A Key Value Store that Supports Strict SLAs and the Applications that Need it

Опубликовано 28 июля 2016, 0:33

Emerging datacenter workloads differ from traditional e-commerce workloads in two ways. First, emerging big-science workloads can access networked storage thousands of times for 1 simulation. Even a few slow accesses can slowdown the entire simulation. Second, emerging niche workloads in green computing have non-technical competitive advantages. Providers of these workloads want to keep technical costs as low as possible while barely meeting their user's performance needs. Both of these workloads point to an emerging technical need, managing performance-oriented service level agreements (SLAs). In this talk, I will present Zoolander, a key value store that can complete a very high percentage of storage accesses (i.e., a service level) within tight, time constraints. The key idea behind Zoolander is to revisit replication for predictability, an old but seldom-used approach to mask the effects of uncommon events. Zoolander mixes replication for predictability, partitioning, and traditional replication to scale efficiently, meeting strict SLAs while using half as many cloud nodes. I will present results where Zoolander can complete 99.99% of its storage accesses within a strict, 15ms latency bound. It does this by reducing 99th percentile latencies by 78%. Zoolander provides write-order consistency, survives failures, and has been scaled in our tests to 32 nodes.