How do you choose a garbage collector, and what do you know about ZGC?

G1 is the default and suits most services: region based, generational, with a configurable pause target in the tens of milliseconds. ZGC is the choice when pause time dominates, since it does almost all work concurrently and keeps pauses under a millisecond even on very large heaps; it has been generational by default since Java 23. Parallel GC still wins on raw throughput for batch jobs where pauses do not matter.

I start by leaving it alone, because G1 with a sensible heap size handles most services, and it already balances throughput against a pause goal you can set. I look at the garbage collection logs before touching anything: allocation rate, pause distribution, how much is promoted, and whether the heap is simply too small. Most of the time what looks like a collector problem is an application allocating far too much, or a heap sized so tightly that it is collecting constantly. If pauses genuinely drive my tail latency, for example a trading or bidding service with a large heap, I move to ZGC, which does marking and relocation concurrently with load barriers and keeps pauses well under a millisecond. It has been generational since Java 23, which fixed the throughput penalty that made people avoid it. The tradeoff is extra CPU and memory overhead for those barriers. For a nightly batch job I go the other way and use Parallel, since a two second pause is irrelevant and raw throughput is what I want.