This metrics-bigquery job generates metrics that summarize data in our Bigquery
test result database. Each metric is defined with a config file that is consumed
by the metrics-bigquery periodic prow job. Each metric config is a yaml file
like the following:
# Metric name
metric: failures
# BigQuery query
query: |
#standardSQL
select /* find the most recent time each job passed (may not be this week) */
job,
max(started) latest_pass
from `k8s-gubernator.build.all`
where
result = 'SUCCESS'
group by job
# JQ filter to make daily results from raw query results
jqfilter: |
[(.[] | select((.latest_pass|length) > 0)
| {(.job): {
latest_pass: (.latest_pass)
}})] | add- build-stats - number of daily builds and pass rate
- presubmit-health - presubmit failure rate and timing across PRs
- failures - find jobs that have been failing the longest
- flakes - find the flakiest jobs this week (and the flakiest tests in each job).
- flakes-daily - find flakes from the previous day. Similar to
flakes, but creates more granular results. - flakes-experiment - Trial run of
flakesusing new repo field in BQ (#19209) - job-health - compute daily health metrics for jobs (runs, tests, failure rate for each, duration percentiles)
- job-flakes - compute consistency of all jobs
- pr-consistency - calculate PR flakiness for the previous day.
- weekly-consistency - compute overall weekly consistency for PRs
- istio-job-flakes - compute overall weekly consistency for postsubmits
To add a new metric, create a PR that adds a new yaml config file
specifying the metric name (metric), the bigquery query to execute (query), and a
jq filter to filter the data for the daily and latest files (jqfilter).
Run ./bigquery.py --config configs/my-new-config.yaml and verify that the
output is what you expect.
Add the new metric to the list above.
After merging, find the new metric on GCS within 24 hours.
Each query is run every 24 hours to produce a json
file containing the complete raw query results named with the format
raw-yyyy-mm-dd.json. The raw file is then filtered with the associated
jq filter and the results are stored in daily-yyyy-mm-dd.json. These
files are stored in the k8s-metrics GCS bucket in a directory named with
the metric name and persist for a year after their creation. Additionally,
the latest filtered results for a metric are stored in the root of the
k8s-metrics bucket and named with the format METRICNAME-latest.json.
If a config specifies the optional jq filter used to create influxdb timeseries data points, then the job will use the filter to generate timeseries points from the raw query results.
At one point, these points were uploaded to a system called velodrome, which had an influxdb instance where they can be used to create graphs and tables, but velodrome is no longer in existence. This may be revised in the future.
The query is written in Standard SQL which is really BigQuery Standard SQL that allows for working with arrays/repeated fields. Each sub-query, from the most indented out, will build a subtable that the outer query runs against. Any one of the sub query blocks can be run independently from the BigQuery console or opionally added to a test query config and run via the same bigquery.py line above.
Consistency means the test, job, pr always produced the same answer. For example suppose we run a build of a job 5 times at the same commit:
- 5 passing runs, 0 failing runs: consistent
- 0 passing runs, 5 failing runs: consistent
- 1-4 passing runs, 1-4 failing runs: inconsistent aka flaked