Thanks a lot for the comments, Till and Fabian.
The RemoteEnvrionment does provide a way to specify jar files at
construction, but we want the jar files to be specified dynamically in the
user code, e.g. in a DDL statement, and the jar files might be in a remote
DFS. As we discussed, I think there are 2 approaches:
1) add new interface env.registerJarFile(jarFiles...), which ships the JAR
files using JobGraph.addJar(). In this case, all jars will be loaded by
default at runtime. This approach will be the same as how SQL client ship
UDF jars now.
2) add new interface env.registerJarFile(name, jarFiles...). It will do
similar things as env.registerCachedFile(), which will register a set of
Jar files with a key name, and we can add a new interface in
RuntimeContext as Fabian suggests, i.e.,
RuntimeContext.getClassloaderWithJar(<key name>). Now user will be able to
load the functions in remote jar dynamically using the returned ClassLoader.
Comparing the 2 approaches:
- Approach 1) will be simpler for user to use.
- Approach 2) will allow us to use different versions of a class in the
same code, and might solve some dependency conflict issues. Also in 2), we
can load Jars on demand, while in 1) all jars will be loaded by default.
I think we can support both interfaces. On the SQL DDL implementation, both
will work and approach 2) will be more complicated, but with some nice
benefit as stated above. However, the implementation choice should be
transparent to the end user. Also, I am wondering outside of the SQL DDL,
will these new functionality/interface be helpful in other scenarios?
Maybe, that will help make the interface better and more generic. Thanks a
lot.
Shuyi
On Tue, Jun 5, 2018 at 1:47 AM Fabian Hueske <fhue...@gmail.com> wrote:
> We could also offer a feature that users can request classloaders with
> additional jars.
> This could work as follows:
>
> 1) Users register jar files in the ExecutionEnvironment (similar to cached
> files) with a name, e.g., env.registerJarFile("~/myJar.jar", "myName");
> 2) In a function, the user can request a user classloader with the
> additional classes, e.g., RuntimeContext.getClassloaderWithJar("myName");
> This could also support to load multiple jar files in the same classloader.
>
> IMO, the interesting part of Shuyi's proposal is to be able to dynamically
> load code from remote locations without fetching it to the client first.
>
> Best, Fabian
>
>
> 2018-05-29 12:42 GMT+02:00 Till Rohrmann <trohrm...@apache.org>:
>
> > I see Shuyi's point that it would nice to allow adding jar files which
> > should be part of the user code classloader programmatically. Actually,
> we
> > expose this functionality in the `RemoteEnvironment` where you can
> specify
> > additional jars which shall be shipped to the cluster in the
> constructor. I
> > assume that is exactly the functionality you are looking for. In that
> > sense, it might be an API inconsistency that we allow it for some cases
> and
> > for others not.
> >
> > But I could also see that the whole functionality of dynamically loading
> > jars at runtime could also perfectly live in the `UdfSqlOperator`. This,
> of
> > course, would entail that one has to take care of clean up of the
> > downloaded resources. But it should be possible to first download the
> > resources and create a custom URLClassLoader at startup and then use this
> > class loader when calling into the UDF.
> >
> > Cheers,
> > Till
> >
> > On Wed, May 16, 2018 at 9:28 PM, Shuyi Chen <suez1...@gmail.com> wrote:
> >
> > > Hi Aljoscha, Fabian, Rong, Ted and Timo,
> > >
> > > Thanks a lot for the feedback. Let me clarify the usage scenario in a
> bit
> > > more detail. The context is that we want to add support for SQL DDL to
> > load
> > > UDF from external JARs located either in local filesystem or HDFS or a
> > HTTP
> > > endpoint in Flink SQL. The local FS option is more for debugging
> purpose
> > > for user to submit the job jar locally, and the later 2 are for
> > production
> > > uses. Below is an example User application with the *CREATE FUNCTION*
> DDL
> > > (Note: grammar and interface not finalized yet).
> > >
> > > ------------------------------------------------------------
> > > -------------------------------------
> > >
> > >
> > >
> > >
> > > *val env = StreamExecutionEnvironment.getExecutionEnvironmentval tEnv =
> > > TableEnvironment.getTableEnvironment(env)// setup the
> > DataStream//......*
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > > *// register the DataStream under the name
> > > "OrderAtEnv.registerDataStream("OrderA", orderA, 'user, 'product,
> > > 'amount)tEnv.sqlDDL( "create function helloFunc as
> > > 'com.example.udf.HelloWorld' using jars
> > > ('hdfs:///users/david/libraries/my-udf-1.0.1-SNAPSHOT.jar')")val result
> > =
> > > tEnv.sqlQuery( "SELECT user, helloFunc(product), amount FROM OrderA
> > WHERE
> > > amount > 2")result.toAppendStream[Order].print()env.execute()*
> > > ------------------------------------------------------------
> > > -------------------------------------
> > >
> > > The example application above does the following:
> > > 1) it registers a DataStream as a Calcite table(
> > > *org.apache.calcite.schema.Table*) under name "OrderA", so SQL can
> > > reference the DataStream as table "OrderA".
> > > 2) it uses the SQL *CREATE FUNCTION* DDL (grammar and interface not
> > > finalized yet) to create a SQL UDF called *helloFunc* from a JAR
> located
> > in
> > > a remote HDFS path.
> > > 3) it issues a sql query that uses the *helloFunc* UDF defined above
> and
> > > generate a Flink table (*org.apache.flink.table.api.Table*)
> > > 4) it convert the Flink table back to a DataStream and print it.
> > >
> > > Step 1), 3), and 4) are already implemented. To implement 2), we need
> to
> > do
> > > the following to implement the *tEnv.sqlDDL()* function.
> > >
> > > a) parse the DDL into a SqlNode to extract the UDF *udfClasspath*, UDF
> > > remote path *udfUrls[]* and UDF SQL name *udfName*.
> > > b) use the URLClassLoader to load the JARs specified in *udfUrls[]*,
> and
> > > register the SQL UDF using the {Batch/Stream/}TableEnvironment
> > > registerFunction methods using* udfClasspath* under name *udfName.*
> > > c) register the JARs *udfUrls[]* through the
> > {Stream}ExecutionEnvironment,
> > > so that the JARs can be distributed to all the TaskManagers during
> > runtime.
> > >
> > >
> > > Since the CREATE FUNCTION DDL is executed within the user application,
> I
> > > dont think we have access to the ClusterClient at the point when
> > > *tEnv.sqlDDL()* is executed. Also the JARs can be in a remote
> filesystem
> > > (which is the main usage scenarios), so the user can't really prepare
> the
> > > jar somehow in advance statically.
> > >
> > > For normal user application, I think {Stream}ExecutionEnvironment is
> the
> > > right place for the functionality, since it provides methods to control
> > the
> > > job execution and to interact with the outside world, and also, it
> > actually
> > > already does similar things provided through the *registerCachedFile*
> > > interface.
> > >
> > > However, in such case, SQL FUNCTION DDL and SQL client will use 2
> > different
> > > routes to register UDF jars, one through *JobGraph.jobConfiguration*
> and
> > > the other through *JobGraph.userJars*. So *maybe we can, as Fabian
> > > suggests, add **registerUserJarFile()/getUserJarFiles() interfaces
> > > in {Stream}ExecutionEnvironment, which stores the jars internally in a
> > > List, and when generating JobGraph, copy the jars to the JobGraph using
> > > the {Stream}ExecutionEnvironment.getUserJarFiles() and
> > > JobGraph.addJar()* (Note,
> > > streaming and batch implementations might vary). In such case, both SQL
> > > FUNCTION DDL and SQL client will use *JobGraph.userJars* to ship the
> UDF
> > > jars.
> > >
> > > Hope that clarifies better. What do you guys think? Thanks a lot.
> > >
> > > Cheers!
> > > Shuyi
> > >
> > > On Wed, May 16, 2018 at 9:45 AM, Rong Rong <walter...@gmail.com>
> wrote:
> > >
> > > > I think the question here is whether registering Jar files (or other
> > > > executable files) during job submission is sufficient for @shuyi's
> use
> > > > case.
