If not configured properly networking could be a real pain in Hadoop. All nodes in an Hadoop cluster need to see each other and they requires DNS (with reverse) and NTP. You can choose to deploy your cluster inside "cluster" network and use a multi-homed edge node as a bridge between "cluster" and "public" network. But you need to understand how you would access your cluster data (e.g.: JDBC thru Hive, WebHDFS for HDFS files and so on...). An edge nodes doesn't grant you access to Ambari Web UI, Ambari API, etc.. so if you deploy such a config you need to open specific TCP/IP ports in order to grant users on the "public" network to access such a service (e.g. Ambari Views). On the edge node you can deploy all clients (HDFS, YARN, Oozie, Hive, etc...) and let users to access edge node using SSH. It really depends how you want to manage access to Hadoop services and which services you need to give access to your end users. You can use NFS gateway or Knox on the edge node but what are your needs? Please take a look at this links: Hadoop TCP/IP ports: http://docs.hortonworks.com/HDPDocuments/HDP2/HDP-... Hadoop IDC and firewalls: http://docs.hortonworks.com/HDPDocuments/HDP2/HDP-... NFS Gateway: http://docs.hortonworks.com/HDPDocuments/HDP2/HDP-... Knox: http://docs.hortonworks.com/HDPDocuments/HDP2/HDP-... ... View more

We had some performance issue with a low profile config (4 vCores, 8 GB RAM), expecially with Oozie. Right now we reccomend at least 4 vCores and 24 GB or RAM). If you're planning a IaaS deployment on Azure using as metadata repository SQL Azure start directly with a S2/S3 instance: if you use Oozie it's the minimum requirement. Pay also attention to Ranger: it's OK to Ranger admin and users but for audits you need to look carefully at DB sizes: it can easly grow up faster. Use a script to truncate the table or use a different instance. ... View more

I think it's also a good starting point to use Availability sets for master nodes and worker nodes. Another good point is about using one storage account for every nodes in the cluster in order to bypass IOPS limits for multiple VMs on the same Storage Account. You can also try to use Azure Data Lake Store (with adl://) in order to check the performance on the new Azure service. You also need to remember the maintenance windows of every Azure region according to your customers: some regions could be a good choice for new service availability (e.g.: US East 2) but not from a maintenance point of view (expecially for european customers). We also verified great differences between IaaS performance and PaaS (HDInsight) performance due to low read/write performance of the Blob Storage: with the first one (configured correctly) you can achieve best performance. ... View more