IGFS was developed with the thinking that it would be a solution for Hadoop acceleration. However, it was discovered in practice that the performance benefits it gives are insignificant for production deployments. For instance, based on my experience IGFS combined with Hadoop Accelerator might show 20%, 30% sometimes 2x or even zero performance acceleration. Plus, it required notable integration efforts.
IGFS failed to show *consistent* order of magnitude performance gains since it's not enough just to put data in RAM. Your RAM-based storage has to be tightly coupled with APIs used by the applications. With IGFS the storage was Ignite while the APIs were developed by Hive, Impala, Pig, MapReduce, etc.
That's why for Hadoop offloading use cases and real-time analytics deploy Ignite in one of its standard configurations - Ignite with native persistence enabled. Use Ignite SQL, compute grid, ML for the data located in Ignite and Hadoop frameworks for HDFS data sets. Consider Spark as a generic API that can go and merge data stored in both databases. More details are in this discussion: