thinking-margin-of-safety

Margin of Safety

Core rule: commit a buffered number when residual uncertainty plus breach cost can cause ruin or costly failure. Margin covers estimation error; it is not free slack for laziness.

When to Use

• Capacity, timeouts, pool sizes, queue depths, storage, or SLA numbers under load uncertainty
• Timeline or budget commitments where underestimation is costly
• Architecture headroom when scaling or recovery is slow
• Any public or production commitment where being wrong has asymmetric downside

When NOT to Use

• Breach is detected immediately and fixed near-zero cost (fast auto-scale, live-tunable limit)—static fat buffers waste resources
• cost(buffer) > P(breach) × cost(breach)—right-size or measure instead of maxing out
• Uncertainty is eliminable by measurement or lookup—get the real number first
• Stopping-criterion / search budget problems—use bounded rationality, not padding
• Fully known parameters and low stakes where edge optimization is the goal

Procedure

1. Point estimate without padding. State the base requirement in explicit units (RPS, weeks, GB, ms) and confidence (high / medium / low).
2. Uncertainty and breach cost. List drivers (spike, growth, unknowns, dependency variance). State what fails if undershot: outage, missed launch, data loss, reputation—and whether failure is recoverable or ruinous.
3. Size the buffer. Multiply or add margin to residual uncertainty and stakes—not to vanity. Typical bands (adjust with evidence):
   • Well-known, low consequence: ~1.2–1.5×
   • Familiar with unknowns / reversible: ~1.5–2×
   • New domain, external deps, SLA, irreversible: ~2–3× Cap or cut when buffer cost exceeds expected breach cost.
4. Ruin constraint. If a breach can cause irreversible harm, size so the worst plausible miss still stays above the failure threshold; if that buffer is unaffordable, change the design (shed load, degrade, stage) rather than pretend precision.
5. Strongest countercase. Challenge both under-buffering ("we will scale later") and over-buffering (idle cost, complexity). Prefer measure-then-trim when history exists.
6. Commit and monitor. Publish the buffered commitment, the failure threshold, and the metric that would prove margin excessive or thin. Stop when the number is set and monitorable.