[velo]

What

Adds an integration test, JemallocArm64OomIT, that reproduces arm64 TaskManagers being OOMKilled. It also adds a jemalloc-arm64-test CI job on ubuntu-24.04-arm that runs only this test.

This test is expected to fail on arm64 until the allocator issue is fixed. The job is not in the needs of maven-deploy, build-and-push-docker-image or build, so it doesn't block deploys.

Root cause

The flink:2.3.0-java17 base image ships Ubuntu's libjemalloc2 5.3.0. On arm64 that build uses 64 KiB pages (Page size: 65536), but Graviton and Bottlerocket kernels use 4 KiB pages. The page size is fixed when jemalloc is compiled, so MALLOC_CONF can't change it.

After the startup allocation burst, amd64 gives about 500 MiB back to the OS and arm64 doesn't.

How the test works

• Cluster: one JobManager and one TaskManager, both from the image Maven builds.
  • The TaskManager is pinned to cpuset 0-3, so jemalloc creates 16 arenas, the same as a 4-vCPU runner.
  • The job is 70 RocksDB-backed GROUP BY and dedup operators, with 4 slots and process.size: 4000m.
• Warm-up: the TaskManager runs for 3 minutes under a loose 6000 MiB container limit.
• Shrink: the test then lowers the limit in place to 3800 MiB with docker update.
• Pass condition: the TaskManager must stay alive for another 3 minutes.
• Skipped unless the Docker daemon is aarch64.

Why warm up and then shrink, instead of using a fixed limit? Measured anon memory:

	Startup peak	Steady state
amd64	~3975 MiB	3350–3550 MiB
arm64	—	~3900 MiB

amd64's startup peak is above arm64's steady state, so no fixed limit separates them. At a fixed 3000 MiB both architectures were OOMKilled.

Results

Environment	Result
arm64 (r6gd.xlarge, 4 vCPU), 3 runs	Failed every time: OOMKilled=true exitCode=137 about 3m07s after the job started, ~228 s per test
amd64, same test with the arch gate lifted	Passed, 401 s

A long-running FlinkDeployment on staging showed the same pattern:

• arm64 TaskManagers were OOMKilled repeatedly at ~3730–3850 MiB RSS.
• amd64 stayed steady at ~3300 MiB for more than an hour.

The ubuntu-24.04-arm CI job hasn't run yet. This PR is its first run.