Maria,
Are you just trying to understand the code or planning to make some customization to this? If the latter, then please open a jira for that. Would be good to understand your scenario. Thanks Bikas From: Siddharth Seth [mailto:[email protected]] Sent: Tuesday, April 5, 2016 10:39 AM To: [email protected] Subject: Re: Re: why only need to configure "the client node hadoop classpath"? Yes. It depends upon the YARN distributed cache. The relevant jars (tez.lib.uris, tez.aux.uris) are localized on each node using YARN local resources (distributed cache). Configs are constructed on the client node and then written out to HDFS, and localized again via the distributed cache. If you're looking for the implementation - TezClientUtils.createApplicationSubmissionContext. If additional jars need to be made available as part of a running DAG - that is done by 1) setting tez.aux.uris, or 2) programmatically using APIs on TezClient / during dag construction. On Tue, Apr 5, 2016 at 12:47 AM, Maria <[email protected] <mailto:[email protected]> > wrote: Much thanks for your quick reply. Siddharth~ :） OK, I got it. But another question arises：How can it propagate the libraries and configuration to other nodes? Does it depends on hadoop distributed cache? I still can not find the code logic. :( At 2016-04-05 10:27:55, "Siddharth Seth" <[email protected] <mailto:[email protected]> > wrote: Tez will run tasks throughout the cluster. However, it takes care of propagating the libraries and configuration to other nodes - which is why only the client needs to be configured. On Mon, Apr 4, 2016 at 6:51 PM, Maria <[email protected] <mailto:[email protected]> > wrote: Hi,all. Is it just because tez is a client-side plug-in, so we can only need to configure tez-libraries in the hadoop client node hadoop classpath,and do not need to send and configure them to any datanode? Is there any other reasons? Thanks for any reply. Maria.
