flink-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Matthias J. Sax" <mj...@informatik.hu-berlin.de>
Subject Re: Package multiple jobs in a single jar
Date Fri, 22 May 2015 20:09:19 GMT
Makes sense to me. :)

One more thing: What about extending the "ProgramDescription" interface
to have multiple methods as Flavio suggested (with the config(...)
method that should be handle by the ParameterTool)

> public interface FlinkJob {
> 
> /** The name to display in the job submission UI or shell */
> //e.g. "My Flink HelloWorld"
> String getDisplayName();
> //e.g. "This program does this and that etc.."
> String getDescription();
> //e.g. <0,Integer,"An integer representing my first param">, <1,String,"An string
representing my second param">
> List<Tuple3<Integer, TypeInfo, String>> paramDescription;
> /** Set up the flink job in the passed ExecutionEnvironment */
> ExecutionEnvironment config(ExecutionEnvironment env);
> }

Right now, the interface is used only a couple of times in Flink's code
base, so it would not be a problem to update those classes. However, it
could break external code that uses the interface already (even if I
doubt that the interface is well known and used often [or at all]).

I personally don't think, that "getDiplayName()" to too helpful.
Splitting the program description and the parameter description seems to
be useful. For example, if wrong parameters are provided, the parameter
description can be included in the error message. If program+parameter
description is given in a single string, this is not possible. But this
is only a minor issue of course.

Maybe, we should also add the interface to the current Flink examples,
to make people more aware of it. Is there any documentation on the web site.


-Matthias



On 05/22/2015 09:43 PM, Robert Metzger wrote:
> Thank you for working on this.
> My responses are inline below:
> 
> (Flavio)
> 
>> My suggestion is to create a specific Flink interface to get also
>> description of a job and standardize parameter passing.
> 
> 
> I've recently merged the ParameterTool which is solving the "standardize
> parameter passing" problem (at least it presents a best practice) :
> http://ci.apache.org/projects/flink/flink-docs-master/apis/best_practices.html#parsing-command-line-arguments-and-passing-them-around-in-your-flink-application
> 
> Regarding the description: Maybe we can use the "ProgramDescription"
> interface for getting a string describing the program in the web frontend.
> 
> (Matthias)
> 
>> I don't want to start working on it, before it's clear that it has a
>> chance to be
>> included in Flink.
> 
> 
> I think the changes discussed here won't change the current behavior, but
> they add new functionality which
> can make the life of our users easier, so I'll vote to include your changes
> (given they meet our quality standards)
> 
> 
> If multiple classes implement "Program" interface an exception should be
>> through (I think that would make sense). However, I am not sure was
>> "good" behavior is, if a single "Program"-class is found and an
>> additional main-method class.
>>   - should "Program"-class be executed (ie, "overwrite" main-method class)
>>   - or, better to through an exception ?
> 
> 
> I would give a class implementing "Program" priority over a random main()
> method in a random class.
> Maybe printing a WARN log message informing the user that the "Program"
> class has been choosen.
> 
> 
> If no "Program"-class is found, but a single main-method class, Flink
>> could execute using main method. But I am not sure either, if this is
>> "good" behavior. If multiple main-method classes are present, throwing
>> and exception is the only way to got, I guess.
> 
> 
> I think the best effort approach "one class with main() found" is good. In
> case of multiple main methods, a helpful exception is the best approach in
> my opinion.
> 
> 
>  If the manifest contains "program-class" or "Main-Class" entry,
>> should we check the jar file right away if the specified class is there?
>> Right now, no check is performed and an error occurs if the user tries
>> to execute the job.
> 
> 
> I'd say the current approach is sufficient. There is no need to have a
> special code path which is doing the check.
> I think the error message will be pretty similar in both cases and I fear
> that this additional code could also introduce new bugs ;)
> 
> 
> 
> 
> On Fri, May 22, 2015 at 9:06 PM, Matthias J. Sax <
> mjsax@informatik.hu-berlin.de> wrote:
> 
>> Hi,
>>
>> two more thoughts to this discussion:
>>
>>  1) looking at the commit history of "CliFrontend", I found the
>> following closed issue and the closing pull request
>>     * https://issues.apache.org/jira/browse/FLINK-1095
>>     * https://github.com/apache/flink/pull/238
>> It stand in opposite of Flavio's request to have a job description. Any
>> comment on this? Should a removed feature be re-introduced? If not, I
>> would suggest to remove the "ProgramDescription" interface completely.
>>
>>  2) If the manifest contains "program-class" or "Main-Class" entry,
>> should we check the jar file right away if the specified class is there?
>> Right now, no check is performed and an error occurs if the user tries
>> to execute the job.
>>
>>
>> -Matthias
>>
>>
>> On 05/22/2015 12:06 PM, Matthias J. Sax wrote:
>>> Thanks for your feedback.
>>>
>>> I agree on the main method "problem". For scanning and listing all stuff
>>> that is found it's fine.
>>>
>>> The tricky question is the automatic invocation mechanism, if "-c" flag
>>> is not used, and no manifest program-class or Main-Class entry is found.
>>>
>>> If multiple classes implement "Program" interface an exception should be
>>> through (I think that would make sense). However, I am not sure was
>>> "good" behavior is, if a single "Program"-class is found and an
>>> additional main-method class.
>>>   - should "Program"-class be executed (ie, "overwrite" main-method
>> class)
>>>   - or, better to through an exception ?
>>>
>>> If no "Program"-class is found, but a single main-method class, Flink
>>> could execute using main method. But I am not sure either, if this is
>>> "good" behavior. If multiple main-method classes are present, throwing
>>> and exception is the only way to got, I guess.
>>>
>>> To sum up: Should Flink consider main-method classes for automatic
>>> invocation, or should it be required for main-method classes to either
>>> list them in "program-class" or "Main-Class" manifest parameter (to
>>> enable them for automatic invocation)?
>>>
>>>
>>> -Matthias
>>>
>>>
>>>
>>>
>>> On 05/22/2015 09:56 AM, Maximilian Michels wrote:
>>>> Hi Matthias,
>>>>
>>>> Thank you for taking the time to analyze Flink's invocation behavior. I
>>>> like your proposal. I'm not sure whether it is a good idea to scan the
>>>> entire JAR for main methods. Sometimes, main methods are added solely
>> for
>>>> testing purposes and don't really serve any practical use. However, if
>>>> you're already going through the JAR to find the ProgramDescription
>>>> interface, then you might look for main methods as well. As long as it
>> is
>>>> just a listing without execution, that should be fine.
>>>>
>>>> Best regards,
>>>> Max
>>>>
>>>> On Thu, May 21, 2015 at 3:43 PM, Matthias J. Sax <
>>>> mjsax@informatik.hu-berlin.de> wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> I had a look into the current Workflow of Flink with regard to the
>>>>> progressing steps of a jar file.
>>>>>
>>>>> If I got it right it works as follows (not sure if this is documented
>>>>> somewhere):
>>>>>
>>>>> 1) check, if "-c" flag is used to set program entry point
>>>>>    if yes, goto 4
>>>>> 2) try to extract "program-class" property from manifest
>>>>>    (if found goto 4)
>>>>> 3) try to extract "Main-Class" property from manifest
>>>>>    -> if not found through exception (this happens also, if no manifest
>>>>> file is found at all)
>>>>>
>>>>> 4) check if entry point class implements "Program" interface
>>>>>    if yes, goto 6
>>>>> 5) check if entry point class provided "public static void
>> main(String[]
>>>>> args)" method
>>>>>    -> if not, through exception
>>>>>
>>>>> 6) execute program (ie, show plan/info or really run it)
>>>>>
>>>>>
>>>>> I also "discovered" the interface "ProgramDescription" with a single
>>>>> method "String getDescription()". Even if some examples implement this
>>>>> interface (and use it in the example itself), Flink basically ignores
>>>>> it... From the CLI there is no way to get this info, and the WebUI does
>>>>> actually get it if present, however, doesn't show it anywhere...
