Premium

Peak Load

Basic System Design
≈ 3 min read

Average load is rarely the most dangerous load

Users do not distribute their activity perfectly evenly across the day. A service may have evening rush hours, spikes after notifications, sales events, the start of a live stream, or another moment when request volume becomes much higher than average.

If a system is designed only for average RPS, it may work perfectly for most of the day and fail regularly at exactly the moment when it is needed most.

Peak factor

For a quick estimate, it is convenient to use a peak factor: how many times the busiest period is above the average. If average load is 1,000 RPS and a reasonable peak assumption is four times higher, then the request path needs to be designed for roughly 4,000 RPS. 

average_RPS = 1,000
peak_factor = 4
peak_RPS = 1,000 × 4 = 4,000

The factor itself is not a universal constant. One product may have a mild peak, while another has short spikes tens of times above average. If historical metrics are available, use them. If they are not, state the assumption explicitly.

Peak and headroom are not the same thing

Peak load describes the expected maximum in the model. Capacity headroom protects against forecasting error, product growth, degraded nodes, or periods when the system is temporarily operating below ideal efficiency. 

estimated_peak = 4,000 RPS
headroom = 25%
target_capacity ≈ 5,000 RPS

Do not turn headroom into a magic number. Too little makes the system fragile, while too much can mean continuously paying for resources that are almost never used. This is another architecture trade-off.

A short spike and a sustained peak are different problems

Two systems with the same maximum RPS can experience very different workloads. If 10,000 RPS lasts for five seconds, that is one scenario. If the same level lasts for two hours, the requirement for sustained performance is much higher.

That is why, in addition to the peak value, it is useful to understand its duration and frequency. This tells you whether the spike is a rare anomaly or a normal part of the working day.

The average answers “how much usually,” while the peak answers “how much must survive.” For sizing a system, the second number is often more important than the first.
Back Next