from __future__ import division
import pandas
import numpy as np
import scipy
import statsmodels.api as sm
import traceback
import logging
from time import time
import os.path
import sys
from os import getpid
sys.path.append(os.path.join(os.path.dirname(os.path.realpath(__file__)), os.pardir))
sys.path.insert(0, os.path.dirname(__file__))
from settings import (
MIRAGE_ALGORITHMS,
MIRAGE_CONSENSUS,
MIRAGE_DATA_FOLDER,
MIRAGE_ENABLE_SECOND_ORDER,
PANDAS_VERSION,
RUN_OPTIMIZED_WORKFLOW,
SKYLINE_TMP_DIR,
REDIS_SOCKET_PATH,
REDIS_PASSWORD,
)
from algorithm_exceptions import *
skyline_app = 'mirage'
skyline_app_logger = '%sLog' % skyline_app
logger = logging.getLogger(skyline_app_logger)
# @added 20180519 - Feature #2378: Add redis auth to Skyline and rebrow
if MIRAGE_ENABLE_SECOND_ORDER:
from redis import StrictRedis
from msgpack import unpackb, packb
if REDIS_PASSWORD:
redis_conn = StrictRedis(password=REDIS_PASSWORD, unix_socket_path=REDIS_SOCKET_PATH)
else:
redis_conn = StrictRedis(unix_socket_path=REDIS_SOCKET_PATH)
"""
This is no man's land. Do anything you want in here,
as long as you return a boolean that determines whether the input timeseries is
anomalous or not.
The key here is to return a True or False boolean.
You should use the pythonic except mechanism to ensure any excpetions do not
cause things to halt and the record_algorithm_error utility can be used to
sample any algorithm errors to log.
To add an algorithm, define it here, and add its name to settings.MIRAGE_ALGORITHMS.
"""
[docs]def tail_avg(timeseries, second_order_resolution_seconds):
"""
This is a utility function used to calculate the average of the last three
datapoints in the series as a measure, instead of just the last datapoint.
It reduces noise, but it also reduces sensitivity and increases the delay
to detection.
"""
try:
t = (timeseries[-1][1] + timeseries[-2][1] + timeseries[-3][1]) / 3
return t
except IndexError:
return timeseries[-1][1]
[docs]def grubbs(timeseries, second_order_resolution_seconds):
"""
A timeseries is anomalous if the Z score is greater than the Grubb's score.
"""
try:
series = scipy.array([x[1] for x in timeseries])
stdDev = scipy.std(series)
# Issue #27 - Handle z_score agent.py RuntimeWarning - https://github.com/earthgecko/skyline/issues/27
# This change avoids spewing warnings on agent.py tests:
# RuntimeWarning: invalid value encountered in double_scalars
# If stdDev is 0 division returns nan which is not > grubbs_score so
# return False here
if stdDev == 0:
return False
mean = np.mean(series)
tail_average = tail_avg(timeseries, second_order_resolution_seconds)
z_score = (tail_average - mean) / stdDev
len_series = len(series)
threshold = scipy.stats.t.isf(.05 / (2 * len_series), len_series - 2)
threshold_squared = threshold * threshold
grubbs_score = ((len_series - 1) / np.sqrt(len_series)) * np.sqrt(threshold_squared / (len_series - 2 + threshold_squared))
return z_score > grubbs_score
except:
traceback_format_exc_string = traceback.format_exc()
algorithm_name = str(get_function_name())
record_algorithm_error(algorithm_name, traceback_format_exc_string)
return None
[docs]def first_hour_average(timeseries, second_order_resolution_seconds):
"""
Calcuate the simple average over one hour, second order resolution seconds ago.
A timeseries is anomalous if the average of the last three datapoints
are outside of three standard deviations of this value.
"""
try:
last_hour_threshold = time() - (second_order_resolution_seconds - 3600)
series = pandas.Series([x[1] for x in timeseries if x[0] < last_hour_threshold])
mean = (series).mean()
stdDev = (series).std()
t = tail_avg(timeseries, second_order_resolution_seconds)
return abs(t - mean) > 3 * stdDev
except:
traceback_format_exc_string = traceback.format_exc()
algorithm_name = str(get_function_name())
record_algorithm_error(algorithm_name, traceback_format_exc_string)
return None
[docs]def stddev_from_average(timeseries, second_order_resolution_seconds):
"""
A timeseries is anomalous if the absolute value of the average of the latest
three datapoint minus the moving average is greater than three standard
deviations of the average. This does not exponentially weight the MA and so
is better for detecting anomalies with respect to the entire series.
"""
try:
series = pandas.Series([x[1] for x in timeseries])
mean = series.mean()
stdDev = series.std()
t = tail_avg(timeseries, second_order_resolution_seconds)
return abs(t - mean) > 3 * stdDev
except:
traceback_format_exc_string = traceback.format_exc()
algorithm_name = str(get_function_name())
record_algorithm_error(algorithm_name, traceback_format_exc_string)
return None
[docs]def stddev_from_moving_average(timeseries, second_order_resolution_seconds):
"""
A timeseries is anomalous if the absolute value of the average of the latest
three datapoint minus the moving average is greater than three standard
deviations of the moving average. This is better for finding anomalies with
respect to the short term trends.
