How to make a generic key for groupBy

[LINZ, Arnaud]

Hello,

 

After a quite successful benchmark yesterday (Flink being about twice faster than Spark on my use cases), I’ve turned instantly from spark-fan to flink-fan – great job, committers!

So I’ve decided to port my existing Spark tools to Flink. Happily, most of the difficulty was renaming classes, packages and variables with “spark” in them to something more neutral J

 

However there is one easy thing in Spark I’m still wondering how to do in Flink : generic keys.

 

I’m trying to make a framework on which my applications are built. That framework thus manipulate “generic types” representing the data, inheriting from an abstract class with a common contract, let’s call it “Bean”.

 

Among other things Bean exposes an abstract method

public Key getKey();   

 

Key being one of my core types used in several java algorithms.

 

Let’s say I have the class :

public class Framework<T extends Bean> implements Serializable {

 

public DataSet<T> doCoolStuff(final DataSet<T> inputDataset) {

        // Group lines according to a key

        final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, Key>() {

            @Override

            public Key getKey(T record)  {

                return record.getKey();

            }

        });

             (…)

       }

}

 

With Spark, a mapToPair works fine because all I have to do is implements correctly hashCode() and equals() on my Key type.

With Flink, Key is not recognized as a POJO object (well it is not) and that does not work.

 

I have tried to expose something like public Tuple getKeyAsTuple(); in Key but Flink does not accept generic Tuples. I’ve tried to parameterize my Tuple but Flink does not know how to infer

the generic type value.

 

So I’m wondering what is the best way to implement it.

For now I have exposed something like public String getKeyAsString(); and turned my generic treatment into :

final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, String>() {

            @Override

            public String getKey(T record)  {

                return record.getKey().getKeyAsString();

            }

        });

But that “ASCII” representation is suboptimal.

 

I thought of passing a key to tuple conversion lambda upon creation of the Framework class but that would be boiler-plate code on the user’s end, which I’m not fond of.

 

So my questions are :

-          Is there a smarter way to do this ?

-          What kind of objects can be passed as a Key ? Is there an Interface to respect ?

-          In the worst case, is byte[]  ok as a Key ? (I can code the serialization on the framework side…)

 

 

Best regards,

Arnaud

 

---

L'intégrité de ce message n'étant pas assurée sur internet, la société expéditrice ne peut être tenue responsable de son contenu ni de ses pièces jointes. Toute utilisation ou diffusion non autorisée est interdite. Si vous n'êtes pas destinataire de ce message, merci de le détruire et d'avertir l'expéditeur.

The integrity of this message cannot be guaranteed on the Internet. The company that sent this message cannot therefore be held liable for its content nor attachments. Any unauthorized use or dissemination is prohibited. If you are not the intended recipient of this message, then please delete it and notify the sender.

[Soumitra Kumar]

Will you elaborate on your use case? It would help to find out where Flink shines. IMO, its a great project, but needs more differentiation from Spark.

On Thu, Apr 23, 2015 at 7:25 AM, LINZ, Arnaud <[hidden email]> wrote:

Hello,

 

After a quite successful benchmark yesterday (Flink being about twice faster than Spark on my use cases), I’ve turned instantly from spark-fan to flink-fan – great job, committers!

So I’ve decided to port my existing Spark tools to Flink. Happily, most of the difficulty was renaming classes, packages and variables with “spark” in them to something more neutral J

 

However there is one easy thing in Spark I’m still wondering how to do in Flink : generic keys.

 

I’m trying to make a framework on which my applications are built. That framework thus manipulate “generic types” representing the data, inheriting from an abstract class with a common contract, let’s call it “Bean”.

 

Among other things Bean exposes an abstract method

public Key getKey();   

 

Key being one of my core types used in several java algorithms.

 

Let’s say I have the class :

public class Framework<T extends Bean> implements Serializable {

 

public DataSet<T> doCoolStuff(final DataSet<T> inputDataset) {

        // Group lines according to a key

        final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, Key>() {

            @Override

            public Key getKey(T record)  {

                return record.getKey();

            }

        });

             (…)

       }

}

 

With Spark, a mapToPair works fine because all I have to do is implements correctly hashCode() and equals() on my Key type.

With Flink, Key is not recognized as a POJO object (well it is not) and that does not work.

 

I have tried to expose something like public Tuple getKeyAsTuple(); in Key but Flink does not accept generic Tuples. I’ve tried to parameterize my Tuple but Flink does not know how to infer

the generic type value.

 

So I’m wondering what is the best way to implement it.

For now I have exposed something like public String getKeyAsString(); and turned my generic treatment into :

final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, String>() {

            @Override

            public String getKey(T record)  {

                return record.getKey().getKeyAsString();

            }

        });

But that “ASCII” representation is suboptimal.

 

I thought of passing a key to tuple conversion lambda upon creation of the Framework class but that would be boiler-plate code on the user’s end, which I’m not fond of.

 

So my questions are :

-          Is there a smarter way to do this ?

-          What kind of objects can be passed as a Key ? Is there an Interface to respect ?

-          In the worst case, is byte[]  ok as a Key ? (I can code the serialization on the framework side…)

 

 

Best regards,

Arnaud

 

---

L'intégrité de ce message n'étant pas assurée sur internet, la société expéditrice ne peut être tenue responsable de son contenu ni de ses pièces jointes. Toute utilisation ou diffusion non autorisée est interdite. Si vous n'êtes pas destinataire de ce message, merci de le détruire et d'avertir l'expéditeur.

The integrity of this message cannot be guaranteed on the Internet. The company that sent this message cannot therefore be held liable for its content nor attachments. Any unauthorized use or dissemination is prohibited. If you are not the intended recipient of this message, then please delete it and notify the sender.

[Stephan Ewen]

Hi Arnaud!

Thank you for the warm words! Let's find a good way to get this to work...

As a bit of background:

In Flink, the API needs to now a bit about the types that go through the functions, because Flink pre-generates and configures serializers, and validates that things fit together.

It is also important that keys are exposed rather specifically, because Flink internally tries to work on serialized data (that makes it in-memory operations predictable and robust).

If you expose a key as a "String", or "long" or "double", then Flink knows how to work on it in a binary fashion.

Also, if you expose a key as a POJO, then Flink interprets the key as a combination of the fields, and can again work on the serialized data.

If you only expose "Comparable" (which is the bare minimum for a key), you experience performance degradation (most notably for sorts), because every key operation involves serialization and deserialization.

So the goal would be to expose the key properly. We can always hint to the API what the key type is, precisely for the cases where the inference cannot do it.

  - To understand things a bit better: What is your "Key" type? Is it an abstract class, an interface, a generic parameter?

Greetings,

Stephan

FYI: In Scala, this works actually quite a bit easier, since Scala does preserve generic types. In Java, we built a lot of reflection tooling, but there are cases where it is impossible to infer the types via reflection, like yours.

On Thu, Apr 23, 2015 at 6:35 PM, Soumitra Kumar <[hidden email]> wrote:

Will you elaborate on your use case? It would help to find out where Flink shines. IMO, its a great project, but needs more differentiation from Spark.

On Thu, Apr 23, 2015 at 7:25 AM, LINZ, Arnaud <[hidden email]> wrote:

Hello,

 

After a quite successful benchmark yesterday (Flink being about twice faster than Spark on my use cases), I’ve turned instantly from spark-fan to flink-fan – great job, committers!

So I’ve decided to port my existing Spark tools to Flink. Happily, most of the difficulty was renaming classes, packages and variables with “spark” in them to something more neutral J

 

However there is one easy thing in Spark I’m still wondering how to do in Flink : generic keys.

 

I’m trying to make a framework on which my applications are built. That framework thus manipulate “generic types” representing the data, inheriting from an abstract class with a common contract, let’s call it “Bean”.

 

Among other things Bean exposes an abstract method

public Key getKey();   

 

Key being one of my core types used in several java algorithms.

 

Let’s say I have the class :

public class Framework<T extends Bean> implements Serializable {

 

public DataSet<T> doCoolStuff(final DataSet<T> inputDataset) {

        // Group lines according to a key

        final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, Key>() {

            @Override

            public Key getKey(T record)  {

                return record.getKey();

            }

        });

             (…)

       }

}

 

With Spark, a mapToPair works fine because all I have to do is implements correctly hashCode() and equals() on my Key type.

With Flink, Key is not recognized as a POJO object (well it is not) and that does not work.

 

I have tried to expose something like public Tuple getKeyAsTuple(); in Key but Flink does not accept generic Tuples. I’ve tried to parameterize my Tuple but Flink does not know how to infer

the generic type value.

 

So I’m wondering what is the best way to implement it.

For now I have exposed something like public String getKeyAsString(); and turned my generic treatment into :

final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, String>() {

            @Override

            public String getKey(T record)  {

                return record.getKey().getKeyAsString();

            }

        });

But that “ASCII” representation is suboptimal.

 

I thought of passing a key to tuple conversion lambda upon creation of the Framework class but that would be boiler-plate code on the user’s end, which I’m not fond of.

 

So my questions are :

-          Is there a smarter way to do this ?

-          What kind of objects can be passed as a Key ? Is there an Interface to respect ?

-          In the worst case, is byte[]  ok as a Key ? (I can code the serialization on the framework side…)

 

 

Best regards,

Arnaud

 

---

L'intégrité de ce message n'étant pas assurée sur internet, la société expéditrice ne peut être tenue responsable de son contenu ni de ses pièces jointes. Toute utilisation ou diffusion non autorisée est interdite. Si vous n'êtes pas destinataire de ce message, merci de le détruire et d'avertir l'expéditeur.

The integrity of this message cannot be guaranteed on the Internet. The company that sent this message cannot therefore be held liable for its content nor attachments. Any unauthorized use or dissemination is prohibited. If you are not the intended recipient of this message, then please delete it and notify the sender.

[LINZ, Arnaud]

Hi,

I see. My Key class is an abstract class, which subclasses are Key1<?>, Key2<?,?> etc, so it’s very like a tuple. It is heavily used in “non-distributed” hash maps once the dataset is reduced to fit on a single JVM.

It exposes the common contract that I need (such as getHeadKey(), getLastl(), or makeKey(Key,Object)) to “navigate” in the key space, and a cached hash code to make hash maps faster. My generic algorithms do not need to know how many fields are exposed in the Key, but they need to be able to construct another key from two keys.

