def test_group_all_by_test(self):
"""
Tests that summarize_by() correctly accumulates all unique combinations of (test,).
"""
report = test_failures.Report(self.ENTRIES)
summed_entries = report.summarize_by(test_failures.Report.TEST)
self.assertEqual(2, len(summed_entries))
self.assertEqual(
summed_entries[0],
self.ENTRY._replace(
task=test_failures.Wildcard("tasks"),
variant=test_failures.Wildcard("variants"),
distro=test_failures.Wildcard("distros"),
start_date=datetime.date(2017, 6, 3),
end_date=datetime.date(2017, 6, 17),
num_pass=3,
num_fail=3,
))
self.assertEqual(
summed_entries[1],
self.ENTRY._replace(
test="jstests/core/all2.js",
start_date=datetime.date(2017, 6, 10),
end_date=datetime.date(2017, 6, 10),
num_pass=1,
num_fail=0,
))
def test_obeys_reliable_time_period(self):
"""
Tests that update_tags() ignores passes from before 'reliable_time_period'.
"""
config = self.CONFIG._replace(
test_fail_rates=self.CONFIG.test_fail_rates._replace(
acceptable=0.9),
task_fail_rates=self.CONFIG.task_fail_rates._replace(
acceptable=0.9),
variant_fail_rates=self.CONFIG.variant_fail_rates._replace(
acceptable=0.9),
distro_fail_rates=self.CONFIG.distro_fail_rates._replace(
acceptable=0.9))
initial_tags = collections.OrderedDict()
lifecycle = ci_tags.TagsConfig.from_dict(
dict(selector=dict(js_test=copy.deepcopy(initial_tags))))
summary_lifecycle = update_test_lifecycle.TagsConfigWithChangelog(
lifecycle)
self.assertEqual(initial_tags,
self.assert_has_only_js_tests(lifecycle))
tests = ["jstests/core/all.js"]
report = test_failures.Report([
self.ENTRY._replace(start_date=(self.ENTRY.start_date -
datetime.timedelta(days=1)),
end_date=(self.ENTRY.end_date -
datetime.timedelta(days=1)),
num_pass=1,
num_fail=0),
self.ENTRY._replace(start_date=(self.ENTRY.start_date -
datetime.timedelta(days=2)),
end_date=(self.ENTRY.end_date -
datetime.timedelta(days=2)),
num_pass=1,
num_fail=0),
self.ENTRY._replace(num_pass=0, num_fail=1),
self.ENTRY._replace(num_pass=0, num_fail=1),
self.ENTRY._replace(num_pass=0, num_fail=1, task="jsCore"),
self.ENTRY._replace(num_pass=0,
num_fail=1,
variant="linux-64-debug"),
self.ENTRY._replace(num_pass=0, num_fail=1, distro="rhel55"),
])
update_test_lifecycle.validate_config(config)
update_test_lifecycle.update_tags(summary_lifecycle, config, report,
tests)
updated_tags = self.assert_has_only_js_tests(lifecycle)
self.assertEqual(
updated_tags,
collections.OrderedDict([
("jstests/core/all.js", [
"unreliable",
"unreliable|jsCore_WT",
"unreliable|jsCore_WT|linux-64",
"unreliable|jsCore_WT|linux-64|rhel62",
]),
]))
def test_non_running_at_all_is_reliable(self):
"""
Tests that tests that are tagged as unreliable but no longer running (either during the
reliable or the unreliable period) have their tags removed.
"""
config = self.CONFIG
tests = ["jstests/core/all.js", "jstests/core/all2.js"]
initial_tags = collections.OrderedDict([
("jstests/core/all2.js", [
"unreliable",
"unreliable|jsCore_WT",
"unreliable|jsCore_WT|linux-64",
"unreliable|jsCore_WT|linux-64|rhel62",
]),
])
lifecycle = ci_tags.TagsConfig.from_dict(
dict(selector=dict(js_test=copy.deepcopy(initial_tags))))
summary_lifecycle = update_test_lifecycle.TagsConfigWithChangelog(
lifecycle)
self.assertEqual(initial_tags,
self.assert_has_only_js_tests(lifecycle))
# all2.js did not run at all
report = test_failures.Report([self.ENTRY])
update_test_lifecycle.validate_config(config)
update_test_lifecycle.update_tags(summary_lifecycle, config, report,
tests)
updated_tags = self.assert_has_only_js_tests(lifecycle)
# The tags for variant and distro have been removed.
self.assertEqual(updated_tags, collections.OrderedDict([]))
def transition_from_unreliable_to_reliable(self, config, initial_tags):
"""
Tests that update_tags() untags a formerly unreliable combination after it has become
reliable again.
