jan/openreplay
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
openreplay/api/chalicelib/core/significance.py

577 lines
24 KiB
Python
Raw Blame History

import logging
import schemas
from chalicelib.core import events, metadata
from chalicelib.utils import sql_helper as sh
"""
todo: remove LIMIT from the query
"""
from typing import List
import math
import warnings
from collections import defaultdict
from psycopg2.extras import RealDictRow
from chalicelib.utils import pg_client, helper
logger = logging.getLogger(__name__)
SIGNIFICANCE_THRSH = 0.4
# Taha: the value 24 was estimated in v1.15
T_VALUES = {1: 12.706, 2: 4.303, 3: 3.182, 4: 2.776, 5: 2.571, 6: 2.447, 7: 2.365, 8: 2.306, 9: 2.262, 10: 2.228,
11: 2.201, 12: 2.179, 13: 2.160, 14: 2.145, 15: 2.13, 16: 2.120, 17: 2.110, 18: 2.101, 19: 2.093, 20: 2.086,
21: 2.080, 22: 2.074, 23: 2.069, 24: 2.067, 25: 2.064, 26: 2.060, 27: 2.056, 28: 2.052, 29: 2.045,
30: 2.042}
def get_stages_and_events(filter_d: schemas.CardSeriesFilterSchema, project_id) -> List[RealDictRow]:
"""
Add minimal timestamp
:param filter_d: dict contains events&filters&...
:return:
"""
stages: [dict] = filter_d.events
filters: [dict] = filter_d.filters
filter_issues = []
# TODO: enable this if needed by an endpoint
# filter_issues = filter_d.get("issueTypes")
# if filter_issues is None or len(filter_issues) == 0:
# filter_issues = []
stage_constraints = ["main.timestamp <= %(endTimestamp)s"]
first_stage_extra_constraints = ["s.project_id=%(project_id)s", "s.start_ts >= %(startTimestamp)s",
"s.start_ts <= %(endTimestamp)s"]
filter_extra_from = []
n_stages_query = []
values = {}
if len(filters) > 0:
meta_keys = None
for i, f in enumerate(filters):
if len(f.value) == 0:
continue
f.value = helper.values_for_operator(value=f.value, op=f.operator)
# filter_args = _multiple_values(f["value"])
op = sh.get_sql_operator(f.operator)
filter_type = f.type
f_k = f"f_value{i}"
values = {**values,
**sh.multi_values(f.value, value_key=f_k)}
is_not = False
if sh.is_negation_operator(f.operator):
is_not = True
if filter_type == schemas.FilterType.user_browser:
first_stage_extra_constraints.append(
sh.multi_conditions(f's.user_browser {op} %({f_k})s', f.value, is_not=is_not, value_key=f_k))
elif filter_type in [schemas.FilterType.user_os, schemas.FilterType.user_os_ios]:
first_stage_extra_constraints.append(
sh.multi_conditions(f's.user_os {op} %({f_k})s', f.value, is_not=is_not, value_key=f_k))
elif filter_type in [schemas.FilterType.user_device, schemas.FilterType.user_device_ios]:
first_stage_extra_constraints.append(
sh.multi_conditions(f's.user_device {op} %({f_k})s', f.value, is_not=is_not, value_key=f_k))
elif filter_type in [schemas.FilterType.user_country, schemas.FilterType.user_country_ios]:
first_stage_extra_constraints.append(
sh.multi_conditions(f's.user_country {op} %({f_k})s', f.value, is_not=is_not, value_key=f_k))
elif filter_type == schemas.FilterType.duration:
if len(f.value) > 0 and f.value[0] is not None:
first_stage_extra_constraints.append(f's.duration >= %(minDuration)s')
values["minDuration"] = f.value[0]
if len(f["value"]) > 1 and f.value[1] is not None and int(f.value[1]) > 0:
first_stage_extra_constraints.append('s.duration <= %(maxDuration)s')
values["maxDuration"] = f.value[1]
elif filter_type == schemas.FilterType.referrer:
# events_query_part = events_query_part + f"INNER JOIN events.pages AS p USING(session_id)"
filter_extra_from = [f"INNER JOIN {events.EventType.LOCATION.table} AS p USING(session_id)"]
first_stage_extra_constraints.append(
sh.multi_conditions(f"p.base_referrer {op} %({f_k})s", f.value, is_not=is_not, value_key=f_k))
elif filter_type == events.EventType.METADATA.ui_type:
if meta_keys is None:
meta_keys = metadata.get(project_id=project_id)
meta_keys = {m["key"]: m["index"] for m in meta_keys}
if f.source in meta_keys.keys():
first_stage_extra_constraints.append(
sh.multi_conditions(
f's.{metadata.index_to_colname(meta_keys[f.source])} {op} %({f_k})s', f.value,
is_not=is_not, value_key=f_k))
# values[f_k] = helper.string_to_sql_like_with_op(f["value"][0], op)
elif filter_type in [schemas.FilterType.user_id, schemas.FilterType.user_id_ios]:
first_stage_extra_constraints.append(
sh.multi_conditions(f's.user_id {op} %({f_k})s', f.value, is_not=is_not, value_key=f_k))
# values[f_k] = helper.string_to_sql_like_with_op(f["value"][0], op)
elif filter_type in [schemas.FilterType.user_anonymous_id,
schemas.FilterType.user_anonymous_id_ios]:
first_stage_extra_constraints.append(
sh.multi_conditions(f's.user_anonymous_id {op} %({f_k})s', f.value, is_not=is_not, value_key=f_k))
# values[f_k] = helper.string_to_sql_like_with_op(f["value"][0], op)
elif filter_type in [schemas.FilterType.rev_id, schemas.FilterType.rev_id_ios]:
first_stage_extra_constraints.append(
sh.multi_conditions(f's.rev_id {op} %({f_k})s', f.value, is_not=is_not, value_key=f_k))
# values[f_k] = helper.string_to_sql_like_with_op(f["value"][0], op)
i = -1
for s in stages:
if s.operator is None:
s.operator = schemas.SearchEventOperator._is
if not isinstance(s.value, list):
s.value = [s.value]
is_any = sh.isAny_opreator(s.operator)
if not is_any and isinstance(s.value, list) and len(s.value) == 0:
continue
i += 1
if i == 0:
extra_from = filter_extra_from + ["INNER JOIN public.sessions AS s USING (session_id)"]
else:
extra_from = []
op = sh.get_sql_operator(s.operator)
# event_type = s["type"].upper()
event_type = s.type
if event_type == events.EventType.CLICK.ui_type:
next_table = events.EventType.CLICK.table
next_col_name = events.EventType.CLICK.column
elif event_type == events.EventType.INPUT.ui_type:
next_table = events.EventType.INPUT.table
next_col_name = events.EventType.INPUT.column
elif event_type == events.EventType.LOCATION.ui_type:
next_table = events.EventType.LOCATION.table
next_col_name = events.EventType.LOCATION.column
elif event_type == events.EventType.CUSTOM.ui_type:
next_table = events.EventType.CUSTOM.table
next_col_name = events.EventType.CUSTOM.column
# IOS --------------
elif event_type == events.EventType.CLICK_IOS.ui_type:
next_table = events.EventType.CLICK_IOS.table
next_col_name = events.EventType.CLICK_IOS.column
elif event_type == events.EventType.INPUT_IOS.ui_type:
next_table = events.EventType.INPUT_IOS.table
next_col_name = events.EventType.INPUT_IOS.column
elif event_type == events.EventType.VIEW_IOS.ui_type:
next_table = events.EventType.VIEW_IOS.table
next_col_name = events.EventType.VIEW_IOS.column
elif event_type == events.EventType.CUSTOM_IOS.ui_type:
next_table = events.EventType.CUSTOM_IOS.table
next_col_name = events.EventType.CUSTOM_IOS.column
else:
logging.warning(f"=================UNDEFINED:{event_type}")
continue
values = {**values, **sh.multi_values(helper.values_for_operator(value=s.value, op=s.operator),
value_key=f"value{i + 1}")}
if sh.is_negation_operator(s.operator) and i > 0:
op = sh.reverse_sql_operator(op)
main_condition = "left_not.session_id ISNULL"
extra_from.append(f"""LEFT JOIN LATERAL (SELECT session_id
FROM {next_table} AS s_main
WHERE
{sh.multi_conditions(f"s_main.{next_col_name} {op} %(value{i + 1})s",
values=s.value, value_key=f"value{i + 1}")}
AND s_main.timestamp >= T{i}.stage{i}_timestamp
AND s_main.session_id = T1.session_id) AS left_not ON (TRUE)""")
else:
if is_any:
main_condition = "TRUE"
else:
main_condition = sh.multi_conditions(f"main.{next_col_name} {op} %(value{i + 1})s",
values=s.value, value_key=f"value{i + 1}")
n_stages_query.append(f"""
(SELECT main.session_id,
{"MIN(main.timestamp)" if i + 1 < len(stages) else "MAX(main.timestamp)"} AS stage{i + 1}_timestamp
FROM {next_table} AS main {" ".join(extra_from)}
WHERE main.timestamp >= {f"T{i}.stage{i}_timestamp" if i > 0 else "%(startTimestamp)s"}
{f"AND main.session_id=T1.session_id" if i > 0 else ""}
AND {main_condition}
{(" AND " + " AND ".join(stage_constraints)) if len(stage_constraints) > 0 else ""}
{(" AND " + " AND ".join(first_stage_extra_constraints)) if len(first_stage_extra_constraints) > 0 and i == 0 else ""}
GROUP BY main.session_id)
AS T{i + 1} {"ON (TRUE)" if i > 0 else ""}
""")
n_stages = len(n_stages_query)
if n_stages == 0:
return []
n_stages_query = " LEFT JOIN LATERAL ".join(n_stages_query)
n_stages_query += ") AS stages_t"
n_stages_query = f"""
SELECT stages_and_issues_t.*, sessions.user_uuid
FROM (
SELECT * FROM (
SELECT T1.session_id, {",".join([f"stage{i + 1}_timestamp" for i in range(n_stages)])}
FROM {n_stages_query}
LEFT JOIN LATERAL
( SELECT ISS.type as issue_type,
ISE.timestamp AS issue_timestamp,
COALESCE(ISS.context_string,'') as issue_context,
ISS.issue_id as issue_id
FROM events_common.issues AS ISE INNER JOIN issues AS ISS USING (issue_id)
WHERE ISE.timestamp >= stages_t.stage1_timestamp
AND ISE.timestamp <= stages_t.stage{i + 1}_timestamp
AND ISS.project_id=%(project_id)s
AND ISE.session_id = stages_t.session_id
AND ISS.type!='custom' -- ignore custom issues because they are massive
{"AND ISS.type IN %(issueTypes)s" if len(filter_issues) > 0 else ""}
LIMIT 10 -- remove the limit to get exact stats
) AS issues_t ON (TRUE)
) AS stages_and_issues_t INNER JOIN sessions USING(session_id);
"""
# LIMIT 10000
params = {"project_id": project_id, "startTimestamp": filter_d.startTimestamp,
"endTimestamp": filter_d.endTimestamp,
"issueTypes": tuple(filter_issues), **values}
with pg_client.PostgresClient() as cur:
query = cur.mogrify(n_stages_query, params)
logging.debug("---------------------------------------------------")
logging.debug(query)
logging.debug("---------------------------------------------------")
try:
cur.execute(query)
rows = cur.fetchall()
except Exception as err:
logging.warning("--------- FUNNEL SEARCH QUERY EXCEPTION -----------")
logging.warning(query.decode('UTF-8'))
logging.warning("--------- PAYLOAD -----------")
logging.warning(filter_d.model_dump_json())
logging.warning("--------------------")
raise err
return rows
def pearson_corr(x: list, y: list):
n = len(x)
if n != len(y):
raise ValueError(f'x and y must have the same length. Got {len(x)} and {len(y)} instead')
if n < 2:
warnings.warn(f'x and y must have length at least 2. Got {n} instead')
return None, None, False
# If an input is constant, the correlation coefficient is not defined.
if all(t == x[0] for t in x) or all(t == y[0] for t in y):
warnings.warn("An input array is constant; the correlation coefficent is not defined.")
return None, None, False
if n == 2:
return math.copysign(1, x[1] - x[0]) * math.copysign(1, y[1] - y[0]), 1.0, True
xmean = sum(x) / len(x)
ymean = sum(y) / len(y)
xm = [el - xmean for el in x]
ym = [el - ymean for el in y]
normxm = math.sqrt((sum([xm[i] * xm[i] for i in range(len(xm))])))
normym = math.sqrt((sum([ym[i] * ym[i] for i in range(len(ym))])))
threshold = 1e-8
if normxm < threshold * abs(xmean) or normym < threshold * abs(ymean):
# If all the values in x (likewise y) are very close to the mean,
# the loss of precision that occurs in the subtraction xm = x - xmean
# might result in large errors in r.
warnings.warn("An input array is constant; the correlation coefficent is not defined.")
