[morgo]

Morgan from the TiDB team here.

I agree that column and row store have very different characteristics, but what I think is worth mentioning is that some hybrid solutions actually store as both row and columnar and have a query optimizer that can pick between them. For example: Oracle DB In-Memory, SQL Server Columnstore index.

At the same event as this announcement, we also announced that we are working on TiFlash which will do similar. Stay tuned for a blog post with more details :-)

[georgewfraser]

As described in that paper, it’s not sufficient to simply store a second, columnar projection of the data to get good performance. You also need a block-oriented execution engine, which means you effectively have two separate databases operating side-by-side. This is a huge challenge and it’s not clear if it’s really worth it, since for logistical reasons you will nearly always operate a separate data warehouse doing mostly OLAP and production database doing mostly OLTP.