"""
lifecycle = ci_tags.TagsConfig.from_dict(
dict(selector=dict(js_test=copy.deepcopy(initial_tags))))
summary_lifecycle = update_test_lifecycle.TagsConfigWithChangelog(
lifecycle)
self.assertEqual(initial_tags,
self.assert_has_only_js_tests(lifecycle))
tests = ["jstests/core/all.js"]
report = test_failures.Report([
self.ENTRY._replace(num_pass=1, num_fail=0),
self.ENTRY._replace(num_pass=1, num_fail=0, task="jsCore"),
self.ENTRY._replace(num_pass=1,
num_fail=0,
variant="linux-64-debug"),
self.ENTRY._replace(num_pass=0, num_fail=1),
self.ENTRY._replace(num_pass=1, num_fail=0, distro="rhel55"),
])
update_test_lifecycle.validate_config(config)
update_test_lifecycle.update_tags(summary_lifecycle, config, report,
tests)
updated_tags = self.assert_has_only_js_tests(lifecycle)
self.assertEqual(updated_tags, collections.OrderedDict())
def test_non_running_in_reliable_period_is_reliable(self):
"""
Tests that tests that have a failure rate above the unacceptable rate during the unreliable
period but haven't run during the reliable period are marked as reliable.
"""
# Unreliable period is 2 days: 2017-06-03 to 2017-06-04.
# Reliable period is 1 day: 2016-06-04.
reliable_period_date = datetime.date(2017, 6, 4)
config = self.CONFIG._replace(
test_fail_rates=self.CONFIG.test_fail_rates._replace(
unacceptable=0.1),
task_fail_rates=self.CONFIG.task_fail_rates._replace(
unacceptable=0.1),
variant_fail_rates=self.CONFIG.variant_fail_rates._replace(
unacceptable=0.1),
distro_fail_rates=self.CONFIG.distro_fail_rates._replace(
unacceptable=0.1),
unreliable_time_period=datetime.timedelta(days=2))
tests = ["jstests/core/all.js"]
initial_tags = collections.OrderedDict([
("jstests/core/all.js", [
"unreliable",
"unreliable|jsCore_WT",
"unreliable|jsCore_WT|linux-64",
"unreliable|jsCore_WT|linux-64|rhel62",
]),
])
lifecycle = ci_tags.TagsConfig.from_dict(
dict(selector=dict(js_test=copy.deepcopy(initial_tags))))
summary_lifecycle = update_test_lifecycle.TagsConfigWithChangelog(
lifecycle)
self.assertEqual(initial_tags,
self.assert_has_only_js_tests(lifecycle))
# The test did not run on the reliable period on linux-64.
report = test_failures.Report([
# Failing.
self.ENTRY._replace(num_pass=0, num_fail=2),
# Passing on a different variant.
self.ENTRY._replace(start_date=reliable_period_date,
end_date=reliable_period_date,
num_pass=3,
num_fail=0,
variant="linux-alt",
distro="debian7"),
])
update_test_lifecycle.validate_config(config)
update_test_lifecycle.update_tags(summary_lifecycle, config, report,
tests)
updated_tags = self.assert_has_only_js_tests(lifecycle)
# The tags for variant and distro have been removed.
self.assertEqual(
updated_tags,
collections.OrderedDict([
("jstests/core/all.js", ["unreliable", "unreliable|jsCore_WT"])
]))
def test_group_all_by_test_task_variant_distro(self):
"""
Tests that summarize_by() correctly accumulates all unique combinations of
(test, task, variant, distro).
