By Franck Pachot
.
People often think of indexes as a way to optimize row filtering (“get item” faster and cheaper). But indexes are also about columns (“attribute projection”) like some kind of vertical partitioning. In relational (“SQL”) databases we often add more columns to the indexed key. This is called “covering” or “including” indexes, to avoid reading the whole row. The same is true in NoSQL. I’ll show in this post how, even when an index is not required to filter the items, because the primary key partitioning is sufficient, we may have to create a secondary index to reduce the cost of partial access to the item. Here is an example with AWS DynamoDB where the cost depends on I/O throughput.