Re: Handling custom types with Kryo

2015-01-20 Thread Arvid Heise
An alternative way that may not be applicable in your case: For Sopremo, all types implemented a common interface. When a package is loaded, the Sopremo package manager scans the jar and looks for classes implementing the interfaces (quite fast, because not the entire class must be loaded). All ty

Re: Drafting a roadmap for Flink

2015-01-20 Thread Max Michels
Hi Kostas, Thanks for putting this into the wiki. I added the JIRA link for the off-heap memory. Now the wiki displays: > Error rendering macro 'jira' : > com.atlassian.confluence.macro.MacroExecutionException: > java.lang.RuntimeException: Not Found If persistent, we should report this. Is t

Re: Handling custom types with Kryo

2015-01-20 Thread Robert Metzger
@Stephan: Yes, you are summarizing it correctly. I'll assign FLINK-1417 to myself and implement it as discussed here (once I have resolved the other issues assigned to me) There is one additional point we forgot in the discussion so far: We are initializing Kryo with twitter/chill's "ScalaKryoInst

Re: Handling custom types with Kryo

2015-01-20 Thread Till Rohrmann
I like the idea to automatically figure out which types are used by a program and to register them at Kryo. Thus, +1 for this idea. On Tue, Jan 20, 2015 at 11:34 AM, Robert Metzger wrote: > @Stephan: Yes, you are summarizing it correctly. > I'll assign FLINK-1417 to myself and implement it as di

Re: Handling custom types with Kryo

2015-01-20 Thread Alexander Alexandrov
+1 for program analysis from me too... Should be doable also on a lower level (e.g. analysis of compiled *.class files) with some off-the-shelf libraries, right? 2015-01-20 11:39 GMT+01:00 Till Rohrmann : > I like the idea to automatically figure out which types are used by a > program and to re

Re: Handling custom types with Kryo

2015-01-20 Thread Ufuk Celebi
On 20 Jan 2015, at 11:51, Alexander Alexandrov wrote: > +1 for program analysis from me too... > > Should be doable also on a lower level (e.g. analysis of compiled *.class > files) with some off-the-shelf libraries, right? Yes. There was a prototypical implementation using byte code analysis

Re: Handling custom types with Kryo

2015-01-20 Thread Timo Walther
Are we talking about the types for the input/output of operators or also types that are used inside UDFs? Operator I/O type classes are known, so we don't need static code analysis for that. For types inside UDFs I can add that requirement to FLINK-1319. On 20.01.2015 11:51, Alexander Alexand

Re: Handling custom types with Kryo

2015-01-20 Thread Alexander Alexandrov
I think we are talking about the Operator I/O types, as types used internally only by the UDFs should not be serialized. 2015-01-20 12:27 GMT+01:00 Timo Walther : > Are we talking about the types for the input/output of operators or also > types that are used inside UDFs? > Operator I/O type clas

Re: Handling custom types with Kryo

2015-01-20 Thread Aljoscha Krettek
Yes, that sounds very reasonable. On Jan 20, 2015 6:40 AM, "Stephan Ewen" wrote: > Yes, I agree that the Avro serializer should be available by default. That > is one case of a typical type that should work out of the box, given that > we support Avro file formats. > > Let me summarize how I unde

Representing Scala base types in the Flink RT

2015-01-20 Thread Alexander Alexandrov
Hi there, I cannot figure out how the Scala base types (e.g. scala.Int, scala.Double, etc.) are mapped to the Flink runtime. It seems that there are not treated the same as their Java counterparts (e.g. java.lang.Integer, java.lang.Double). For example, if I write the following code: val inputFo

Re: Representing Scala base types in the Flink RT

2015-01-20 Thread Alexander Alexandrov
Just to clarify in order to spare us some time in the discussion. I *deliberately* want to use Flink Java API from Scala with Scala core types. 2015-01-20 18:53 GMT+01:00 Alexander Alexandrov < alexander.s.alexand...@gmail.com>: > Hi there, > > I cannot figure out how the Scala base types (e.g. s

Keeping around temp datasets

2015-01-20 Thread Alexander Alexandrov
Hi there, I have to implement some generic fallback strategy on top of a more abstract DSL in order to keep datasets in a temp space (e.g. Tachyon). My implementation is based on the 0.8 release. At the moment I am undecided between three options: - BinaryInputFormat / BinaryOutputFormat -

[jira] [Created] (FLINK-1422) Missing usage example for "withParameters"

2015-01-20 Thread Alexander Alexandrov (JIRA)
Alexander Alexandrov created FLINK-1422: --- Summary: Missing usage example for "withParameters" Key: FLINK-1422 URL: https://issues.apache.org/jira/browse/FLINK-1422 Project: Flink Issue

Implementing a list accumulator

2015-01-20 Thread Max Michels
Hi everyone, I'm running into some problems implementing a Accumulator for returning a list of a DataSet. https://github.com/mxm/flink/tree/count/collect Basically, it works fine in this test case: ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment(); Integer[] in

Re: Keeping around temp datasets

2015-01-20 Thread Robert Metzger
I would not recommend using the AvroInput/Output format because its meant to be used with Avro types (usually POJOs generated from an Avro schema). I would use the TypeSerializerInputFormat / OutputFormat. Then you can be sure that its able to read/write all types supported by our system. On Tue,

Master not building and how to notice it faster in the future

2015-01-20 Thread Robert Metzger
Hi, it seems that our master is currently not building. See: https://travis-ci.org/apache/flink/jobs/47689754 We need to come up with a good solution to notify dev@flink when builds on Travis are failing. We also had unstable builds recently due to too short akka timeouts and it took some time t

Re: Master not building and how to notice it faster in the future

2015-01-20 Thread Ufuk Celebi
On 20 Jan 2015, at 23:57, Robert Metzger wrote: > There are certainly ways to fix this. Right now, the best approach is > probably setting up a REST 2 e-mail service somewhere which is mailing to > our dev@ list ( > http://docs.travis-ci.com/user/notifications/#Webhook-notification). That sound

Re: Master not building and how to notice it faster in the future

2015-01-20 Thread Robert Metzger
I think its just a missing import. Maybe we can use Google AppEngine for that. It seems that their free offering is sufficient for our purpose: https://cloud.google.com/pricing/#app-engine. It also allows sending emails. I guess its hard to get the token for the "apache" user. Maybe there is is a

Re: Master not building and how to notice it faster in the future

2015-01-20 Thread Ufuk Celebi
On 21 Jan 2015, at 00:19, Robert Metzger wrote: > I think its just a missing import. Yes. > Maybe we can use Google AppEngine for that. It seems that their free > offering is sufficient for our purpose: > https://cloud.google.com/pricing/#app-engine. It also allows sending emails. > I guess it

[jira] [Created] (FLINK-1423) No Tags for new release on the github repo

2015-01-20 Thread Carsten Brandt (JIRA)
Carsten Brandt created FLINK-1423: - Summary: No Tags for new release on the github repo Key: FLINK-1423 URL: https://issues.apache.org/jira/browse/FLINK-1423 Project: Flink Issue Type: Task

[jira] [Created] (FLINK-1424) bin/flink run does not recognize -c parameter anymore

2015-01-20 Thread Carsten Brandt (JIRA)
Carsten Brandt created FLINK-1424: - Summary: bin/flink run does not recognize -c parameter anymore Key: FLINK-1424 URL: https://issues.apache.org/jira/browse/FLINK-1424 Project: Flink Issue T