What is the difference between load, stress, and soak testing, and what would you measure?

Load testing runs expected traffic to confirm the system meets its targets. Stress testing pushes past capacity to find the breaking point and check that it degrades gracefully rather than collapsing. Soak testing runs sustained moderate load for hours to expose leaks and resource exhaustion. Measure latency percentiles, throughput, error rate, and saturation of CPU, memory, connections, and the database.

They differ by intent. Load testing answers whether the system meets its target at expected traffic, so I need that target agreed before I start, something like p95 under 300 milliseconds at two thousand concurrent users with an error rate below a tenth of a percent. Stress testing deliberately goes past that to find where it breaks and, more importantly, how. Graceful degradation with queuing and clear errors is acceptable; data corruption or a cascade that takes down neighboring services is not. Soak testing runs a moderate load for eight or twelve hours, which is the only way to catch slow leaks, connection pools that never return, and log volumes that fill a disk. I once found a memory leak that only appeared after about six hours and would never have surfaced in a thirty minute run. On measurement I report percentiles, because an average hides everything that matters, so p50, p95, and p99 alongside throughput and error rate. And I watch saturation at the same time, so CPU, memory, connection pool usage, and database locks, because response time tells you something is wrong and saturation tells you what.