What 2020 taught us about capacity planning
Every capacity model in Europe broke last year. Grocery APIs took Christmas load on a Tuesday in March. Government services built for thousands of daily users met millions. Video, logistics, curbside collection: the load curves of 2020 resembled nothing any planning exercise had drawn, and they arrived without notice.
A year on, with the postmortems written, a pattern stands out. The estates that held were not, as a rule, the over-provisioned ones. They were the ones that could see clearly and shed deliberately.
Seeing: the first advantage
Teams that watched consumer-level traffic in real time knew within minutes which callers were surging and which services were straining. Teams that watched only infrastructure metrics learned about problems from their users. The difference was not tooling budget; it was whether observability was attached to identity, who is calling, not just how much.
The distinction sounds academic until you are inside a surge. "CPU is at ninety percent" tells you that you have a problem; it does not tell you what to do about it. "This partner's traffic has grown twentyfold in an hour and is consuming the capacity your checkout needs" tells you exactly what to do. Identity-aware visibility turns a wall of alarming but undifferentiated metrics into a set of actionable decisions, and during 2020 that translation was often the whole difference between a controlled response and a scramble.
Shedding: the second advantage
No estate survives a tenfold surge by scaling alone. The estates that stayed up made choices: batch jobs deferred, non-essential consumers throttled, degraded modes that kept checkout alive while recommendations went quiet. Where quotas and priorities already existed at the gateway, those choices were policy changes made in minutes. Where they did not, they were emergency code changes made in fear.
Resilience turned out to be less about how much you could handle and more about how gracefully you could choose what not to.
The crucial point is that graceful shedding is not something you can improvise mid-crisis. Deciding, under load, which consumers matter most and which functions can be sacrificed is a business judgment, and making business judgments in a panic at three in the morning produces bad ones. The estates that shed well had made those decisions in advance, encoded as quotas and priorities, so that the crisis required only executing a plan rather than inventing one.
Keeping the lesson
The temptation now is to over-provision and move on. Better to keep what actually worked: per-consumer visibility, limits that encode priorities before the crisis, and degradation paths that are designed rather than improvised. Over-provisioning is a bet that the next surge will be no larger than the headroom you bought, and 2020 was a thorough lesson in how badly that bet can lose. Nobody knows what shape the next surge takes. The estates that met the last one calmly had simply decided, in advance, what mattered most.