The follow up to this is what happens when the node reboots and tries to
rejoin. :) Node on fire is "easy".

On Thu, Jan 15, 2015 at 8:49 AM, Eric Tune notifications@github.com wrote:

Write an e2e test that has a pod and a replication controller and multiple
nodes. Delete the node that the pod is on, and see that the absence of the
pod/node is detected by the replication controller, and that a replacement
pod is created.

—
Reply to this email directly or view it on GitHub
#3520.

This could also be accomplished in an integration test.

As a variation of my serve_hostnames soak/reliability test I intend to make a "cauldron" version of it which uses a replication controller and which every so often kills or add pods and checks to make sure the expected number of pods are up over a given window. Would that meet the requirements of the issue?

/cc @gmarek

@satnam6502 That test would certainly be better than nothing.

I'm of the opinion that focused, standalone e2e tests have a lot of value too, and if it were up to me, I'd wrote one of those before embedding the test into a test with a larger scope (reliability). Basically, I think each test should have a purpose which can be described in like one sentence, without use of conjunctions. But that's just my opinion.

Agreed, coherent focused e2e tests are of great value. I can take this one unless someone else is super keen to do it. Back to back meetings today in Seattle but I expect I can have it done on Tuesday/Wednesday if that's not too late for you.

not super keen.

Un-assiging temporarily while I look at issues with our network e2e test. No worries if someone else wants to pick this up before I can get back to it.

@jszczepkowski - does your restart tests cover this?

This exact test case is covered by Nodes.Resize and Nodes.Network. Closing