There are two definitions of latency: 1. How long before the writer gets an acknowledgement for their write. This depends on the acks setting the producer has as Guozhang says. If acks = 1 we wait just on the leader, if acks=-1 we wait on all "in sync" brokers (i.e. alive brokers). 2. How long before the consumer gets the message. This always blocks on full acknowledgement by "in sync" brokers.
So both can be dependent on the latency of the followers, however if a followers falls behind it will be removed from the in sync set, so there is a bound. -Jay On Tue, Feb 4, 2014 at 11:57 AM, Michael Popov <mpo...@microsoft.com> wrote: > Hi, > > I am trying to understand what influences message latency in Kafka. > > According to the Kafka Replication Design document ( > https://linkedin.jira.com/secure/attachment/10050/kafka_replication_highlevel_design.pdf > ): > > 1. Writes: Once the leader receives the acknowledgment from all > replicas in ISR, the message is committed. The leader advances the HW and > sends an acknowledgment to the client (page 4) > > 2. Reads: Only messages up to the HW are exposed to the reader (page > 5) > > It means message latency depends on how fast followers pulled data from > the leader and reported successful data writes. > Is this correct? > > Thank you, > Michael Popov > >