"""
report = test_failures.Report(self.ENTRIES)
summed_entries = report.summarize_by(
test_failures.Report.TEST_TASK_VARIANT_DISTRO)
self.assertEqual(5, len(summed_entries))
self.assertEqual(
summed_entries[0],
self.ENTRY._replace(
task="jsCore",
start_date=datetime.date(2017, 6, 5),
end_date=datetime.date(2017, 6, 5),
num_pass=0,
num_fail=1,
))
self.assertEqual(
summed_entries[1],
self.ENTRY._replace(
distro="rhel55",
start_date=datetime.date(2017, 6, 10),
end_date=datetime.date(2017, 6, 10),
num_pass=0,
num_fail=1,
))
self.assertEqual(
summed_entries[2],
self.ENTRY._replace(
start_date=datetime.date(2017, 6, 3),
end_date=datetime.date(2017, 6, 10),
num_pass=3,
num_fail=0,
))
self.assertEqual(
summed_entries[3],
self.ENTRY._replace(
variant="linux-64-debug",
start_date=datetime.date(2017, 6, 17),
end_date=datetime.date(2017, 6, 17),
num_pass=0,
num_fail=1,
))
self.assertEqual(
summed_entries[4],
self.ENTRY._replace(
test="jstests/core/all2.js",
start_date=datetime.date(2017, 6, 10),
end_date=datetime.date(2017, 6, 10),
num_pass=1,
num_fail=0,
))
def test_group_weekly_by_test_starting_on_date(self):
"""
Tests that summarize_by() correctly accumulates by week when the beginning of the week is
specified as a datetime.date() value.
"""
date = datetime.date(2017, 6, 7)
self.assertEqual(2, date.weekday(), "2017 Jun 07 is a Wednesday")
report = test_failures.Report(self.ENTRIES)
summed_entries = report.summarize_by(
test_failures.Report.TEST,
time_period=test_failures.Report.WEEKLY,
start_day_of_week=date)
self.assertEqual(4, len(summed_entries))
self.assertEqual(
summed_entries[0],
self.ENTRY._replace(
task=test_failures.Wildcard("tasks"),
start_date=datetime.date(2017, 6, 3),
end_date=datetime.date(2017, 6, 6),
num_pass=1,
num_fail=1,
))
self.assertEqual(
summed_entries[1],
self.ENTRY._replace(
distro=test_failures.Wildcard("distros"),
start_date=datetime.date(2017, 6, 7),
end_date=datetime.date(2017, 6, 13),
num_pass=2,
num_fail=1,
))
self.assertEqual(
summed_entries[2],
self.ENTRY._replace(
variant="linux-64-debug",
start_date=datetime.date(2017, 6, 14),
end_date=datetime.date(2017, 6, 17),
num_pass=0,
num_fail=1,
))
self.assertEqual(
summed_entries[3],
self.ENTRY._replace(
test="jstests/core/all2.js",
start_date=datetime.date(2017, 6, 7),
end_date=datetime.date(2017, 6, 13),
num_pass=1,
num_fail=0,
))
def test_group_weekly_by_test_starting_on_monday(self):
"""
Tests that summarize_by() correctly accumulates by week when the beginning of the week is
specified as the string "monday".
"""
report = test_failures.Report(self.ENTRIES)
summed_entries = report.summarize_by(
test_failures.Report.TEST,
time_period=test_failures.Report.WEEKLY,
start_day_of_week=test_failures.Report.MONDAY)
self.assertEqual(4, len(summed_entries))
self.assertEqual(
summed_entries[0],
self.ENTRY._replace(
start_date=datetime.date(2017, 6, 3),
end_date=datetime.date(2017, 6, 4),
num_pass=1,
num_fail=0,
))
self.assertEqual(
summed_entries[1],
self.ENTRY._replace(
task=test_failures.Wildcard("tasks"),
distro=test_failures.Wildcard("distros"),
start_date=datetime.date(2017, 6, 5),
end_date=datetime.date(2017, 6, 11),
num_pass=2,
num_fail=2,
))
self.assertEqual(
summed_entries[2],
self.ENTRY._replace(
variant="linux-64-debug",
start_date=datetime.date(2017, 6, 12),
end_date=datetime.date(2017, 6, 17),
num_pass=0,
num_fail=1,
))
self.assertEqual(
summed_entries[3],
self.ENTRY._replace(
test="jstests/core/all2.js",
start_date=datetime.date(2017, 6, 5),
end_date=datetime.date(2017, 6, 11),
num_pass=1,
num_fail=0,
))
def test_remain_unreliable(self):
"""
Tests that update_tags() preserves the tags for unreliable combinations.
