[ttraub]

If network latency is a significant factor, and server storage and processing resources are sufficient, then the author's example of an 800 column table needs redesign.

Break that monster into smaller, more manageable tables and architect your database such that you can obtain exactly the data you need with well indexed joins, cached lookups, etc.

[jerzyt]

800 columns in an OLTP is insane, that was my reaction when I read the OP. But in analytics it's not so unusual, although at that point a SQL DB is probably not the right solution anyway, and the data scientists probably do need all the columns.