ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Franciscus, Naden" <Naden.Francis...@team.telstra.com>
Subject Re: Spark/YARN
Date Tue, 03 May 2016 02:00:37 GMT

Just as a follow up to this:

IgniteConfiguration has properties e.g. CountdownLatch which aren't serializable and HDFS
Path is not serializable either so there's no clean way to
construct a IgniteContext without breaking backwards compatibility.

Maybe update this page:

To mention that you can't use an IgniteContext within any Spark map, filter etc. So code such
as this e.g. Using BinaryObjects will not work.

val pairRdd = rdd.map(x => {
  val builder = igniteConext.ignite.binary.builder("DT1")
  builder.setField("id", x.toString)
  builder.setField("name", "test-" + x.toString)
   val binObj = builder.build
}).zipWithIndex.map(r => (r._2, r._1))

Shame that Ignite doesn’t really work on Spark. Hopefully one day !

From: Naden Franciscus <naden.franciscus@team.telstra.com<mailto:naden.franciscus@team.telstra.com>>
Date: Monday, 2 May 2016 at 4:32 PM
To: "dev@ignite.apache.org<mailto:dev@ignite.apache.org>" <dev@ignite.apache.org<mailto:dev@ignite.apache.org>>
Subject: Spark/YARN

Hey guys,

So it looks like using Spark/Ignite on YARN together simply doesn't work.

Many of us have Hadoop appliances where we aren't allowed to install anything on the nodes.
So the only option is YARN which
barring a few bugs seems to work okay. But the IgniteContext within Spark doesn't allow you
to read configuration files from YARN.

So since you allow users to pass in an IgniteConfiguration we have tried to manually set configuration
on the POJOs:

But during any Spark distributed operation it will attempt to serialise this which is not
possible since most of the classes contained
within IgniteConfiguration e.g. TcpDiscoverySpi are not serializable.

I am going to go through and see how many classes will need to be marked serializable (could
be dozens) but a call will need to be made:

 1.  Mark everything within IgniteConfiguration as serializable.
 2.  Force ALL users of IgniteContext to either read config from HDFS or from a Local Filesystem.
Both will go through Spring layer.

What's the best way to get a decision on this ?


View raw message