias='count'),
Sum([Property('a')], alias='sum'),
Min([Property('a')], alias='min'),
Max([Property('a')], alias='max'),
Avg([Property('a')], alias='avg')
]).to_dict())
self.assertEqual(len(events), 200 / 50)
for event in events:
self.assertEqual(event[constants.TIMESTAMP_FIELD] % 50, 0)
self.assertEqual(event['count'], 50)
self.assertEqual(event['min'], 0)
self.assertEqual(event['max'], 2)
self.assertEqual(event['sum'],
sums[event[constants.TIMESTAMP_FIELD]])
self.assertTrue(event['avg'] * 50 > event['sum'] - 0.1)
self.assertTrue(event['avg'] * 50 < event['sum'] + 0.1)
events = self.query(
Aggregate(
KronosSource('kronos', self.stream, 0, 1000),
GroupBy(
Floor([Property(constants.TIMESTAMP_FIELD),
Constant(50)],
alias=constants.TIMESTAMP_FIELD)),
[Count([], alias='count')]).to_dict())
self.assertEqual(len(events), 200 / 50)
def aggregate(query_plan, operands):
aggregates = []
for agg in operands['aggregates']:
aggregates.append(agg_op(agg['agg_type'], cpf(agg['agg_on']),
agg['alias']))
groups = []
for group in operands['groups']:
groups.append(cpf(group['field'], group['alias']))
group_by = GroupBy(groups)
return Aggregate(query_plan, group_by, aggregates)
def aggregate(query_plan, operands):
aggregates = []
for agg in operands['aggregates']:
cpf_type = agg['agg_on']['cpf_type']
property_name = agg['agg_on'].get('property_name')
constant_value = agg['agg_on'].get('constant_value')
empty = (cpf_type == 'property' and not property_name
or cpf_type == 'constant' and not constant_value)
if empty:
agg_on_cpf = None
else:
agg_on_cpf = cpf(agg['agg_on'])
aggregates.append(agg_op(agg['agg_type'], agg_on_cpf, agg['alias']))
groups = []
for group in operands['groups']:
groups.append(cpf(group['field'], group['alias']))
group_by = GroupBy(groups)
return Aggregate(query_plan, group_by, aggregates)
def cohort_queryplan(plan):
"""
Input:
{
'source': 'kronos', # Name of data source from settings
'cohort':
{'stream': CohortTest.EMAIL_STREAM, # Kronos stream to define cohort from.
'transform': lambda x: x, # Transformations on the kstream.
'start': date.now(), # The day of the first cohort.
'unit': DateUnit.XX, # Users are in the same cohort
# if they are in the same day/week.
'cohorts': 5 # How many cohorts (days/weeks/months)
# to track.
'grouping_key': 'user'}, # What key in an event should we tie
# to a key in the action stream?
'action':
{'stream': CohortTest.FRONTPAGE_STREAM, # Stream users take actions on.
'transform': lambda x: x # Transformations on the stream.
'unit': DateUnit.XX, # Track events in day/week/months.
'repetitions': 14 # How many days/weeks/months to track.
'grouping_key': 'user_id'} # What key in an event should we tie
# to a key in the action stream?
}
Output:
A metis-compatible query plan to return a cohort analysis.
"""
cohort = plan['cohort']
action = plan['action']
source = plan['source']
# Calculate the start and end dates, in Kronos time, of the
# beginning and end of the cohort and action streams that will be
# relevant.
cohort_start = datetime_to_kronos_time(_date_to_datetime(cohort['start']))
cohort_span = timedelta(**{cohort['unit']: cohort['cohorts']})
cohort_end = cohort['start'] + cohort_span
action_span = timedelta(**{action['unit']: action['repetitions']})
action_end = cohort_end + action_span
cohort_end = datetime_to_kronos_time(_date_to_datetime(cohort_end)) + 1
action_end = datetime_to_kronos_time(_date_to_datetime(action_end)) + 1
left = _cohort_stream_transform(source,
cohort['stream'], cohort_start, cohort_end,
cohort.get('transform'),
cohort['grouping_key'], cohort['unit'])
right = _cohort_stream_transform(source,
action['stream'], cohort_start, action_end,
action.get('transform'),
action['grouping_key'], action['unit'])
additional_action_time = (DateUnit.unit_to_kronos_time(action['unit']) *
action['repetitions'])
left.alias = 'cohort'
right.alias = 'action'
joined = Join(left,
right,
(Condition(Condition.Op.EQ,
Property('cohort.%s' % cohort['grouping_key']),
Property('action.%s' % action['grouping_key'])) &
Condition(Condition.Op.GTE,
Property('action.%s' % TIMESTAMP_FIELD),
Property('cohort.%s' % TIMESTAMP_FIELD)) &
Condition(Condition.Op.LT,
Property('action.%s' % TIMESTAMP_FIELD),
Add([Property('cohort.%s' % TIMESTAMP_FIELD),
Constant(additional_action_time)]))))
user_aggregated = Aggregate(
joined,
GroupBy([Property('cohort.date', alias=TIMESTAMP_FIELD),
Property('cohort.%s' % cohort['grouping_key'], alias='group'),
Floor([Subtract([Property('action.%s' % TIMESTAMP_FIELD),
Property('cohort.%s' % TIMESTAMP_FIELD)]),
Constant(DateUnit.unit_to_kronos_time(action['unit']))],
alias='action_step')]),
[Count([], alias='count')]
)
aggregated = Aggregate(
user_aggregated,
GroupBy([Property(TIMESTAMP_FIELD, alias=TIMESTAMP_FIELD),
Property('action_step', alias='action_step')]),
[Count([], alias='cohort_actions')])
# TODO(marcua): Also sum up the cohort sizes, join with the plan.
return aggregated.to_dict()