> > > >
> > > > If I understand correctly regarding the part of dynamic distribution
> of
> > > the
> > > > external libraries in runtime. This is used to deal with DDL/DSL such
> > as:
> > > > CREATE FUNCTION my_fun FROM url://<some_remote_jar>
> > > > during execution. Correct me if I am wrong @shuyi, The basic
> assumption
> > > > that "we can locate and ship all executable JARs during job
> submission"
> > > no
> > > > longer holds for your use case right?
> > > >
> > > > I guess we are missing details here regarding the "distribution of
> > > external
> > > > libraries in runtime" part. Maybe you can share more example of this
> > use
> > > > case. Would this be included in the design doc @Timo?
> > > >
> > > > --
> > > > Rong
> > > >
> > > > On Wed, May 16, 2018 at 5:41 AM, Timo Walther <twal...@apache.org>
> > > wrote:
> > > >
> > > > > Yes, we are using the addJar functionionality of the JobGraph as
> well
> > > for
> > > > > the SQL Client.
> > > > >
> > > > > I think the execution environment is not the right place to specify
> > > jars.
> > > > > The location of the jars depends on the submission method. If a
> local
> > > > path
> > > > > is specified in the main() method of a packaged Flink jar, it would
> > not
> > > > > work when such a program is submitted through the REST API.
> > > > >
> > > > > Regards,
> > > > > Timo
> > > > >
> > > > > Am 16.05.18 um 14:32 schrieb Aljoscha Krettek:
> > > > >
> > > > > I think this functionality is already there, we just have to expose
> > it
> > > in
> > > > >> the right places: ClusterClient.submitJob() takes a JobGraph,
> > JobGraph
> > > > has
> > > > >> method addJar() for adding jars that need to be in the classloader
> > for
> > > > >> executing a user program.
> > > > >>
> > > > >> On 16. May 2018, at 12:34, Fabian Hueske <fhue...@gmail.com>
> wrote:
> > > > >>>
> > > > >>> Hi Ted,
> > > > >>>
> > > > >>> The design doc is in late draft status and proposes support for
> SQL
> > > DDL
> > > > >>> statements (CREATE TABLE, CREATE FUNCTION, etc.).
> > > > >>> The question about registering JARs came up because we need a way
> > to
> > > > >>> distribute JAR files that contain the code of user-defined
> > functions.
> > > > >>>
> > > > >>> The design doc will soon be shared on the dev mailing list to
> > gather
> > > > >>> feedback from the community.
> > > > >>>
> > > > >>> Best, Fabian
> > > > >>>
> > > > >>> 2018-05-16 10:45 GMT+02:00 Ted Yu <yuzhih...@gmail.com>:
> > > > >>>
> > > > >>> bq. In a design document, Timo mentioned that we can ship
> multiple
> > > JAR
> > > > >>>> files
> > > > >>>>
> > > > >>>> Mind telling us where the design doc can be retrieved ?
> > > > >>>>
> > > > >>>> Thanks
> > > > >>>>
> > > > >>>> On Wed, May 16, 2018 at 1:29 AM, Fabian Hueske <
> fhue...@gmail.com
> > >
> > > > >>>> wrote:
> > > > >>>>
> > > > >>>> Hi,
> > > > >>>>>
> > > > >>>>> I'm not sure if we need to modify the existing method.
> > > > >>>>> What we need is a bit different from what registerCachedFile()
> > > > >>>>> provides.
> > > > >>>>> The method ensures that a file is copied to each TaskManager
> and
> > > can
> > > > be
> > > > >>>>> locally accessed from a function's RuntimeContext.
> > > > >>>>> In our case, we don't need to access the file but would like to
> > > make
> > > > >>>>> sure
> > > > >>>>> that it is loaded into the class loader.
> > > > >>>>> So, we could also just add a method like registerUserJarFile().
> > > > >>>>>
> > > > >>>>> In a design document, Timo mentioned that we can ship multiple
> > JAR
> > > > >>>>> files
> > > > >>>>> with a job.
> > > > >>>>> So, we could also implement the UDF shipping logic by loading
> the
> > > Jar
> > > > >>>>> file(s) to the client and distribute them from there.
> > > > >>>>> In that case, we would not need to add new method to the
> > execution
> > > > >>>>> environment.
