In workloads with:

• Many more threads than CPU cores
• Enough cache pressure that the cache remains 100% full
• Application threads are required to contribute to eviction
• Transactions span multiple operations

Some threads can be tasked with doing more eviction than is fair which leads to occasional operations with longer than expected latencies.

We have gone backwards and forwards on the best approach for sharing eviction when there is lots of cache pressure and many threads.

WT-2702 was needed to fix the code, so memory usage can't grow above 100%.
WT-2114 was a change to improve the fairness of the code to avoid long latency operations.