How do you reason about garbage-collection latency, throughput, and tuning?

I account for tail latency across the whole request path rather than attributing every spike to stop-the-world time. Collector work competes for CPU, and a low-pause goal can reduce application throughput if the environment lacks headroom. Any tuning change gets a rollback path and comparison against service-level metrics, not only GC logs. Long-term improvements usually combine right-sized capacity, bounded retention, controlled allocation, current JDK fixes, and a collector suited to the actual workload.