Some workloads create cache pressure at a faster rate than WiredTiger's background eviction threads can handle. While we can improve efficiency, with our current architecture there will always be some situations that cause the cache to reach the "trigger" point where application operations stall to perform eviction.
As a backstop, have applications stop adding to cache pressure when background threads aren't keeping up. Since most of the cases we have seen involve large scans, this be similar to automatically turning on the "read-won't-need" flag in WiredTiger.
- causes
-
SERVER-40433 Mongodump extremely slow on 4.0.8
- Closed
-
WT-4881 Soften the restrictions on re-entering reconciliation
- Closed
-
WT-5218 Improve eviction to differentiate between clean and dirty pages with WT_CACHE_EVICT_NOKEEP readgen
- Closed
- is related to
-
WT-4502 Assertion checking hazard pointers on page discard is too strong
- Closed
- related to
-
WT-4893 Fix a race between internal page child-page eviction checks and cursors in the tree
- Closed
-
WT-4957 Revert part of a change about when pages are queued for urgent eviction
- Closed
-
WT-5050 Assertion failure during urgent eviction of metadata page
- Closed
-
WT-4898 Don't allow the eviction server to reconcile if it's busy
- Closed