def MakeChangePoint(series, split_index):
"""Makes a ChangePoint object for the given series at the given point.
Args:
series: A list of (x, y) pairs.
split_index: Index of the first point after the split.
Returns:
A ChangePoint object.
"""
assert 0 <= split_index < len(series)
x_values, y_values = zip(*series)
left, right = y_values[:split_index], y_values[split_index:]
left_median, right_median = math_utils.Median(left), math_utils.Median(
right)
ttest_results = ttest.WelchsTTest(left, right)
return ChangePoint(
x_value=x_values[split_index],
median_before=left_median,
median_after=right_median,
size_before=len(left),
size_after=len(right),
window_start=x_values[0],
window_end=x_values[-1], # inclusive bound
relative_change=math_utils.RelativeChange(left_median, right_median),
std_dev_before=math_utils.StandardDeviation(left),
t_statistic=ttest_results.t,
degrees_of_freedom=ttest_results.df,
p_value=ttest_results.p)
def _PassesThresholds(values, split_index, min_segment_size,
min_absolute_change, min_relative_change, min_steppiness,
multiple_of_std_dev):
"""Checks whether a point in a series appears to be an change point.
Args:
values: A list of numbers.
split_index: An index in the list of numbers.
min_segment_size: Threshold for size of segments before or after a point.
min_absolute_change: Minimum absolute median change threshold.
min_relative_change: Minimum relative median change threshold.
min_steppiness: Threshold for how similar to a step a change point must be.
multiple_of_std_dev: Threshold for change as multiple of std. deviation.
Returns:
A tuple of (bool, string) where the bool indicates whether the split index
passes the thresholds and the string being the reason it did not.
"""
left, right = values[:split_index], values[split_index:]
left_median, right_median = math_utils.Median(left), math_utils.Median(
right)
# 1. Segment size filter.
if len(left) < min_segment_size or len(right) < min_segment_size:
return (False, 'min_segment_size')
# 2. Absolute change filter.
absolute_change = abs(left_median - right_median)
if absolute_change < min_absolute_change:
return (False, 'min_absolute_change')
# 3. Relative change filter.
relative_change = math_utils.RelativeChange(left_median, right_median)
if relative_change < min_relative_change:
return (False, 'min_relative_change')
# 4. Multiple of standard deviation filter.
min_std_dev = min(math_utils.StandardDeviation(left),
math_utils.StandardDeviation(right))
if absolute_change < multiple_of_std_dev * min_std_dev:
return (False, 'min_std_dev')
# 5. Steppiness filter.
steppiness = find_step.Steppiness(values, split_index)
if steppiness < min_steppiness:
return (False, 'min_steppiness')
# Passed all filters!
return (True, 'passed')
def testGetSimilarHistoricalTimings_Same(self):
now = datetime.datetime.now()
self._RecordTiming(
{
'configuration': 'linux',
'benchmark': 'foo',
'story': 'bar1'
}, now - datetime.timedelta(minutes=1), now)
median = math_utils.Median([i for i in range(0, 10)])
std_dev = math_utils.StandardDeviation([i for i in range(0, 10)])
p90 = math_utils.Percentile([i for i in range(0, 10)], 0.9)
for i in range(0, 10):
j = self._RecordTiming(
{
'configuration': 'linux',
'benchmark': 'foo',
'story': 'bar2'
}, now - datetime.timedelta(seconds=i), now)
timings, tags = timing_record.GetSimilarHistoricalTimings(j)
self.assertEqual(['try', 'linux', 'foo', 'bar2'], tags)
self.assertClose(median, timings[0].total_seconds())
self.assertClose(std_dev, timings[1].total_seconds())
self.assertClose(p90, timings[2].total_seconds())
def testZeroMedian_ResultProperties(self):
nums = [3.4, 8, 100.2, 78, 3, -4, 12, 3.14, 1024]
zeroed_nums = find_change_points._ZeroMedian(nums)
# The output of _ZeroMedian has the same standard deviation as the input.
self.assertEqual(math_utils.StandardDeviation(nums),
math_utils.StandardDeviation(zeroed_nums))
# Also, the median of the output is always zero.
self.assertEqual(0, math_utils.Median(zeroed_nums))
def FindMagnitudeBetweenCommits(test_key, start_commit, end_commit):
start_commit = _GitHashToCommitPosition(start_commit)
end_commit = _GitHashToCommitPosition(end_commit)
test = test_key.get()
num_points = anomaly_config.GetAnomalyConfigDict(test).get(
'min_segment_size', find_change_points.MIN_SEGMENT_SIZE)
start_rows = graph_data.GetRowsForTestBeforeAfterRev(
test_key, start_commit, num_points, 0)
end_rows = graph_data.GetRowsForTestBeforeAfterRev(test_key, end_commit, 0,
num_points)
if not start_rows or not end_rows:
return None
median_before = math_utils.Median([r.value for r in start_rows])
median_after = math_utils.Median([r.value for r in end_rows])
return median_after - median_before
def _Estimate(tags, completed_before=None):
records = _QueryTimingRecords(tags, completed_before)
if not records:
if tags:
return _Estimate(tags[:-1])
return None
times = [(r.completed - r.started).total_seconds() for r in records]
median = math_utils.Median(times)
std_dev = math_utils.StandardDeviation(times)
p90 = math_utils.Percentile(times, 0.9)
timings = Timings(datetime.timedelta(seconds=median),
datetime.timedelta(seconds=std_dev),
datetime.timedelta(seconds=p90))
return EstimateResult(timings, tags)
def _ZeroMedian(values):
"""Subtracts the median value in the list from all values in the list."""
median = math_utils.Median(values)
return [val - median for val in values]
def testMedian_EvenLengthList_UsesMeanOfMiddleTwoValues(self):
self.assertEqual(10.0, math_utils.Median([1, 4, 16, 145]))
def testMedian_OddLengthList_UsesMiddleValue(self):
self.assertEqual(4.0, math_utils.Median([1, 4, 16]))
def testMedian_OneValue(self):
self.assertEqual(3.0, math_utils.Median([3]))
def testMedian_EmptyInput_ReturnsNan(self):
self.assertTrue(math.isnan(math_utils.Median([])))