Writing good e2e tests for Kubernetes

Patterns and Anti-Patterns

Goals of e2e tests

Beyond the obvious goal of providing end-to-end system test coverage, there are a few less obvious goals that you should bear in mind when designing, writing and debugging your end-to-end tests. In particular, “flaky” tests, which pass most of the time but fail intermittently for difficult-to-diagnose reasons are extremely costly in terms of blurring our regression signals and slowing down our automated merge queue. Up-front time and effort designing your test to be reliable is very well spent. Bear in mind that we have hundreds of tests, each running in dozens of different environments, and if any test in any test environment fails, we have to assume that we potentially have some sort of regression. So if a significant number of tests fail even only 1% of the time, basic statistics dictates that we will almost never have a “green” regression indicator. Stated another way, writing a test that is only 99% reliable is just about useless in the harsh reality of a CI environment. In fact it’s worse than useless, because not only does it not provide a reliable regression indicator, but it also costs a lot of subsequent debugging time, and delayed merges.


If your test fails, it should provide as detailed as possible reasons for the failure in its output. “Timeout” is not a useful error message. “Timed out after 60 seconds waiting for pod xxx to enter running state, still in pending state” is much more useful to someone trying to figure out why your test failed and what to do about it. Specifically, assertion code like the following generates rather useless errors:


Rather annotate your assertion with something like this:

Expect(err).NotTo(HaveOccurred(), "Failed to create %d foobars, only created %d", foobarsReqd, foobarsCreated)

On the other hand, overly verbose logging, particularly of non-error conditions, can make it unnecessarily difficult to figure out whether a test failed and if so why? So don’t log lots of irrelevant stuff either.

Ability to run in non-dedicated test clusters

To reduce end-to-end delay and improve resource utilization when running e2e tests, we try, where possible, to run large numbers of tests in parallel against the same test cluster. This means that:

  1. you should avoid making any assumption (implicit or explicit) that your test is the only thing running against the cluster. For example, making the assumption that your test can run a pod on every node in a cluster is not a safe assumption, as some other tests, running at the same time as yours, might have saturated one or more nodes in the cluster. Similarly, running a pod in the system namespace, and assuming that that will increase the count of pods in the system namespace by one is not safe, as some other test might be creating or deleting pods in the system namespace at the same time as your test. If you do legitimately need to write a test like that, make sure to label it “[Serial]“ so that it’s easy to identify, and not run in parallel with any other tests.
  2. You should avoid doing things to the cluster that make it difficult for other tests to reliably do what they’re trying to do, at the same time. For example, rebooting nodes, disconnecting network interfaces, or upgrading cluster software as part of your test is likely to violate the assumptions that other tests might have made about a reasonably stable cluster environment. If you need to write such tests, please label them as “[Disruptive]“ so that it’s easy to identify them, and not run them in parallel with other tests.
  3. You should avoid making assumptions about the Kubernetes API that are not part of the API specification, as your tests will break as soon as these assumptions become invalid. For example, relying on specific Events, Event reasons or Event messages will make your tests very brittle.

Speed of execution

We have hundreds of e2e tests, some of which we run in serial, one after the other, in some cases. If each test takes just a few minutes to run, that very quickly adds up to many, many hours of total execution time. We try to keep such total execution time down to a few tens of minutes at most. Therefore, try (very hard) to keep the execution time of your individual tests below 2 minutes, ideally shorter than that. Concretely, adding inappropriately long ‘sleep’ statements or other gratuitous waits to tests is a killer. If under normal circumstances your pod enters the running state within 10 seconds, and 99.9% of the time within 30 seconds, it would be gratuitous to wait 5 minutes for this to happen. Rather just fail after 30 seconds, with a clear error message as to why your test failed (“e.g. Pod x failed to become ready after 30 seconds, it usually takes 10 seconds”). If you do have a truly legitimate reason for waiting longer than that, or writing a test which takes longer than 2 minutes to run, comment very clearly in the code why this is necessary, and label the test as “[Slow]“, so that it’s easy to identify and avoid in test runs that are required to complete timeously (for example those that are run against every code submission before it is allowed to be merged). Note that completing within, say, 2 minutes only when the test passes is not generally good enough. Your test should also fail in a reasonable time. We have seen tests that, for example, wait up to 10 minutes for each of several pods to become ready. Under good conditions these tests might pass within a few seconds, but if the pods never become ready (e.g. due to a system regression) they take a very long time to fail and typically cause the entire test run to time out, so that no results are produced. Again, this is a lot less useful than a test that fails reliably within a minute or two when the system is not working correctly.

