In order for you to be able to search through Elasticsearch, the data must be stored in the Elasticsearch indices. The ES-Hadoop connector allows transfer of data between Hadoop and Elasticsearch, but Elasticsearch does not directly access Hadoop. If you require near real-time access to a subset of your data the best way is likely to feed it to Elasticsearch at the same time it is fed to Hadoop. Hadoop will still hold all data and be your primary data store.
If this is not possible and you need to write the data to Hadoop first, it is likely there will be a delay. How long this is depends on how you do the indexing, and using MapReduce jobs can as you initially pointed out be slow.