WT-12218 cache_eviction_target_strategy Statistics bug fix and Optimized macro definition #10061
Conversation
Record how many pages we want and how many pages we have queued so far in a tree, then resume the walk next iteration. This avoids a single tree with a target larger than the queue size being walked completely before the eviction server moves on to the next tree. (cherry picked from commit fca6a8d)
…s. (wiredtiger#3826) (cherry picked from commit 2f489ea)
There's trickiness in the page-pinned check. By definition a remove operation leaves a cursor positioned if it's initially positioned. However, if every item on the page is deleted and we unpin the page, eviction might delete the page and our search will re-instantiate an empty page for us. Cursor remove returns not-found whether or not that eviction/deletion happens and it's OK unless cursor-overwrite is configured (which means we return success even if there's no item to delete). In that case, we'll fail when we try to point the cursor at the key on the page to satisfy the positioned requirement. It's arguably safe to simply leave the key initialized in the cursor (as that's all a positioned cursor implies), but it's probably safer to avoid page eviction entirely in the positioned case. (cherry picked from commit 2ac616e)
…dtiger#4017) This reverts commit 565496e.
|
Hi @y123456yz, thank you for your submission! |
korteland
changed the title
| } else | ||
| } | ||
|
|
||
| if (F_ISSET(cache, WT_CACHE_EVICT_DIRTY) && F_ISSET(cache, WT_CACHE_EVICT_UPDATES)) | ||
| WT_STAT_CONN_INCR(session, cache_eviction_target_strategy_both_clean_and_dirty); |
There was a problem hiding this comment.
@y123456yz thanks for this PR.
Could you help me understand the reasoning behind these changes?
|
@y123456yz let me know your thoughts on #10061 (comment) |
|
|
@y123456yz thanks, but I do not see the current check as a bug. Could you please explain the reasoning behind this change. |
hi, @Siddhartha8899 Since dirty contains updates, it can be simplified to: Our disagreement should be whether to consider WT_CACHE_EVICT_UPDATES？ |
|
Hi @y123456yz we are currently reviewing this specific code base internally and would like to hold off making any changes until we have determined how to proceed. So we would like to close this PR without merging at this time. There is currently no deadline for the current code review, but we will endeavor to update the corresponding Jira ticket when appropriate. |
cache_eviction_target_strategy Statistics bug fix and Optimized macro definition