Capacity
Capacity is a quality attribute that describes the maximum amount of work a system can handle at once while still meeting its other quality requirements, such as latency and availability. It is most often expressed as a ceiling on concurrent requests, concurrent users, or concurrent sessions, rather than as a sustained rate over time.
Capacity is distinct from its near neighbours. Throughput is the amount of work completed per unit of time, a rate rather than a concurrent ceiling. Scalability is the ability to grow that ceiling by adding resources, whereas capacity is the ceiling as it stands today. A system with high capacity is not necessarily scalable, and a scalable system may be running well below its capacity.
The relationship between capacity and latency is not linear. As load approaches the system’s capacity, queues lengthen and latency rises, often dramatically. For this reason capacity is usually stated alongside a target latency or service level, not as a bare number. A system that "handles 10,000 requests per second" is only meaningfully specified once the latency at that load is also known.
Production systems are rarely run at full capacity. A margin of headroom is held in reserve so that traffic spikes, component failures, and the startup lag of auto-scaling do not immediately degrade service. Running close to capacity is a common cause of incidents. When one node in a load-balanced fleet fails, the surviving nodes must absorb its share, and a fleet with no headroom will then saturate. This is why capacity and redundancy are sized together, eg. the N+1 rule in active-active clusters. The ceiling itself is established through load testing, which progressively increases demand to find where performance degrades. Stress testing goes one step further, pushing past that ceiling to find where the system actually breaks.