[
https://issues.apache.org/jira/browse/KAFKA-6414?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16307606#comment-16307606
]
Ted Yu commented on KAFKA-6414:
-------------------------------
Interesting.
Since the proposal changes the semantics of replica failover handling, does
this need a KIP ?
> Inverse replication for replicas that are far behind
> ----------------------------------------------------
>
> Key: KAFKA-6414
> URL: https://issues.apache.org/jira/browse/KAFKA-6414
> Project: Kafka
> Issue Type: Bug
> Components: replication
> Reporter: Ivan Babrou
>
> Let's suppose the following starting point:
> * 1 topic
> * 1 partition
> * 1 reader
> * 24h retention period
> * leader outbound bandwidth is 3x of inbound bandwidth (1x replication + 1x
> reader + 1x slack = total outbound)
> In this scenario, when replica fails and needs to be brought back from
> scratch, you can catch up at 2x inbound bandwidth (1x regular replication +
> 1x slack used).
> 2x catch-up speed means replica will be at the point where leader is now in
> 24h / 2x = 12h. However, in 12h the oldest 12h of the topic will fall out of
> retention cliff and will be deleted. There's absolutely to use for this data,
> it will never be read from the replica in any scenario. And this not even
> including the fact that we still need to replicate 12h more of data that
> accumulated since the time we started.
> My suggestion is to refill sufficiently out of sync replicas backwards from
> the tip: newest segments first, oldest segments last. Then we can stop when
> we hit retention cliff and replicate far less data. The lower the ratio of
> catch-up bandwidth to inbound bandwidth, the higher the returns would be.
> This will also set a hard cap on retention time: it will be no higher than
> retention period if catch-up speed if >1x (if it's less, you're forever out
> of ISR anyway).
> What exactly "sufficiently out of sync" means in terms of lag is a topic for
> a debate. The default segment size is 1GiB, I'd say that being >1 full
> segments behind probably warrants this.
> As of now, the solution for slow recovery appears to be to reduce retention
> to speed up recovery, which doesn't seem very friendly.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
