fix(runner): add 10s cooldown between concurrent and serial search ph…#801
fix(runner): add 10s cooldown between concurrent and serial search ph…#801zpatronus wants to merge 1 commit into
Conversation
…ases The serial search phase measures single-query latency under idle conditions. Previously it launched immediately after the concurrent phase, while the end-to-end path may still be in a saturated state. This significantly affects the accuracy of serial latency, especially p99/p95, because the first few queries experience an environment similar to the concurrent test rather than the intended single-user condition.
|
[APPROVALNOTIFIER] This PR is NOT APPROVED This pull-request has been approved by: zpatronus The full list of commands accepted by this bot can be found here. DetailsNeeds approval from an approver in each of these files:Approvers can indicate their approval by writing |
|
/assign @XuanYang-cn |
|
@zpatronus I'd suggest closing this. The PR assumes serial latency should reflect a single-user / idle condition, and a cooldown brings it closer to that. But serial isn't measuring idle — it measures warm best-case latency. That's why serial runs after concurrent on purpose: the concurrent phase replays the same queries to warm the server, and serial then measures that warm state. So post-concurrent isn't contamination to recover from — it's exactly the intended condition. A cooldown only pushes the server back toward cold, away from the target. Also, |
|
@XuanYang-cn Fair point on the warm state. The cooldown isn't meant to evict data caches or similar thing, just to drain transient things like in-flight requests, network buffers, etc. |
|
#803 implements the configurable parameter. |
3 participants
…ases
The serial search phase measures single-query latency under idle conditions. Previously it launched immediately after the concurrent phase, while the end-to-end path may still be in a saturated state. This significantly affects the accuracy of serial latency, especially p99/p95, because the first few queries experience an environment similar to the concurrent test rather than the intended single-user condition.
This improves serial latency accuracy for all database backends.