flaky-test-reporter

README

flaky-test-reporter

flaky-test-reporter is a tool that identifies flaky tests by retrospectively analyzing continuous flows, tracks flaky tests with Github issues, and sends summary of flaky tests to Slack channels.

Basic Usage

Flags for this tool are:

• --service-account specifies the path of file containing service account for GCS access.
• --github-account specifies the path of file containing Github token for Github API calls.
• --slack-account specifies the path of file containing Slack token for Slack web API calls.
• --dry-run enables dry-run mode.

IMPORTANT: This tool is NOT intended to run locally, as this could interfere with real Github issues and potentially flood Knative Slack channels

For debugging purpose it's highly recommended to start with --dry-run flag, by passing this flag it will only collect information from GCS/Github/Slack, and all the manipulations of Github/Slack resources are omitted.

Prow Job

There is a prow job that triggers this tool at 4:00/5:00AM(Day light saving) everyday.

Considerations

Criterias for a test to be considered flaky/passed

This tool scans latest 10 runs. A test is considered flaky if it failed in some but not all runs. For a test to be considered pass, it has to pass in all runs. Exceptions are test being ignored or omitted, these may be results of bad runs or test being omitted for any reason, which is tolerized for up to 2 runs. For example, if a test passed 8 times and skipped/omitted 2 times, it's still considered pass.

Logics for Github issue to be created/closed/reopened

See diagram below

Github Issue deduplication

Efficient deduplication is crucial for the sustainability of this tool, and this can be achieved by associate issue with test by unique identifiers. There are 2 identifiers associated with each flaky issue:

• auto:flaky label. This tool adds this label when creating issues, and only manages issues with this label
• [DONT_MODIFY_TEST_IDENTIFIER]...[DONT_MODIFY_TEST_IDENTIFIER] section in issue body. Issue with auto:flaky label but not this identifier is considered abnormal and this tool will stop at information collection phase.

Minimize Noise

Too many flaky tests identified

When there are too many tests found to be flaky, most likely something abnormal is going on, and we don't want to create Github issues for all of them, or list all of them in Slack notifications. There are thresholds defined in constants.go, if the flaky rate went over a threshold there will be only 1 Github issue created, and Slack notification will not list all flaky tests.

Github issue updates

This tool runs everyday, and leaves comments on existing issues everyday, to avoid flooding the issue with comments, all histories are written into the first comment by prepending it in a collapsible format, like example below:

PLACEHOLDER for contents of latest results

>Click to see older results