def test_assert_duration(self):
metric_source = metrics_pb2.MetricSource(
literals=metrics_pb2.LiteralSource(
assertions={
"duration":
metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=100,
upper_bound=200,
),
inclusive_bounds=False,
)
}))
event = metrics_pb2.TestCompletedEvent(
benchmark_id="test_benchmark",
duration=duration_pb2.Duration(seconds=150),
metric_collection_config=metrics_pb2.MetricCollectionConfig(
sources=[metric_source]))
collector = literal_collector.LiteralCollector(
event=event, raw_source=metric_source)
points = collector.metric_points()
self.assertLen(points, 1)
self.assertEqual(points[0].metric_key, 'duration')
self.assertEqual(points[0].metric_value, 150)
self.assertEqual(points[0].bounds, utils.Bounds(100, 200, False))
def test_create_test_completed_event(self, succeeded_count, failed_count, conditions, expected_status):
job = _job_from_dict({
'metadata': {
'name': 'job-name',
'namespace': 'namespace',
'labels': {
'benchmarkId': 'test-job',
},
},
'status': {
'startTime': _START_TIME,
'succeeded': succeeded_count,
'failed': failed_count,
'conditions': [
{
'status': True,
'reason': reason,
'type': cond_type,
'lastTransitionTime': _END_TIME,
}
for cond_type, reason in conditions
]
}
})
actual_event = event_publisher.create_test_completed_event(
job,
model_output_bucket='gs://fake-bucket',
cluster_name='cluster-name',
cluster_location='cluster-location',
project='project-id'
)
start_time = timestamp_pb2.Timestamp()
start_time.FromDatetime(_START_TIME)
duration = duration_pb2.Duration()
duration.FromTimedelta(_END_TIME - _START_TIME)
expected_event = metrics_pb2.TestCompletedEvent(
benchmark_id='test-job',
output_path='gs://fake-bucket/job-name',
status=metrics_pb2.TestCompletedEvent.TestStatus.Value(expected_status),
num_attempts=succeeded_count + failed_count,
start_time=start_time,
duration=duration,
labels={'benchmarkId': 'test-job'},
debug_info=metrics_pb2.DebugInfo(
logs_link='https://console.cloud.google.com/logs?project=project-id&advancedFilter=resource.type%3Dk8s_container%0Aresource.labels.project_id%3Dproject-id%0Aresource.labels.cluster_name%3Dcluster-name%0Aresource.labels.namespace_name%3Dnamespace%0Aresource.labels.pod_name%3Ajob-name%0Aresource.labels.location%3Acluster-location%0A',
details_link=f'https://console.cloud.google.com/kubernetes/job/cluster-location/cluster-name/namespace/job-name?project=project-id'
),
metric_collection_config=metrics_pb2.MetricCollectionConfig(),
)
self.assertProtoEqual(expected_event, actual_event)
def test_aggregate_metrics_include_all_strategies(self):
metric_source = metrics_pb2.MetricSource(
tensorboard=metrics_pb2.TensorBoardSource(include_tags=[
metrics_pb2.TensorBoardSource.TagStrategy(
tag_pattern="*",
strategies=[
metrics_pb2.TensorBoardSource.FINAL,
metrics_pb2.TensorBoardSource.MAX,
metrics_pb2.TensorBoardSource.MIN,
metrics_pb2.TensorBoardSource.AVERAGE,
metrics_pb2.TensorBoardSource.MEDIAN,
])
]))
event = metrics_pb2.TestCompletedEvent(
benchmark_id="test_benchmark",
output_path=self.temp_dir,
metric_collection_config=metrics_pb2.MetricCollectionConfig(
sources=[metric_source]))
collector = tensorboard_collector.TensorBoardCollector(
event=event, raw_source=metric_source)
points = list(collector.metric_points())
metric_to_value = {key: value for key, value, _ in points}
self.assertDictEqual(
metric_to_value, {
'foo_final': 2,
'foo_min': 1,
'foo_max': 2,
'foo_average': 1.5,
'foo_median': 1.5,
'eval/accuracy_final': .25,
'eval/accuracy_min': .125,
'eval/accuracy_max': .5,
'eval/accuracy_average': np.mean([.125, .25, .5]),
'eval/accuracy_median': np.median([.125, .25, .5]),
'train/bar_final': 100,
'train/bar_min': 10,
'train/bar_max': 100,
'train/bar_average': np.mean([10, 100, 100]),
'train/bar_median': np.median([10, 100, 100]),
})
for _, _, bounds in points:
self.assertEqual(bounds, utils.NO_BOUNDS)
def test_aggregate_metrics_with_assertion(self):
metric_source = metrics_pb2.MetricSource(
tensorboard=metrics_pb2.TensorBoardSource(
include_tags=[
metrics_pb2.TensorBoardSource.TagStrategy(
tag_pattern="eval/*",
strategies=[
metrics_pb2.TensorBoardSource.FINAL,
