KubeCon + CloudNativeCon + Open Source Summit + AI_dev China 2024

As a distributed kv database, TiKV supports no-transactional operations and transactional operations. In the transactional mode, there would be many kv versions if there are many UPDATE operations. Scanning such a range containing a large number of kv versions, the read latency is unpredictable. In order to achieve stable low latency for TiKV, we introduced a "In-Memory Engine" to reduce the read latency in such case. Meanwhile the "In-Memory Engine" also can improve the overall performance when there are some hot ranges with heavy read workload.

作为一个分布式kv数据库，TiKV支持非事务操作和事务操作。在事务模式下，如果有很多UPDATE操作，就会有很多kv版本。扫描包含大量kv版本的范围时，读取延迟是不可预测的。为了实现TiKV的稳定低延迟，我们引入了一个“内存引擎”来减少这种情况下的读取延迟。同时，“内存引擎”还可以在存在一些热门范围和大量读取工作负载时提高整体性能。