[
https://issues.apache.org/jira/browse/KAFKA-1835?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14628122#comment-14628122
]
David Hay commented on KAFKA-1835:
----------------------------------
It seems to me that there are two aspects to sending a message: (a) selecting a
partition and (b) sending the message to the partition leader. Currently, (b)
is performed asynchronously, (a) is not, and contains a potentially blocking
call to get partition metadata (or refresh it).
Maybe I'm missing something, but why couldn't the producer just queue the
ProducerRecord and have a separate thread take care of querying for broker
metadata, serializing the key and message, selecting a partition and pushing
the result onto the send queue? In other words, put the current implementation
of {{send()}} in a separate thread? This is essentially what we've had to do
in order to work around the possibility of {{send()}} blocking when we don't
want it to. In addition, it puts the serialization process in a separate
thread, which is important when we need to send a message and respond to user
input as quickly as possible.
> Kafka new producer needs options to make blocking behavior explicit
> -------------------------------------------------------------------
>
> Key: KAFKA-1835
> URL: https://issues.apache.org/jira/browse/KAFKA-1835
> Project: Kafka
> Issue Type: Improvement
> Components: clients
> Affects Versions: 0.8.2.0, 0.8.3, 0.9.0
> Reporter: Paul Pearcy
> Fix For: 0.8.3
>
> Attachments: KAFKA-1835-New-producer--blocking_v0.patch,
> KAFKA-1835.patch
>
> Original Estimate: 504h
> Remaining Estimate: 504h
>
> The new (0.8.2 standalone) producer will block the first time it attempts to
> retrieve metadata for a topic. This is not the desired behavior in some use
> cases where async non-blocking guarantees are required and message loss is
> acceptable in known cases. Also, most developers will assume an API that
> returns a future is safe to call in a critical request path.
> Discussing on the mailing list, the most viable option is to have the
> following settings:
> pre.initialize.topics=x,y,z
> pre.initialize.timeout=x
>
> This moves potential blocking to the init of the producer and outside of some
> random request. The potential will still exist for blocking in a corner case
> where connectivity with Kafka is lost and a topic not included in pre-init
> has a message sent for the first time.
> There is the question of what to do when initialization fails. There are a
> couple of options that I'd like available:
> - Fail creation of the client
> - Fail all sends until the meta is available
> Open to input on how the above option should be expressed.
> It is also worth noting more nuanced solutions exist that could work without
> the extra settings, they just end up having extra complications and at the
> end of the day not adding much value. For instance, the producer could accept
> and queue messages(note: more complicated than I am making it sound due to
> storing all accepted messages in pre-partitioned compact binary form), but
> you're still going to be forced to choose to either start blocking or
> dropping messages at some point.
> I have some test cases I am going to port over to the Kafka producer
> integration ones and start from there. My current impl is in scala, but
> porting to Java shouldn't be a big deal (was using a promise to track init
> status, but will likely need to make that an atomic bool).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
