Predictable single-digit millisecond latency

[zhangjinpeng87]

Background

As the general storage layer of other mission critical RDMBS database like TiDB, predictable single-digit millisecond latency makes a lot of sense. But the latency of TiKV can be spike under some cases like temporarily slow IO, TiKV node restarted, data range (region) leader transferred to another TiKV node, there are some background jobs like import, data range replica re-balanced, scale-in some TiKV nodes, etc.

Development Task

• Slow disk IO impact the write latency
• Too many MVCC versions impact the read latency
• Scale-in TiKV nodes cause read/write latency spike
• Background jobs impact read/write latency
• Restart a TiKV node cause read latency spike
• Leader transfer cause read latency spike