mutable hashmap outside of stream, does it get snapshotted ?

[Bart van Deenen]

Hi all

I'm having a datastream transformation, that updates a mutable
hashmap that exists outside of the stream.

So it's something like

object FlinkJob {
  val uriLookup = mutable.HashMap.empty[String, Int]


  def main(args: Array[String]) {

    val stream: DataStream = ...

    stream.keybBy(1).timeWindow(..).fold(..)
    .window(..)
    .map(..).fold(..)
    .addSink(..)
  }
}

where the uriLookup hashmap gets updated inside the stream
transformation,
and is serialized in the step before the addSink

It works fine, however

Does the snapshotting mechanism in case of a node failure actually
serialize this map?

And out of curiousity, can I actually see what data exists inside the
snapshot data?

Thanks.

Bart

[stefanobaghino]

Hi Bart,

to make sure that the state is checkpointed you have to:

1. configure your Flink installation with a reliable state backend (optional for development, you can read more about it here)
2. explicitly enable checkpointing in your program (see how here — it's just a couple of lines of code)
3. extend your operators so that they checkpoint data, by implementing the `Checkpointed` interface or using an instance field (the semantics of the two approaches are slightly different, you can read more about it here)

When your data is checkpointed you can access the state if your operator implements the `RichFunction` interface (via an abstract class that wraps the operator you need to implement, like `RichMapFunction`).

For your need in particular, I don't know a way to checkpoint state shared between different operators; perhaps you can you refactor your code so that the state is encapsulated in an operator implementation and then moved through your pipeline as a parameter of the following operators. Would that work?

I apologize for just providing pointers to the docs in my reply but checkpointing deserves a good explanation and I feel the docs get the job done pretty well. I will gladly help you if you have any doubt.

Hope I've been of some help.

On Thu, Apr 7, 2016 at 3:41 PM, Bart van Deenen <[hidden email]> wrote:

Hi all

I'm having a datastream transformation, that updates a mutable
hashmap that exists outside of the stream.

So it's something like

object FlinkJob {
  val uriLookup = mutable.HashMap.empty[String, Int]


  def main(args: Array[String]) {

    val stream: DataStream = ...

    stream.keybBy(1).timeWindow(..).fold(..)
    .window(..)
    .map(..).fold(..)
    .addSink(..)
  }
}

where the uriLookup hashmap gets updated inside the stream
transformation,
and is serialized in the step before the addSink

It works fine, however

Does the snapshotting mechanism in case of a node failure actually
serialize this map?

And out of curiousity, can I actually see what data exists inside the
snapshot data?

Thanks.

Bart

--

[Aljoscha Krettek]

Hi,

good explanation and pointers!

I just want to add that the uriLookup table in your example is not really shared between your operator instances in a distributed setting. When serializing your transformations the current state of the HashMap is serialized with them because it is in the closure of the transformations. Then, on the cluster, the HashMap is serialized and all parallel instances basically work on their now local copy of the empty HashMap.

Cheers,

Aljoscha

On Thu, 7 Apr 2016 at 18:30 Stefano Baghino <[hidden email]> wrote:

Hi Bart,

to make sure that the state is checkpointed you have to:

1. configure your Flink installation with a reliable state backend (optional for development, you can read more about it here)
2. explicitly enable checkpointing in your program (see how here — it's just a couple of lines of code)
3. extend your operators so that they checkpoint data, by implementing the `Checkpointed` interface or using an instance field (the semantics of the two approaches are slightly different, you can read more about it here)

When your data is checkpointed you can access the state if your operator implements the `RichFunction` interface (via an abstract class that wraps the operator you need to implement, like `RichMapFunction`).

For your need in particular, I don't know a way to checkpoint state shared between different operators; perhaps you can you refactor your code so that the state is encapsulated in an operator implementation and then moved through your pipeline as a parameter of the following operators. Would that work?

I apologize for just providing pointers to the docs in my reply but checkpointing deserves a good explanation and I feel the docs get the job done pretty well. I will gladly help you if you have any doubt.

Hope I've been of some help.

On Thu, Apr 7, 2016 at 3:41 PM, Bart van Deenen <[hidden email]> wrote:

Hi all

I'm having a datastream transformation, that updates a mutable
hashmap that exists outside of the stream.

So it's something like

object FlinkJob {
  val uriLookup = mutable.HashMap.empty[String, Int]


  def main(args: Array[String]) {

    val stream: DataStream = ...

    stream.keybBy(1).timeWindow(..).fold(..)
    .window(..)
    .map(..).fold(..)
    .addSink(..)
  }
}

where the uriLookup hashmap gets updated inside the stream
transformation,
and is serialized in the step before the addSink

It works fine, however

Does the snapshotting mechanism in case of a node failure actually
serialize this map?

And out of curiousity, can I actually see what data exists inside the
snapshot data?

Thanks.

Bart

--

BR,

Stefano Baghino

Software Engineer @ Radicalbit

[Bart van Deenen]

Thanks all!

I was under the mistaken impression that Flink automagically did the snapshotting for me.

The info is really clear, I'll have no trouble implementing it.

 

Bart

 

 

On Thu, Apr 7, 2016, at 18:40, Aljoscha Krettek wrote:

Hi,

good explanation and pointers!

 

I just want to add that the uriLookup table in your example is not really shared between your operator instances in a distributed setting. When serializing your transformations the current state of the HashMap is serialized with them because it is in the closure of the transformations. Then, on the cluster, the HashMap is serialized and all parallel instances basically work on their now local copy of the empty HashMap.

 

Cheers,

Aljoscha

 

On Thu, 7 Apr 2016 at 18:30 Stefano Baghino <[hidden email]> wrote:

Hi Bart,

 

to make sure that the state is checkpointed you have to:

1. configure your Flink installation with a reliable state backend (optional for development, you can read more about it here)
   
2. explicitly enable checkpointing in your program (see how here — it's just a couple of lines of code)
   
3. extend your operators so that they checkpoint data, by implementing the `Checkpointed` interface or using an instance field (the semantics of the two approaches are slightly different, you can read more about it here)
   

When your data is checkpointed you can access the state if your operator implements the `RichFunction` interface (via an abstract class that wraps the operator you need to implement, like `RichMapFunction`).

 

For your need in particular, I don't know a way to checkpoint state shared between different operators; perhaps you can you refactor your code so that the state is encapsulated in an operator implementation and then moved through your pipeline as a parameter of the following operators. Would that work?

 

I apologize for just providing pointers to the docs in my reply but checkpointing deserves a good explanation and I feel the docs get the job done pretty well. I will gladly help you if you have any doubt.

 

Hope I've been of some help.

 

 

On Thu, Apr 7, 2016 at 3:41 PM, Bart van Deenen <[hidden email]> wrote:

Hi all

 

I'm having a datastream transformation, that updates a mutable

hashmap that exists outside of the stream.

 

So it's something like

 

object FlinkJob {

val uriLookup = mutable.HashMap.empty[String, Int]

 

 

def main(args: Array[String]) {

 

val stream: DataStream = ...

 

stream.keybBy(1).timeWindow(..).fold(..)

.window(..)

.map(..).fold(..)

.addSink(..)

}

}

 

where the uriLookup hashmap gets updated inside the stream

transformation,

and is serialized in the step before the addSink

 

It works fine, however

 

Does the snapshotting mechanism in case of a node failure actually

serialize this map?

 

And out of curiousity, can I actually see what data exists inside the

snapshot data?

 

Thanks.

Bart

 

 

 

--

BR,

Stefano Baghino

Software Engineer @ Radicalbit