I dont think there is an easy solution for your case. Here are my ideas:
1. Copy source data from cluster B to cluster A and ingest it normally (f.e: DistCp).
2. Mount hdfs from cluster B locally on middlemanagers and use 'local' inputSource in ingestion spec
3. If cluster B has webhdfs then maybe use HTTP inputSource
4. Deploy separately middleManagers for cluster B, but this will split deepstorage, idk.
5. Probably there is a way to override peon classpath/config