in prediction_generation/old-code/cpdbench_mozilla_rep_with_restart_on_alert.py [0:0]
def main():
logger = logging.getLogger(__name__)
args = parse_args()
data, mat = load_dataset(args.input)
raw_data = data.copy()
start_time = time.time()
with open(args.signatures_attributes, 'r') as file:
signatures_attributes = json.load(file)
signature_id = os.path.splitext(os.path.basename(args.input))[0]
signature_attributes = signatures_attributes[signature_id]
Signature = namedtuple('Signature', signature_attributes.keys())
signature = Signature(**signature_attributes)
raw_args = copy.deepcopy(args)
#try:
series = data['series'][0]['raw']
push_timestamp = data['time']['raw']
push_timestamp = [datetime.strptime(ts, "%Y-%m-%d %H:%M:%S") for ts in push_timestamp]
#unique_push_timestamp = sorted(set(datetime.strptime(ts, "%Y-%m-%d %H:%M:%S") for ts in push_timestamp))
unique_push_timestamp = sorted(set(push_timestamp))
grouped_data = defaultdict(list)
for ts, value in zip(push_timestamp, series):
grouped_data[ts].append(value)
data = [RevisionDatum(ts, ts, grouped_data[ts]) for ts in grouped_data]
# data_sorted = sorted(data)
# These values are the default taken from the Mozilla code, Note that min_back_window, max_back_window, and fore_window come from class Performancesignature, I did not find them on record in the signatures data we have o we will be using the defaults
min_back_window=12
max_back_window=24
fore_window=12
alert_threshold=2
locations = []
selected_data = data
alert_detected_in_prev_iteration = True
while alert_detected_in_prev_iteration:
temp_locations = []
temp_timestamps = []
analyzed_series = detect_changes(
selected_data,
min_back_window=min_back_window,
max_back_window=max_back_window,
fore_window=fore_window,
)
for prev, cur in zip(analyzed_series, analyzed_series[1:]):
if cur.change_detected:
prev_value = cur.historical_stats["avg"]
new_value = cur.forward_stats["avg"]
alert_properties = get_alert_properties(
prev_value, new_value, signature.lower_is_better
)
noise_profile = "N/A"
try:
# Gather all data up to the current data point that
# shows the regression and obtain a noise profile on it.
# This helps us to ignore this alert and others in the
# calculation that could influence the profile.
noise_data = []
for point in analyzed_series:
if point == cur:
break
noise_data.append(geomean(point.values))
noise_profile, _ = moz_measure_noise.deviance(noise_data)
if not isinstance(noise_profile, str):
raise Exception(
f"Expecting a string as a noise profile, got: {type(noise_profile)}"
)
except Exception:
# Fail without breaking the alert computation
newrelic.agent.notice_error()
logger.error("Failed to obtain a noise profile.")
# ignore regressions below the configured regression threshold
# ALERT_PCT, ALERT_ABS, and ALERT_CHANGE_TYPES come from the PerformanceSignature class in the Treeherder code
ALERT_PCT = 0
ALERT_ABS = 1
ALERT_CHANGE_TYPES = ((ALERT_PCT, "percentage"), (ALERT_ABS, "absolute"))
if (
(
signature.alert_change_type is None
or signature.alert_change_type == ALERT_PCT
)
and alert_properties.pct_change < alert_threshold
) or (
signature.alert_change_type == ALERT_ABS
and abs(alert_properties.delta) < alert_threshold
):
continue
# django/mysql doesn't understand "inf", so just use some
# arbitrarily high value for that case
t_value = cur.t
if t_value == float("inf"):
t_value = 1000
# This is where we create the alert aka append its index in the locations list
temp_locations += [i for i, ts in enumerate(unique_push_timestamp) if ts == cur.push_timestamp]
temp_timestamps += [ts for i, ts in enumerate(unique_push_timestamp) if ts == cur.push_timestamp]
if len(temp_locations) == 0:
alert_detected_in_prev_iteration = False
else:
earliest_alert_timestamp = sorted(temp_timestamps)[0]
earliest_alert_index = sorted(temp_locations)[0]
selected_data = [rev for rev in selected_data if rev.push_timestamp > earliest_alert_timestamp]
locations += [earliest_alert_index]
stop_time = time.time()
runtime = stop_time - start_time
exit_success(raw_data, raw_args, vars(args), locations, runtime, __file__)