Amey-Thakur Capacity planning
Capacity planning answers "will we fall over, and when" before the answer arrives as an outage. It is load forecasting plus limit discovery plus a headroom policy: guessing at any of the three is how Black Friday becomes a postmortem.
Method
- Model load from a driver, not a vibe. Tie resource demand to a business metric (requests per active user, compute per order: see product-metrics) and forecast the driver (organic growth, marketing events, seasonality: see product-launch's spike). Per-hop capacity flows from the load model through the request path; a plan built on "traffic will grow" without a number is not a plan.
- Discover limits empirically. Load-test to find where each component saturates and how it fails (graceful degradation vs cliff: see load-testing, backpressure): the bottleneck is rarely where intuition points, and the failure mode matters as much as the number. Test the whole path; the database connection pool or a downstream rate limit usually caps you before CPU does.
- Set headroom from failure cost and lead time. Enough buffer to absorb spikes within your scale-out actuation time (see autoscaling-policies' lag math) plus a margin for the forecast being wrong; more headroom for hard-to-scale stateful tiers (see sharding-partitioning) than for elastic stateless ones. Headroom is insurance priced against downtime cost.
- Distinguish elastic from fixed capacity. Autoscaling handles variance within provisioned limits (see autoscaling-policies), but the limits themselves (account quotas, database size, license seats, IP ranges, third-party rate caps) are capacity planning's job: the outage is often hitting a ceiling nobody tracked, not running out of compute.
- Pre-provision for known events, verify by test. Marketing spikes, launches, seasonal peaks: raise quotas ahead (cloud quota increases take days), pre-scale (see autoscaling-policies' scheduled scaling), load-test at projected peak, and rehearse the failure (game-day the spike: see chaos-gameday). Reactive scaling pays the actuation lag exactly when you cannot afford it.
- Track cost against capacity continuously. Cost per unit of the business driver (see cloud-cost-optimization's unit economics): capacity and spend are the same curve, and a plan that ignores cost provisions a fleet nobody will pay for. Review utilization trends monthly; pinned-at-max autoscaling is capacity planning paging you (see autoscaling-policies).
Boundaries
- Forecasts are wrong; the plan's value is the headroom and the discovered limits, not the precision of the prediction (see estimation-techniques' error-bar honesty). Re-forecast on a cadence.
- Capacity planning sizes for expected and reasonable-peak load; genuine overload beyond the plan is handled by shedding and degradation (see backpressure), not by provisioning for the tail forever.
- Multi-region and DR capacity is a separate axis (can one region absorb another's load on failover: see multi-region-design, cloud-disaster-recovery); steady-state planning that ignores it under-provisions the failure case.