When supporting rolling updates, we cannot use the same fixed socket paths for
old and new pod. With the revised API, the caller no longer specifies the full
socket paths, only directories. The logic about how to name sockets then can be
in the helper.
While at it, avoid passing a context to the gRPC helper code when
all that the helper code needs is a logger. That leads to confusion
about whether cancellation has an effect.
The goal is to simplify writing DRA drivers. This is also a first step towards
supporting seamless upgrades.
DRA drivers no longer need to implement the kubelet plugin API
directly. Instead, the helper wraps an implementation of an interface. The
helper then provides common functionality:
- retrieve and validate ResourceClaims
- serialize gRPC calls (enabled by default, can be opted out)
- gRPC logging
The definition of that interface is meant to be comprehensive enough that a
correct DRA driver can be implemented by following the documentation of the
package, without having to cross-reference KEPs.
The DRAPlugin interface used to be the abstract API of the helper. Now it's
what the DRA driver kubelet plugin needs to implement. The helper is a concrete
Server struct with no exported fields. It only exports the methods that
drivers need when using the helper.
While at it, support for the v1alpha4 API gets removed from the helper, which
implies removing the corresponding E2E tests. The kubelet implementation will
be dropped separately.
Noticed that cache might not nesseserly observe the write causing test
to flake. Fixed that changing the logic to require LessOrEqual of
writeRV instead of equal to writeRV. Also added comments explaining
edge cases.
The image puller's PullImage() method should be just a dumb pull
without any further logic. Make it accept everything it needs
to pull an image and defer any other magic to the image manager.
add missing metric about uncore / L3 / Last-Level cache alignment,
plus its e2e tests.
Exposing uncore alignment requires a bit of refactoring in the static
policy implementation because, differently from full PCPUs alignment
and NUMA alignment, can't be easily and safely inferred by construction.
The main reason for this is that uncore cache alignment is preferred,
not mandatory, thus the cpu allocator can legally use cross-uncore
allocation. Because of that, the final cpuset union step can
create a final cpuset which is not uncore-aligned even though all
its parts are uncore-aligned.
The safest way seems thus to run just a final uncore-alignment check
once the final cpuset is computed.
Signed-off-by: Francesco Romani <fromani@redhat.com>
There's no need to use the generic reflect.DeepEqual,
we can use the cpuset Equals method in tests,
which is more idiomatic and a tad clearer.
Signed-off-by: Francesco Romani <fromani@redhat.com>
abort the test in the unlikely case we fail to create
the Policy object. Exactly because this is unlikely
we should fail loudly.
In my case this happened because unrelated bad parameters,
which was hard to catch
Signed-off-by: Francesco Romani <fromani@redhat.com>
+k8s:optional should be used everywhere +optional is.
This does not change the vailidation of the field, but it
is a good practice, and code generator recognizes the fields
tags and outputs this comment above the field validation
as a result: "optional value-type fields with zero-value
defaults are purely documentation".
