Workloads that insert a large value, and then do many WT_CURSOR->modify operations to make many small updates to the value while holding the oldest transaction/timestamp pinned can generate very large pages in cache, and force us to regularly re-write the page via lookaside.
Which leads to effectively pinning the cache full.
We could do better by decreasing how often we instantiate a full value in such workloads, so there is less memory pinned in cache.
- causes
-
WT-4115 Valgrind error in est_wt4105_large_doc_small_upd
- Closed
- is depended on by
-
SERVER-33191 Cache-full hangs on 3.6
- Closed
- related to
-
SERVER-36221 [3.6] Performance regression on small updates to large documents
- Closed
-
WT-4107 Reduce lookaside usage in update heavy workloads where history is pinned
- Closed