We were dealing with terabytes of data and 10's of billions of writes per month. Most tables had 4 to 6 indexes, resulting in multiple updates for each write. If your data sets and indexes are huge there is little locality of data. Each row returned may have several disk hits. If you can keep your table sizes small enough that the indexes can remain in memory, performance is mush better.

Even worse were the autogenerated queries that joined a dozen tables together. I saw some that were 1500 lines long.