Modern Apache NiFi Load Balancing
Join the DZone community and get the full member experience.Join For Free
In today's Apache NiFi, there is a new and improved means of load balancing data between nodes in a cluster. With the introduction of NiFi 1.8.0, connection load balancing has been added between every processor in any connection. You now have an easy to set option for automatically load balancing between your nodes.
The legacy days of using Remote Process Groups to distribute the load between Apache NiFi nodes is over. For maximum flexibility, performance and ease, please make sure you upgrade your existing flows to use the built-in Connection Load Balancing.
If you are running a newer version of Apache NiFI or Cloudera Flow Management (CFM), you have had a better way of distributing processing between processors and servers. This is for versions of Apache NiFi 1.8.0 and higher, including the newest version 1.10.0.
Note: Remote Process Groups are no longer necessary for load balancing! Use an actual load-balanced connection instead! Remote Process Groups should only be used for distributing to other clusters.
You may also like: Apache NiFi 1.0.0: Zero-Master Clustering.
Apache NiFi Load Balancing
Since 2018, it's been an awesome feature: https://blogs.apache.org/nifi/entry/load-balancing-across-the-cluster.
We have a few options for Load Balancing, including "Round Robin," that, during failure conditions, allows data to be rebalanced to another node. This can rebalance thousands of flow files per second or more, depending on the flow file size. This is done to give a node the chance to reconnect and continue processing.
Data Distribution Strategies
Another option is to “Partition by Attribute” and “Single Node,” which will queue up data until that single node or partitioned node returns. You cannot pick which node in the cluster does that processing for portability purposes.
We need to be dynamic and elastic, so it just needs to be one node. This allows for “like data” to be sent to the same node in a cluster that may be necessary for certain use cases.
Using a custom Attribute Name for this routing can be powerful for Merges in table loading use cases. We can also choose to not load balance at all.
Elastic Scaling for Apache NiFi
An important new feature that was added to NiFi is to allow nodes to be decommissioned and disconnected from the cluster and all of their data offloaded. This is important for Kubernetes and dynamic scaling for elasticity.
Elastic Scaling is important for workloads that differ during the day or year like once an hour loads or weekly jobs — scale up to meet SLAs and deadlines, but scale down when possible to save cloud spend! Now, NiFi not only solves data problems but saves you money.
Apache NiFi Node Affinity
Remote Process Groups do not support node affinity. Node affinity is supported in our Partition by Attribute strategy and has many uses.
Remote Process Groups
To replace the former big use case, we used Remote Process Groups. We have a better solution, for the first connection, like ListSFTP that runs on one node, and the connections can then be "Round Robin."
Important Use Case
This load balancing feature of Apache NiFi shows the power of distributing a large dataset or unstructured data capture at the edge or other datacenter, split and transfer, then use attribute affinity to a node to reconstitute the data in a particular order.
So what happens is sometimes you have a large bulk data export from a system like a relational database dump in one multiple terabyte file. We need one NiFi node to load this file and then split it up into chunks, transfer it, and send it to nodes to process. Sometimes, ordering records will require you to use an attribute to keep related chunks (say the same Table) together on one node.
We also see this with a large zip file containing many files of many types. Often, there will be hundreds of files of the multiple types, and we may want to route to the same node based on filename root. That way, one NiFi node will be processing all the same file types or tables.
This is how trivial it is to implement and easy for any NiFi user to examine and see what is going on in an ETL process.
Opinions expressed by DZone contributors are their own.