Memory consumption optimized from 2.7G down to 1.5G on full dataset:
* Record index is not necessary in runtime storage since it it
needed for update operations only
* Store compact records as tuples instead of dicts
Co-author: Yuriy Taraday <yorik.sar@gmail.com>
Change-Id: I9a361e9373b5ed4fada527959a604cf09a02e5b9