I guess he used client model and the local Spark version is 1.5.2 but the standalone Spark version is 1.5.1. In other words, he used a 1.5.2 driver to talk with 1.5.1 executors.

On Mon, Feb 1, 2016 at 2:08 PM, Holden Karau <holden@pigscanfly.ca> wrote:
So I'm a little confused to exactly how this might have happened - but one quick guess is that maybe you've built an assembly jar with Spark core, can you mark it is a provided and or post your build file?

On Fri, Jan 29, 2016 at 7:35 AM, Ted Yu <yuzhihong@gmail.com> wrote:
I logged SPARK-13084

For the moment, please consider running with 1.5.2 on all the nodes.

On Fri, Jan 29, 2016 at 5:29 AM, Jason Plurad <pluradj@gmail.com> wrote:
I agree with you, Ted, if RDD had a serial version UID this might not be an issue. So that could be a JIRA to submit to help avoid version mismatches in future Spark versions, but that doesn't help my current situation between 1.5.1 and 1.5.2.

Any other ideas? Thanks.
On Thu, Jan 28, 2016 at 5:06 PM Ted Yu <yuzhihong@gmail.com> wrote:
I am not Scala expert.

RDD extends Serializable but doesn't have @SerialVersionUID() annotation.
This may explain what you described.

One approach is to add @SerialVersionUID so that RDD's have stable serial version UID.

Cheers

On Thu, Jan 28, 2016 at 1:38 PM, Jason Plurad <pluradj@gmail.com> wrote:
I've searched through the mailing list archive. It seems that if you try to run, for example, a Spark 1.5.2 program against a Spark 1.5.1 standalone server, you will run into an exception like this:

WARN  org.apache.spark.scheduler.TaskSetManager  - Lost task 0.0 in stage 0.0 (TID 0, 192.168.14.103): java.io.InvalidClassException: org.apache.spark.rdd.RDD; local class incompatible: stream classdesc serialVersionUID = -3343649307726848892, local class serialVersionUID = -3996494161745401652

If my application is using a library that builds against Spark 1.5.2, does that mean that my application is now tied to that same Spark standalone server version?

Is there a recommended way for that library to have a Spark dependency but keep it compatible against a wider set of versions, i.e. any version 1.5.x?

Thanks!





--