Enhanecement request to have file connector accept a list of hosts, and not just a single host name
Client uses orc and Avro HDFS formats, which in turn use File Connector stage instead of Big Data File stage.
The file connector leverages WebHDFS and has a â€œhostâ€ property that must be explicitly
set to point to an active name node, in order for it to function.
Since any one of two (or more) name nodes may be active on this client's EDL at any
given time, the workaround implemented was to poll the available name nodes and seeing which one is active, just prior to launching jobs.
The client has various flavors of this scheme, with all having the following commonality:
a. They are somewhat fragile since the name node can always fail over during, or right after the poll
b. There is a balancing act between polling more often and cluttering your job stream with repeated polling calls and polling less often and inviting job failures when the name node fails over and the next job uses the no-longer-active name node
Datastage cannot dynamically find the active namenode, and Datastage sequences will fail due to name node failover in the middle of the sequence.
What the client needs is for the file connector to accept a list of hosts, and not just a
single host name, in order to handle failover scenarios appropriately. The expectation is that this scenario is common in a real-world Hadoop implementation, and do so â€œout of the boxâ€.
Bank of Nova Scotia