 

Arnaud

 

De : [hidden email] [mailto:[hidden email]] De la part de Stephan Ewen
Envoyé : vendredi 24 avril 2015 11:14
À : [hidden email]
Objet : Re: How to make a generic key for groupBy

 

Hi Arnaud!

 

Thank you for the warm words! Let's find a good way to get this to work...

 

As a bit of background:

In Flink, the API needs to now a bit about the types that go through the functions, because Flink pre-generates and configures serializers, and validates that things fit together.

 

It is also important that keys are exposed rather specifically, because Flink internally tries to work on serialized data (that makes it in-memory operations predictable and robust).

 

If you expose a key as a "String", or "long" or "double", then Flink knows how to work on it in a binary fashion.

Also, if you expose a key as a POJO, then Flink interprets the key as a combination of the fields, and can again work on the serialized data.

 

If you only expose "Comparable" (which is the bare minimum for a key), you experience performance degradation (most notably for sorts), because every key operation involves serialization and deserialization.

 

So the goal would be to expose the key properly. We can always hint to the API what the key type is, precisely for the cases where the inference cannot do it.

  - To understand things a bit better: What is your "Key" type? Is it an abstract class, an interface, a generic parameter?

 

 

Greetings,

Stephan

 

 

FYI: In Scala, this works actually quite a bit easier, since Scala does preserve generic types. In Java, we built a lot of reflection tooling, but there are cases where it is impossible to infer the types via reflection, like yours.

 

 

 

On Thu, Apr 23, 2015 at 6:35 PM, Soumitra Kumar <[hidden email]> wrote:

Will you elaborate on your use case? It would help to find out where Flink shines. IMO, its a great project, but needs more differentiation from Spark.

 

On Thu, Apr 23, 2015 at 7:25 AM, LINZ, Arnaud <[hidden email]> wrote:

Hello,

 

After a quite successful benchmark yesterday (Flink being about twice faster than Spark on my use cases), I’ve turned instantly from spark-fan to flink-fan – great job, committers!

So I’ve decided to port my existing Spark tools to Flink. Happily, most of the difficulty was renaming classes, packages and variables with “spark” in them to something more neutral J

 

However there is one easy thing in Spark I’m still wondering how to do in Flink : generic keys.

 

I’m trying to make a framework on which my applications are built. That framework thus manipulate “generic types” representing the data, inheriting from an abstract class with a common contract, let’s call it “Bean”.

 

Among other things Bean exposes an abstract method

public Key getKey();   

 

Key being one of my core types used in several java algorithms.

 

Let’s say I have the class :

public class Framework<T extends Bean> implements Serializable {

 

public DataSet<T> doCoolStuff(final DataSet<T> inputDataset) {

        // Group lines according to a key

        final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, Key>() {

            @Override

            public Key getKey(T record)  {

                return record.getKey();

            }

        });

             (…)

       }

}

 

With Spark, a mapToPair works fine because all I have to do is implements correctly hashCode() and equals() on my Key type.

With Flink, Key is not recognized as a POJO object (well it is not) and that does not work.

 

I have tried to expose something like public Tuple getKeyAsTuple(); in Key but Flink does not accept generic Tuples. I’ve tried to parameterize my Tuple but Flink does not know how to infer

the generic type value.

 

So I’m wondering what is the best way to implement it.

For now I have exposed something like public String getKeyAsString(); and turned my generic treatment into :

final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, String>() {

            @Override

            public String getKey(T record)  {

                return record.getKey().getKeyAsString();

            }

        });

But that “ASCII” representation is suboptimal.

 

I thought of passing a key to tuple conversion lambda upon creation of the Framework class but that would be boiler-plate code on the user’s end, which I’m not fond of.

 

So my questions are :

-          Is there a smarter way to do this ?

-          What kind of objects can be passed as a Key ? Is there an Interface to respect ?

-          In the worst case, is byte[]  ok as a Key ? (I can code the serialization on the framework side…)

 

 

Best regards,

Arnaud

 

 

---

L'intégrité de ce message n'étant pas assurée sur internet, la société expéditrice ne peut être tenue responsable de son contenu ni de ses pièces jointes. Toute utilisation ou diffusion non autorisée est interdite. Si vous n'êtes pas destinataire de ce message, merci de le détruire et d'avertir l'expéditeur.

The integrity of this message cannot be guaranteed on the Internet. The company that sent this message cannot therefore be held liable for its content nor attachments. Any unauthorized use or dissemination is prohibited. If you are not the intended recipient of this message, then please delete it and notify the sender.

 

 

[Stephan Ewen]

Hi!

I have some ideas, let me see if I can make them concrete until tomorrow...

Greetings,

Stephan

On Mon, Apr 27, 2015 at 5:29 PM, LINZ, Arnaud <[hidden email]> wrote:

Hi,

I see. My Key class is an abstract class, which subclasses are Key1<?>, Key2<?,?> etc, so it’s very like a tuple. It is heavily used in “non-distributed” hash maps once the dataset is reduced to fit on a single JVM.

It exposes the common contract that I need (such as getHeadKey(), getLastl(), or makeKey(Key,Object)) to “navigate” in the key space, and a cached hash code to make hash maps faster. My generic algorithms do not need to know how many fields are exposed in the Key, but they need to be able to construct another key from two keys.

