To give a very broad example of one report: take 10,000 latest time series records from geographic location X. Now for each of those records load the latest 100 time series records from location Y and Z. Multiply all records from Y and Z together then divide by the value at location X. Throw in lots of business rules at every one of these steps so we have no chance of representing this as a tidy join operation and so that we can make no assumptions about the data we'll need to load to generate the report :)
Edit - returning multiple result sets from the server is always an option too.
Sorry, I probably edited my post while you were typing. The amount of data to load is theoretically unbounded (could load the entire data set for a big report) and it's very difficult to estimate ahead of time which data needs to be loaded. This is because the data in the time series dictates which secondary sets of data will be loaded on a record-by-record basis.
Yeah, this is the unfortunate conclusion that I arrived at too. I'm looking at making these reports go through a job queue which e-mails the result to the user when it's done. This would allow me to develop much more complex, but much slower queries in C#.