eugenebolotin / redis-replica

Redis failover and master election manager

Introduction

People! We are all waiting for production version of Redis Cluster.
But we want to use Redis in production now and we need some failover in it.

A year ago Redis Sentinel was not able to return falled master into cluster as slave.
Then I decided to write a small and dumb Redis failover to use in production.

Redis-replica

So, here it is. Redis-replica is a failover and master election manager for Redis.
It continuously checks availability of specified Redis nodes.
Each redis-replica daemon is connected to other redis-replica daemons and exchanges information about pings.

All redis-replica nodes act equally. So if multiple redis-replica nodes detect a problem all of them try to solve it.
Redis can handle multiple similar SLAVE OF-commands, so in our case above strategy is normal.

Redis nodes that are visible from majority of redis-replica nodes are called Quorum-Nodes.
The counterpart is called Nonquorum-Nodes.

Several conditions are checked:

1. No master at all.
2. Multiple masters.
3. Slaves are connected to different masters.
4. Split-brain condition - all nonquorum master turned into slaves.

In the cases 2 and 3 master with more used_memory is elected. It's our production case but can be easily changed.

Usage

./redis-replica.py < config >

Third-party

Redis first of all.

redis-replica depends on two cool python libraries:

• python-redis
• python-httplib2