• 垃圾回收器选错了我的 Java 服务内存炸了！

If you’re running a high‑concurrency Java service and suddenly see memory usage climb toward heap limit,Full GCs erupt every few minutes,and response times spike from milliseconds to seconds — you’re probably dealing with a mis‑chosen garbage collector.

This article walks through a real production incident,explains *** default collector caused memory runaway。and shows how we switched to a 娱乐ter fit.

主体

背景这方面，事故现象与初步排查

The service was built on JDK 11 using G1 by default. During a major promotion it exhibited:

• The heap grew steadily toward –Xmx .
• Total pause time rose from roughly one Full GC per hour to several per minute.
• Averaged latency ballooned from ~50 ms to several seconds;some nodes became unresponsive.
Using `jstat -gcutil`,we saw Old Generation utilization consistently above 90%,while Mixed GCs barely reclaimed old objects. It was clear that garbage collection had become a bottleneck. Note的观点是，In this section we’ll keep focus on what *you* will see in your logs rar than just oretical numbers. For reference here’s a quick snapshot of our stack trace during one of those spikes: GcInfo{count=12345,timeMillis=,lastGcTimeMillis=,lastGcTimeNs=,gcType=FULL_GC,totalMemoryBytes=,usedMemoryBytes=}

---

问题根源的观点是，G1 的适用场景与局限性

G1 aims for “balanced throughput and latency”,targeting large heaps and low latency requirements . Its core assumptions are:
1. The allocation rate is moderate so that Young collections free short‑lived objects quickly.The proportion of long‑lived objects in Old Gen stays modest so that Mixed GCs can reclaim m efficiently. But our workload broke both assumptions:
   1. A lot of short‑term objects survive slightly longer than expected,causing early promotion into Old Gen.A large historical cache occupies over half of heap,inflating Old Gen usage. Because G1’s Mixed phase relies on “mark–compact”,heavy fragmentation dramatically reduces reclamation efficiency—eventually forcing anor Full GC.

      ---

      对比实验的观点是。Parallel GC vs CMS vs ZGC

      We ran three identical workloads under each collector:

      从最终方案来看，Shenandoah GC 的救场

      JDK 12 introduced Shenandoah—a low‑pause collector that compresses concurrently like ZGC but with higher throughput . Switching simply required adding `-XX:+UseShenandoahGC`>`. After tuning we observed:
      • Heap usage dropped from ~16 GB down to ~12 GB—old gen shrank thanks to concurrent compaction.
      • Maximum pause time fell from ~2000 ms down to under 10 ms.
      • Full GCs vanished entirely,leaving only brief mixed pauses during occasional compaction cycles.

      These changes stabilized response times at ir target range .

      ---

      从通用选型教程来看，怎么选 Garbage Collector？

      ‏	Parallel GC	CMS	ZGC
      Status after stress test ’
      Sustained Throughput ’	% high     —
      Pausable Stalls per Minute ’	*Measured via jcmd VM.native_memory summary after each run.
      Key takeaways:
      Throughput> High – Parallel – best overall when latency is non-critical.

      堆大小	延迟需求	吞吐需求	推荐 Collector
      <4 GB	⬇️ Low	➕ High	Serial / Parallel
      4–8 GB	⬇️ ≤200 ms	➕ High	CMS / G1
      ≥8 GB	⬇️ ≤10 ms	➕ Medium	Shenandoah / ZGC

      场景建议

      • 大量短期对象 → 避免 G1； 优先 Parallel 或 ZGC。
      • 长期缓存热点 → 避免 CMS；优先 Shenandoah 或 G1。

      ---

      & 接下来行动计划：

      垃圾回收器不是“一成不变”的默认值，而是需动态挑选。

      🔍 建议： * 在生产环境启用 gc.log 并结合 JMC 或 AsyncProfiler 持续监控。* 定期复盘日志，检查是否出现频繁 Full 或 Mixed 扰乱。*，

      如果你正面临类似“内存爆炸、停机频繁”的痛点，这套经验可以方便你定位并修复。按理说，祝你部署顺利 🚀