 

Arnaud

 

De : [hidden email] [mailto:[hidden email]] De la part de Stephan Ewen
Envoyé : vendredi 24 avril 2015 11:14
À : [hidden email]
Objet : Re: How to make a generic key for groupBy

 

Hi Arnaud!

 

Thank you for the warm words! Let's find a good way to get this to work...

 

As a bit of background:

In Flink, the API needs to now a bit about the types that go through the functions, because Flink pre-generates and configures serializers, and validates that things fit together.

 

It is also important that keys are exposed rather specifically, because Flink internally tries to work on serialized data (that makes it in-memory operations predictable and robust).

 

If you expose a key as a "String", or "long" or "double", then Flink knows how to work on it in a binary fashion.

Also, if you expose a key as a POJO, then Flink interprets the key as a combination of the fields, and can again work on the serialized data.

 

If you only expose "Comparable" (which is the bare minimum for a key), you experience performance degradation (most notably for sorts), because every key operation involves serialization and deserialization.

 

So the goal would be to expose the key properly. We can always hint to the API what the key type is, precisely for the cases where the inference cannot do it.

  - To understand things a bit better: What is your "Key" type? Is it an abstract class, an interface, a generic parameter?

 

 

Greetings,

Stephan

 

 

FYI: In Scala, this works actually quite a bit easier, since Scala does preserve generic types. In Java, we built a lot of reflection tooling, but there are cases where it is impossible to infer the types via reflection, like yours.

 

 

 

On Thu, Apr 23, 2015 at 6:35 PM, Soumitra Kumar <[hidden email]> wrote:

Will you elaborate on your use case? It would help to find out where Flink shines. IMO, its a great project, but needs more differentiation from Spark.

 

On Thu, Apr 23, 2015 at 7:25 AM, LINZ, Arnaud <[hidden email]> wrote:

Hello,

 

After a quite successful benchmark yesterday (Flink being about twice faster than Spark on my use cases), I’ve turned instantly from spark-fan to flink-fan – great job, committers!

So I’ve decided to port my existing Spark tools to Flink. Happily, most of the difficulty was renaming classes, packages and variables with “spark” in them to something more neutral J

 

However there is one easy thing in Spark I’m still wondering how to do in Flink : generic keys.

 

I’m trying to make a framework on which my applications are built. That framework thus manipulate “generic types” representing the data, inheriting from an abstract class with a common contract, let’s call it “Bean”.

 

Among other things Bean exposes an abstract method

public Key getKey();   

 

Key being one of my core types used in several java algorithms.

 

Let’s say I have the class :

public class Framework<T extends Bean> implements Serializable {

 

public DataSet<T> doCoolStuff(final DataSet<T> inputDataset) {

        // Group lines according to a key

        final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, Key>() {

            @Override

            public Key getKey(T record)  {

                return record.getKey();

            }

        });

             (…)

       }

}

 

With Spark, a mapToPair works fine because all I have to do is implements correctly hashCode() and equals() on my Key type.

With Flink, Key is not recognized as a POJO object (well it is not) and that does not work.

 

I have tried to expose something like public Tuple getKeyAsTuple(); in Key but Flink does not accept generic Tuples. I’ve tried to parameterize my Tuple but Flink does not know how to infer

the generic type value.

 

So I’m wondering what is the best way to implement it.

For now I have exposed something like public String getKeyAsString(); and turned my generic treatment into :

final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, String>() {

            @Override

            public String getKey(T record)  {

                return record.getKey().getKeyAsString();

            }

        });

But that “ASCII” representation is suboptimal.

 

I thought of passing a key to tuple conversion lambda upon creation of the Framework class but that would be boiler-plate code on the user’s end, which I’m not fond of.

 

So my questions are :

-          Is there a smarter way to do this ?

-          What kind of objects can be passed as a Key ? Is there an Interface to respect ?

-          In the worst case, is byte[]  ok as a Key ? (I can code the serialization on the framework side…)

 

 

Best regards,

Arnaud

 

 

---

L'intégrité de ce message n'étant pas assurée sur internet, la société expéditrice ne peut être tenue responsable de son contenu ni de ses pièces jointes. Toute utilisation ou diffusion non autorisée est interdite. Si vous n'êtes pas destinataire de ce message, merci de le détruire et d'avertir l'expéditeur.

The integrity of this message cannot be guaranteed on the Internet. The company that sent this message cannot therefore be held liable for its content nor attachments. Any unauthorized use or dissemination is prohibited. If you are not the intended recipient of this message, then please delete it and notify the sender.

 

 

[Stephan Ewen]

Hey Arnaud!

I have made a quick sample implementation of how you can very efficiently support generic keys, like yours. I put the code in this repository: https://github.com/StephanEwen/flink-generic-keys

It implements a special key selector. You can use that to do what you used to do, and it internally does a bit of magic to expose type information to the Flink pre-flight phase:

https://github.com/StephanEwen/flink-generic-keys/blob/master/src/main/java/com/dataartisans/flink/TupleKeySelector.java

Here is a quick intro to what this does:

  - It exposes keys as Flink tuples, by wrapping your Keys in a Tuple in a generic way. Tuples are the fastest keys in Flink, because they support the best "on-binary-data" operations.

  - It figures out which types are going to be in the tuples (and in your generic keys), by analyzing the Method signature of your concrete "getKey()" method implementation.

