At Storage Field Day last month, Jon Klaus saw a presentation from SNIA, where they talked about tail latency. These are tiny spikes in latency, 2-10x longer than higher. These can have big impacts, effecting up to 2% of all IO, which is magnified when seen at a hyperscale level. Smaller organizations might simply over provision to lower overall latency so that these spikes still fall within acceptable ranges, but at scale this approach isn’t practical. SNIA presented several ways to overcome tail latency, and Jon shares his thoughts on what he saw.
