Replication Latency Case (1) - Eventual Consistency

Tolerating node failures is only one reason to use replication.Other reasons include:

• Scalability, using multiple nodes to handle more requests
• Low latency, allowing replicas to be geographically closer to users

Master-slave replication requires that all write requests are processed by the master node, and the slave node can only process it.This is a good choice for more read and less write scenarios: create multiple slave nodes, spread read requests to all slave nodes, thereby reducing the load on the master node and allowing read requests to be sent to the nearest replica.

With this scalable structure, only adding more slave nodes can improve the service throughput of read requests.But this is only suitable for asynchronous replication. If you try to replicate to all slave nodes synchronously, a single node failure or network outage will make the entire system unable to write.And the more nodes, the higher the probability of failure, so the fully synchronized configuration is very unreliable.

2.1 Eventual Consistency

If the application happens to be reading from an asynchronous slave node that lags behind the master node, it may see outdated data, causing inconsistencies in the database: since not all writes are reflected in the slave node, ifAt the same time, the same query is issued to the master and slave nodes, and different results may be obtained.This inconsistency is only a temporary state. If you stop writing the DB and wait for a while, the slave node will eventually catch up and be consistent with the master node.Not only NoSQL databases are eventually consistent: asynchronous replica followers in relational databases have the same feature.

The word "final" is intentionally vague, and theoretically, there is no limit to how far a copy can fall behind.In normal operation, the time delay (replication lag) between the completion of write operations on the master node and the slave node may be less than 1s, which usually does not cause much impact in practice.But if the system is operating near its limit or there is a problem with the network, the delay can easily exceed a few seconds or even minutes.