I've just observed something quite odd.
I have created a table with data for routing, and this table contains data for a graph over six different hour conditions. So when performing shortest_path, I select only the segments that corresponds to the "hour" column that I currently want.
I have created an index on the hour column but not on the source/target columns, and afterwards performed an analyze.
But when querying for shortest_path, I was NOT getting the shortest path, but a path where the route was lead through edges with extremely high weights (1,000,000 (created to tell not to go this way) where other weights are in the range 1-100).
After a lot of debugging (and blaming both my own graphs, pgrouting and everything else), I tried to perform a "Vacuum Analyze" on the table, and suddenly the results are correct.
So it seems like PostgreSQL had some problems that lead to pgRouting determining the wrong shortest path.
Now I ask you, if you have seen something like this before, and how to make sure this does not happen? The creation task of the graph table is automatized and I can see that analyze was actually performed after creation of the table.