metrics_pb2.TensorBoardSource.MAX,
metrics_pb2.TensorBoardSource.MIN,
])
],
aggregate_assertions=[
metrics_pb2.TensorBoardSource.AggregateAssertion(
tag='eval/accuracy',
strategy=metrics_pb2.TensorBoardSource.MAX,
assertion=metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=.4,
upper_bound=1.0,
),
inclusive_bounds=True,
))
]))
event = metrics_pb2.TestCompletedEvent(
benchmark_id="test_benchmark",
output_path=self.temp_dir,
metric_collection_config=metrics_pb2.MetricCollectionConfig(
sources=[metric_source]))
collector = tensorboard_collector.TensorBoardCollector(
event=event, raw_source=metric_source)
points = list(collector.metric_points())
self.assertCountEqual(
points,
[
utils.MetricPoint('eval/accuracy_max', .5,
utils.Bounds(.4, 1.0, True)),
utils.MetricPoint('eval/accuracy_min', .125, utils.NO_BOUNDS),
utils.MetricPoint('eval/accuracy_final', .25, utils.NO_BOUNDS),
],
)
def test_include_and_exclude(self):
metric_source = metrics_pb2.MetricSource(
tensorboard=metrics_pb2.TensorBoardSource(
include_tags=[
metrics_pb2.TensorBoardSource.TagStrategy(
tag_pattern="*",
strategies=[
metrics_pb2.TensorBoardSource.FINAL,
])
],
exclude_tags=[
'foo',
'train/*',
],
aggregate_assertions=[
metrics_pb2.TensorBoardSource.AggregateAssertion(
tag='foo',
strategy=metrics_pb2.TensorBoardSource.MIN,
assertion=metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=0.,
upper_bound=2.,
)))
]))
event = metrics_pb2.TestCompletedEvent(
benchmark_id="test_benchmark",
output_path=self.temp_dir,
metric_collection_config=metrics_pb2.MetricCollectionConfig(
sources=[metric_source]))
collector = tensorboard_collector.TensorBoardCollector(
event=event, raw_source=metric_source)
points = list(collector.metric_points())
self.assertCountEqual(
points,
[
utils.MetricPoint('eval/accuracy_final', .25, utils.NO_BOUNDS),
utils.MetricPoint('foo_min', 1, utils.Bounds(0., 2., False)),
],
)
def test_metric_collection_config(self, gcs_subdir):
job = _job_from_dict({
'metadata': {
'name': 'job-name',
'namespace': 'namespace',
'labels': {
'benchmarkId': 'test-job',
},
'annotations': {
'ml-testing-accelerators/metric-config': json.dumps({
'sources': [{
'literals': {
'assertions': {
'duration': {
'within_bounds': {
'lower_bound': 1,
'upper_bound': 2,
}
}
}
}
}]
})
}
},
'status': {
'startTime': _START_TIME,
'completionTime': _END_TIME,
'succeeded': 1,
'conditions': [
{
'status': True,
'type': 'Complete',
'lastTransitionTime': _END_TIME,
}
]
}
})
if gcs_subdir:
job.metadata.annotations['ml-testing-accelerators/gcs-subdir'] = gcs_subdir
actual_event = event_publisher.create_test_completed_event(
job,
model_output_bucket='gs://fake-bucket',
cluster_name='cluster-name',
cluster_location='cluster-location',
project='project-id'
)
actual_mcc = actual_event.metric_collection_config
expected_mcc = metrics_pb2.MetricCollectionConfig(
sources=[
metrics_pb2.MetricSource(
literals=metrics_pb2.LiteralSource(
assertions={
'duration': metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=1,
upper_bound=2,
)
)
}
)
)
]
)
self.assertEqual(actual_event.output_path, os.path.join('gs://fake-bucket', gcs_subdir or '', 'job-name'))
self.assertProtoEqual(expected_mcc, actual_mcc)
def test_get_metrics_from_perfzero_summary(self):
temp_dir = self.create_tempdir().full_path
summary_dir = os.path.join(temp_dir, 'date_and_time')
pathlib.Path(summary_dir).mkdir(parents=True, exist_ok=True)
summary_path = os.path.join(summary_dir, 'perfzero_summary.json')
with open(summary_path, 'w') as f:
json.dump(
{
"execution_id": "execution_id",
"execution_timestamp": 1234567890.1,
"benchmark_result": {
"wall_time":
1234,
"metrics": [{
"name": "exp_per_second",
"value": 1.1,
}, {
"name": "avg_exp_per_second",
"value": 2.2,
}, {
"name": "startup_time",
"value": 3.3
}],
},
"benchmark_info": {
"not": "important",
},
"setup_info": {},
"ml_framework_info": {
"not": "important",
},
"system_info": {
"not": "important"
},
"process_info": {
"max_rss": 4.4,
"max_vms": 5.5,
"max_cpu_percent": 6.6,
}
}, f)
metric_source = metrics_pb2.MetricSource(
perfzero=metrics_pb2.PerfZeroSource(
assertions={
'total_wall_time':
metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=1230,
upper_bound=1240,
)),
'exp_per_second':
metrics_pb2.Assertion(
within_bounds=metrics_pb2.Assertion.WithinBounds(
lower_bound=1,
upper_bound=100,
), )
}))
event = metrics_pb2.TestCompletedEvent(
benchmark_id="test_benchmark",
output_path=temp_dir,
metric_collection_config=metrics_pb2.MetricCollectionConfig(
sources=[metric_source]))
collector = perfzero_collector.PerfZeroCollector(
event=event, raw_source=metric_source)
points = list(collector.metric_points())
self.assertCountEqual(
points,
{
utils.MetricPoint("total_wall_time", 1234,
utils.Bounds(1230., 1240., False)),
utils.MetricPoint("exp_per_second", 1.1,
utils.Bounds(1., 100., False)),
utils.MetricPoint("avg_exp_per_second", 2.2, utils.NO_BOUNDS),
utils.MetricPoint("startup_time", 3.3, utils.NO_BOUNDS),
utils.MetricPoint("process_info/max_rss", 4.4,
utils.NO_BOUNDS),
utils.MetricPoint("process_info/max_vms", 5.5,
utils.NO_BOUNDS),
utils.MetricPoint("process_info/max_cpu_percent", 6.6,
utils.NO_BOUNDS),
},
)
def create_test_completed_event(
job: kubernetes.client.V1Job, model_output_bucket: str,
cluster_name: str, cluster_location: str,
project: str) -> metrics_pb2.TestCompletedEvent:
"""Returns a TestCompletedEvent to publish to PubSub.
Args:
job: A Kubernetes Job resource.
model_output_bucket: Path to GCS bucket with model outputs.
cluster_name: Name of the current Kubernetes cluster.
cluster_location: Location (region or zone) of the current Kubernetes cluster.
project: The project ID of the current project.
Returns:
A TestCompletedEvent with the information from job.
"""
if len(job.status.conditions) == 1:
condition = job.status.conditions[0]
# job.status.conditions _usually_ has length 1, but it can have both passing and failing conditions.
# Give precedence to failing conditions.
elif len(job.status.conditions) == 0:
logging.error('Job %s has no conditions.', job.metadata.name)
return
else:
condition = next(
(c for c in job.status.conditions if c.type == 'Failed'), None)
if not condition:
logging.error('This should never happen. Conditions: %s',
str(job.status.conditions))
return
elif condition.reason == 'DeadlineExceeded':
job_status = metrics_pb2.TestCompletedEvent.TIMEOUT
elif condition.reason == 'BackoffLimitExceeded':
job_status = metrics_pb2.TestCompletedEvent.FAILED
elif condition.type == 'Complete':
job_status = metrics_pb2.TestCompletedEvent.COMPLETED
else:
logging.error('Unknown condition for Job %s: %s', job.metadata.name,
str(condition))
return
annotations = job.metadata.annotations or {}
gcs_subdir = annotations.get('ml-testing-accelerators/gcs-subdir', '')
output_path = os.path.join(model_output_bucket, gcs_subdir,
job.metadata.name)
metric_config = metrics_pb2.MetricCollectionConfig()
mcc_json = annotations.get('ml-testing-accelerators/metric-config', '{}')
json_format.Parse(mcc_json, metric_config)
stackdriver_query = textwrap.dedent(f"""\
resource.type=k8s_container
resource.labels.project_id={project}
resource.labels.cluster_name={cluster_name}
resource.labels.namespace_name={job.metadata.namespace}
resource.labels.pod_name:{job.metadata.name}
resource.labels.location:{cluster_location}
""")
stackdriver_link = "https://console.cloud.google.com/logs?{}".format(
urllib.parse.urlencode({
'project': project,
'advancedFilter': stackdriver_query
}))
start_time = timestamp_pb2.Timestamp()
start_time.FromDatetime(job.status.start_time)
duration = duration_pb2.Duration()
duration.FromTimedelta(condition.last_transition_time -
job.status.start_time)
return metrics_pb2.TestCompletedEvent(
benchmark_id=job.metadata.labels['benchmarkId'],
output_path=output_path,
status=job_status,
num_attempts=(job.status.succeeded or 0) + (job.status.failed or 0),
start_time=start_time,
duration=duration,
metric_collection_config=metric_config,
labels=job.metadata.labels,
debug_info=metrics_pb2.DebugInfo(
logs_link=stackdriver_link,
# TODO: fix hard-coded region and cluster name
details_link=
f'https://console.cloud.google.com/kubernetes/job/{cluster_location}/{cluster_name}/{job.metadata.namespace}/{job.metadata.name}?project={project}'
))