sd/eureka: tests are incredibly flaky on CI

[peterbourgon]

Cmd-F for FAIL: on each of these:

• https://circleci.com/gh/go-kit/kit/1222
• https://circleci.com/gh/go-kit/kit/1223
• https://circleci.com/gh/go-kit/kit/1224

I can't reproduce the failures locally. I suspect it is a nondeterminsm/timing issue related to the underlying Fargo library. It started happening after we merged the new SD PR, but I'm not sure that was the cause so much as the trigger. Could someone with more domain knowledge take a look?

@yurishkuro @martinbaillie @seh

[yurishkuro]

I'll take a look. Interestingly, it's the unit test that fails, not integration tests, but the latter could use some improvement as they use unconditional sleep().

[seh]

I finally got CircleCI to display the failure, so I'm looking into it now too.

[seh]

The most mystifying part I've found so far is why this unit test expects only one instance to be present. It looks to me like expecting two is more reasonable, given that the application has two instances.

[seh]

@yurishkuro, it looks like the one is a holdover in commit 432e292. The test could pass mistakenly if it retrieves the cache's state before the goroutine feeding the updates to the cache gets a chance to run for the first time. That's why there's an "await" parameter for fargo.ScheduleAppUpdates: to allow for coordination like this.

[yurishkuro]

@seh unless there is an error retrieving the instances in NewInstancer(), the state should be initialized before NewInstancer returns

[yurishkuro]

But the following update could add 2 instances instead of one given to connection at construction, and that's certainly a race condition, the instancer could get either of those two states

[seh]

eureka.NewInstancer starts a goroutine here. That goroutine may not get a chance to run before NewInstancer completes and the test goes on to probe the cache.

[seh]

I could see adding a means to eureka.testConnection to wait until its first update has been delivered via (*testConnection).ScheduleAppUpdates. We'd have to add a mutex and boolean field.

[seh]

If you're amenable, I can submit a proposal, but I know that this is mostly your code, so I understand if you'd like to fix it yourself. Please let me know how to proceed.

[yurishkuro]

Yeah. Strictly speaking the main test should be checkin two conditions, that the Instancer gets the initial state when being constructed, and that it gets an update once something changes in the discovery.

[yurishkuro]

Please, feel free to submit a PR. It's not really my code, I only adapted it to the change in the discovery API.

[seh]

Can we change the signature of eureka.NewInstancer, or must we keep it as is to retain backward compatibility?

I ask because I can in a few things to allow the tests to work, but I suspect that callers of eureka.NewInstancer might also want a way to wait until at least one update has arrived before proceeding.

[yurishkuro]

I am not sure it makes sense to me to have that kind of control for the constructor. Under normal circumstances the discovery system is already running somewhere by itself, and you have no guarantee as to when it would send an update. That's why the constructor synchronously calls getInstances(), so once the Instancer is created it has the current state of the discovery system.

[seh]

The test eureka.TestInstancerScheduleUpdates is also racy, but it gets by due to sleeping for 50 milliseconds being enough, apparently, to first beat the goroutine updating the cache and then fall in behind its first update.