Closed hzxa21 closed 4 days ago
This is a long existing issue. We should first make sure all executors are not evicting cache on barrier basis.
This is a long existing issue. We should first make sure all executors are not evicting cache on barrier basis.
I think we already did? Just checked hash join and hash agg.
This is a long existing issue. We should first make sure all executors are not evicting cache on barrier basis.
It had been done before the sequence-based operator cache eviction.
After #16087, we have switched to using sequence-based memory eviction policy instead of an epoch-based policy. This means we no longer need to rely on barrier to increase the watermark. However, it seems that we are still use
barrier_interval_ms
as the ticker interval here. This may result in delay in cache eviction whenbarrier_interval_ms
is increased.