Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
Kafka >> mail # user >> Purgatory


Copy link to this message
-
Re: Purgatory
Marc, thanks for writing that up. I think it is worth adding some
details on the request-purgatory on a wiki (Jay had started a wiki
page for kafka internals [1] a while ago, but we have not had time to
add much to it since.) Your write-up could be reviewed and added
there. Do you have edit permissions on the wiki?

As for the purge interval config - yes the documentation can be
improved a bit. It's one of those "internal" configs that generally
don't need to be modified by users. The reason we added that was as
follows:
- We found that for low-volume topics, replica fetch requests were
getting expired but sitting around in purgatory
- This was because we were expiring them from the delay queue (used to
track when requests should expire), but they were still sitting in the
watcherFor map - i.e., they would get purged when the next producer
request to that topic/partition arrived, but for low volume topics
this could be a long time (or never in the worst case) and we would
eventually run into an OOME.
- So we needed to periodically go through the entire watcherFor map
and explicitly remove those requests that had expired.
- More details on this are in KAFKA-664.

Thanks,

Joel

[1] https://cwiki.apache.org/confluence/display/KAFKA/Kafka+Internals

On Fri, Nov 1, 2013 at 12:33 PM, Marc Labbe <[EMAIL PROTECTED]> wrote:
> Guozhang,
>
> I have to agree with Priya the doc isn't very clear. Although the
> configuration is documented, it is simply rewording the name of the config,
> which isn't particularly useful if you want more information about what the
> purgatory is. I searched the whole wiki and doc and could not find anything
> very useful as opposed looking a the code. In this case,
> kafka.server.KafkaApis and kafka.server.RequestPurgatory will be your
> friends.
>
> I'll try to add to Joe's answer here, mostly just reporting what's
> available in the Scala doc from the project. I am doing this to understand
> the mechanics myself btw.
>
> As Joe said, messages are not dropped by the purgatory but simply removed
> from the purgatory when they are satisfied. Satisfaction conditions are
> different for both fetch and produce requests and this is implemented in
> their respective DelayedRequest implementation (DelayedFetch and
> DelayedProduce).
>
> Requests purgatories are defined as follow in the code:
>  - ProducerRequestPurgatory: A holding pen for produce requests waiting to
> be satisfied.
>  - FetchRequestPurgatory: A holding pen for fetch requests waiting to be
> satisfied
>
> Each request purgatory runs a thread (ExpiredRequestReaper). This thread
> will first try to find an expired delayed request. When one if found, it
> will run the purgatory's expire method to handle the delayed request
> expiration. In both produce and fetch cases, it sends a response to the
> client. An expired request will be a satisfied request. The next step of
> the thread's loop is when it checks for the configuration parameters you
> asked for initially (purgatory.purge.interval.requests). When the number of
> delayed requests given to watch by the purgatory reaches this value, it
> goes through all previously queued requests and removes those which are
> marked as satisfied. Because of that, it is really an interval more than it
> is a threshold since it doesn't really care about the amount of satisfied
> requests or the size of the queue.
>
> Producer request
> - When is it added to purgatory (delayed)?:
>   * when it uses ack=-1 (actually, the code tells me anything but 0 or 1);
> Producer config: request.required.acks
>   * partitions have more than one replica (in this case, ack=-1 isn't
> different to ack=1 and it doesn't make much sense to use a delayed request)
>   * not all partitions are in error
> - When does it expire? when it reaches the timeout defined in the produce
> request (ackTimeoutMs). Translates from producer config request.timeout.ms.
> - What happens (on the broker) when it expires? Sends a response to the
> client. Response content depends on the request of course.