The requirement for the second part is that your concrete types declare the concrete key in their signature. Below is an example. Note that the "SomeType" class does not declare "Key" as the return type of "getKey()", but very specifically "Key2<String, Integer>". That way, the key selector can pick up the types (String and Integer) and expose it to Flink, so that Flink can generate the key comparators to do efficient binary operations for the keys.

public abstract class TypeBase {

public abstract Key getKey();

}

public class SomeType extends TypeBase {

public String someString;

public int anotherInteger;

public Date aDate;

@Override

public Key2<String, Integer> getKey() {

return new Key2<String, Integer>(someString, anotherInteger);

}

}

The good thing about exposing this information to Flink in the pre-flight phase (before the job runs in parallel) is that it pre-checks many things to prevent most memory and serialization surprises at runtime.

Let me know if you have questions!

Greetings,

Stephan

On Mon, Apr 27, 2015 at 6:38 PM, Stephan Ewen <[hidden email]> wrote:

Hi!

I have some ideas, let me see if I can make them concrete until tomorrow...

Greetings,

Stephan

On Mon, Apr 27, 2015 at 5:29 PM, LINZ, Arnaud <[hidden email]> wrote:

Hi,

I see. My Key class is an abstract class, which subclasses are Key1<?>, Key2<?,?> etc, so it’s very like a tuple. It is heavily used in “non-distributed” hash maps once the dataset is reduced to fit on a single JVM.

It exposes the common contract that I need (such as getHeadKey(), getLastl(), or makeKey(Key,Object)) to “navigate” in the key space, and a cached hash code to make hash maps faster. My generic algorithms do not need to know how many fields are exposed in the Key, but they need to be able to construct another key from two keys.

 

Arnaud

 

De : [hidden email] [mailto:[hidden email]] De la part de Stephan Ewen
Envoyé : vendredi 24 avril 2015 11:14
À : [hidden email]
Objet : Re: How to make a generic key for groupBy

 

Hi Arnaud!

 

Thank you for the warm words! Let's find a good way to get this to work...

 

As a bit of background:

In Flink, the API needs to now a bit about the types that go through the functions, because Flink pre-generates and configures serializers, and validates that things fit together.

 

It is also important that keys are exposed rather specifically, because Flink internally tries to work on serialized data (that makes it in-memory operations predictable and robust).

 

If you expose a key as a "String", or "long" or "double", then Flink knows how to work on it in a binary fashion.

Also, if you expose a key as a POJO, then Flink interprets the key as a combination of the fields, and can again work on the serialized data.

 

If you only expose "Comparable" (which is the bare minimum for a key), you experience performance degradation (most notably for sorts), because every key operation involves serialization and deserialization.

 

So the goal would be to expose the key properly. We can always hint to the API what the key type is, precisely for the cases where the inference cannot do it.

  - To understand things a bit better: What is your "Key" type? Is it an abstract class, an interface, a generic parameter?

 

 

Greetings,

Stephan

 

 

FYI: In Scala, this works actually quite a bit easier, since Scala does preserve generic types. In Java, we built a lot of reflection tooling, but there are cases where it is impossible to infer the types via reflection, like yours.

 

 

 

On Thu, Apr 23, 2015 at 6:35 PM, Soumitra Kumar <[hidden email]> wrote:

Will you elaborate on your use case? It would help to find out where Flink shines. IMO, its a great project, but needs more differentiation from Spark.

 

On Thu, Apr 23, 2015 at 7:25 AM, LINZ, Arnaud <[hidden email]> wrote:

Hello,

 

After a quite successful benchmark yesterday (Flink being about twice faster than Spark on my use cases), I’ve turned instantly from spark-fan to flink-fan – great job, committers!

So I’ve decided to port my existing Spark tools to Flink. Happily, most of the difficulty was renaming classes, packages and variables with “spark” in them to something more neutral J

 

However there is one easy thing in Spark I’m still wondering how to do in Flink : generic keys.

 

I’m trying to make a framework on which my applications are built. That framework thus manipulate “generic types” representing the data, inheriting from an abstract class with a common contract, let’s call it “Bean”.

 

Among other things Bean exposes an abstract method

public Key getKey();   

 

Key being one of my core types used in several java algorithms.

 

Let’s say I have the class :

public class Framework<T extends Bean> implements Serializable {

 

public DataSet<T> doCoolStuff(final DataSet<T> inputDataset) {

        // Group lines according to a key

        final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, Key>() {

            @Override

            public Key getKey(T record)  {

                return record.getKey();

            }

        });

             (…)

       }

}

 

With Spark, a mapToPair works fine because all I have to do is implements correctly hashCode() and equals() on my Key type.

With Flink, Key is not recognized as a POJO object (well it is not) and that does not work.

 

I have tried to expose something like public Tuple getKeyAsTuple(); in Key but Flink does not accept generic Tuples. I’ve tried to parameterize my Tuple but Flink does not know how to infer

the generic type value.

 

So I’m wondering what is the best way to implement it.

For now I have exposed something like public String getKeyAsString(); and turned my generic treatment into :

final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, String>() {

            @Override

            public String getKey(T record)  {

                return record.getKey().getKeyAsString();

            }

        });

But that “ASCII” representation is suboptimal.

 

I thought of passing a key to tuple conversion lambda upon creation of the Framework class but that would be boiler-plate code on the user’s end, which I’m not fond of.