Resilience to relatively rare, temporary infrastructure glitches or delays

Remember that your test will be run many thousands of times, at different times of day and night, probably on different cloud providers, under different load conditions. And often the underlying state of these systems is stored in eventually consistent data stores. So, for example, if a resource creation request is theoretically asynchronous, even if you observe it to be practically synchronous most of the time, write your test to assume that it’s asynchronous (e.g. make the “create” call, and poll or watch the resource until it’s in the correct state before proceeding). Similarly, don’t assume that API endpoints are 100% available. They’re not. Under high load conditions, API calls might temporarily fail or time-out. In such cases it’s appropriate to back off and retry a few times before failing your test completely (in which case make the error message very clear about what happened, e.g. “Retried http://… 3 times - all failed with xxx”. Use the standard retry mechanisms provided in the libraries detailed below.

Some concrete tools at your disposal

Obviously most of the above goals apply to many tests, not just yours. So we’ve developed a set of reusable test infrastructure, libraries and best practices to help you to do the right thing, or at least do the same thing as other tests, so that if that turns out to be the wrong thing, it can be fixed in one place, not hundreds, to be the right thing.

Here are a few pointers:

  • E2e Framework: Familiarise yourself with this test framework and how to use it. Amongst others, it automatically creates uniquely named namespaces within which your tests can run to avoid name clashes, and reliably automates cleaning up the mess after your test has completed (it just deletes everything in the namespace). This helps to ensure that tests do not leak resources. Note that deleting a namespace (and by implication everything in it) is currently an expensive operation. So the fewer resources you create, the less cleaning up the framework needs to do, and the faster your test (and other tests running concurrently with yours) will complete. Your tests should always use this framework. Trying other home-grown approaches to avoiding name clashes and resource leaks has proven to be a very bad idea.
  • E2e utils library: This handy library provides tons of reusable code for a host of commonly needed test functionality, including waiting for resources to enter specified states, safely and consistently retrying failed operations, usefully reporting errors, and much more. Make sure that you’re familiar with what’s available there, and use it. Likewise, if you come across a generally useful mechanism that’s not yet implemented there, add it so that others can benefit from your brilliance. In particular pay attention to the variety of timeout and retry related constants at the top of that file. Always try to reuse these constants rather than try to dream up your own values. Even if the values there are not precisely what you would like to use (timeout periods, retry counts etc), the benefit of having them be consistent and centrally configurable across our entire test suite typically outweighs your personal preferences.
  • Follow the examples of stable, well-written tests: Some of our existing end-to-end tests are better written and more reliable than others. A few examples of well-written tests include: Replication Controllers, Services, Reboot.
  • Ginkgo Test Framework: This is the test library and runner upon which our e2e tests are built. Before you write or refactor a test, read the docs and make sure that you understand how it works. In particular be aware that every test is uniquely identified and described (e.g. in test reports) by the concatenation of its Describe clause and nested It clauses. So for example Describe("Pods",...).... It(""should be scheduled with cpu and memory limits") produces a sane test identifier and descriptor Pods should be scheduled with cpu and memory limits, which makes it clear what’s being tested, and hence what’s not working if it fails. Other good examples include:
   CAdvisor should be healthy on every node


   Daemon set should run and stop complex daemon

On the contrary (these are real examples), the following are less good test descriptors:

   KubeProxy should test kube-proxy


Nodes [Disruptive] Network when a node becomes unreachable
[replication controller] recreates pods scheduled on the
unreachable node AND allows scheduling of pods on a node after
it rejoins the cluster

An improvement might be

Unreachable nodes are evacuated and then repopulated upon rejoining [Disruptive]

Note that opening issues for specific better tooling is welcome, and code implementing that tooling is even more welcome :-).