Skip to content

Batch Window Planner

Converts a volume and a deadline into the throughput and worker count that make it, with the setup time and the safety margin counted.

Workers required to make the deadline
109

Your 60 workers finish in 10.7 h — 4.7 h past the deadline.

Usable window after setup
5.8 h
Required throughput
95.2 items/s
Required throughput, per minute
5,714 items/min
One worker delivers
1.25 items/s
Workers at 100% utilisation
76.2
Workers at 70% utilisation
109
Throughput of the planned fleet
52.5 items/s
Finish time
10.7 h
Slack against the deadline
−4.7 h
What this assumes: items are independent and uniform in cost, so the fleet keeps every worker busy for the whole window — real batches have a tail, and a long-running item at the end idles everything else. Service time is wall clock per item including the network, not CPU time. The utilisation figure is the safety margin doing its job: it is not a measured efficiency, it is the fraction of the theoretical rate you are willing to commit to. No retries are counted; add them to the item count.

Why the required number is the one with headroom in it

The arithmetic of a batch window is simple enough to do in your head and it is still worth writing down, because the naive version leaves out the two terms that make jobs miss deadlines. Items divided by the window gives a required rate. Service time inverted gives a per-worker rate. Divide one by the other and you have the number of workers — at which point people provision exactly that many and are surprised when it does not finish.

The first missing term is the usable window, which is shorter than the window. Spin-up, model loading, the first-request penalty on a cold pool, and the final write all come out of it, and on a six-hour window a ten-minute overhead is 3% of your capacity gone before an item is processed.

The second is utilisation. Planning at 100% of theoretical throughput means every provider slowdown, every retry and every straggler eats directly into the deadline, because there is nowhere for the lost time to go. Planning at 70% costs 43% more workers and turns a missed deadline into a late finish. This is the same trade as a queueing system running hot — the closer to saturation you plan, the more violently small perturbations move the finish time. If the workers are rented by the hour the extra ones cost nothing extra overall: the same total work over a shorter time. The headroom is close to free, which is why it is odd how rarely it is bought.

Batch Window Planner · Multigrid