On one hand I liked Rdio because it was cheap and had a lot of niche music I liked. On the other hand their algorithms were shit.

E.g. I once indicated I liked some electronica by Marek Biliński, a Polish musician, and it took hours to stop the ensuing assault on my ears of 70's Polish pop music.

In general it took very little for their algorithm to start playing music I had no interest in, and a lot of effort to make it stop playing stuff I found awful.

Algorithmic music selection in general still seem to be in a very poor state - I've yet to find any that do great, presumably because a lot of them seem to treat it as a question of solving "people who listened to A also likes B". But it isn't. It's more a matter of "people who listened to some sequence of A,B and C, also listened to D relatively shortly afterwards without skipping/downvoting it" coupled with a lot of other checks (e.g. I might skip D even if it fits in perfectly with both my mood and A,B and C if I have heard it too often lately, or if it's a very loud/intensive track and it's late a night, and any number of othr factors).

I wonder how many of these systems try to take more factors into consideration but just aren't very good at it. And also how problematic the lack of more contextual knowledge is (e.g. I may make different selections when alone, vs. when a romantic partner vs. when with friends, even if the starting type is the same type of tracks - it doesn't mean my taste has changed). Many recommendation systems makes me vary about indicating a preference because they're not clear about how it'll be used and the ones I've used so far gives me little confidence they'll do the right thing...

rdio did not exist outside the US and Grooveshark was never really a legitimate service.