SyncKnownPods began triggering UpdatePod() for pods that have been
orphaned by desired config to ensure pods run to termination. This
test reads a mutex protected value while pod workers are running
in the background and as a consequence triggers a data race.
Wait for the workers to stabilize before reading the value. Other
tests validate that the correct sync events are triggered (see
kubelet_pods_test.go#TestKubelet_HandlePodCleanups for full
verification of this behavior).
It is slightly concerning that I was unable to recreate the race
locally even under stress testing, but I cannot identify why.
If a CRI error occurs during the terminating phase after a pod is
force deleted (API or static) then the housekeeping loop will not
deliver updates to the pod worker which prevents the pod's state
machine from progressing. The pod will remain in the terminating
phase but no further attempts to terminate or cleanup will occur
until the kubelet is restarted.
The pod worker now maintains a store of the pods state that it is
attempting to reconcile and uses that to resync unknown pods when
SyncKnownPods() is invoked, so that failures in sync methods for
unknown pods no longer hang forever.
The pod worker's store tracks desired updates and the last update
applied on podSyncStatuses. Each goroutine now synchronizes to
acquire the next work item, context, and whether the pod can start.
This synchronization moves the pending update to the stored last
update, which will ensure third parties accessing pod worker state
don't see updates before the pod worker begins synchronizing them.
As a consequence, the update channel becomes a simple notifier
(struct{}) so that SyncKnownPods can coordinate with the pod worker
to create a synthetic pending update for unknown pods (i.e. no one
besides the pod worker has data about those pods). Otherwise the
pending update info would be hidden inside the channel.
In order to properly track pending updates, we have to be very
careful not to mix RunningPods (which are calculated from the
container runtime and are missing all spec info) and config-
sourced pods. Update the pod worker to avoid using ToAPIPod()
and instead require the pod worker to directly use
update.Options.Pod or update.Options.RunningPod for the
correct methods. Add a new SyncTerminatingRuntimePod to prevent
accidental invocations of runtime only pod data.
Finally, fix SyncKnownPods to replay the last valid update for
undesired pods which drives the pod state machine towards
termination, and alter HandlePodCleanups to:
- terminate runtime pods that aren't known to the pod worker
- launch admitted pods that aren't known to the pod worker
Any started pods receive a replay until they reach the finished
state, and then are removed from the pod worker. When a desired
pod is detected as not being in the worker, the usual cause is
that the pod was deleted and recreated with the same UID (almost
always a static pod since API UID reuse is statistically
unlikely). This simplifies the previous restartable pod support.
We are careful to filter for active pods (those not already
terminal or those which have been previously rejected by
admission). We also force a refresh of the runtime cache to
ensure we don't see an older version of the state.
Future changes will allow other components that need to view the
pod worker's actual state (not the desired state the podManager
represents) to retrieve that info from the pod worker.
Several bugs in pod lifecycle have been undetectable at runtime
because the kubelet does not clearly describe the number of pods
in use. To better report, add the following metrics:
kubelet_desired_pods: Pods the pod manager sees
kubelet_active_pods: "Admitted" pods that gate new pods
kubelet_mirror_pods: Mirror pods the kubelet is tracking
kubelet_working_pods: Breakdown of pods from the last sync in
each phase, orphaned state, and static or not
kubelet_restarted_pods_total: A counter for pods that saw a
CREATE before the previous pod with the same UID was finished
kubelet_orphaned_runtime_pods_total: A counter for pods detected
at runtime that were not known to the kubelet. Will be
populated at Kubelet startup and should never be incremented
after.
Add a metric check to our e2e tests that verifies the values are
captured correctly during a serial test, and then verify them in
detail in unit tests.
Adds 23 series to the kubelet /metrics endpoint.
When adding functionality to the kubelet package and a test file, is
kind of painful to run unit tests today locally.
We usually can't run specifying the test file, as if xx_test.go and
xx.go use the same package, we need to specify all the dependencies. As
soon as xx.go uses the Kuebelet type (we need to do that to fake a
kubelet in the unit tests), this is completely impossible to do in
practice.
So the other option is to run the unit tests for the whole package or
run only a specific funtion. Running a single function can work in some
cases, but it is painful when we want to test all the functions we
wrote. On the other hand, running the test for the whole package is very
slow.
Today some unit tests try to connect to the API server (with retries)
create and list lot of pods/volumes, etc. This makes running the unit
test for the kubelet package slow.
This patch tries to make running the unit test for the whole package
more palatable. This patch adds a skip if the short version was
requested (go test -short ...), so we don't try to connect
to the API server or skip other slow tests.
Before this patch running the unit tests took in my computer (I've run
it several times so the compilation is already done):
$ time go test -v
real 0m21.303s
user 0m9.033s
sys 0m2.052s
With this patch it takes ~1/3 of the time:
$ time go test -short -v
real 0m7.825s
user 0m9.588s
sys 0m1.723s
Around 8 seconds is something I can wait to run the tests :)
Signed-off-by: Rodrigo Campos <rodrigoca@microsoft.com>
Other components must know when the Kubelet has released critical
resources for terminal pods. Do not set the phase in the apiserver
to terminal until all containers are stopped and cannot restart.
As a consequence of this change, the Kubelet must explicitly transition
a terminal pod to the terminating state in the pod worker which is
handled by returning a new isTerminal boolean from syncPod.
Finally, if a pod with init containers hasn't been initialized yet,
don't default container statuses or not yet attempted init containers
to the unknown failure state.
- Allow a podWorker to start if it is blocked by a pod that has been
terminated before starting
- When a pod can't start AND has already been terminated, exit cleanly
- Add a unit test that exercises race conditions in pod workers
If a pod is killed (no longer wanted) and then a subsequent create/
add/update event is seen in the pod worker, assume that a pod UID
was reused (as it could be in static pods) and have the next
SyncKnownPods after the pod terminates remove the worker history so
that the config loop can restart the static pod, as well as return
to the caller the fact that this termination was not final.
The housekeeping loop then reconciles the desired state of the Kubelet
(pods in pod manager that are not in a terminal state, i.e. admitted
pods) with the pod worker by resubmitting those pods. This adds a
small amount of latency (2s) when a pod UID is reused and the pod
is terminated and restarted.
Fixes two issues with how the pod worker refactor calculated the
pods that admission could see (GetActivePods() and
filterOutTerminatedPods())
First, completed pods must be filtered from the "desired" state
for admission, which arguably should be happening earlier in
config. Exclude the two terminal pods states from GetActivePods()
Second, the previous check introduced with the pod worker lifecycle
ownership changes was subtly wrong for the admission use case.
Admission has to include pods that haven't yet hit the pod worker,
which CouldHaveRunningContainers was filtering out (because the
pod worker hasn't seen them). Introduce a weaker check -
IsPodKnownTerminated() - that returns true only if the pod is in
a known terminated state (no running containers AND known to pod
worker). This weaker check may only be called from components that
need admitted pods, not other kubelet subsystems.
This commit does not fix the long standing bug that force deleted
pods are omitted from admission checks, which must be fixed by
having GetActivePods() also include pods "still terminating".
If a pod is already in terminated and the housekeeping loop sees an
out of date cache entry for a running container, the pod worker
should ignore that running pod termination request. Once the worker
completes, a subsequent housekeeping invocation will then invoke
terminating because the worker is no longer processing any pod
with that UID.
This does leave the possibility of syncTerminatedPod being blocked
if a container in the pod is started after killPod successfully
completes but before syncTerminatedPod can exit successfully,
perhaps because the terminated flow (detach volumes) is blocked on
that running container. A future change will address that issue.
A number of race conditions exist when pods are terminated early in
their lifecycle because components in the kubelet need to know "no
running containers" or "containers can't be started from now on" but
were relying on outdated state.
Only the pod worker knows whether containers are being started for
a given pod, which is required to know when a pod is "terminated"
(no running containers, none coming). Move that responsibility and
podKiller function into the pod workers, and have everything that
was killing the pod go into the UpdatePod loop. Split syncPod into
three phases - setup, terminate containers, and cleanup pod - and
have transitions between those methods be visible to other
components. After this change, to kill a pod you tell the pod worker
to UpdatePod({UpdateType: SyncPodKill, Pod: pod}).
Several places in the kubelet were incorrect about whether they
were handling terminating (should stop running, might have
containers) or terminated (no running containers) pods. The pod worker
exposes methods that allow other loops to know when to set up or tear
down resources based on the state of the pod - these methods remove
the possibility of race conditions by ensuring a single component is
responsible for knowing each pod's allowed state and other components
simply delegate to checking whether they are in the window by UID.
Removing containers now no longer blocks final pod deletion in the
API server and are handled as background cleanup. Node shutdown
no longer marks pods as failed as they can be restarted in the
next step.
See https://docs.google.com/document/d/1Pic5TPntdJnYfIpBeZndDelM-AbS4FN9H2GTLFhoJ04/edit# for details
This change removes RuntimeCache in the pod workers and the syncPod() function.
Note that it doesn't deprecate RuntimeCache completely as other components
still rely on the cache.
Currently kubelet syncs all pods every 10s. This is not preferred because
* Some pods may have been sync'd recently.
* This may cause all the pods to be sync'd at once, causing undesirable
CPU spikes.
This PR replaces the global syncs with independent, periodic pod syncs. At the
end of syncing, each pod worker will enqueue itslef with a future timestamp (
current time + sync interval), when it will be due for another sync.
* If the pod worker encoutners an sync error, it may requeue with a different
timestamp to retry sooner.
* If a sync is triggered by the update channel (events or spec changes), the
pod worker would enqueue a new sync time.
This change is necessary for moving to long or no periodic sync period once pod
lifecycle event generator is completed. We will still rely on the mechanism to
requeue the pod on sync error.
This change also makes sure that if a sync does not succeed (either due to
real error or the per-container backoff mechanism), an error would be propagated
back to the pod worker, which is responsible for requeuing.
