flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Robert Metzger <rmetz...@apache.org>
Subject Re: Flink questions
Date Wed, 11 Mar 2015 17:34:48 GMT
So regarding your last question, the internal data format:

Flink is internally (for sending data from one operator to another) using
its own data serialization framework, which is very efficient for
Tuples/Case classes etc.
So you as a user do not need to care about that much. If other components
of your infrastructure (application servers for example) rely on generated
classes (for example created by protobuf or Avro), you can use these
classes with Flink as well.

If you are planning to write some of the data you're streaming to a
persistent storage (for example HDFS or a local file system), then I would
recommend a format like Avro.

Regarding the incoming data: JSON is indeed a bit space inefficient when
represented as a string. If you are able to change the format of the
incoming data, you can also write your own data source (implementing the
SourceFunction interface)




On Wed, Mar 11, 2015 at 4:37 PM, Márton Balassi <mbalassi@apache.org> wrote:

> Dear Emmanuel,
>
> I'm Marton, one of the Flink Streaming developers - Robert forwarded your
> issue to me. Thanks for trying out our project.
>
> 1) Debugging: TaskManager logs are currently not forwarded to the UI, but
> you can find them on the taskmanager machines in the log folder of your
> Flink distribution. We have this issue on our agenda in the very near
> future - they need to be accessible from the UI.
>
> 2) Output to socket: Currently we do not have a preimplemented sink for
> sockets (although we offer a socket source and sinks writing to Apache
> Kafka, Flume and RabbitMQ). You can easily implement a socket sink by
> extending the abstract RichSinkFunction class though. [1]
>
> For using that you can simply say dataStream.addSink(MySinkFunction()) -
> in that you can bring up a socket or any other service. You would create a
> socket in the open function and then in the invoke method you would write
> every value out to it.
>
> I do agree that this is a nice tool to have so I have opened a JIRA ticket
> for it. [2]
>
> 3) Internal data format: Robert was kind enough to offer a more detailed
> answer on this issue. In general streaming sinks support any file output
> that is supported by batch Flink including Avro. You can use this
> functionality by dataStream.addSink(new FileSinkFunction<>(OutputFormat)).
>
> [1]
> http://ci.apache.org/projects/flink/flink-docs-master/streaming_guide.html#connecting-to-the-outside-world
> [2] https://issues.apache.org/jira/browse/FLINK-1688
>
> Best,
>
> Marton
>
> *From:* Emmanuel <eleroy@msn.com>
> *Date:* 11. März 2015 14:59:31 MEZ
> *To:* Robert Metzger <rmetzger@apache.org>, Henry Saputra <
> hsaputra@apache.org>
> *Subject:* *Flink questions*
>
> Hello,
>
>
>
> Thanks again for the help yesterday: the simple things go a long way to
> get me moving...
>
> I have more questions i hope I can get your opinion and input about:
>
> *Debugging:*
> What's the preferred or recommended way to proceed?
> I have been using some System.out.println() statements in my simple test
> code, and the results are confusing:
> First, in the UI, the logs are for the jobmanager.out, but there is never
> anything there; wherever i see output in a log it's on the taskmanager.out
> file
> Also, even more confusing is the fact that often times I just get no log
> at all... the UI says the topology is running, but nothing get printed
> out...
> Is there a process you'd recommend to follow to debug properly with logs?
>
> *Output to socket*
> Ideally I'd like to print out to a socket/stream and read from another
> machine so as not to choke the node with disk I/Os when testing
> performances. Not sure how to do that.
>
> *Internal Data format*
> Finally, a practical question about data format: we ingest JSON, which is
> not convenient, and uses a lot of space. Internally Java/Scala prefers
> Tuples, and we were thinking of using ProtoBuffs.
> There is also Avro that could do this as I understand it... What would be
> the recommended way to format data internally?
>
> Thanks for your input.
>
> Cheers
> Emmanuel
>
>
>
>

Mime
View raw message