When would you use Parallel.ForEachAsync instead of Task.WhenAll?

In production I choose based on workload shape and dependency capacity. Task.WhenAll works well for a small fixed fan-out; Parallel.ForEachAsync or a channel/queue is better for large collections where I need backpressure and controlled concurrency. I also care about cancellation, partial failure handling, retries, and result aggregation. For CPU-heavy work, I consider CPU cores and isolation; for I/O-heavy work, I consider downstream limits and rate policies. The pattern should make overload behavior predictable. In production I would bound the work explicitly and observe it: active operations, queue length, cancellation rate, dependency latency, and failure rate. I would avoid designs where every request can create unlimited work, because that turns a small dependency slowdown into a process-wide outage.