Turned out that datediff(millisecond, ... can only return durations for
up to less than 4 weeks, causing some problems. Turning it into datediff_big.
Diagnosing this was difficult because the insert-exec together with try-catch-rollback
ended up obscuring the original error from logs. We had this problem during
release 3 too. To avoid this in the future, we copied the code in
sweep() into sweep_loop() instead of doing a procedure call. This seemed
like overall the cleanest approach towards users (in MSSQL, you can't
return tables from functions and any other approach would somehow be more
invasive.
To compensate (and be able to test), if you pass duration=0 to sweep_loop()
it should always do exactly 1 iteration. Perhaps in the future we just
remove sweep() from the library and require that you use sweep_loop(duration=0)
instead, but keeping it for now.
Fixed a bug in a join in sweep()