build-runtime-config was being called in verify-prereqs, which didn't
match how GCE called it, and didn't seem to actually work.
Instead call it just before the master configuration is built. Also
call it just before the node configuration is built, even though the
nodes don't _currently_ require the runtime_config.
According to AWS, the ELB healthy threshold is "Number of consecutive health check successes before declaring an EC2 instance healthy." It has an unusual interaction with Kubernetes, since all nodes will enter either an unhealthy state or a healthy state together depending on the service's healthiness as a whole.
We have observed that if our service goes down for the unhealthy threshold (which is 2 checks at 30 second intervals = 60 seconds), then the ELB will stop serving traffic to all nodes in the cluster, and will wait for the healthy threshold (currently 10 * 30 = 300 seconds) AFTER the service is restored to add back the cluster nodes, meaning it remains unreachable for an extra 300 seconds.
With the new settings, the ELB will continue to timeout dead nodes after 60 seconds, but will restore healthy nodes after 20 seconds. The minimum value for healthyThreshold is 2, and the minimum value for interval is 5 seconds. I went for 10 seconds instead of the minimum sort of arbitrarily because I was not sure how much this value may affect the scalability of clusters in EC2, as it does put some extra load on the kube-proxy.
We decided to remove this test, as there's no way to get an upper bound
on its running time. Etcd restart behavior should be tested in
integration or e2e tests.
When job.spec.completions is nil, only
one task needs to succeed for the job to succeed,
and parallelism can be scaled freely during runtime.
Added tests.
Release Note:
This causes two minor changes to the API.
First, unset parallelism previously was defaulted to be
equal to completions. Now it always defaults to 1 if unset.
Second, having parallelism=N and completions unset would previously
be defaulted to 1 completion and N parallelism.
(this is not something we expect people to do, though)
Now, no defaulting occurs in that case, and the job's
behavior is different (any completion causes success).
We don't cope well if a PD is in multiple zones, but this is actually
fairly easy to detect. This is probably justified purely on the basis
that we never want to delete the wrong volume (DeleteDisk), but also
because this means that we now warn on creation if a disk is in multiple
zones (with the labeling admission controller).
This also means that with the scheduling predicate in place, that many
of our volume problems "go away" in practice: you still can't create or
delete a volume when it is ambiguous, but thereafter the volume will be
labeled with the zone, that will match it only to nodes with the same
zone, and then we query for the volume in that zone when we
attach/detach it.