>>>>>
>>>>>
>>>>> I think it would be nice, if we would extend the following functions:
>>>>>
>>>>>  - extend the possibility to specify multiple entry classes in
>>>>> "program-class" or "Main-Class" -> in this case, the user needs to
use
>>>>> "-c" flag to pick program to run every time
>>>>>
>>>>>  - add a CLI option that allows the user to see what entry point
>> classes
>>>>> are available
>>>>>    for this, consider
>>>>>      a) "program-class" entry
>>>>>      b) "Main-Class" entry
>>>>>      c) if neither is found, scan jar-file for classes implementing
>>>>> "Program" interface
>>>>>      d) if still not found, scan jar-file for classes with "main"
>> method
>>>>>
>>>>>  - if user looks for entry point classes via CLI, check for
>>>>> "ProgramDesciption" interface and show info
>>>>>
>>>>>  - extend WebUI to show all available entry-classes (pull request
>>>>> already there, for multiple entries in "program-class")
>>>>>
>>>>>  - extend WebUI to show "ProgramDescription" info
>>>>>
>>>>>
>>>>> What do you think? I am not too sure about the "auto scan" of the jar
>>>>> file if no manifest entry is provided. We might get some "fat jars" and
>>>>> scanning might take some time.
>>>>>
>>>>>
>>>>> -Matthias
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On 05/19/2015 10:44 AM, Stephan Ewen wrote:
>>>>>> We actually has an interface like that before ("Program"). It is
still
>>>>>> supported, but in all new programs we simply use the Java main method.
>>>>> The
>>>>>> advantage is that
>>>>>> most IDEs can create executable JARs automatically, setting the JAR
>>>>>> manifest attributes, etc.
>>>>>>
>>>>>> The "Program" interface still works, though. Most tool classes (like
>>>>>> "PackagedProgram") have a way to figure out whether the code uses
>>>>> "main()"
>>>>>> or implements "Program"
>>>>>> and calls the right method.
>>>>>>
>>>>>> You can try and extend the program interface. If you want to
>> consistently
>>>>>> support multiple programs in one JAR file, you may need to adjust
the
>>>>> util
>>>>>> classes as
>>>>>> well to deal with that.
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Tue, May 19, 2015 at 10:10 AM, Matthias J. Sax <
>>>>>> mjsax@informatik.hu-berlin.de> wrote:
>>>>>>
>>>>>>> Supporting an interface like this seems to be a nice idea. Any
other
>>>>>>> opinions on it?
>>>>>>>
>>>>>>> It seems to be some more work to get it done right. I don't want
to
>>>>>>> start working on it, before it's clear that it has a chance to
be
>>>>>>> included in Flink.
>>>>>>>
>>>>>>> @Flavio: I moved the discussion to dev mailing list (user list
is not
>>>>>>> appropriate for this discussion). Are you subscribed to it or
should
>> I
>>>>>>> cc you in each mail?
>>>>>>>
>>>>>>>
>>>>>>> -Matthias
>>>>>>>
>>>>>>>
>>>>>>> On 05/19/2015 09:39 AM, Flavio Pompermaier wrote:
>>>>>>>> Nice feature Matthias!
>>>>>>>> My suggestion is to create a specific Flink interface to
get also
>>>>>>>> description of a job and standardize parameter passing.
>>>>>>>> Then, somewhere (e.g. Manifest) you could specify the list
of
>> packages
>>>>>>> (or
>>>>>>>> also directly the classes) to inspect with reflection to
extract the
>>>>> list
>>>>>>>> of available Flink jobs.
>>>>>>>> Something like:
>>>>>>>>
>>>>>>>> public interface FlinkJob {
>>>>>>>>
>>>>>>>> /** The name to display in the job submission UI or shell
*/
>>>>>>>> //e.g. "My Flink HelloWorld"
>>>>>>>> String getDisplayName();
>>>>>>>>  //e.g. "This program does this and that etc.."
>>>>>>>> String getDescription();
>>>>>>>>  //e.g. <0,Integer,"An integer representing my first param">,
>>>>>>> <1,String,"An
>>>>>>>> string representing my second param">
>>>>>>>> List<Tuple3<Integer, TypeInfo, String>> paramDescription;
>>>>>>>>  /** Set up the flink job in the passed ExecutionEnvironment
*/
>>>>>>>> ExecutionEnvironment config(ExecutionEnvironment env);
>>>>>>>> }
>>>>>>>>
>>>>>>>> What do you think?
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On Sun, May 17, 2015 at 10:38 PM, Matthias J. Sax <
>>>>>>>> mjsax@informatik.hu-berlin.de> wrote:
>>>>>>>>
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> I like the idea that Flink's WebClient can show different
plans for
>>>>>>>>> different jobs within a single jar file.
