On a YCSB test, I saw a 33% performance increase, both on the max latency and on the throughput. I'm convinced enough that this value is better that I think it makes sense to change it on 0.98 as well.
More fundamentally, but outside of the scope of this patch, I think this parameter should be changed to something at the region server level: today, we have:
- global memstore check: if we're other 40%, we flush the biggest memstore
- local: no more than 2 (proposed: 4) memstore size per region.
But if we have enough memory and a spike on a region, there is no reason for not taking the write.