Re: Distribute DataSet to subset of nodes

Posted by Fabian Hueske-2 on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/Distribute-DataSet-to-subset-of-nodes-tp2814p2843.html

Hi Stefan,

the problem is that you cannot directly influence the scheduling of tasks to nodes to ensure that you can read the data that you put in the local filesystems of your nodes. HDFS gives a shared file system which means that each node can read data from anywhere in the cluster.

I assumed the data is small enough to broadcast because you want to keep it in memory.

Regarding your question. It is not guaranteed that two different tasks, each with parallelism 5, will be distributed to all 10 nodes (even if you have only 10 processing slots).

What would work is to have one map task with parallelism 10 and a Flink setup with 10 task managers on 10 machines with only one processing slot per TM. However, you won't be able to replicate the data to both sets of maps because you cannot know which task instance will be executed on which machine (you cannot distinguish the tasks of both task sets).

As I said, reading from local file system in a cluster and forcing task scheduling to specific nodes is quite tricky.

Cheers, Fabian

2015-09-15 23:15 GMT+02:00 Stefan Bunk <[hidden email]>:

Hi Fabian,

I think we might have a misunderstanding here. I have already copied the first file to five nodes, and the second file to five other nodes, outside of Flink. In the open() method of the operator, I just read that file via normal Java means. I do not see, why this is tricky or how HDFS should help here.

Then, I have a normal Flink DataSet, which I want to run through the operator (using the previously read data in the flatMap implementation). As I run the program several times, I do not want to broadcast the data every time, but rather just copy it on the nodes, and be fine with it.

Can you answer my question from above? If the setParallelism-method works and selects five nodes for the first flatMap and five _other_ nodes for the second flatMap, then that would be fine for me if there is no other easy solution.

Thanks for your help!

Best

Stefan

On 14 September 2015 at 22:28, Fabian Hueske <[hidden email]> wrote:

Hi Stefan,

forcing the scheduling of tasks to certain nodes and reading files from the local file system in a multi-node setup is actually quite tricky and requires a bit understanding of the internals.

It is possible and I can help you with that, but would recommend to use a shared filesystem such as HDFS if that is possible.

Best, Fabian

2015-09-14 19:16 GMT+02:00 Stefan Bunk <[hidden email]>:

Hi,

actually, I am distributing my data before the program starts, without using broadcast sets.

However, the approach should still work, under one condition:

DataSet mapped1 = data.flatMap(yourMap).withBroadcastSet(smallData1,"data").setParallelism(5);
DataSet mapped2 = data.flatMap(yourMap).withBroadcastSet(smallData2,"data").setParallelism(5);

Is it guaranteed, that this selects a disjoint set of nodes, i.e. five nodes for mapped1 and five other nodes for mapped2?

Is there any way of selecting the five nodes concretely? Currently, I have stored the first half of the data on nodes 1-5 and the second half on nodes 6-10. With this approach, I guess, nodes are selected randomly so I would have to copy both halves to all of the nodes.

Best,

Stefan