Glossary term
Latency
The elapsed time between a cause and its observable effect in a computing, communication, measurement, or control system.
Definition
metricLatency is the time delay between an input, request, event, or command and the corresponding response or observable effect.
Latency appears in networks, processors, storage, data buses, sensors, actuators, control loops, user interfaces, cloud systems, and real-time software. It may include propagation delay, queueing delay, processing time, scheduling delay, conversion time, buffering, serialization, filtering, and actuation delay. Low average latency is not enough for many systems; worst-case latency, jitter, determinism, and deadline misses often matter more.
Latency is elapsed time between an event and the response that depends on it. In a web service, it may be the time between a request and a response. In a control loop, it may be the time between measuring a state and applying the corresponding actuator command. In a sensor chain, it may include sampling, filtering, conversion, communication, computation, and output update.
Sources
Latency is often a sum of several delays. Communication systems include propagation delay, serialization delay, routing delay, buffering, retransmission, and queueing. Computers include scheduling delay, cache misses, memory access, operating-system overhead, interrupt handling, and processing time. Measurement systems include sensor response, anti-alias filtering, analog-to-digital conversion, digital filtering, packetization, and timestamping. Actuator systems include command processing, drive delay, mechanical response, and plant dynamics.
Latency versus bandwidth and jitter
Bandwidth describes rate or frequency range. Latency describes delay. A system can have high bandwidth and still high latency if buffering or scheduling delays are large. Jitter is variation in latency or event timing. For real-time systems, a predictable 5 ms delay may be easier to compensate than an average 2 ms delay with occasional 50 ms spikes.
Round-trip time is common in networks, while one-way latency is often more relevant for control, audio, video, trading systems, and distributed measurement. The two are not interchangeable unless paths are symmetric and clocks are synchronized.
Engineering consequences
Latency reduces phase margin in feedback control, degrades teleoperation, makes user interfaces feel sluggish, increases packet buffering, worsens synchronization, and limits closed-loop bandwidth. In safety-critical systems, deadline misses can be more important than average delay. For user-facing systems, tail latency can dominate perceived quality because rare slow responses are noticed strongly.
Common mistakes
A common mistake is reporting only average latency. Percentiles, worst-case bounds, deadline misses, and jitter are often more important. Another mistake is measuring latency at a convenient internal point rather than end-to-end. Good latency reporting states start event, stop event, timing source, load condition, distribution, percentile, sample size, and whether the measurement includes network, buffering, rendering, actuation, or user-interface delay.