G1, ZGC, Parallel, Serial — how do you choose and tune a garbage collector for a service?

🎤 Say this first

Choose by SLO, not fashion: Parallel maximizes throughput and is fine for batch; G1 (default) balances pause vs throughput for typical services; ZGC (generational since JDK 21) delivers sub-millisecond pauses for latency-critical or very large heaps, at some throughput/memory cost. Tune by setting a pause goal and heap size first, measuring with GC logs/JFR — not by copying flag lists off blogs.

The full picture

Collector	Design goal	Pause profile	Pick when
Serial	Smallest footprint	Long, single-threaded	Tiny containers, CLI tools
Parallel	Max throughput	Long but efficient STW	Batch/ETL where p99 latency is irrelevant
G1 (default)	Balanced, region-based	Target via `MaxGCPauseMillis` (~200ms default)	Most services, heaps 4–32GB
ZGC (generational)	Ultra-low latency, concurrent	<1ms typical, heap-size independent	Trading, gateways, multi-100GB heaps
Shenandoah	Low latency (Red Hat lineage)	Few ms, concurrent compaction	Similar niche to ZGC

A sane tuning workflow: (1) fix the heap size — in containers via MaxRAMPercentage; (2) pick the collector by SLO; (3) enable -Xlog:gc* and observe allocation rate, promotion rate, pause distribution; (4) only then touch advanced flags. 90% of 'GC problems' are actually allocation-rate problems or heap-sized-too-small problems — fixing the code (object churn, caching giant graphs) beats flag alchemy.

🎤 Say this first

The full picture

🔄 Likely follow-up questions