How do you size a thread pool?

For CPU-bound tasks, more threads than cores only adds context switches: availableProcessors(), or one less if a thread is needed elsewhere. For blocking tasks the threads are mostly asleep, so the useful count scales with how much of the time is spent waiting: N = cores * utilisation * (1 + W/S). A task that waits 90 ms per 10 ms of compute wants roughly ten threads per core.

Say the rest of it, because that is what distinguishes an engineer from a formula: the number is a starting point to be measured, separate pools for separate workloads so a slow dependency cannot starve fast ones, and availableProcessors() respects cgroup limits in a container only on a reasonably modern JVM, so check what the pod actually reports.

Virtual threads change this: for blocking work the answer becomes "do not pool at all".