> > > > >>>>>
> > > > >>>>> Best,
> > > > >>>>> Fabian
> > > > >>>>>
> > > > >>>>> 2018-05-15 3:50 GMT+02:00 Rong Rong <walter...@gmail.com>:
> > > > >>>>>
> > > > >>>>> +1. This could be very useful for "dynamic" UDF.
> > > > >>>>>>
> > > > >>>>>> Just to clarify, if I understand correctly, we are tying to
> use
> > an
> > > > >>>>>> ENUM
> > > > >>>>>> indicator to
> > > > >>>>>> (1) Replace the current Boolean isExecutable flag.
> > > > >>>>>> (2) Provide additional information used by
> ExecutionEnvironment
> > to
> > > > >>>>>>
> > > > >>>>> decide
> > > > >>>>
> > > > >>>>> when/where to use the DistributedCached file.
> > > > >>>>>>
> > > > >>>>>> In this case, DistributedCache.CacheType or
> > > > DistributedCache.FileType
> > > > >>>>>> sounds more intuitive, what do you think?
> > > > >>>>>>
> > > > >>>>>> Also, I was wondering is there any other useful information
> for
> > > the
> > > > >>>>>>
> > > > >>>>> cached
> > > > >>>>>
> > > > >>>>>> file to be passed to runtime.
> > > > >>>>>> If we are just talking about including the library to the
> > > > classloader,
> > > > >>>>>>
> > > > >>>>> can
> > > > >>>>>
> > > > >>>>>> we directly extend the interface with
> > > > >>>>>>
> > > > >>>>>> public void registerCachedFile(
> > > > >>>>>> String filePath,
> > > > >>>>>> String name,
> > > > >>>>>> boolean executable,
> > > > >>>>>> boolean includeInClassLoader)
> > > > >>>>>>
> > > > >>>>>>
> > > > >>>>>> Thanks,
> > > > >>>>>> Rong
> > > > >>>>>>
> > > > >>>>>> On Sun, May 13, 2018 at 11:14 PM, Shuyi Chen <
> > suez1...@gmail.com>
> > > > >>>>>>
> > > > >>>>> wrote:
> > > > >>>>
> > > > >>>>> Hi Flink devs,
> > > > >>>>>>>
> > > > >>>>>>> In an effort to support loading external libraries and
> creating
> > > > UDFs
> > > > >>>>>>>
> > > > >>>>>> from
> > > > >>>>>
> > > > >>>>>> external libraries using DDL in Flink SQL, we want to use
> > Flink’s
> > > > >>>>>>>
> > > > >>>>>> Blob
> > > > >>>>
> > > > >>>>> Server to distribute the external libraries in runtime and load
> > > those
> > > > >>>>>>> libraries into the user code classloader automatically.
> > > > >>>>>>>
> > > > >>>>>>> However, the current [Stream]ExecutionEnvironment.
> > > > registerCachedFile
> > > > >>>>>>> interface limits only to registering executable or
> > non-executable
> > > > >>>>>>>
> > > > >>>>>> blobs.
> > > > >>>>>
> > > > >>>>>> It’s not possible to tell in runtime if the blob files are
> > > libraries
> > > > >>>>>>>
> > > > >>>>>> and
> > > > >>>>>
> > > > >>>>>> should be loaded into the user code classloader in
> > RuntimeContext.
> > > > >>>>>>> Therefore, I want to propose to add an enum called *BlobType*
> > > > >>>>>>>
> > > > >>>>>> explicitly
> > > > >>>>>
> > > > >>>>>> to
> > > > >>>>>>
> > > > >>>>>>> indicate the type of the Blob file being distributed, and the
> > > > >>>>>>>
> > > > >>>>>> following
> > > > >>>>
> > > > >>>>> interface in [Stream]ExecutionEnvironment to support it. In
> > > general,
> > > > >>>>>>>
> > > > >>>>>> I
> > > > >>>>
> > > > >>>>> think the new BlobType information can be used by Flink runtime
> > to
> > > > >>>>>>> preprocess the Blob files if needed.
> > > > >>>>>>>
> > > > >>>>>>> */***
> > > > >>>>>>> ** Registers a file at the distributed cache under the given
> > > name.