 

So my questions are :

-          Is there a smarter way to do this ?

-          What kind of objects can be passed as a Key ? Is there an Interface to respect ?

-          In the worst case, is byte[]  ok as a Key ? (I can code the serialization on the framework side…)

 

 

Best regards,

Arnaud

 

 

---

L'intégrité de ce message n'étant pas assurée sur internet, la société expéditrice ne peut être tenue responsable de son contenu ni de ses pièces jointes. Toute utilisation ou diffusion non autorisée est interdite. Si vous n'êtes pas destinataire de ce message, merci de le détruire et d'avertir l'expéditeur.

The integrity of this message cannot be guaranteed on the Internet. The company that sent this message cannot therefore be held liable for its content nor attachments. Any unauthorized use or dissemination is prohibited. If you are not the intended recipient of this message, then please delete it and notify the sender.

 

 

[LINZ, Arnaud]

Hi Stephen,

Thanks a lot, that solves my problem. I am truly amazed and grateful by the time you spent to help me on this subject.

Greetings,
Arnaud

---

De : [hidden email]
Envoyé : ‎29/‎04/‎2015 17:19
À : [hidden email]
Objet : Re: How to make a generic key for groupBy

Hey Arnaud!

I have made a quick sample implementation of how you can very efficiently support generic keys, like yours. I put the code in this repository: https://github.com/StephanEwen/flink-generic-keys

It implements a special key selector. You can use that to do what you used to do, and it internally does a bit of magic to expose type information to the Flink pre-flight phase:

https://github.com/StephanEwen/flink-generic-keys/blob/master/src/main/java/com/dataartisans/flink/TupleKeySelector.java

Here is a quick intro to what this does:

  - It exposes keys as Flink tuples, by wrapping your Keys in a Tuple in a generic way. Tuples are the fastest keys in Flink, because they support the best "on-binary-data" operations.

  - It figures out which types are going to be in the tuples (and in your generic keys), by analyzing the Method signature of your concrete "getKey()" method implementation.

The requirement for the second part is that your concrete types declare the concrete key in their signature. Below is an example. Note that the "SomeType" class does not declare "Key" as the return type of "getKey()", but very specifically "Key2<String, Integer>". That way, the key selector can pick up the types (String and Integer) and expose it to Flink, so that Flink can generate the key comparators to do efficient binary operations for the keys.

public abstract class TypeBase {

public abstract Key getKey();

}

public class SomeType extends TypeBase {

public String someString;

public int anotherInteger;

public Date aDate;

@Override

public Key2<String, Integer> getKey() {

return new Key2<String, Integer>(someString, anotherInteger);

}

}

The good thing about exposing this information to Flink in the pre-flight phase (before the job runs in parallel) is that it pre-checks many things to prevent most memory and serialization surprises at runtime.

Let me know if you have questions!

Greetings,

Stephan

On Mon, Apr 27, 2015 at 6:38 PM, Stephan Ewen <[hidden email]> wrote:

Hi!

I have some ideas, let me see if I can make them concrete until tomorrow...

Greetings,

Stephan

On Mon, Apr 27, 2015 at 5:29 PM, LINZ, Arnaud <[hidden email]> wrote:

Hi,

I see. My Key class is an abstract class, which subclasses are Key1<?>, Key2<?,?> etc, so it’s very like a tuple. It is heavily used in “non-distributed” hash maps once the dataset is reduced to fit on a single JVM.

It exposes the common contract that I need (such as getHeadKey(), getLastl(), or makeKey(Key,Object)) to “navigate” in the key space, and a cached hash code to make hash maps faster. My generic algorithms do not need to know how many fields are exposed in the Key, but they need to be able to construct another key from two keys.

 

Arnaud

 

De : [hidden email] [mailto:[hidden email]] De la part de Stephan Ewen
Envoyé : vendredi 24 avril 2015 11:14
À : [hidden email]
Objet : Re: How to make a generic key for groupBy

 

Hi Arnaud!

 

Thank you for the warm words! Let's find a good way to get this to work...

 

As a bit of background:

In Flink, the API needs to now a bit about the types that go through the functions, because Flink pre-generates and configures serializers, and validates that things fit together.

 

It is also important that keys are exposed rather specifically, because Flink internally tries to work on serialized data (that makes it in-memory operations predictable and robust).

 

If you expose a key as a "String", or "long" or "double", then Flink knows how to work on it in a binary fashion.

Also, if you expose a key as a POJO, then Flink interprets the key as a combination of the fields, and can again work on the serialized data.

 

If you only expose "Comparable" (which is the bare minimum for a key), you experience performance degradation (most notably for sorts), because every key operation involves serialization and deserialization.

 

So the goal would be to expose the key properly. We can always hint to the API what the key type is, precisely for the cases where the inference cannot do it.

  - To understand things a bit better: What is your "Key" type? Is it an abstract class, an interface, a generic parameter?

 

 

Greetings,

Stephan

 

 

FYI: In Scala, this works actually quite a bit easier, since Scala does preserve generic types. In Java, we built a lot of reflection tooling, but there are cases where it is impossible to infer the types via reflection, like yours.

 

 

 

On Thu, Apr 23, 2015 at 6:35 PM, Soumitra Kumar <[hidden email]> wrote:

Will you elaborate on your use case? It would help to find out where Flink shines. IMO, its a great project, but needs more differentiation from Spark.

