Separate Input, Retrieval, and Concurrency Limits for E-commerce Product Search

A single query limit is too blunt for product search, since long pasted messages, broad retrieval requests, and bursts of small searches each strain different parts of the system. The article recommends setting separate admission limits on input size, candidate breadth, and concurrent work, then rejecting or degrading requests before they consume the latency budget. It warns that reducing query length or candidate counts can silently drop decisive details, so teams should track a labeled query set alongside latency metrics. Capacity reviews should examine latency distributions and the heaviest accepted request class, especially after catalog imports change document size. Rejections should return stable, documented responses with reason-labeled metrics, so clients can tell invalid input apart from temporary capacity limits.