flaky-tests.md

Flaky tests

Any test that fails occasionally is "flaky". Since our merges only proceed when all tests are green, and we have a number of different CI systems running the tests in various combinations, even a small percentage of flakes results in a lot of pain for people waiting for their PRs to merge.

Therefore, it's very important that we write tests defensively. Situations that "almost never happen" happen with some regularity when run thousands of times in resource-constrained environments. Since flakes can often be quite hard to reproduce while still being common enough to block merges occasionally, it's additionally important that the test logs be useful for narrowing down exactly what caused the failure.

Note that flakes can occur in unit tests, integration tests, or end-to-end tests, but probably occur most commonly in end-to-end tests.

Filing issues for flaky tests

Because flakes may be rare, it's very important that all relevant logs be discoverable from the issue.

1. Search for the test name. If you find an open issue and you're 90% sure the flake is exactly the same, add a comment instead of making a new issue.
2. If you make a new issue, you should title it with the test name, prefixed by "e2e/unit/integration flake:" (whichever is appropriate)
3. Reference any old issues you found in step one. Also, make a comment in the old issue referencing your new issue, because people monitoring only their email do not see the backlinks github adds. Alternatively, tag the person or people who most recently worked on it.
4. Paste, in block quotes, the entire log of the individual failing test, not just the failure line.
5. Link to durable storage with the rest of the logs. This means (for all the tests that Google runs) the GCS link is mandatory! The Jenkins test result link is nice but strictly optional: not only does it expire more quickly, it's not accessible to non-Googlers.

Expectations when a flaky test is assigned to you

Note that we won't randomly assign these issues to you unless you've opted in or you're part of a group that has opted in. We are more than happy to accept help from anyone in fixing these, but due to the severity of the problem when merges are blocked, we need reasonably quick turn-around time on test flakes. Therefore we have the following guidelines:

1. If a flaky test is assigned to you, it's more important than anything else you're doing unless you can get a special dispensation (in which case it will be reassigned). If you have too many flaky tests assigned to you, or you have such a dispensation, then it's still your responsibility to find new owners (this may just mean giving stuff back to the relevant Team or SIG Lead).
2. You should make a reasonable effort to reproduce it. Somewhere between an hour and half a day of concentrated effort is "reasonable". It is perfectly reasonable to ask for help!
3. If you can reproduce it (or it's obvious from the logs what happened), you should then be able to fix it, or in the case where someone is clearly more qualified to fix it, reassign it with very clear instructions.
4. PRs that fix or help debug flakes may have the P0 priority set to get them through the merge queue as fast as possible.
5. Once you have made a change that you believe fixes a flake, it is conservative to keep the issue for the flake open and see if it manifests again after the change is merged.
6. If you can't reproduce a flake: don't just close it! Every time a flake comes back, at least 2 hours of merge time is wasted. So we need to make monotonic progress towards narrowing it down every time a flake occurs. If you can't figure it out from the logs, add log messages that would have help you figure it out. If you make changes to make a flake more reproducible, please link your pull request to the flake you're working on.
7. If a flake has been open, could not be reproduced, and has not manifested in 3 months, it is reasonable to close the flake issue with a note saying why.

Reproducing unit test flakes

Try the stress command.

Just

$ go install golang.org/x/tools/cmd/stress

Then build your test binary

$ go test -c -race

Then run it under stress

$ stress ./package.test -test.run=FlakyTest

It runs the command and writes output to /tmp/gostress-* files when it fails. It periodically reports with run counts. Be careful with tests that use the net/http/httptest package; they could exhaust the available ports on your system!

Hunting flaky unit tests in Kubernetes

Sometimes unit tests are flaky. This means that due to (usually) race conditions, they will occasionally fail, even though most of the time they pass.

We have a goal of 99.9% flake free tests. This means that there is only one flake in one thousand runs of a test.

Running a test 1000 times on your own machine can be tedious and time consuming. Fortunately, there is a better way to achieve this using Kubernetes.

Note: these instructions are mildly hacky for now, as we get run once semantics and logging they will get better

There is a testing image brendanburns/flake up on the docker hub. We will use this image to test our fix.

Create a replication controller with the following config:

apiVersion: v1
kind: ReplicationController
metadata:
  name: flakecontroller
spec:
  replicas: 24
  template:
    metadata:
      labels:
        name: flake
    spec:
      containers:
      - name: flake
        image: brendanburns/flake
        env:
        - name: TEST_PACKAGE
          value: pkg/tools
        - name: REPO_SPEC
          value: https://github.com/kubernetes/kubernetes

Note that we omit the labels and the selector fields of the replication controller, because they will be populated from the labels field of the pod template by default.

kubectl create -f ./controller.yaml

This will spin up 24 instances of the test. They will run to completion, then exit, and the kubelet will restart them, accumulating more and more runs of the test.

You can examine the recent runs of the test by calling docker ps -a and looking for tasks that exited with non-zero exit codes. Unfortunately, docker ps -a only keeps around the exit status of the last 15-20 containers with the same image, so you have to check them frequently.

You can use this script to automate checking for failures, assuming your cluster is running on GCE and has four nodes:

echo "" > output.txt
for i in {1..4}; do
  echo "Checking kubernetes-node-${i}"
  echo "kubernetes-node-${i}:" >> output.txt
  gcloud compute ssh "kubernetes-node-${i}" --command="sudo docker ps -a" >> output.txt
done
grep "Exited ([^0])" output.txt

Eventually you will have sufficient runs for your purposes. At that point you can delete the replication controller by running:

kubectl delete replicationcontroller flakecontroller

If you do a final check for flakes with docker ps -a, ignore tasks that exited -1, since that's what happens when you stop the replication controller.

Happy flake hunting!