Date:

Estimated Time:less than a minute

Yarn Reflexions

Here are some Apache Yarn consideratoins.

Yarn is so great

Queue modifications

Some application needs extra cores/memory and need to be boosted somehow? Yarns allows on the fly adjustement of queues. In our case, we have nightly ETL and online users the day. It's then easy to dispatch the resources where needed, and without turning off running application.

GPU management

Hadoop 3 will provide a new dimension in yarn responsibility. Yarn will be able to allocate GPU resources for users. This will be of incredible help for machine learning programs and users to share and optimize access to them.

Yarn has some drawbacks

Applications logs

They are stored locally. Same case for temporary files on the local filesystem such map results. And there is no trivial way to store point to hdfs. This is an essential point when configuring the cluster, because the disk allocated for storing yarn user files (hive, spark....) is a real bottleneck. Depending on the cluster workload, keep in mind having a minimum of 2TO for this stuff, per machine.

Yarn superuser

There is now way to kill a process if you are not the owner of the process. This makes things a bit more complex for an admin to supervise such platform.

This page was last modified: