fix: pruner OOM — skip input fetching, fix slice race, auto-configure GOMEMLIMIT

[freemans13]

Summary

• Skip inputs bin + parsing when skipParentUpdates=true: The existing skipParentUpdates setting already skipped the final flush of parent updates, but the pruner was still fetching the inputs bin from Aerospike, parsing every input into bt.Input objects, and accumulating them into allParentUpdates — only to discard everything. For billion-record prune cycles at 1.3M records/sec, this means billions of wasted allocations and significant network I/O for the largest per-record bin. Now the setting also skips the upstream fetch/parse/accumulate, eliminating the allocation pressure at its source.

• Fix slice reuse data race in partitionWorker: chunk = chunk[:0] resets the slice length but keeps the same backing array, which the main loop then overwrites via append while goroutines are still reading from it. Replaced with chunk = make([]*aerospike.Result, 0, s.chunkSize) to allocate a fresh backing array per chunk.

• Auto-configure GOMEMLIMIT from cgroup limits: Reads cgroup v2/v1 memory limits at daemon startup and sets debug.SetMemoryLimit to 90% of the container limit. Combined with the existing GOGC=200, this lets Go allocate fast but tighten GC as it approaches the soft limit — preventing OOM kills while maintaining throughput. No-op on local dev (no cgroup files). Benefits all services, not just the pruner.

Test plan

• make build-teranode compiles cleanly
• make lint passes with 0 issues
• go test -v -race ./services/pruner/... passes
• go test -v -race ./daemon/... passes
• Deploy to staging and monitor pruner memory during billion-record prune cycle — should stay within 2Gi
• Verify -race flag in CI catches no data races in pruner path

🤖 Generated with Claude Code

[github-actions]

🤖 Claude Code Review

Status: Complete

---

Findings:

This PR introduces three important improvements: skipping unnecessary input fetching when skipParentUpdates=true, fixing a slice data race, and auto-configuring GOMEMLIMIT from cgroup limits. The race fix and memory limit improvements look solid.

However, there is a logic bug introduced by the skipParentUpdates changes:

In processRecordChunk() around line 950, the code checks if len(inputs) > 0 to determine the file type for external files. When skipParentUpdates=true, the inputs variable is declared inside the conditional block (line 918-919) and will not be initialized, causing this check to always evaluate to false.

Suggested fix: Move var inputs []*bt.Input to line 917 (before the if !s.skipParentUpdates block) so it is always initialized to an empty slice, ensuring the len(inputs) > 0 check works correctly in both modes.

The memory limit and race fix changes are excellent and address real production issues.

[sonarqubecloud]

Quality Gate failed

Failed conditions
71.9% Coverage on New Code (required ≥ 80%)

See analysis details on SonarQube Cloud