Use coarse time when calculating metrics refresh interval #1381

Activity

1. andygrove

   added

   enhancementNew feature or requestNew feature or request

   on Feb 8, 2025
2. andygrove

   added this to the 0.7.0 milestone on Feb 8, 2025
3. andygrove

   modified the milestones: 0.7.0, 0.8.0 on Mar 3, 2025
4. andygrove

   modified the milestones: 0.8.0, 0.9.0 on Apr 17, 2025
5. andygrove

   removed this from the 0.9.0 milestone on Jun 5, 2025
6. andygrove commented on Sep 28, 2026

   

   andygrove

   on Sep 28, 2026

   MemberAuthor

   More actions

   Closing this, since the loop it was about is gone. The interval check's Instant::now() used to run on every spin of the busy-poll loop in executePlan. #2940 cut that to once per 100 polls, and #6219 replaced the busy loop with one that parks while it waits on native I/O. What's left is one clock read per output batch and one per pending poll. Each of those sits next to a JNI handoff, a pull from the JVM or a wake-up that costs far more than the read. Most DataFusion operators also read the clock twice per batch for their own metrics, and coarsetime can't reach those, so switching this one call would add a dependency without a measurable gain. Happy to reopen if a profile says otherwise.
7. andygrove

   closed this as not plannedon Sep 28, 2026