[
https://issues.apache.org/jira/browse/KAFKA-2063?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14952925#comment-14952925
]
Tom Lee commented on KAFKA-2063:
--------------------------------
This sounds familiar. Am I right in thinking something similar could impact
broker replica fetchers given a large number of partitions per broker?
We've observed symptoms of behavior like this in a somewhat badly configured
0.8.1.1 cluster, with almost 1k partitions per broker where the leader/follower
split is roughly 50/50, replica.fetch.max.bytes is ~20MB & there are some 20
replica fetcher threads.
"Steady state" is more or less fine with an 8GB heap and the live set is
generally minimal, but we see OOMEs during startup as the broker plays "catch
up" on partitions its following and heavy GC during replica reassignments. With
a much larger ~20GB heap the OOMEs go away, but GC is still *very* heavy for
several minutes during startup and we see regular promotion failures despite a
huge amount of free space in old gen, (implying large promotions).
>From what I've seen of the 0.8.1.1 code, I have a very strong suspicion a
>similar scenario to what you're describing here would also occur in
>ReplicaFetcherThread/AbstractFetcherThread as each thread might be pulling
>data back from dozens of partitions at a time. With
>replica.fetch.max.bytes=20MB, all it takes is us being a little behind on say
>50 partitions & all of a sudden we need 1GB for a single fetch.
Even with saner values for replica.fetch.max.bytes (say ~5MB), it doesn't get a
whole lot better given the number of partitions involved should a broker fall a
little behind due to e.g. unexpected GC pauses.
Anyway, perhaps it's worth opening a separate ticket for this but I suspect
fixing this issue at the protocol level will let us fix the issue for replica
fetchers / brokers too (if it hasn't already been fixed for replica fetchers in
0.8.2+ by some other means).
> Bound fetch response size
> -------------------------
>
> Key: KAFKA-2063
> URL: https://issues.apache.org/jira/browse/KAFKA-2063
> Project: Kafka
> Issue Type: Improvement
> Reporter: Jay Kreps
>
> Currently the only bound on the fetch response size is
> max.partition.fetch.bytes * num_partitions. There are two problems:
> 1. First this bound is often large. You may chose
> max.partition.fetch.bytes=1MB to enable messages of up to 1MB. However if you
> also need to consume 1k partitions this means you may receive a 1GB response
> in the worst case!
> 2. The actual memory usage is unpredictable. Partition assignment changes,
> and you only actually get the full fetch amount when you are behind and there
> is a full chunk of data ready. This means an application that seems to work
> fine will suddenly OOM when partitions shift or when the application falls
> behind.
> We need to decouple the fetch response size from the number of partitions.
> The proposal for doing this would be to add a new field to the fetch request,
> max_bytes which would control the maximum data bytes we would include in the
> response.
> The implementation on the server side would grab data from each partition in
> the fetch request until it hit this limit, then send back just the data for
> the partitions that fit in the response. The implementation would need to
> start from a random position in the list of topics included in the fetch
> request to ensure that in a case of backlog we fairly balance between
> partitions (to avoid first giving just the first partition until that is
> exhausted, then the next partition, etc).
> This setting will make the max.partition.fetch.bytes field in the fetch
> request much less useful and we should discuss just getting rid of it.
> I believe this also solves the same thing we were trying to address in
> KAFKA-598. The max_bytes setting now becomes the new limit that would need to
> be compared to max_message size. This can be much larger--e.g. setting a 50MB
> max_bytes setting would be okay, whereas now if you set 50MB you may need to
> allocate 50MB*num_partitions.
> This will require evolving the fetch request protocol version to add the new
> field and we should do a KIP for it.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
