Faced with unprecedented growth in earth data volume and demand, NASA has developed the Earth Data Analytic Services (EDAS) framework, a high performance big data analytics and machine learning framework. This framework enables scientists to execute data processing workflows combining common analysis and forecast operations close to the massive data stores at NASA. The data is accessed in standard (NetCDF, HDF, etc.) formats in a POSIX file system and processed using vetted tools of earth data science, e.g. ESMF, CDAT, NCO, Keras, Tensorflow, etc. EDAS utilizes high performance parallel data access, a custom distributed array framework, and a streaming parallel in-memory workflow for efficiently processing huge datasets within limited memory spaces with interactive response times. EDAS services are accessed via a WPS API being developed in collaboration with the ESGF Compute Working Team to support server-side analytics for ESGF. The API can be accessed using direct web service calls, a Python script, a Unix-like shell client, or a JavaScript-based web application. New analytic operations can be developed in Python, Java, or Scala (with support for other languages planned). Client packages in Python, Java/Scala, or JavaScript contain everything needed to build and submit EDAS requests. The EDAS architecture brings together the tools, data storage, and high-performance computing required for timely analysis of large-scale data sets, where the data resides, to ultimately produce societal benefits. It is currently deployed at NASA in support of the Collaborative REAnalysis Technical Environment (CREATE) project, which centralizes numerous global reanalysis datasets onto a single advanced data analytics platform. This service enables decision makers to compare multiple reanalysis datasets and investigate trends, variability, and anomalies in earth system dynamics around the globe. EDAS services include configurable high performance neural network learning modules designed to operate on the products of EDAS workflows. As a science technology driver we have explored the capabilities of these services for long-range forecasting of the interannual variation of important regional scale seasonal cycles. Neural networks were trained to forecast All-India Summer Monsoon Rainfall (AISMR) one year in advance using (as input) the top 8-64 principal components of the global surface temperature and 200 hPa geopotential height fields from NASA’s MERRA2 and NOAA’s Twentieth Century Reanalyses. The promising results from these investigations illustrate the power of easily accessible machine learning services coupled to huge repositories of earth science data.