No. Queueing, skew, pauses, and bursts affect the tail.

Estimate peak reads/writes per second, mean service time, in-flight work, storage growth, bandwidth, and hot-key concentration. Little’s Law L=λW applies to stable averages with consistent system boundaries. For workers handling one request at a time, a first estimate is peak_rate×mean_service_time/target_utilization, rounded up; async multiplexing and resource-specific bottlenecks require a different model. Separate CPU, database connections, I/O, memory, and network limits. Add explicit failure headroom and verify with representative load tests rather than copying internet capacity numbers.

What the function promises: The model estimates single-request worker count and raw replicated payload storage using stated assumptions. It excludes indexes, compression, metadata, and operational overhead.

from math import ceil

def estimate_capacity(peak_rps, mean_seconds, utilization,
                      writes_per_day, bytes_per_write, days, replicas):
    if (peak_rps < 0 or mean_seconds < 0 or not 0 < utilization < 1
            or min(writes_per_day, bytes_per_write, days) < 0 or replicas < 1):
        raise ValueError("Invalid assumptions")
    return {
        "workers": ceil(peak_rps * mean_seconds / utilization),
        "raw_bytes": writes_per_day * bytes_per_write * days * replicas,
    }

result = estimate_capacity(100, 0.2, 0.5, 1000, 200, 30, 3)
assert result == {"workers": 40, "raw_bytes": 18000000}

Work and storage: Constant-sized arithmetic, followed by real measurement work outside the model.

When to recognize it: Estimate bytes per write, retention, replication, indexes, and overhead. Also estimate read bandwidth and the hottest key, not only total requests.

Check before applying it: Capacity arithmetic is a hypothesis with units and assumptions, not a substitute for load tests.

No. Measure the hottest shard, key, tenant, and operation. A lightly loaded cluster can still contain one saturated partition.