>>>>>>>>>
>>>>>>>>> I prepared a prototype for this feature. You can find
it here:
>>>>>>>>> https://github.com/mjsax/flink/tree/multipleJobsWebUI
>>>>>>>>>
>>>>>>>>> To test the feature, you need to prepare a jar file,
that contains
>> the
>>>>>>>>> code of multiple programs and specify each entry class
in the
>> manifest
>>>>>>>>> file as comma separated values in "program-class" line.
>>>>>>>>>
>>>>>>>>> Feedback is welcome. :)
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> -Matthias
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On 05/08/2015 03:08 PM, Flavio Pompermaier wrote:
>>>>>>>>>> Thank you all for the support!
>>>>>>>>>> It will be a really nice feature if the web client
could be able
>> to
>>>>>>> show
>>>>>>>>>> me the list of Flink jobs within my jar..
>>>>>>>>>> it should be sufficient to mark them with a special
annotation and
>>>>>>>>>> inspect the classes within the jar..
>>>>>>>>>>
>>>>>>>>>> On Fri, May 8, 2015 at 3:03 PM, Malte Schwarzer <ms@mieo.de
>>>>>>>>>> <mailto:ms@mieo.de>> wrote:
>>>>>>>>>>
>>>>>>>>>>     Hi Flavio,
>>>>>>>>>>
>>>>>>>>>>     you also can put each job in a single class and
use the –c
>>>>>>> parameter
>>>>>>>>>>     to execute jobs separately:
>>>>>>>>>>
>>>>>>>>>>     /bin/flink run –c com.myflinkjobs.JobA
>>>>>>> /path/to/jar/multiplejobs.jar
>>>>>>>>>>     /bin/flink run –c com.myflinkjobs.JobB
>>>>>>> /path/to/jar/multiplejobs.jar
>>>>>>>>>>     …
>>>>>>>>>>
>>>>>>>>>>     Cheers
>>>>>>>>>>     Malte
>>>>>>>>>>
>>>>>>>>>>     Von: Robert Metzger <rmetzger@apache.org <mailto:
>>>>>>> rmetzger@apache.org
>>>>>>>>>>>
>>>>>>>>>>     Antworten an: <user@flink.apache.org <mailto:
>>>>> user@flink.apache.org
>>>>>>>>>
>>>>>>>>>>     Datum: Freitag, 8. Mai 2015 14:57
>>>>>>>>>>     An: "user@flink.apache.org <mailto:user@flink.apache.org>"
>>>>>>>>>>     <user@flink.apache.org <mailto:user@flink.apache.org>>
>>>>>>>>>>     Betreff: Re: Package multiple jobs in a single
jar
>>>>>>>>>>
>>>>>>>>>>     Hi Flavio,
>>>>>>>>>>
>>>>>>>>>>     the pom from our quickstart is a good
>>>>>>>>>>     reference:
>>>>>>>>>
>>>>>>>
>>>>>
>> https://github.com/apache/flink/blob/master/flink-quickstart/flink-quickstart-java/src/main/resources/archetype-resources/pom.xml
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>     On Fri, May 8, 2015 at 2:53 PM, Flavio Pompermaier
>>>>>>>>>>     <pompermaier@okkam.it <mailto:pompermaier@okkam.it>>
wrote:
>>>>>>>>>>
>>>>>>>>>>         Ok, get it.
>>>>>>>>>>         And is there a reference pom.xml for shading
my
>> application
>>>>>>> into
>>>>>>>>>>         one fat-jar? which flink dependencies can
I exclude?
>>>>>>>>>>
>>>>>>>>>>         On Fri, May 8, 2015 at 1:05 PM, Fabian Hueske
<
>>>>>>> fhueske@gmail.com
>>>>>>>>>>         <mailto:fhueske@gmail.com>> wrote:
>>>>>>>>>>
>>>>>>>>>>             I didn't say that the main should return
the
>>>>>>>>>>             ExecutionEnvironment.
>>>>>>>>>>             You can define and execute as many programs
in a main
>>>>>>>>>>             function as you like.
>>>>>>>>>>             The program can be defined somewhere
else, e.g., in a
>>>>>>>>>>             function that receives an ExecutionEnvironment
and
>>>>> attaches
>>>>>>>>>>             a program such as
>>>>>>>>>>
>>>>>>>>>>             public void buildMyProgram(ExecutionEnvironment
env) {
>>>>>>>>>>               DataSet<String> lines = env.readTextFile(...);
>>>>>>>>>>               // do something
>>>>>>>>>>               lines.writeAsText(...);
>>>>>>>>>>             }
>>>>>>>>>>
>>>>>>>>>>             That method could be invoked from main():
>>>>>>>>>>
>>>>>>>>>>             psv main() {
>>>>>>>>>>               ExecutionEnv env = ...
>>>>>>>>>>
>>>>>>>>>>               if(...) {
>>>>>>>>>>                 buildMyProgram(env);
>>>>>>>>>>               }
>>>>>>>>>>               else {
>>>>>>>>>>                 buildSomeOtherProg(env);
>>>>>>>>>>               }
>>>>>>>>>>
>>>>>>>>>>               env.execute();
>>>>>>>>>>
>>>>>>>>>>               // run some more programs
>>>>>>>>>>             }
>>>>>>>>>>
>>>>>>>>>>             2015-05-08 12:56 GMT+02:00 Flavio Pompermaier
>>>>>>>>>>             <pompermaier@okkam.it <mailto:pompermaier@okkam.it>>:
>>>>>>>>>>
>>>>>>>>>>                 Hi Fabian,
>>>>>>>>>>                 thanks for the response.
>>>>>>>>>>                 So my mains should be converted in
a method
>> returning
>>>>>>>>>>                 the ExecutionEnvironment.
>>>>>>>>>>                 However it think that it will be
very nice to
>> have a
>>>>>>>>>>                 syntax like the one of the Hadoop
ProgramDriver to
>>>>>>>>>>                 define jobs to invoke from a single
root class.
>>>>>>>>>>                 Do you think it could be useful?
>>>>>>>>>>
>>>>>>>>>>                 On Fri, May 8, 2015 at 12:42 PM,
Fabian Hueske
>>>>>>>>>>                 <fhueske@gmail.com <mailto:fhueske@gmail.com>>
>>>>> wrote:
>>>>>>>>>>
>>>>>>>>>>                     You easily have multiple Flink
programs in a
>>>>> single
>>>>>>>>>>                     JAR file.
>>>>>>>>>>                     A program is defined using an
>>>>> ExecutionEnvironment
>>>>>>>>>>                     and executed when you call
>>>>>>>>>>                     ExecutionEnvironment.exeucte().
>>>>>>>>>>                     Where and how you do that does
not matter.
>>>>>>>>>>
>>>>>>>>>>                     You can for example implement
a main function
>>>>> such
>>>>>>>>> as:
>>>>>>>>>>
>>>>>>>>>>                     public static void main(String...
args) {
>>>>>>>>>>
>>>>>>>>>>                       if (today == Monday) {
>>>>>>>>>>                         ExecutionEnvironment env
= ...
>>>>>>>>>>                         // define Monday prog
>>>>>>>>>>                         env.execute()
>>>>>>>>>>                       }
>>>>>>>>>>                       else {
>>>>>>>>>>                         ExecutionEnvironment env
= ...
>>>>>>>>>>                         // define other prog
>>>>>>>>>>                         env.execute()
>>>>>>>>>>                       }
>>>>>>>>>>                     }
>>>>>>>>>>
>>>>>>>>>>                     2015-05-08 11:41 GMT+02:00 Flavio
Pompermaier
>>>>>>>>>>                     <pompermaier@okkam.it <mailto:
>>>>> pompermaier@okkam.it
>>>>>>>>>>> :
>>>>>>>>>>
>>>>>>>>>>                         Hi to all,
>>>>>>>>>>                         is there any way to keep
multiple jobs in
>> a
>>>>> jar
>>>>>>>>>>                         and then choose at runtime
the one to
>> execute
>>>>>>>>>>                         (like what ProgramDriver
does in Hadoop)?
>>>>>>>>>>
>>>>>>>>>>                         Best,
>>>>>>>>>>                         Flavio
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>
>>>
>>
>>
> 


Mime
View raw message