The way you can consume the data is, in an incremental approach from the learning curve point of view:
Working with the raw data, either "locally" (i.e. logging into the Head Node of the cluster) by using the Hadoop commands, either "remotely" by using the WebHDFS/HttpFS REST API. Please observe within this approach you have to implement whichever analyzing logic you need, since Cosmos only allows you to manage, as said, raw data.
Working with Hive in order to query the data in a SQL-like approach. Again, you can do it locally by invoking the Hive CLI, or remotely by implementing your own Hive client in Java (there are some other languages) using the Hive libraries.
Working with MapReduce (MR) in order to implement strong analysis. In order to do this, you'll have to create your own MR-based application (typically in Java) and run it locally. Once you are done with the local run of the MR app, you can go with Oozie, which allows you to run such MR apps in a remote way.
My advice is you start with Hive (the step 1 is easy but does not provide any analyzing capabilities), first locally trying to execute some Hive queries, then remotely implementing your own client. If this kind of analysis is not enough for you, then move to MapReduce and Oozie.
All the documentation regarding Cosmos can be found in the FI-WARE Catalogue of enablers. Within this documentation, I would highlight: