Trinkets

Decide

What's worth doing, and in what order.

FMEA & Risk PriorityRank what could go wrong by Severity × Occurrence × Detection — the classic RPN, where a moderate failure nobody would catch in time outranks an obvious one. Fix the top of the list first.Open →

Forecast

When work will land, and how it flows.

Connection Pool SizingSize a connection or thread pool from Little's Law: throughput × latency is how many are busy on average, but never size to the average — add headroom so a burst doesn't pin the pool at 100% and queue. Read off the pool size and the load it absorbs.Open →
Queue CliffWhy you can't run a server — or a team — at 100%: watch wait time explode as utilization climbs, solve for the busy ceiling a target wait buys, and see how extra workers sharing one queue flatten the cliff. M/M/c queueing math.Open →
FIFO vs. LIFOPast capacity, the fair FIFO queue spends itself on work that's already too late — on-time delivery collapses while “unfair” LIFO keeps shipping fresh work. Race the disciplines on the same simulated arrivals, and see what a short queue that turns work away buys.Open →
Power of Two ChoicesRandom load balancing builds hotspots — one unlucky queue sets your p99. Peek at just two random servers and take the shorter, and the hotspots collapse almost to the full-scan ideal. One extra look buys nearly all the information.Open →
Rush-Hour BacklogThe rush ends; the queue doesn't. A surge past capacity builds a backlog that drains only at your spare capacity — so a team at 95% busy takes ten times the surge to dig out of it. Headroom is recovery speed.Open →
Erlang-B BlockingWhen every line is busy, callers aren't queued — they're turned away. Erlang's 1917 formula sizes phone trunks, support desks and chargers: lines in, busy-signal odds out, or pick a target and read off the lines it takes. And why big pools get to run hotter at the same risk.Open →
Cache EffectivenessEffective latency or cost as a function of hit rate — an Amdahl-shaped curve where the misses you can't eliminate cap the win, however fast the cache.Open →
Retry StormRetries pile on load exactly when a service is already failing — naive retry multiplies offered load by the retry count, and only backoff with jitter flattens the thundering herd back down.Open →
Timeout & Retry BudgetSet a timeout against your latency curve and see what it costs: a tighter timeout abandons more requests, and retries multiply that into extra load. Read off the per-attempt timeout rate, the retry amplification, and the failure rate that survives.Open →
Token BucketRefill rate sets the long-run cap; bucket depth is the burst headroom. Pick both plus an offered rate, and read off how long the burst lasts before throttling — and the idle gap to burst again.Open →

Measure

Whether the result is real, and how sure you can be.

Operate

Keeping a live system healthy, and how much headroom is left.