"""
try:
series = pandas.Series([x[1] for x in timeseries])
if PANDAS_VERSION < '0.18.0':
expAverage = pandas.stats.moments.ewma(series, com=50)
stdDev = pandas.stats.moments.ewmstd(series, com=50)
else:
expAverage = pandas.Series.ewm(series, ignore_na=False, min_periods=0, adjust=True, com=50).mean()
stdDev = pandas.Series.ewm(series, ignore_na=False, min_periods=0, adjust=True, com=50).std(bias=False)
if PANDAS_VERSION < '0.17.0':
return abs(series.iget(-1) - expAverage.iget(-1)) > 3 * stdDev.iget(-1)
else:
return abs(series.iat[-1] - expAverage.iat[-1]) > 3 * stdDev.iat[-1]
# http://stackoverflow.com/questions/28757389/loc-vs-iloc-vs-ix-vs-at-vs-iat
except:
traceback_format_exc_string = traceback.format_exc()
algorithm_name = str(get_function_name())
record_algorithm_error(algorithm_name, traceback_format_exc_string)
return None
[docs]def mean_subtraction_cumulation(timeseries, second_order_resolution_seconds):
"""
A timeseries is anomalous if the value of the next datapoint in the
series is farther than three standard deviations out in cumulative terms
after subtracting the mean from each data point.
"""
try:
series = pandas.Series([x[1] if x[1] else 0 for x in timeseries])
series = series - series[0:len(series) - 1].mean()
stdDev = series[0:len(series) - 1].std()
# @modified 20180910 - Task #2588: Update dependencies
# This expAverage is unused
# if PANDAS_VERSION < '0.18.0':
# expAverage = pandas.stats.moments.ewma(series, com=15)
# else:
# expAverage = pandas.Series.ewm(series, ignore_na=False, min_periods=0, adjust=True, com=15).mean()
if PANDAS_VERSION < '0.17.0':
return abs(series.iget(-1)) > 3 * stdDev
else:
return abs(series.iat[-1]) > 3 * stdDev
except:
traceback_format_exc_string = traceback.format_exc()
algorithm_name = str(get_function_name())
record_algorithm_error(algorithm_name, traceback_format_exc_string)
return None
[docs]def least_squares(timeseries, second_order_resolution_seconds):
"""
A timeseries is anomalous if the average of the last three datapoints
on a projected least squares model is greater than three sigma.
"""
try:
x = np.array([t[0] for t in timeseries])
y = np.array([t[1] for t in timeseries])
A = np.vstack([x, np.ones(len(x))]).T
# @modified 20180910 - Task #2588: Update dependencies
# This results and residual are unused
# results = np.linalg.lstsq(A, y)
# residual = results[1]
# @modified 20180910 - Task #2588: Update dependencies
# Changed in version numpy 1.14.0 - see full comments in
# analyzer/algorithms.py under least_squares np.linalg.lstsq
# m, c = np.linalg.lstsq(A, y)[0]
m, c = np.linalg.lstsq(A, y, rcond=-1)[0]
errors = []
# Evaluate append once, not every time in the loop - this gains ~0.020 s
# on every timeseries potentially @earthgecko #1310
append_error = errors.append
# Further a question exists related to performance and accruracy with
# regards to how many datapoints are in the sample, currently all datapoints
# are used but this may not be the ideal or most efficient computation or
# fit for a timeseries... @earthgecko is checking graphite...
for i, value in enumerate(y):
projected = m * x[i] + c
error = value - projected
# errors.append(error) # @earthgecko #1310
append_error(error)
if len(errors) < 3:
return False
std_dev = scipy.std(errors)
t = (errors[-1] + errors[-2] + errors[-3]) / 3
return abs(t) > std_dev * 3 and round(std_dev) != 0 and round(t) != 0
except:
traceback_format_exc_string = traceback.format_exc()
algorithm_name = str(get_function_name())
record_algorithm_error(algorithm_name, traceback_format_exc_string)
return None
[docs]def histogram_bins(timeseries, second_order_resolution_seconds):
"""
A timeseries is anomalous if the average of the last three datapoints falls
into a histogram bin with less than 20 other datapoints (you'll need to
tweak that number depending on your data)
Returns: the size of the bin which contains the tail_avg. Smaller bin size
means more anomalous.
