Ok got it. How are partitions determined? Is this something that
producer is responsible for or can it be automatically handled by the
broker?
On 11/6/11 11:13 AM, Neha Narkhede wrote:
Ok so the partitioning is done on the hadoop side during importing and has
nothing to do with Kafka partitions.
That's right.
Kafka partitions help scale consumption by allowing multiple consumer
processes pull data for a topic in parallel. The parallelism factor is
limited by the total number of Kafka partitions. For example, if a
topic has 2 partitions, 2 Hadoop mappers can pull data for the entire
topic in parallel. If another topic has 8 partitions, the parallelism
factor increased by 4x. Now 8 mappers can pull all the data for this
topic at the same time.
Thanks,
Neha
On Sun, Nov 6, 2011 at 11:00 AM, Mark<static.void....@gmail.com> wrote:
Ok so the partitioning is done on the hadoop side during importing and has
nothing to do with Kafka partitions. Would you mind explaining what Kafka
partitions are used for and when one should use them?
On 11/6/11 10:52 AM, Neha Narkhede wrote:
We use Avro serialization for the message data and use Avro schemas to
convert event objects into Kafka message payload on the producers. On
the Hadoop side, we use Avro schemas to deserialize Kafka message
payload back into an event object. Each such event object has a
timestamp field that the Hadoop job uses to put the message into its
hourly and daily partition. So if the Hadoop job runs every 15 mins,
it will run 4 times to collect data into the current hour's partition.
Very soon, the plan is to open-source this Avro-Hadoop pipeline as well.
Thanks,
Neha
On Sun, Nov 6, 2011 at 10:37 AM, Mark<static.void....@gmail.com> wrote:
"At LinkedIn we use the InputFormat provided in contrib/hadoop-consumer
to
load the data for
topics in daily and hourly partitions."
Sorry for my ignorance but what exactly do you mean by loading the data
in
daily and hourly partitions?
On 11/6/11 10:26 AM, Neha Narkhede wrote:
There should be no changes to the way you create topics to achieve
this kind of HDFS data load for Kafka. At LinkedIn we use the
InputFormat provided in contrib/hadoop-consumer to load the data for
topics in daily and hourly partitions. These Hadoop jobs run every 10
mins or so. So the maximum delay of data being available from
producer->Hadoop is around 10 mins.
Thanks,
Neha
On Sun, Nov 6, 2011 at 8:45 AM, Mark<static.void....@gmail.com>
wrote:
This is more of a general design question but what is the preferred way
of
importing logs from Kafka to HDFS when you want your data segmented by
hour
or day? Is there anyway to say "Import only this {hour|day} of logs" or
does
one need to create their topics around the way they would like to
import
them.. ie Topic: "search_logs/2011/11/06". If its the latter, is there
any
documentation/best practices on topic/key design?
Thanks
