The uptake of Pandora and Spotify show that listeners really like a
"just hit play" experience: start with some songs you like, hit play,
and Good Stuff comes out the speakers. Sometimes music you know,
sometimes music you don't. Sometimes it plays the same track multiple
times in an hour, other times you hear a track once and never again.
It's like a radio with infinite stations. You get the picture.
Now, the primary reason they made their service this way was for cost
savings: the license to do that is way cheaper (though still cripplingly
expensive). It wasn't a technical reason that drove their design, it
was a financial reason.
However, recognizing that the result is acceptable and popular with
users, why not exploit this fact to address P2P's #1 shortcoming:
download start times?
P2P can and should be faster and more reliable than HTTP streams.
However, its complexity causes it to "start" downloading slower (because
connection setup often involves finding peers, doing NAT penetration,
onionskin routing, etc). Furthermore, even though P2P can sustain very
high throughputs, because it's coming from multiple sources the actual
stream is "jittery". This means if you want to do true streaming, you
need large buffers -- again, delaying startup time. For these reasons,
P2P has typically ceded the entire "on demand" streaming experience to
webservers, instead focusing on downloading (where out-of-order delivery
can be employed to maximize throughput and swarm health).
But by using a Pandora/Spotify-like "dynamic playlist" experience, the
application would alleviate the requirement for instantly playing any
specific song and just instead focus on instantly playing whatever's
available -- while going out and getting more in the background.
I see (at least) two layers:
1) Player layer. All it does is look at the songs on your hard drive
and decide:
- What song should I play next?
- What songs would I like to play, but don't have?
It could do this by calling central services, or by checking a DHT, or
whatever. Its recommendation engine would likely be populated
explicitly by users clicking "thumbs up/down" on given songs, and
implicitly by recording which songs uses skip versus allow to pay to
completion.
2) Transport layer. All it does is look at the "what would I like to
play?" file output by layer 1, go download it, and dump it onto the hard
drive in a place layer 1 can find it.
Layer 1 is completely legal. All it does is assemble a dynamic playlist
of the songs you own, on your hard drive. It works perfectly well with
MP3s ripped from your CDs, purchased from Amazon or iTunes, etc.
Layer 2 is less clearly legal. It just automatically downloads whatever
is suggested by Layer 1 as "boy, I wish I could play these songs..."
There'd need to be some way to convert song titles into magnet links,
which could then be pulled off the standard uTorrent/Azereus DHTs.
The result is a totally free equivalent to Pandora or Spotify.
Why doesn't someone build this?
-david
_______________________________________________
p2p-hackers mailing list
[email protected]
http://lists.zooko.com/mailman/listinfo/p2p-hackers