 

On Thu, Apr 23, 2015 at 7:25 AM, LINZ, Arnaud <[hidden email]> wrote:

Hello,

 

After a quite successful benchmark yesterday (Flink being about twice faster than Spark on my use cases), I’ve turned instantly from spark-fan to flink-fan – great job, committers!

So I’ve decided to port my existing Spark tools to Flink. Happily, most of the difficulty was renaming classes, packages and variables with “spark” in them to something more neutral J

 

However there is one easy thing in Spark I’m still wondering how to do in Flink : generic keys.

 

I’m trying to make a framework on which my applications are built. That framework thus manipulate “generic types” representing the data, inheriting from an abstract class with a common contract, let’s call it “Bean”.

 

Among other things Bean exposes an abstract method

public Key getKey();   

 

Key being one of my core types used in several java algorithms.

 

Let’s say I have the class :

public class Framework<T extends Bean> implements Serializable {

 

public DataSet<T> doCoolStuff(final DataSet<T> inputDataset) {

        // Group lines according to a key

        final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, Key>() {

            @Override

            public Key getKey(T record)  {

                return record.getKey();

            }

        });

             (…)

       }

}

 

With Spark, a mapToPair works fine because all I have to do is implements correctly hashCode() and equals() on my Key type.

With Flink, Key is not recognized as a POJO object (well it is not) and that does not work.

 

I have tried to expose something like public Tuple getKeyAsTuple(); in Key but Flink does not accept generic Tuples. I’ve tried to parameterize my Tuple but Flink does not know how to infer

the generic type value.

 

So I’m wondering what is the best way to implement it.

For now I have exposed something like public String getKeyAsString(); and turned my generic treatment into :

final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, String>() {

            @Override

            public String getKey(T record)  {

                return record.getKey().getKeyAsString();

            }

        });

But that “ASCII” representation is suboptimal.

 

I thought of passing a key to tuple conversion lambda upon creation of the Framework class but that would be boiler-plate code on the user’s end, which I’m not fond of.

 

So my questions are :

-          Is there a smarter way to do this ?

-          What kind of objects can be passed as a Key ? Is there an Interface to respect ?

-          In the worst case, is byte[]  ok as a Key ? (I can code the serialization on the framework side…)

 

 

Best regards,

Arnaud

 

 

---

L'intégrité de ce message n'étant pas assurée sur internet, la société expéditrice ne peut être tenue responsable de son contenu ni de ses pièces jointes. Toute utilisation ou diffusion non autorisée est interdite. Si vous n'êtes pas destinataire de ce message, merci de le détruire et d'avertir l'expéditeur.

The integrity of this message cannot be guaranteed on the Internet. The company that sent this message cannot therefore be held liable for its content nor attachments. Any unauthorized use or dissemination is prohibited. If you are not the intended recipient of this message, then please delete it and notify the sender.

 

 

[Stephan Ewen]

My pleasure, happy to hear that it solved your problem.

I was wondering: Do more people have a similar structure in their types and utilities?

If yes, it may make sense to add a generic version of these utilities to the Fink codebase.

On Fri, May 1, 2015 at 11:51 AM, LINZ, Arnaud <[hidden email]> wrote:

Hi Stephen,

Thanks a lot, that solves my problem. I am truly amazed and grateful by the time you spent to help me on this subject.

Greetings,
Arnaud

---

De : [hidden email]
Envoyé : ‎29/‎04/‎2015 17:19

À : [hidden email]
Objet : Re: How to make a generic key for groupBy

Hey Arnaud!

I have made a quick sample implementation of how you can very efficiently support generic keys, like yours. I put the code in this repository: https://github.com/StephanEwen/flink-generic-keys

It implements a special key selector. You can use that to do what you used to do, and it internally does a bit of magic to expose type information to the Flink pre-flight phase:

https://github.com/StephanEwen/flink-generic-keys/blob/master/src/main/java/com/dataartisans/flink/TupleKeySelector.java

Here is a quick intro to what this does:

  - It exposes keys as Flink tuples, by wrapping your Keys in a Tuple in a generic way. Tuples are the fastest keys in Flink, because they support the best "on-binary-data" operations.

  - It figures out which types are going to be in the tuples (and in your generic keys), by analyzing the Method signature of your concrete "getKey()" method implementation.

The requirement for the second part is that your concrete types declare the concrete key in their signature. Below is an example. Note that the "SomeType" class does not declare "Key" as the return type of "getKey()", but very specifically "Key2<String, Integer>". That way, the key selector can pick up the types (String and Integer) and expose it to Flink, so that Flink can generate the key comparators to do efficient binary operations for the keys.

public abstract class TypeBase {

public abstract Key getKey();

}

public class SomeType extends TypeBase {

public String someString;

public int anotherInteger;

public Date aDate;

@Override

public Key2<String, Integer> getKey() {

return new Key2<String, Integer>(someString, anotherInteger);

}

}

The good thing about exposing this information to Flink in the pre-flight phase (before the job runs in parallel) is that it pre-checks many things to prevent most memory and serialization surprises at runtime.

Let me know if you have questions!

Greetings,

Stephan

On Mon, Apr 27, 2015 at 6:38 PM, Stephan Ewen <[hidden email]> wrote:

Hi!

