This updates the EndpointSlice controller to make use of the
EndpointSlice tracker to identify when expected changes are not present
in the cache yet. If this is detected, the controller will wait to sync
until all expected updates have been received. This should help avoid
race conditions that would result in duplicate EndpointSlices or failed
attempts to update stale EndpointSlices. To simplify this logic, this
also moves the EndpointSlice tracker from relying on resource versions
to generations.
This is to avoid unnecessary GCE API calls done by getInstanceByName
helper, which is iterating over all zones to find in which zone the
VM exists.
ProviderID already contains all the information - it's in the form:
gce://<VM URL> (VM URL contains project, zone, VM name).
ProviderID is propagated by Kubelet on node registration and in case
of bugs backfilled by node-controller.
When a pod is deleted, it is given a deletion timestamp. However the
pod might still run for some time during graceful shutdown. During
this time it might still produce CPU utilization metrics and be in a
Running phase.
Currently the HPA replica calculator attempts to ignore deleted pods
by skipping over them. However by not adding them to the ignoredPods
set, their metrics are not removed from the average utilization
calculation. This allows pods in the process of shutting down to drag
down the recommmended number of replicas by producing near 0%
utilization metrics.
In fact the ignoredPods set is misnomer. Those pods are not fully
ignored. When the replica calculator recommends to scale up, 0%
utilization metrics are filled in for those pods to limit the scale
up. This prevents overscaling when pods take some time to startup. In
fact, there should be 4 sets considered (readyPods, unreadyPods,
missingPods, ignoredPods) not just 3.
This change renames ignoredPods as unreadyPods and leaves the scaleup
limiting semantics. Another set (actually) ignoredPods is added to
which delete pods are added instead of being skipped during
grouping. Both ignoredPods and unreadyPods have their metrics removed
from consideration. But only unreadyPods have 0% utilization metrics
filled in upon scaleup.
EndpointController was accidentally requiring all headless services to
be IPv4-only in clusters with IPv6DualStack enabled.
This still leaves "legacy" (ie, IPFamily-less) headless services as
always IPv4-only because the controller doesn't currently have easy
access to the information that would allow it to fix that.
(EndpointSliceController had the same problem already, and still
does.) This can be fixed, if needed, by manually setting IPFamily,
and the proposed API for 1.20 will handle this situation better.
Rewrite some of the test helpers to better support single-stack IPv4
vs single-stack IPv6 vs dual-stack IPv4 primary vs dual-stack IPv6
primary, and update TestPodToEndpointAddressForService to test some
more cases.
The endpoint controllers responded to Pod changes by trying to figure
out if the generated endpoint resource would change, rather than just
checking if the Pod had changed, but since the set of Pod fields that
need to be checked depend on the Service and Node as well, the code
ended up only checking for a subset of the changes it should have.
In particular, EndpointSliceController ended up only looking at IPv4
Pod IPs when processing Pod update events, so when a Pod went from
having no IP to having only an IPv6 IP, EndpointSliceController would
think it hadn't changed.
Most of these could have been refactored automatically but it wouldn't
have been uglier. The unsophisticated tooling left lots of unnecessary
struct -> pointer -> struct transitions.
This is gross but because NewDeleteOptions is used by various parts of
storage that still pass around pointers, the return type can't be
changed without significant refactoring within the apiserver. I think
this would be good to cleanup, but I want to minimize apiserver side
changes as much as possible in the client signature refactor.
