[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Limit on number of files to read for Dataset

Thanks all for your responses. I am now much more clearer on this.


On Tue, Aug 14, 2018 at 9:46 AM, Fabian Hueske <fhueske@xxxxxxxxx> wrote:

Flink InputFormats generate their InputSplits sequentially on the JobManager.
These splits are stored in the heap of the JM process and handed out to SourceTasks when they request them lazily.
Split assignment is done by a InputSplitAssigner, that can be customized. FileInputFormats typically use a LocatableInputSplitAssigner which tries to assign splits based on locality.

I see three potential problems:
1) InputSplit generation might take a long while. The JM is blocked until splits are generated.
2) All InputSplits need to be stored on the JM heap. You might need to assign more memory to the JM process.
3) Split assignment might take a while depending on the complexity of the InputSplitAssigner. You can implement a custom assigner to make this more efficient (from an assignment point of view).

Best, Fabian

2018-08-14 8:19 GMT+02:00 Jörn Franke <jornfranke@xxxxxxxxx>:
It causes more overhead (processes etc) which might make it slower. Furthermore if you have them stored on HDFS then the bottleneck is the namenode which will have to answer millions of requests. 
The latter point will change in future Hadoop versions with http://ozone.hadoop.apache.org/

On 13. Aug 2018, at 21:01, Darshan Singh <darshan.meel@xxxxxxxxx> wrote:

Hi Guys,

Is there a limit on number of files flink dataset can read? My question is will there be any sort of issues if I have say millions of files to read to create single dataset.