Hi Xiaowei,
thanks for sharing this proposal. How would fault tolerance work with the
BatchFunction? Since the batch function seems to manage its own buffer,
users would also have to make sure that in-flight elements which are
buffered but not yet processed are checkpointed, wouldn't they?
Cheers,
Till
On Thu, Oct 20, 2016 at 9:50 AM, Xiaowei Jiang <xiaow...@gmail.com> wrote:
> Very often, it's more efficient to process a batch of records at once
> instead of processing them one by one. We can use window to achieve this
> functionality. However, window will store all records in states, which can
> be costly. It's desirable to have an efficient implementation of batch
> operator. The batch operator works per task and behave similarly to aligned
> windows. Here is an example of how the interface looks like to a user.
>
> interface BatchFunction {
> // add the record to the buffer
> // returns if the batch is ready to be flushed
> boolean addRecord(T record);
>
> // process all pending records in the buffer
> void flush(Collector collector) ;
> }
>
> DataStream ds = ...
> BatchFunction func = ...
> ds.batch(func);
>
> The operator calls addRecord for each record. The batch function saves the
> record in its own buffer. The addRecord returns if the pending buffer
> should be flushed. In that case, the operator invokes flush.
>
> Please share your thoughts. The corresponding JIRA is
> https://issues.apache.org/jira/browse/FLINK-4854
>
> Xiaowei
>
