dgraph zero follower memory usage continuously increases when running on k8s

[hankhsutc]

Describe the bug

We've observed that the memory usage of the zero follower pods continuously increased over time.
Eventually the pods got OOMKilled by k8s.

Expected behavior

memory usage should remain stable over time.

Screenshots

Environment

• Version : v25.3.4
• k8s
• cpu request/limit: 2/6
• memory request/limit: 4Gi/4Gi

Additional context

Add any other context about the problem here.

[shiva-istari]

Thanks for the report, @hankhsutc and for the memory graph, it's helpful.

Could you help us gather a bit more data? The heap profile is by far the most useful.

1. Heap profile from a high-memory follower pod. Zero exposes pprof on its HTTP port (default 6080):
   curl 'http://:6080/debug/pprof/heap' -o zero-heap.out
2. Please attach zero-heap.out (or paste the output of go tool pprof -top -inuse_space zero-heap.out).
3. Goroutine dump from the same pod:
   curl 'http://:6080/debug/pprof/goroutine?debug=2' -o zero-goroutines.txt
4. Leader vs. follower: Does memory grow only on followers, or on the leader too? You can identify the leader via curl http://:6080/state.
5. Does restarting a follower reset its memory to baseline? (We expect yes.)
6. Workload + timing: Roughly how many transactions/commits per second does the cluster handle, and how long does a follower run before it gets OOMKilled? We're trying to confirm memory scales with cumulative commit count rather than concurrent transactions.
7. Topology & config: Number of Zeros, number of Alpha groups, and the full dgraph zero command/flags you run with.
8. Logs: Any lines matching Purged below and how frequently Zero snapshots occur would help us rule out a secondary path.