I have some ideas, let me see if I can make them concrete until tomorrow...

Greetings,

Stephan

On Mon, Apr 27, 2015 at 5:29 PM, LINZ, Arnaud <[hidden email]> wrote:

Hi,

I see. My Key class is an abstract class, which subclasses are Key1<?>, Key2<?,?> etc, so it’s very like a tuple. It is heavily used in “non-distributed” hash maps once the dataset is reduced to fit on a single JVM.

It exposes the common contract that I need (such as getHeadKey(), getLastl(), or makeKey(Key,Object)) to “navigate” in the key space, and a cached hash code to make hash maps faster. My generic algorithms do not need to know how many fields are exposed in the Key, but they need to be able to construct another key from two keys.

 

Arnaud

 

De : [hidden email] [mailto:[hidden email]] De la part de Stephan Ewen
Envoyé : vendredi 24 avril 2015 11:14
À : [hidden email]
Objet : Re: How to make a generic key for groupBy

 

Hi Arnaud!

 

Thank you for the warm words! Let's find a good way to get this to work...

 

As a bit of background:

In Flink, the API needs to now a bit about the types that go through the functions, because Flink pre-generates and configures serializers, and validates that things fit together.

 

It is also important that keys are exposed rather specifically, because Flink internally tries to work on serialized data (that makes it in-memory operations predictable and robust).

 

If you expose a key as a "String", or "long" or "double", then Flink knows how to work on it in a binary fashion.

Also, if you expose a key as a POJO, then Flink interprets the key as a combination of the fields, and can again work on the serialized data.

 

If you only expose "Comparable" (which is the bare minimum for a key), you experience performance degradation (most notably for sorts), because every key operation involves serialization and deserialization.

 

So the goal would be to expose the key properly. We can always hint to the API what the key type is, precisely for the cases where the inference cannot do it.

  - To understand things a bit better: What is your "Key" type? Is it an abstract class, an interface, a generic parameter?

 

 

Greetings,

Stephan

 

 

FYI: In Scala, this works actually quite a bit easier, since Scala does preserve generic types. In Java, we built a lot of reflection tooling, but there are cases where it is impossible to infer the types via reflection, like yours.

 

 

 

On Thu, Apr 23, 2015 at 6:35 PM, Soumitra Kumar <[hidden email]> wrote:

Will you elaborate on your use case? It would help to find out where Flink shines. IMO, its a great project, but needs more differentiation from Spark.

 

On Thu, Apr 23, 2015 at 7:25 AM, LINZ, Arnaud <[hidden email]> wrote:

Hello,

 

After a quite successful benchmark yesterday (Flink being about twice faster than Spark on my use cases), I’ve turned instantly from spark-fan to flink-fan – great job, committers!

So I’ve decided to port my existing Spark tools to Flink. Happily, most of the difficulty was renaming classes, packages and variables with “spark” in them to something more neutral J

 

However there is one easy thing in Spark I’m still wondering how to do in Flink : generic keys.

 

I’m trying to make a framework on which my applications are built. That framework thus manipulate “generic types” representing the data, inheriting from an abstract class with a common contract, let’s call it “Bean”.

 

Among other things Bean exposes an abstract method

public Key getKey();   

 

Key being one of my core types used in several java algorithms.

 

Let’s say I have the class :

public class Framework<T extends Bean> implements Serializable {

 

public DataSet<T> doCoolStuff(final DataSet<T> inputDataset) {

        // Group lines according to a key

        final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, Key>() {

            @Override

            public Key getKey(T record)  {

                return record.getKey();

            }

        });

             (…)

       }

}

 

With Spark, a mapToPair works fine because all I have to do is implements correctly hashCode() and equals() on my Key type.

With Flink, Key is not recognized as a POJO object (well it is not) and that does not work.

 

I have tried to expose something like public Tuple getKeyAsTuple(); in Key but Flink does not accept generic Tuples. I’ve tried to parameterize my Tuple but Flink does not know how to infer

the generic type value.

 

So I’m wondering what is the best way to implement it.

For now I have exposed something like public String getKeyAsString(); and turned my generic treatment into :

final UnsortedGrouping<YT> groupe = inputDataset.groupBy(new KeySelector<T, String>() {

            @Override

            public String getKey(T record)  {

                return record.getKey().getKeyAsString();

            }

        });

But that “ASCII” representation is suboptimal.

 

I thought of passing a key to tuple conversion lambda upon creation of the Framework class but that would be boiler-plate code on the user’s end, which I’m not fond of.

 

So my questions are :

-          Is there a smarter way to do this ?

-          What kind of objects can be passed as a Key ? Is there an Interface to respect ?

-          In the worst case, is byte[]  ok as a Key ? (I can code the serialization on the framework side…)

 

 

Best regards,

Arnaud

 

 

---

L'intégrité de ce message n'étant pas assurée sur internet, la société expéditrice ne peut être tenue responsable de son contenu ni de ses pièces jointes. Toute utilisation ou diffusion non autorisée est interdite. Si vous n'êtes pas destinataire de ce message, merci de le détruire et d'avertir l'expéditeur.

The integrity of this message cannot be guaranteed on the Internet. The company that sent this message cannot therefore be held liable for its content nor attachments. Any unauthorized use or dissemination is prohibited. If you are not the intended recipient of this message, then please delete it and notify the sender.