Performance regressions or improvements detected with the Change Point Analysis using edivisive analyzer are automatically filed as Beam GitHub issues with a label perf-alert.
The GitHub issue description will contain the information on the affected test and metric by providing the metric values for N consecutive runs with timestamps before and after the observed change point. Observed change point is pointed as Anomaly in the issue description.
Example: sample perf alert GitHub issue.
If a performance alert is created on a test, a GitHub issue will be created and the GitHub issue metadata such as GitHub issue URL, issue number along with the change point value and timestamp are exported to BigQuery. This data will be used to analyze the next change point observed on the same test to update already created GitHub issue or ignore performance alert by not creating GitHub issue to avoid duplicate issue creation.
The config file defines the structure to run change point analysis on a given test. To add a test to the config file, please follow the below structure.
NOTE: The Change point analysis only supports reading the metric data from Big Query for now.
# the test_1 must be a unique id.
test_1:
test_name: Pytorch image classification on 50k images of size 224 x 224 with resnet 152
test_target: apache_beam.testing.benchmarks.inference.pytorch_image_classification_benchmarks
source: big_query
metrics_dataset: beam_run_inference
metrics_table: torch_inference_imagenet_results_resnet152
project: apache-beam-testing
metric_name: mean_load_model_latency_milli_secs
labels:
- run-inference
min_runs_between_change_points: 3 # optional parameter
num_runs_in_change_point_window: 30 # optional parameter
NOTE: test_target is optional. It is used for identifying the test that was causing the regression.
Note: If the source is BigQuery, the metrics_dataset, metrics_table, project and metric_name should match with the values defined for performance/load tests. The above example uses this test configuration to fill up the values required to fetch the data from source.
min_runs_between_change_points: As the metric data moves across the runs, the change point analysis can place the change point in a slightly different place. These change points refer to the same regression and are just noise. When we find a new change point, we will search up to the min_runs_between_change_points in both directions from the current change point. If an existing change point is found within the distance, then the current change point will be suppressed.
num_runs_in_change_point_window: This defines how many runs to consider from the most recent run to be in change point window. Sometimes, the change point found might be way back in time and could be irrelevant. For a test, if a change point needs to be reported only when it was observed in the last 7 runs from the current run, setting num_runs_in_change_point_window=7 will achieve it.
If a new test needs to be registered for the performance alerting tool, please add the required test parameters to the config file.
All the performance/load tests metrics defined at beam/.test-infra/jenkins are imported to Grafana dashboards for visualization. Please find the alerted test dashboard to find a spike in the metric values.
For example, for the below configuration,
apache_beam.testing.benchmarks.inference.pytorch_image_classification_benchmarksmean_load_model_latency_milli_secsGrafana dashboard can be found at http://metrics.beam.apache.org/d/ZpS8Uf44z/python-ml-runinference-benchmarks?orgId=1&viewPanel=7
If the dashboard for a test is not found, you can use the notebook analyze_metric_data.ipynb to generate a plot for the given test, metric_name.
If you confirm there is a change in the pattern of the values for a test, find the timestamp of when that change happened and use that timestamp to find possible culprit commit.
If the performance alert is a false positive, close the issue as Close as not planned.