sodonnel commented on pull request #1291: URL: https://github.com/apache/hadoop-ozone/pull/1291#issuecomment-671532820
I wanted to add a few more thoughts on this issue as I have discussed it with others. I believe there is a bug (which this PR is intended to fix), as on a cluster with many racks and nodes all healthy, it is currently possible for the only remaining nodes with capacity for more pipelines to be on a single rack. This means that the vast majority of pipelines will be rack aware, but to fill the cluster pipeline capacity, a few non-rack aware pipelines can get created. That should not be allowed to happen, as it will give additional work to replication manager for all the 'bad' pipelines. Then consider the badly configured clusters I mentioned in the opening comment. 10 nodes on 1 rack and 1 node on the other rack. My proposal is that the pipeline capacity of the cluster is limited by the single node on the second rack. If we don't do this, and limit the pipeline capacity, then all the non-rack-aware piplines will create containers all on rack 1. Then replication manager will need to replicate them all to the only possible destination on rack 2. Almost every closed container will then need to be replicated putting a lot of pressure on the cluster and the single node on rack2 will fill up. This really is a corner case for a small 2 rack cluster which is badly configured. Note that HDFS has similar problems to this - racks need to be fairly well balanced in terms of node count, otherwise the nodes on the smaller racks tend to fill up faster. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: ozone-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: ozone-issues-h...@hadoop.apache.org
