The --contention-profiling flag enables block profiling by calling
SetBlockProfileRate(). It, however does not call SetMutexProfileFraction
which enables mutex profiling.
Mutex profiling gives us information about lock contention whereas
block profiling on the other hand gives us information on gorotuines
being blocked on sync primitives.
Updating the docs to "block profiling" in order to make it more accurate.
Eventhough a block profile may reflect contention points on locks to some
extent, it also talks about other sources where goroutines may be blocked.
Signed-off-by: Madhav Jivrajani <madhav.jiv@gmail.com>
Kubernetes-commit: f236ab24dfcb84906a1d453877a79bd4a94f897d
ContextForChannel uses a goroutine to transform a channel close to
a context cancel. However, this exposes a synchronization issue if
we want to unify the underlying implementation between contextless
and with context - a ConditionFunc that closes the channel today
expects the behavior that no subsequent conditions will be invoked
(we have a test in wait_test.go TestUntilReturnsImmediately that
verifies this expectation). We can't unify the implementation
without ensuring this property holds.
To do that this commit changes from the goroutine propagation to
implementing context.Context and using stopCh as the Done(). We
then implement Err() by returning context.Canceled and stub the
other methods. Since our context cannot be explicitly cancelled
by users, we cease to return the cancelFn and callers that need
that behavior must wrap the context as normal.
This should be invisible to clients - they would already observe
the same behavior from the context, and the existing error
behavior of Poll* is preserved (which ignores ctx.Err()).
As a side effect, one less goroutine is created making it more
efficient.
Kubernetes-commit: 95051a63b323081daf8a3fe55a252eb79f0053aa
This change in a no-op refactor of the encryption at rest code that
primarily changes the wiring to consistently use context for
lifecycle management (instead of a mixture of context and stop
channels).
Signed-off-by: Monis Khan <mok@microsoft.com>
Kubernetes-commit: 70b414b0e5cbe9706e5a4fc0d7299ec6b169d642
The fact that we're building the OpenAPI using the proto.Models is an
implementation detail that we shouldn't have to expose. Since we're
going to change the way this is transformed, let's first hide it behind
the common NewTypeConverter so that the next change is transparent.
This will also enable other clean-ups like hiding the gvkParser which
shouldn't be exposed and prevent some refactoring.
Kubernetes-commit: a7ab6b86db83e31ff599e4d21a065f6845fb93dd
Serving profiling information can leak information or expose the
apiserver to possible DoS attacks. Serving on a UDS is more secure
though slightly less convenient. One can't use `go tool pprof` directly
against the socket since it's not supported, but can either run a proxy
to copy from the socket over to http, or use `curl --unix-socket` to
download the profile and then use `go tool pprof`.
Kubernetes-commit: 667599b0ddfad8ba760d3bbfe006aae0d8f7dec6
so that aggregated-apiservers can also take advantage. discovered by e2e tests with feature enabled
Kubernetes-commit: c9b34884004079ed3f184b475f7408984f9226f4
Also make some design changes exposed in testing and review.
Do not remove the ambiguous old metric
`apiserver_flowcontrol_request_concurrency_limit` because reviewers
though it is too early. This creates a problem, that metric can not
keep both of its old meanings. I chose the configured concurrency
limit.
Testing has revealed a design flaw, which concerns the initialization
of the seat demand state tracking. The current design in the KEP is
as follows.
> Adjustment is also done on configuration change … For a newly
> introduced priority level, we set HighSeatDemand, AvgSeatDemand, and
> SmoothSeatDemand to NominalCL-LendableSD/2 and StDevSeatDemand to
> zero.
But this does not work out well at server startup. As part of its
construction, the APF controller does a configuration change with zero
objects read, to initialize its request-handling state. As always,
the two mandatory priority levels are implicitly added whenever they
are not read. So this initial reconfig has one non-exempt priority
level, the mandatory one called catch-all --- and it gets its
SmoothSeatDemand initialized to the whole server concurrency limit.
From there it decays slowly, as per the regular design. So for a
fairly long time, it appears to have a high demand and competes
strongly with the other priority levels. Its Target is higher than
all the others, once they start to show up. It properly gets a low
NominalCL once other levels show up, which actually makes it compete
harder for borrowing: it has an exceptionally high Target and a rather
low NominalCL.
I have considered the following fix. The idea is that the designed
initialization is not appropriate before all the default objects are
read. So the fix is to have a mode bit in the controller. In the
initial state, those seat demand tracking variables are set to zero.
Once the config-producing controller detects that all the default
objects are pre-existing, it flips the mode bit. In the later mode,
the seat demand tracking variables are initialized as originally
designed.
However, that still gives preferential treatment to the default
PriorityLevelConfiguration objects, over any that may be added later.
So I have made a universal and simpler fix: always initialize those
seat demand tracking variables to zero. Even if a lot of load shows
up quickly, remember that adjustments are frequent (every 10 sec) and
the very next one will fully respond to that load.
Also: revise logging logic, to log at numerically lower V level when
there is a change.
Also: bug fix in float64close.
Also, separate imports in some file
Co-authored-by: Han Kang <hankang@google.com>
Kubernetes-commit: feb42277884bc7cfbd6f0bb1d875cc63b1b6caac
This change enables hot reload of encryption config file when api server
flag --encryption-provider-config-automatic-reload is set to true. This
allows the user to change the encryption config file without restarting
kube-apiserver. The change is detected by polling the file and is done
by using fsnotify watcher. When file is updated it's process to generate
new set of transformers and close the old ones.
Signed-off-by: Nilekh Chaudhari <1626598+nilekhc@users.noreply.github.com>
Kubernetes-commit: 761b7822fca569d475f782b135ef433e5b014147
Co-authored-by: Alexander Zielenski <zielenski@google.com>
Co-authored-by: Joe Betz <jpbetz@google.com>
Kubernetes-commit: e7d83a1fb7b3e4f6a75ed73bc6e410946e12ad9f
This change adds a flag --encryption-provider-config-automatic-reload
which will be used to drive automatic reloading of the encryption
config at runtime. While this flag is set to true, or when KMS v2
plugins are used without KMS v1 plugins, the /healthz endpoints
associated with said plugins are collapsed into a single endpoint at
/healthz/kms-providers - in this state, it is not possible to
configure exclusions for specific KMS providers while including the
remaining ones - ex: using /readyz?exclude=kms-provider-1 to exclude
a particular KMS is not possible. This single healthz check handles
checking all configured KMS providers. When reloading is enabled
but no KMS providers are configured, it is a no-op.
k8s.io/apiserver does not support dynamic addition and removal of
healthz checks at runtime. Reloading will instead have a single
static healthz check and swap the underlying implementation at
runtime when a config change occurs.
Signed-off-by: Monis Khan <mok@microsoft.com>
Kubernetes-commit: 22e540bc48d9bf698c4f381ccb56ed57dea0dae2
Anish Ramasekar
SataQiu
Nilekh Chaudhari
Abu Kashem
Madhav Jivrajani
Richa Banker
Joseph Anttila Hall
Monis Khan
Patrick Ohly
Rita Zhang
Clayton Coleman
Andrew Sy Kim
lengrongfu
Krzysztof Ostrowski
root
Antoine Pelisse
Arda Güçlü
Jordan Liggitt
Tim Allclair
Jefftree
David Ashpole
Cici Huang
Mike Spreitzer
Alexander Zielenski
Paweł Banaszewski
Nick Turner