r = sum(
i[0] * i[1] for i in zip([xm[i] / normxm for i in range(len(xm))], [ym[i] / normym for i in range(len(ym))]))
# Presumably, if abs(r) > 1, then it is only some small artifact of floating point arithmetic.
# However, if r < 0, we don't care, as our problem is to find only positive correlations
r = max(min(r, 1.0), 0.0)
# approximated confidence
if n < 31:
t_c = T_VALUES[n]
elif n < 50:
t_c = 2.02
else:
t_c = 2
if r >= 0.999:
confidence = 1
else:
confidence = r * math.sqrt(n - 2) / math.sqrt(1 - r ** 2)
if confidence > SIGNIFICANCE_THRSH:
return r, confidence, True
else:
return r, confidence, False
# def tuple_or(t: tuple):
# x = 0
# for el in t:
# x |= el # | is for bitwise OR
# return x
#
# The following function is correct optimization of the previous function because t is a list of 0,1
def tuple_or(t: tuple):
for el in t:
if el > 0:
return 1
return 0
def get_transitions_and_issues_of_each_type(rows: List[RealDictRow], all_issues, first_stage, last_stage):
"""
Returns two lists with binary values 0/1:
transitions ::: if transited from the first stage to the last - 1
else - 0
errors ::: a dictionary WHERE the keys are all unique issues (currently context-wise)
the values are lists
if an issue happened between the first stage to the last - 1
else - 0
For a small task of calculating a total drop due to issues,
we need to disregard the issue type when creating the `errors`-like array.
The `all_errors` array can be obtained by logical OR statement applied to all errors by issue
The `transitions` array stays the same
"""
transitions = []
n_sess_affected = 0
errors = {}
for row in rows:
t = 0
first_ts = row[f'stage{first_stage}_timestamp']
last_ts = row[f'stage{last_stage}_timestamp']
if first_ts is None:
continue
elif last_ts is not None:
t = 1
transitions.append(t)
ic_present = False
for error_id in all_issues:
if error_id not in errors:
errors[error_id] = []
ic = 0
row_issue_id = row['issue_id']
if row_issue_id is not None:
if last_ts is None or (first_ts < row['issue_timestamp'] < last_ts):
if error_id == row_issue_id:
ic = 1
ic_present = True
errors[error_id].append(ic)
if ic_present and t:
n_sess_affected += 1
all_errors = [tuple_or(t) for t in zip(*errors.values())]
return transitions, errors, all_errors, n_sess_affected
def get_affected_users_for_all_issues(rows, first_stage, last_stage):
"""
:param rows:
:param first_stage:
:param last_stage:
:return:
"""
affected_users = defaultdict(lambda: set())
affected_sessions = defaultdict(lambda: set())
all_issues = {}
n_affected_users_dict = defaultdict(lambda: None)
n_affected_sessions_dict = defaultdict(lambda: None)
n_issues_dict = defaultdict(lambda: 0)
issues_by_session = defaultdict(lambda: 0)
for row in rows:
# check that the session has reached the first stage of subfunnel:
if row[f'stage{first_stage}_timestamp'] is None:
continue
iss = row['issue_type']
iss_ts = row['issue_timestamp']
# check that the issue exists and belongs to subfunnel:
if iss is not None and (row[f'stage{last_stage}_timestamp'] is None or
(row[f'stage{first_stage}_timestamp'] < iss_ts < row[f'stage{last_stage}_timestamp'])):
if row["issue_id"] not in all_issues:
all_issues[row["issue_id"]] = {"context": row['issue_context'], "issue_type": row["issue_type"]}
n_issues_dict[row["issue_id"]] += 1
if row['user_uuid'] is not None:
affected_users[row["issue_id"]].add(row['user_uuid'])
affected_sessions[row["issue_id"]].add(row['session_id'])
issues_by_session[row[f'session_id']] += 1
if len(affected_users) > 0:
n_affected_users_dict.update({
iss: len(affected_users[iss]) for iss in affected_users
})
if len(affected_sessions) > 0:
n_affected_sessions_dict.update({
iss: len(affected_sessions[iss]) for iss in affected_sessions
})
return all_issues, n_issues_dict, n_affected_users_dict, n_affected_sessions_dict
def count_sessions(rows, n_stages):
session_counts = {i: set() for i in range(1, n_stages + 1)}
for row in rows:
for i in range(1, n_stages + 1):
if row[f"stage{i}_timestamp"] is not None:
session_counts[i].add(row[f"session_id"])
session_counts = {i: len(session_counts[i]) for i in session_counts}
return session_counts
def count_users(rows, n_stages):
users_in_stages = {i: set() for i in range(1, n_stages + 1)}
for row in rows:
for i in range(1, n_stages + 1):
if row[f"stage{i}_timestamp"] is not None:
users_in_stages[i].add(row["user_uuid"])
users_count = {i: len(users_in_stages[i]) for i in range(1, n_stages + 1)}
return users_count
def get_stages(stages, rows):
