+1 (Binding), huge thanks to all the other folks spent time to design, development, test and review this feature.
In addition to what Jonathan mentioned, all exposed APIs / CLIs are marked as unstable so there should not have backward-compatibility issues. Participated initial design and review all patches related to RM/scheduler, CLI/Rest API and documentations. Considering scope of this patch is relatively small (slightly over 200k) and Jonathan and folks from Linkedin has done lots of validations to new logics, I'm supportive to merge this into branch-2/branch-3.0 and trunk. Thanks, Wangda On Mon, Oct 2, 2017 at 11:23 AM, Zhe Zhang <z...@apache.org> wrote: > +1 (binding) > > LinkedIn has been running an internal version of OrgQueue for almost 2 > years. It is an essential feature for running YARN scheduler on clusters > shared by multiple internal organizations. > > I have participated in the YARN-5734 design and have regularly synced with > Jonathan on development and implementation details. I believe the feature > is of decent quality now and is ready to merge into trunk. > > On Mon, Oct 2, 2017 at 11:09 AM Jonathan Hung <jyhung2...@gmail.com> > wrote: > > > Hi all, > > > > From discussion at [1], I'd like to start a vote to merge feature branch > > YARN-5734 to trunk, branch-3.0, and branch-2. Vote will be 7 days, ending > > Monday Oct 9 at 11:00AM PDT. > > > > This branch adds a framework to the scheduler to allow scheduler > > configuration mutation on the fly, including a REST and CLI interface, > and > > an interface for the scheduler configuration backing store. Currently the > > capacity scheduler implements this framework. > > > > Umbrella is here (YARN-5734 > > <https://issues.apache.org/jira/browse/YARN-5734>), jenkins build is > here > > ( > > YARN-7241 <https://issues.apache.org/jira/browse/YARN-7241>). All > required > > tasks for this feature are committed. Since this feature changes RM only, > > we have tested this on a local RM setup with a suite of configuration > > changes with no issue so far. > > > > Key points: > > - The feature is turned off by default, and must be explicitly configured > > to turn on. When turned off, the behavior reverts back to the original > file > > based mechanism for changing scheduler configuration (i.e. yarn rmadmin > > -refreshQueues). > > - The framework was designed in a way to be extendable to other > schedulers > > (most notably FairScheduler). > > - A pluggable ACL policy (YARN-5949 > > <https://issues.apache.org/jira/browse/YARN-5949>) allows admins > > fine-grained control for who can change what configurations. > > - The configuration storage backend is also pluggable. Currently an > > in-memory, leveldb, and zookeeper implementation are supported. > > > > There were 15 subtasks completed for this feature. > > > > Huge thanks to everyone who helped with reviews, commits, guidance, and > > technical discussion/design, including Carlo Curino, Xuan Gong, Subru > > Krishnan, Min Shen, Konstantin Shvachko, Carl Steinbach, Wangda Tan, > Vinod > > Kumar Vavilapalli, Suja Viswesan, Zhe Zhang, Ye Zhou. > > > > [1] > > > > http://mail-archives.apache.org/mod_mbox/hadoop-yarn-dev/201709.mbox/% > 3CCAHzWLgfEAgczjcEOUCg-03ma3ROtO=pkec9dpggyx9rzf3n...@mail.gmail.com%3E > > > > Jonathan Hung > > > -- > Zhe Zhang > Apache Hadoop Committer > http://zhe-thoughts.github.io/about/ | @oldcap >