> > > > >>>>>>>
> > > > >>>>>> The
> > > > >>>>
> > > > >>>>> file
> > > > >>>>>>
> > > > >>>>>>> will be accessible*
> > > > >>>>>>> ** from any user-defined function in the (distributed)
> runtime
> > > > under
> > > > >>>>>>>
> > > > >>>>>> a
> > > > >>>>
> > > > >>>>> local path. Files*
> > > > >>>>>>> ** may be local files (as long as all relevant workers have
> > > access
> > > > to
> > > > >>>>>>>
> > > > >>>>>> it),
> > > > >>>>>>
> > > > >>>>>>> or files in a distributed file system.*
> > > > >>>>>>> ** The runtime will copy the files temporarily to a local
> > cache,
> > > if
> > > > >>>>>>> needed.*
> > > > >>>>>>> ***
> > > > >>>>>>> ** <p>The {@link org.apache.flink.api.common.
> > > > >>>>>>>
> > > > >>>>>> functions.RuntimeContext}
> > > > >>>>
> > > > >>>>> can
> > > > >>>>>>
> > > > >>>>>>> be obtained inside UDFs via*
> > > > >>>>>>> ** {@link
> > > > >>>>>>> org.apache.flink.api.common.functions.RichFunction#
> > > > >>>>>>>
> > > > >>>>>> getRuntimeContext()}
> > > > >>>>>
> > > > >>>>>> and
> > > > >>>>>>> provides access*
> > > > >>>>>>> ** {@link org.apache.flink.api.common.ca
> > > > >>>>>>> <http://org.apache.flink.api.common.ca>che.DistributedCache}
> > > via*
> > > > >>>>>>> ** {@link
> > > > >>>>>>> org.apache.flink.api.common.functions.RuntimeContext#
> > > > >>>>>>> getDistributedCache()}.*
> > > > >>>>>>> ***
> > > > >>>>>>> ** @param filePath The path of the file, as a URI (e.g.
> > > > >>>>>>>
> > > > >>>>>> "file:///some/path"
> > > > >>>>>>
> > > > >>>>>>> or "hdfs://host:port/and/path")*
> > > > >>>>>>> ** @param name The name under which the file is registered.*
> > > > >>>>>>> ** @param blobType indicating the type of the Blob file*
> > > > >>>>>>> **/*
> > > > >>>>>>>
> > > > >>>>>>> *public void registerCachedFile(String filePath, String name,
> > > > >>>>>>> DistributedCache.BlobType blobType) {...}*
> > > > >>>>>>>
> > > > >>>>>>> Optionally, we can add another interface to register UDF Jars
> > > which
> > > > >>>>>>>
> > > > >>>>>> will
> > > > >>>>>
> > > > >>>>>> use the interface above to implement.
> > > > >>>>>>>
> > > > >>>>>>> *public void registerJarFile(String filePath, String name)
> > {...}*
> > > > >>>>>>>
> > > > >>>>>>> The existing interface in the following will be marked
> > > deprecated:
> > > > >>>>>>>
> > > > >>>>>>> *public void registerCachedFile(String filePath, String name,
> > > > boolean
> > > > >>>>>>> executable) {...}*
> > > > >>>>>>>
> > > > >>>>>>> And the following interface will be implemented using the new
> > > > >>>>>>>
> > > > >>>>>> interface
> > > > >>>>
> > > > >>>>> proposed above with a EXECUTABLE BlobType:
> > > > >>>>>>>
> > > > >>>>>>> *public void registerCachedFile(String filePath, String
> name) {
> > > ...
> > > > >>>>>>>
> > > > >>>>>> }*
> > > > >>>>
> > > > >>>>> Thanks a lot.
> > > > >>>>>>> Shuyi
> > > > >>>>>>>
> > > > >>>>>>> "So you have to trust that the dots will somehow connect in
> > your
> > > > >>>>>>>
> > > > >>>>>> future."
> > > > >>>>>
> > > > >>>>
> > > > >
> > > > >
> > > >
> > >
> > >
> > >
> > > --
> > > "So you have to trust that the dots will somehow connect in your
> future."
> > >
> >
>
--
"So you have to trust that the dots will somehow connect in your future."