n_stages = len(stages)
session_counts = count_sessions(rows, n_stages)
users_counts = count_users(rows, n_stages)
stages_list = []
for i, stage in enumerate(stages):
drop = None
if i != 0:
if session_counts[i] == 0:
drop = 0
elif session_counts[i] > 0:
drop = int(100 * (session_counts[i] - session_counts[i + 1]) / session_counts[i])
stages_list.append(
{"value": stage.value,
"type": stage.type,
"operator": stage.operator,
"sessionsCount": session_counts[i + 1],
"drop_pct": drop,
"usersCount": users_counts[i + 1],
"dropDueToIssues": 0
}
)
return stages_list
def get_issues(stages, rows, first_stage=None, last_stage=None, drop_only=False):
"""
:param stages:
:param rows:
:param first_stage: If it's a part of the initial funnel, provide a number of the first stage (starting from 1)
:param last_stage: If it's a part of the initial funnel, provide a number of the last stage (starting from 1)
:return:
"""
n_stages = len(stages)
if first_stage is None:
first_stage = 1
if last_stage is None:
last_stage = n_stages
if last_stage > n_stages:
logging.debug(
"The number of the last stage provided is greater than the number of stages. Using n_stages instead")
last_stage = n_stages
n_critical_issues = 0
issues_dict = {"significant": [],
"insignificant": []}
session_counts = count_sessions(rows, n_stages)
drop = session_counts[first_stage] - session_counts[last_stage]
all_issues, n_issues_dict, affected_users_dict, affected_sessions = get_affected_users_for_all_issues(
rows, first_stage, last_stage)
transitions, errors, all_errors, n_sess_affected = get_transitions_and_issues_of_each_type(rows,
all_issues,
first_stage, last_stage)
del rows
if any(all_errors):
total_drop_corr, conf, is_sign = pearson_corr(transitions, all_errors)
if total_drop_corr is not None and drop is not None:
total_drop_due_to_issues = int(total_drop_corr * n_sess_affected)
else:
total_drop_due_to_issues = 0
else:
total_drop_due_to_issues = 0
if drop_only:
return total_drop_due_to_issues
for issue_id in all_issues:
if not any(errors[issue_id]):
continue
r, confidence, is_sign = pearson_corr(transitions, errors[issue_id])
if r is not None and drop is not None and is_sign:
lost_conversions = int(r * affected_sessions[issue_id])
else:
lost_conversions = None
if r is None:
r = 0
issues_dict['significant' if is_sign else 'insignificant'].append({
"type": all_issues[issue_id]["issue_type"],
"title": helper.get_issue_title(all_issues[issue_id]["issue_type"]),
"affected_sessions": affected_sessions[issue_id],
"unaffected_sessions": session_counts[1] - affected_sessions[issue_id],
"lost_conversions": lost_conversions,
"affected_users": affected_users_dict[issue_id],
"conversion_impact": round(r * 100),
"context_string": all_issues[issue_id]["context"],
"issue_id": issue_id
})
if is_sign:
n_critical_issues += n_issues_dict[issue_id]
# To limit the number of returned issues to the frontend
issues_dict["significant"] = issues_dict["significant"][:20]
issues_dict["insignificant"] = issues_dict["insignificant"][:20]
return n_critical_issues, issues_dict, total_drop_due_to_issues
def get_top_insights(filter_d: schemas.CardSeriesFilterSchema, project_id):
output = []
stages = filter_d.events
if len(stages) == 0:
logging.debug("no stages found")
return output, 0
# The result of the multi-stage query
rows = get_stages_and_events(filter_d=filter_d, project_id=project_id)
if len(rows) == 0:
return get_stages(stages, []), 0
# Obtain the first part of the output
stages_list = get_stages(stages, rows)
# Obtain the second part of the output
total_drop_due_to_issues = get_issues(stages, rows,
first_stage=1,
last_stage=len(filter_d.events),
drop_only=True)
return stages_list, total_drop_due_to_issues
def get_issues_list(filter_d: schemas.CardSeriesFilterSchema, project_id, first_stage=None, last_stage=None):
output = dict({"total_drop_due_to_issues": 0, "critical_issues_count": 0, "significant": [], "insignificant": []})
stages = filter_d.events
# The result of the multi-stage query
rows = get_stages_and_events(filter_d=filter_d, project_id=project_id)
if len(rows) == 0:
return output
# Obtain the second part of the output
n_critical_issues, issues_dict, total_drop_due_to_issues = get_issues(stages, rows, first_stage=first_stage,
last_stage=last_stage)
output['total_drop_due_to_issues'] = total_drop_due_to_issues
# output['critical_issues_count'] = n_critical_issues
output = {**output, **issues_dict}
return output
Reference in a new issue View git blame Copy permalink