kafka-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Evan Huus <evan.h...@shopify.com>
Subject Re: Topicmetadata response miss some partitions information sometimes
Date Mon, 02 Mar 2015 03:11:58 GMT
Which I think is my point - based on my current understanding, there is
*no* way to find out the total number of partitions for a topic besides
hard-coding it or manually reading it from zookeeper. The kafka metadata
API does not reliably expose that information.

Evan

On Sun, Mar 1, 2015 at 10:07 PM, Guozhang Wang <wangguoz@gmail.com> wrote:

> I see.
>
> If you need to make sure messages are going to the same partition during
> broker bouncing / failures, then you should not depend on the partitioner
> to decide the partition id but explicitly set it before calling send().
> For example, you can use the total number of partitions for the topic, not
> the number of available partitions to compute partition id.
>
> Guozhang
>
> On Sun, Mar 1, 2015 at 6:42 PM, Evan Huus <evan.huus@shopify.com> wrote:
>
> > My concern is more with the partitioner that determines the partition of
> > the message. IIRC, it does something like "hash(key) mod #partitions" in
> > the normal case, which means if the # of partitions changes because some
> of
> > them are offline, then certain messages will be sent to the wrong
> (online)
> > partition, no?
> >
> > Evan
> >
> > On Sun, Mar 1, 2015 at 9:36 PM, Guozhang Wang <wangguoz@gmail.com>
> wrote:
> >
> > > Evan,
> > >
> > > In the java producer, partition id of the message is determined in the
> > > send() call and then the data is appended to the corresponding batch
> > buffer
> > > (one buffer for each partition), i.e. the partition id will never
> change
> > > once it is decided. If the partition becomes offline after this, the
> send
> > > call will fail and then retry. In the end the message will either
> exhaust
> > > all retries and be dropped on the floor or the partition becomes online
> > > again and the metadata gets refreshed, and message send retry
> > successfully.
> > > Hence, if some of your partitions becomes offline for too long some
> data
> > > will be lost on the producer side (unless you set infinite retry, of
> > > course).
> > >
> > > Guozhang
> > >
> > > On Sun, Mar 1, 2015 at 5:25 AM, Evan Huus <evan.huus@shopify.com>
> wrote:
> > >
> > > > On Sun, Mar 1, 2015 at 1:46 AM, Guozhang Wang <wangguoz@gmail.com>
> > > wrote:
> > > >
> > > > > Hi Honghai,
> > > > >
> > > > > 1. If a partition has no leader (i.e. all of its replicas are down)
> > it
> > > > will
> > > > > become offline, and hence the metadata response will not have this
> > > > > partition's info.
> > > > >
> > > >
> > > > If I am understanding this correctly, then this is a problem (in
> > certain
> > > > cases) for the producer because of an ambiguity.
> > > >
> > > > If a producer using hash-partitioning receives partitions 1-4 and
> > begins
> > > > producing, then updates its metadata and receives only partitions
> 1-2,
> > > > there are two possible scenarios it cannot distinguish between:
> > > >
> > > > 1. The user reduced the number of partitions for this topic. The
> > producer
> > > > should continue producing and distribute all new messages between the
> > two
> > > > remaining partitions.
> > > > 2. Two of the four partitions are entirely offline. The producer
> should
> > > > continue to distribute messages among all four partitions (to
> maintain
> > > the
> > > > consistency of the hashing) but two of the four partitions will
> simply
> > > > fail.
> > > >
> > > > Whichever behaviour the producer chooses, in the other scenario it
> will
> > > > incorrectly distribute messages among the partitions, thus breaking
> the
> > > > hash-partitioner guarantee.
> > > >
> > > > If all of the replicas are down for a partition, why do metadata
> > requests
> > > > not simply return that partition with LeaderNotAvailable?
> > > >
> > > > Thanks,
> > > > Evan
> > > >
> > > >
> > > > > 2. Any of the brokers cache metadata and hence can handle the
> > metadata
> > > > > request. It's just that their cache are updated asynchronously and
> > > hence
> > > > > when there is a update to the metadata, some brokers may got the
> new
> > > > > metadata value a bit eariler than others.
> > > > >
> > > > > On Thu, Feb 26, 2015 at 7:21 PM, ChenHongHai <
> > > > waldenchenkafka@outlook.com>
> > > > > wrote:
> > > > >
> > > > > > We have one topic with 4 partitions, but sometimes only get
> > metadata
> > > > of 2
> > > > > > partitions, did anyone meet this kind of situation before?If
some
> > > > > partition
> > > > > > has no leader at that moment, will it cause this problem?  
  How
> > to
> > > > make
> > > > > > some partition has no leader?If 6 brokers has some partitions
of
> > the
> > > > > topic,
> > > > > > will they return same result?  Do I need try all of them and
> merge
> > > the
> > > > > > result?
> > > > > >
> > > > > >      SimpleConsumer consumer =
> consumerPool.getConsumer(seed.host,
> > > > > > seed.port, connectionTimeOut, consumerBufferSize,
> > > > > "refreshPartitionMeta");
> > > > > >               List<String> topics = new ArrayList<String>()
{{
> > > > > >       add(topic);                }};
> > > > TopicMetadataResponse
> > > > > > resp = consumer.send(new TopicMetadataRequest(topics));
> > > > > > List<TopicMetadata> metaData = resp.topicsMetadata();
> > > > > >                 for (TopicMetadata item : metaData) {
> > > > > > if(item.errorCode() != kafka.common.ErrorMapping.NoError())
> > > > > >         LOG.error(String.format("Something wrong with topic
> > metadata
> > > > for
> > > > > > topic: %s error code: %d ", item.topic(), item.errorCode() ));
> > > > > >       for (PartitionMetadata part : item.partitionsMetadata())
{
> > > > > >             partitionMeta.put(part.partitionId(), part);
> > > > > >                      }                }
> > > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > --
> > > > > -- Guozhang
> > > > >
> > > >
> > >
> > >
> > >
> > > --
> > > -- Guozhang
> > >
> >
>
>
>
> --
> -- Guozhang
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message