Kafka 'replications.number' of brokers in a Kafka 2.3 cluster

Question

Im trying to optimize a 3-brokers Kafka cluster, and got into a very basic yet unclear issue.

Consider i have this configuration:

kafka.topic.partitions.number=10
kafka.partition.replications.number=3
min.insync.replicas=2

This means that :

10 partition per topic.

Each partition should be replicated on all brokers

Partition is healthy if 2 in-sync-replicas (allows 1 failing broker)

The question is about the necessity of replication=3, is it really necessary ?

the value '2' will also allow:

1 broker to fail

probably faster convergence and recovery

In the case of 2 failing brokers, the cluster will not function anyway.

BTW - my zookeeper installed on the same kafka machines.

Thanks .

H.Ç.T H.Ç.T · Accepted Answer · 2020-01-22T07:41:05

First of all to make it more clear:

2 in-sync-replicas means you must have 2 in-sync-replicas to produce messages, otherwise your producer gets NOT_ENOUGH_REPLICAS exception and producing message will not be possible. (I suppose you set acks=all in producer side)

For your scenario to have topic with replication.factor = 2:

Suppose that your broker ids are 1,2,3. At the beginning you have 3 healhty brokers and when you describe topic you will see something like this:

What if broker with id 2 is crashed? Then you you will have just one in-sycn-replica for partiton 1 and 2 as shown below:

and your producer cannot produce messages because min.insync.replicas = 2

Of course you can reassign-partitions at that point but it is still a problem as you see.

As a result; best practice is to set replication.factor = 3 if your min.insync.replicas = 2 and number of brokers in cluster is 3.

Kafka 'replications.number' of brokers in a Kafka 2.3 cluster

2 Answers