"""
try:
series = scipy.array([x[1] for x in timeseries])
t = tail_avg(timeseries, second_order_resolution_seconds)
h = np.histogram(series, bins=15)
bins = h[1]
for index, bin_size in enumerate(h[0]):
if bin_size <= 20:
# Is it in the first bin?
if index == 0:
if t <= bins[0]:
return True
# Is it in the current bin?
elif t >= bins[index] and t < bins[index + 1]:
return True
return False
except:
traceback_format_exc_string = traceback.format_exc()
algorithm_name = str(get_function_name())
record_algorithm_error(algorithm_name, traceback_format_exc_string)
return None
[docs]def ks_test(timeseries, second_order_resolution_seconds):
"""
A timeseries is anomalous if 2 sample Kolmogorov-Smirnov test indicates
that data distribution for last 10 minutes is different from last hour.
It produces false positives on non-stationary series so Augmented
Dickey-Fuller test applied to check for stationarity.
"""
try:
hour_ago = time() - 3600
ten_minutes_ago = time() - 600
reference = scipy.array([x[1] for x in timeseries if x[0] >= hour_ago and x[0] < ten_minutes_ago])
probe = scipy.array([x[1] for x in timeseries if x[0] >= ten_minutes_ago])
if reference.size < 20 or probe.size < 20:
return False
ks_d, ks_p_value = scipy.stats.ks_2samp(reference, probe)
if ks_p_value < 0.05 and ks_d > 0.5:
adf = sm.tsa.stattools.adfuller(reference, 10)
if adf[1] < 0.05:
return True
return False
except:
traceback_format_exc_string = traceback.format_exc()
algorithm_name = str(get_function_name())
record_algorithm_error(algorithm_name, traceback_format_exc_string)
return None
return False
"""
THE END of NO MAN'S LAND
THE START of UTILITY FUNCTIONS
"""
[docs]def get_function_name():
"""
This is a utility function is used to determine what algorithm is reporting
an algorithm error when the record_algorithm_error is used.
"""
return traceback.extract_stack(None, 2)[0][2]
[docs]def record_algorithm_error(algorithm_name, traceback_format_exc_string):
"""
This utility function is used to facilitate the traceback from any algorithm
errors. The algorithm functions themselves we want to run super fast and
without fail in terms of stopping the function returning and not reporting
anything to the log, so the pythonic except is used to "sample" any
algorithm errors to a tmp file and report once per run rather than spewing
tons of errors into the log.
.. note::
algorithm errors tmp file clean up
the algorithm error tmp files are handled and cleaned up in
:class:`Analyzer` after all the spawned processes are completed.
:param algorithm_name: the algoritm function name
:type algorithm_name: str
:param traceback_format_exc_string: the traceback_format_exc string
:type traceback_format_exc_string: str
:return:
- ``True`` the error string was written to the algorithm_error_file
- ``False`` the error string was not written to the algorithm_error_file
:rtype:
- boolean
"""
current_process_pid = getpid()
algorithm_error_file = '%s/%s.%s.%s.algorithm.error' % (
SKYLINE_TMP_DIR, skyline_app, str(current_process_pid), algorithm_name)
try:
with open(algorithm_error_file, 'w') as f:
f.write(str(traceback_format_exc_string))
return True
except:
return False
[docs]def is_anomalously_anomalous(metric_name, ensemble, datapoint):
"""
This method runs a meta-analysis on the metric to determine whether the
metric has a past history of triggering. TODO: weight intervals based on datapoint
"""
# We want the datapoint to avoid triggering twice on the same data
new_trigger = [time(), datapoint]
# Get the old history
raw_trigger_history = redis_conn.get('mirage_trigger_history.' + metric_name)
if not raw_trigger_history:
redis_conn.set('mirage_trigger_history.' + metric_name, packb([(time(), datapoint)]))
return True
trigger_history = unpackb(raw_trigger_history)
# Are we (probably) triggering on the same data?
if (new_trigger[1] == trigger_history[-1][1] and
new_trigger[0] - trigger_history[-1][0] <= 300):
return False
# Update the history
trigger_history.append(new_trigger)
redis_conn.set('mirage_trigger_history.' + metric_name, packb(trigger_history))
# Should we surface the anomaly?
trigger_times = [x[0] for x in trigger_history]
intervals = [
trigger_times[i + 1] - trigger_times[i]
for i, v in enumerate(trigger_times)
if (i + 1) < len(trigger_times)
]
series = pandas.Series(intervals)
mean = series.mean()
stdDev = series.std()
return abs(intervals[-1] - mean) > 3 * stdDev
[docs]def run_selected_algorithm(timeseries, metric_name, second_order_resolution_seconds):
"""
Run selected algorithms
"""
try:
ensemble = [globals()[algorithm](timeseries, second_order_resolution_seconds) for algorithm in MIRAGE_ALGORITHMS]
threshold = len(ensemble) - MIRAGE_CONSENSUS
if ensemble.count(False) <= threshold:
if MIRAGE_ENABLE_SECOND_ORDER:
if is_anomalously_anomalous(metric_name, ensemble, timeseries[-1][1]):
return True, ensemble, timeseries[-1][1]
else:
return True, ensemble, timeseries[-1][1]
return False, ensemble, timeseries[-1][1]
except:
logger.error('Algorithm error: %s' % traceback.format_exc())
return False, [], 1