แก้ไขปัญหาหน่วยความจำรั่วและแรงกดดันจาก GC ในระบบจริง

Memory issues rarely announce themselves cleanly. Watch for these patterns:

• Slowly rising RSS that never drops
• Increasing latency from longer GC pauses
• Periodic OOM kills and restarts

This lesson covers diagnosing them in production.

Distinguish three failure modes:

• Leak: memory grows unbounded and is never freed
• Bloat: high but stable usage from large caches
• Churn: rapid allocate/free cycles stressing the GC

Each needs a different fix.

Plot memory over time. A leak shows a steadily climbing baseline even after GC. Bloat shows a high but flat line. Healthy services have a sawtooth that resets after each collection.

A heap snapshot captures every live object at a moment in time. Take two snapshots minutes apart and compare: objects that grew between them are your leak suspects.

An object stays alive because something retains it. The retainer chain shows who is holding the reference. The dominator tree shows which single object, if freed, would release the most memory.

Follow retainers to find the unintended reference keeping memory alive.

Most leaks come from a handful of patterns:

• Caches without eviction limits
• Event listeners never unregistered
• Growing global collections
• Closures capturing large objects

For GC churn, you care about allocation rate, not live size. An allocation profiler shows which call sites create the most short-lived objects, which is what keeps the collector busy.

Long GC pauses spike latency. Causes include too-small heaps forcing frequent collection, or huge heaps making each collection slow.

Correlate pause times in GC logs with your latency tracing to confirm GC is the culprit before tuning.

GC tuning (heap size, collector choice) treats symptoms. Reducing allocations or fixing a leak treats the cause. Always prefer the fix; tune only to buy time or smooth a fundamentally healthy workload.

Heap dumps can pause the process and contain sensitive data. Capture on a canary instance pulled from rotation, store dumps securely, and prefer sampling profilers with low overhead for always-on insight.

Putting it together:

• Confirm leak vs bloat vs churn from the memory curve
• Take and diff heap snapshots
• Follow retainer chains to the holding reference
• For churn, use allocation profiling
• Fix the cause; tune GC only as needed

Test your understanding of memory debugging.

You learned to debug memory problems in live services.

• Tell leaks from bloat and churn via the memory curve
• Diff heap snapshots and follow retainers
• Profile allocations for GC churn
• Fix causes; tune GC only to smooth healthy load