spark-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Bin Wang <wbi...@gmail.com>
Subject Re: Get only updated RDDs from or after updateStateBykey
Date Thu, 24 Sep 2015 14:16:58 GMT
Thanks, it seems good, though a little hack.

And here is another question. updateByKey compute on all the data from the
beginning, but in many situation, we just need to update the coming data.
This could be a big improve on speed and resource. Would this to be support
in the future?

Shixiong Zhu <zsxwing@gmail.com>于2015年9月24日周四 下午6:01写道:

> You can create connection like this:
>
>     val updateFunc = (iterator: Iterator[(String, Seq[Int], Option[Int])])
> => {
>       val dbConnection = create a db connection
>       iterator.flatMap { case (key, values, stateOption) =>
>         if (values.isEmpty) {
>           // don't access database
>         } else {
>           // update to new state and save to database
>         }
>         // return new state
>       }
>       TaskContext.get().addTaskCompletionListener(_ => db.disconnect())
>     }
>
>
> Best Regards,
> Shixiong Zhu
>
> 2015-09-24 17:42 GMT+08:00 Bin Wang <wbin00@gmail.com>:
>
>> It seems like a work around. But I don't know how to get the database
>> connection from the working nodes.
>>
>> Shixiong Zhu <zsxwing@gmail.com>于2015年9月24日周四 下午5:37写道:
>>
>>> Could you write your update func like this?
>>>
>>>     val updateFunc = (iterator: Iterator[(String, Seq[Int],
>>> Option[Int])]) => {
>>>       iterator.flatMap { case (key, values, stateOption) =>
>>>         if (values.isEmpty) {
>>>           // don't access database
>>>         } else {
>>>           // update to new state and save to database
>>>         }
>>>         // return new state
>>>       }
>>>     }
>>>
>>> and use this overload:
>>>
>>> def updateStateByKey[S: ClassTag](
>>>       updateFunc: (Seq[V], Option[S]) => Option[S],
>>>       partitioner: Partitioner
>>>     ): DStream[(K, S)]
>>>
>>> There is a JIRA: https://issues.apache.org/jira/browse/SPARK-2629 but
>>> doesn't have a doc now...
>>>
>>>
>>> Best Regards,
>>> Shixiong Zhu
>>>
>>> 2015-09-24 17:26 GMT+08:00 Bin Wang <wbin00@gmail.com>:
>>>
>>>> Data that are not updated should be saved earlier: while the data added
>>>> to the DStream at the first time, it should be considered as updated. So
>>>> save the same data again is a waste.
>>>>
>>>> What are the community is doing? Is there any doc or discussion that I
>>>> can look for? Thanks.
>>>>
>>>>
>>>>
>>>> Shixiong Zhu <zsxwing@gmail.com>于2015年9月24日周四 下午4:27写道:
>>>>
>>>>> For data that are not updated, where do you save? Or do you only want
>>>>> to avoid accessing database for those that are not updated?
>>>>>
>>>>> Besides,  the community is working on optimizing "updateStateBykey"'s
>>>>> performance. Hope it will be delivered soon.
>>>>>
>>>>> Best Regards,
>>>>> Shixiong Zhu
>>>>>
>>>>> 2015-09-24 13:45 GMT+08:00 Bin Wang <wbin00@gmail.com>:
>>>>>
>>>>>> I've read the source code and it seems to be impossible, but I'd
like
>>>>>> to confirm it.
>>>>>>
>>>>>> It is a very useful feature. For example, I need to store the state
>>>>>> of DStream into my database, in order to recovery them from next
redeploy.
>>>>>> But I only need to save the updated ones. Save all keys into database
is a
>>>>>> lot of waste.
>>>>>>
>>>>>> Through the source code, I think it could be add easily: StateDStream
>>>>>> can get prevStateRDD so that it can make a diff. Is there any chance
to add
>>>>>> this as an API of StateDStream? If so, I can work on this feature.
>>>>>>
>>>>>> If not possible, is there any work around or hack to do this by
>>>>>> myself?
>>>>>>
>>>>>
>>>>>
>>>
>

Mime
View raw message