Search latency appears only under concurrent request load

The goal is to determine whether the latency observed only when many search requests run concurrently is caused by search thread‑pool queuing or by other background activities such as translog flushes or refresh operations.

To isolate the effect, index‑level settings that influence refresh and translog behavior (e.g., refresh_interval, translog.flush_threshold_size) must remain unchanged while the concurrent request load is varied. Thread‑pool statistics (queue size and rejected count) should be collected alongside client‑side latency measurements to see if latency correlates with queue growth.

Does the search thread‑pool queue size increase proportionally with observed latency when the number of concurrent requests exceeds the configured thread‑pool size?
Is latency still present when the thread‑pool queue size remains zero despite high concurrent request volume?