"""
config = self.CONFIG._replace(
test_fail_rates=self.CONFIG.test_fail_rates._replace(
unacceptable=0.1),
task_fail_rates=self.CONFIG.task_fail_rates._replace(
unacceptable=0.1),
variant_fail_rates=self.CONFIG.variant_fail_rates._replace(
unacceptable=0.1),
distro_fail_rates=self.CONFIG.distro_fail_rates._replace(
unacceptable=0.1))
initial_tags = collections.OrderedDict([
("jstests/core/all.js", [
"unreliable",
"unreliable|jsCore_WT",
"unreliable|jsCore_WT|linux-64",
"unreliable|jsCore_WT|linux-64|rhel62",
]),
])
lifecycle = ci_tags.TagsConfig.from_dict(
dict(selector=dict(js_test=copy.deepcopy(initial_tags))))
summary_lifecycle = update_test_lifecycle.TagsConfigWithChangelog(
lifecycle)
self.assertEqual(initial_tags,
self.assert_has_only_js_tests(lifecycle))
tests = ["jstests/core/all.js"]
report = test_failures.Report([
self.ENTRY._replace(num_pass=0, num_fail=1),
self.ENTRY._replace(num_pass=0, num_fail=1, task="jsCore"),
self.ENTRY._replace(num_pass=0,
num_fail=1,
variant="linux-64-debug"),
self.ENTRY._replace(num_pass=1, num_fail=0),
self.ENTRY._replace(num_pass=0, num_fail=1, distro="rhel55"),
])
update_test_lifecycle.validate_config(config)
update_test_lifecycle.update_tags(summary_lifecycle, config, report,
tests)
updated_tags = self.assert_has_only_js_tests(lifecycle)
self.assertEqual(updated_tags, initial_tags)
def test_group_4days_by_test(self):
"""
Tests that summarize_by() correctly accumulates by multiple days.
"""
report = test_failures.Report(self.ENTRIES)
summed_entries = report.summarize_by(
test_failures.Report.TEST, time_period=datetime.timedelta(days=4))
self.assertEqual(4, len(summed_entries))
self.assertEqual(
summed_entries[0],
self.ENTRY._replace(
task=test_failures.Wildcard("tasks"),
start_date=datetime.date(2017, 6, 3),
end_date=datetime.date(2017, 6, 6),
num_pass=1,
num_fail=1,
))
self.assertEqual(
summed_entries[1],
self.ENTRY._replace(
distro=test_failures.Wildcard("distros"),
start_date=datetime.date(2017, 6, 7),
end_date=datetime.date(2017, 6, 10),
num_pass=2,
num_fail=1,
))
self.assertEqual(
summed_entries[2],
self.ENTRY._replace(
variant="linux-64-debug",
start_date=datetime.date(2017, 6, 15),
end_date=datetime.date(2017, 6, 17),
num_pass=0,
num_fail=1,
))
self.assertEqual(
summed_entries[3],
self.ENTRY._replace(
test="jstests/core/all2.js",
start_date=datetime.date(2017, 6, 7),
end_date=datetime.date(2017, 6, 10),
num_pass=1,
num_fail=0,
))
def test_obeys_unreliable_min_runs(self):
"""
Tests that update_tags() only considers a test unreliable if it has more than
'unreliable_min_runs'.
