I personally never cared about what minimal logging exists when loading billions of rows from CSV files in SQL Server 2008. I mean, it's large data already and having a large transaction log is going to happen. I too never cared about rolling anything back. The CSV file is the raw file, we just reprocess it in case of fire and the transaction log just gets replaced regardless along with ensuring there is a bulk load of space to fill up. Just the nature of the game. Not a huge deal.

When it comes to NoSQL style tables etc. You have Polybase. You have a window into the NoSQL data layer or NoSQL-Like data layers (i.e.: data buckets). That's all you really need. You can create external tables that you can just read into a physical table. Most of your transactional updates can happen outside of SQL Server where all you need to focus on is the INSERT and NEVER THE UPDATE. At least, that has been my approach most recently. Not going to update the data warehouse. I'm just going to INSERT with at most DROP and REPLACE with fresh new data that has been built elsewhere. That way SQL Server can focus on serving the data and providing a easy layer for analyzing the data with the magic of T-SQL, which is a good language that is extremely powerful outside of coding Python or learning other complicated languages to query.

Some things I would love to see though would be more machine learning integration as well service integration across the cloud based products. Some products have more advances in these than others. Would love to see them all come up to the same level so it's not a constant battle of switching to whatever one has more of that integration than the other. Like for example, would love Azure DB to have the same integration of ML as that of SQL Server 2016. Be happy to pay extra for it as long as we can further enhance what exists etc.