The original limit of 32 seemed sufficient for a single GPU on a node. But for
shared non-local resources it is too low. For example, a ResourceClaim might be
used to allocate an interconnect channel that connects all pods of a workload
running on several different nodes, in which case the number of pods can be
considerably larger.
256 is high enough for currently planned systems. If we need something even
higher in the future, an alternative approach might be needed to avoid
scalability problems.
Normally, increasing such a limit would have to be done incrementally over two
releases. In this case we decided on
Slack (https://kubernetes.slack.com/archives/CJUQN3E4T/p1734593174791519) to
make an exception and apply this change to current master for 1.33 and backport
it to the next 1.32.x patch release for production usage.
This breaks downgrades to a 1.32 release without this change if there are
ResourceClaims with a number of consumers > 32 in ReservedFor. In practice,
this breakage is very unlikely because there are no workloads yet which need so
many consumers and such downgrades to a previous patch release are also
unlikely. Downgrades to 1.31 already weren't supported when using DRA v1beta1.
This had been left out unintentionally earlier. Because theoretically there
might now be existing objects with parameters that are larger than whatever
limit gets enforced now, the limit only gets checked when parameters get
created or modified.
This is similar to the validation of CEL expressions and for consistency, the
same 10 Ki limit as for those is chosen.
Because the limit is not enforced for stored parameters, it can be increased in
the future, with the caveat that users who need larger parameters then depend
on the newer Kubernetes release with a higher limit. Lowering the limit is
harder because creating deployments that worked in older Kubernetes will not
work anymore with newer Kubernetes.
Using the "normal" logic for a feature gated field simplifies the
implementation of the feature gate.
There is one (entirely theoretic!) problem with updating from 1.31: if a claim
was allocated in 1.31 with admin access, the status field was not set because
it didn't exist yet. If a driver now follows the current definition of "unset =
off", then it will not grant admin access even though it should. This is
theoretic because drivers are starting to support admin access with 1.32, so
there shouldn't be any claim where this problem could occur.
The new DRAAdminAccess feature gate has the following effects:
- If disabled in the apiserver, the spec.devices.requests[*].adminAccess
field gets cleared. Same in the status. In both cases the scenario
that it was already set and a claim or claim template get updated
is special: in those cases, the field is not cleared.
Also, allocating a claim with admin access is allowed regardless of the
feature gate and the field is not cleared. In practice, the scheduler
will not do that.
- If disabled in the resource claim controller, creating ResourceClaims
with the field set gets rejected. This prevents running workloads
which depend on admin access.
- If disabled in the scheduler, claims with admin access don't get
allocated. The effect is the same.
The alternative would have been to ignore the fields in claim controller and
scheduler. This is bad because a monitoring workload then runs, blocking
resources that probably were meant for production workloads.
Drivers need to know that because admin access may also grant additional
permissions. The allocator needs to ignore such results when determining which
devices are considered as allocated.
In both cases it is conceptually cleaner to not rely on the content of the
ClaimSpec.
The main purpose is to protect against denial-of-service attacks. Scheduling
time depends a lot on unpredictable factors and expected scheduling time also
varies, so no attempt is made to limit the overall time spent on evaluating CEL
expressions per claim.