"""
config = self.CONFIG._replace(
test_fail_rates=self.CONFIG.test_fail_rates._replace(
unacceptable=0.1),
task_fail_rates=self.CONFIG.task_fail_rates._replace(
unacceptable=0.1),
variant_fail_rates=self.CONFIG.variant_fail_rates._replace(
unacceptable=0.1),
distro_fail_rates=self.CONFIG.distro_fail_rates._replace(
unacceptable=0.1),
unreliable_min_runs=100)
initial_tags = collections.OrderedDict()
lifecycle = ci_tags.TagsConfig.from_dict(
dict(selector=dict(js_test=copy.deepcopy(initial_tags))))
summary_lifecycle = update_test_lifecycle.TagsConfigWithChangelog(
lifecycle)
self.assertEqual(initial_tags,
self.assert_has_only_js_tests(lifecycle))
tests = ["jstests/core/all.js"]
report = test_failures.Report([
self.ENTRY._replace(num_pass=0, num_fail=1),
self.ENTRY._replace(num_pass=0, num_fail=1, task="jsCore"),
self.ENTRY._replace(num_pass=0,
num_fail=1,
variant="linux-64-debug"),
self.ENTRY._replace(num_pass=1, num_fail=0),
self.ENTRY._replace(num_pass=0, num_fail=1, distro="rhel55"),
])
update_test_lifecycle.validate_config(config)
update_test_lifecycle.update_tags(summary_lifecycle, config, report,
tests)
updated_tags = self.assert_has_only_js_tests(lifecycle)
self.assertEqual(updated_tags, initial_tags)
def test_group_all_by_variant_task(self):
"""
Tests that summarize_by() correctly accumulates all unique combinations of (variant, task).
"""
report = test_failures.Report(self.ENTRIES)
summed_entries = report.summarize_by(["variant", "task"])
self.assertEqual(3, len(summed_entries))
self.assertEqual(
summed_entries[0],
self.ENTRY._replace(
task="jsCore",
start_date=datetime.date(2017, 6, 5),
end_date=datetime.date(2017, 6, 5),
num_pass=0,
num_fail=1,
))
self.assertEqual(
summed_entries[1],
self.ENTRY._replace(
test=test_failures.Wildcard("tests"),
distro=test_failures.Wildcard("distros"),
start_date=datetime.date(2017, 6, 3),
end_date=datetime.date(2017, 6, 10),
num_pass=4,
num_fail=1,
))
self.assertEqual(
summed_entries[2],
self.ENTRY._replace(
variant="linux-64-debug",
start_date=datetime.date(2017, 6, 17),
end_date=datetime.date(2017, 6, 17),
num_pass=0,
num_fail=1,
))
def test_group_daily_by_test(self):
"""
Tests that summarize_by() correctly accumulates by day.
"""
report = test_failures.Report(self.ENTRIES)
summed_entries = report.summarize_by(
test_failures.Report.TEST, time_period=test_failures.Report.DAILY)
self.assertEqual(6, len(summed_entries))
self.assertEqual(
summed_entries[0],
self.ENTRY._replace(
start_date=datetime.date(2017, 6, 3),
end_date=datetime.date(2017, 6, 3),
num_pass=1,
num_fail=0,
))
self.assertEqual(
summed_entries[1],
self.ENTRY._replace(
task="jsCore",
start_date=datetime.date(2017, 6, 5),
end_date=datetime.date(2017, 6, 5),
num_pass=0,
num_fail=1,
))
self.assertEqual(
summed_entries[2],
self.ENTRY._replace(
start_date=datetime.date(2017, 6, 9),
end_date=datetime.date(2017, 6, 9),
num_pass=1,
num_fail=0,
))
self.assertEqual(
summed_entries[3],
self.ENTRY._replace(
distro=test_failures.Wildcard("distros"),
start_date=datetime.date(2017, 6, 10),
end_date=datetime.date(2017, 6, 10),
num_pass=1,
num_fail=1,
))
self.assertEqual(
summed_entries[4],
self.ENTRY._replace(
variant="linux-64-debug",
start_date=datetime.date(2017, 6, 17),
end_date=datetime.date(2017, 6, 17),
num_pass=0,
num_fail=1,
))
self.assertEqual(
summed_entries[5],
self.ENTRY._replace(
test="jstests/core/all2.js",
start_date=datetime.date(2017, 6, 10),
end_date=datetime.date(2017, 6, 10),
num_pass=1,
num_fail=0,
))
