The insertion slowdown is typically due to several different factors: * Often the columns are stored sorted so insertions can cause a lot of IO as the data get

The insertion slowdown is typically due to several different factors: * Often the columns are stored sorted so insertions can cause a lot of IO as the data get's shuffled around if the insert needs to happen in the middle of the current data. (This is less of a problem if your writes only ever have to append.) * Inserting a single row for a columnar store sometimes means that you have to write each field into multiple different places. While in a row based stores you can often write all the fields all at once. The upshot is the inserts or updates on a columnar store can result in more io operations for a columnar store than a row store which translates to slower writes to disk. However you get gains in for analytical read workloads for the reasons that others have commented about here.