[bart0sh]

topology-aware forgets its DRA claims when the plugin restarted. Setup() clears the policy's cache entries, and Sync() places every running container again, so containers holding a claim lost its CPUs, and new claims can get them. The kubelet does not prepare a claim again after our restart, so nothing rebuilt a claim once it was gone.

Now Setup() keeps the cache entries and restores the saved claim grants. Container grants are not restored: Sync() allocates every running container again, and finds the claims by then. A container which holds a claim the policy does not know gets a warning in the log.

If the new configuration cannot give a claim its CPUs, because they are reserved or no longer available or the pool is gone, the claim gets other CPUs of the same NUMA nodes, and Sync() moves its containers to them. A claim result is only a number of CPUs of a node, so the kubelet and the scheduler see no change. The claims which keep their CPUs are restored first, so a moved claim cannot take the CPUs of another claim. The start fails only when a node has too few CPUs for a claim. The error names the claim, and the agent writes it into the status of the config CR. Reverting the config lets the plugin start again with the claim restored. A runtime reconfiguration which cannot hold a claim is still refused.

Two fixes on the side: Reserve() took a reserved CPU which is also isolated for the isolated part of a claim, which also affected runtime reconfiguration on main. And the topology-aware policy tests unpacked the sysfs tarball once per test, about 20s each with -race, which made the package hit the 10m go test timeout. With one shared copy the package runs in 42s with -race.

Testing:

• go test -race -count=1 ./... passes with the default 10m timeout.
• golangci-lint v2.11.4 reports no issues.
• e2e on Kubernetes 1.37 with DRAConsumableCapacity, with images built from this branch: test42 (topology-aware, also with DRANodeAllocatableResources) and test01-dra (template) pass. test42 restarts the plugin and checks that running containers keep their claimed CPUs and a new claim gets only free CPUs, that a config which reserves a CPU of a claim moves the claim and its containers to the free CPU of the same node, and that a config which leaves too few CPUs fails the start with the claim named in the config status. test01-dra restarts template with fewer allowed CPUs: the claim keeps its CPUs with a warning, and a new claim does not get them.

[Copilot]

🟡 Changes recommended

Restored claims can overlap reserved workloads when a claimed isolated CPU becomes reserved.

1 open issue

• Reject claimed CPUs overlapping reserved CPUs during restart · New

What changed in this PR

Restores topology-aware DRA claims from cache before running containers are reallocated after a plugin restart.

Changes:

• Preserves cached claims and reports restoration failures.
• Warns when containers reference unknown claims.
• Adds restart tests and updates recovery documentation.

File	Description
test/​e2e/​policies.test-suite/​topology-aware/​n4c16/​test42-dra-devices/​code.var.sh	Tests claim persistence and failed-start recovery.
test/​e2e/​policies.test-suite/​template/​n4c16/​test01-dra/​code.var.sh	Tests restart with fewer allowed CPUs.
docs/​resource-policy/​policy/​topology-aware.md	Documents restart behavior and recovery.
cmd/​plugins/​topology-aware/​policy/​topology-aware-policy.go	Preserves cache and restores claims during setup.
cmd/​plugins/​topology-aware/​policy/​dra.go	Warns about unknown container claims.
cmd/​plugins/​topology-aware/​policy/​dra_test.go	Adds restoration and failure-recovery unit tests.
cmd/​plugins/​topology-aware/​policy/​cache.go	Restores claim grants and tracks unresolved pools.

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

No unresolved issues remain, and regression coverage includes claim persistence, configuration conflicts, and recovery.

No open issues 🎉

1 resolved since last review

• Reject claimed CPUs overlapping reserved CPUs during restart

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Persisted CPU assignment migration and failed-start recovery warrant final human validation.

No open issues 🎉

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

No unresolved blocking issues remain, and regression coverage addresses restoration, relocation, and configuration-failure recovery.

No open issues 🎉

_{Review effort: Balanced}