Bifalgorithm post

Let's design Spotify. At first, it sounds pretty simple. A user opens the app, searches for a song, presses play, and music starts. But behind that one tap, Spotify has to solve a few very different problems at massive scale. First, storing the music. Spotify has millions of tracks, and each one can exist in different audio qualities and formats. So when a song is uploaded, Spotify doesn't just keep one file. It processes that track into multiple encoded versions. Maybe a lower bitrate for slow mobile connections. A higher bitrate for premium users, different codecs for different devices. Then those audio files are stored in large distributed storage. Now let's talk about playback. When you press play, Spotify's backend first handles things like Who are you? Do you have premium? What song did you request? Are you allowed to play it in your country? Where did you stop last time? But the backend should not stream every audio byte itself. That would become a massive bottleneck. Instead, when Spotify knows what you're allowed to play, the actual audio can be delivered from servers much closer to you, usually through a CDN. So the control plane and the media delivery path are separated. The back end decides what should happen. The CDN delivers the heavy audio traffic. Now Spotify has another important problem. Music should start almost instantly. Nobody wants to press play and wait five seconds. So the client aggressively buffers a small amount of audio ahead of time. While you're listening to Second Ten, Spotify might already be downloading Second Twenty or Thirty. That buffer protects playback if your connection briefly becomes unstable. And if your network gets worse, Spotify can switch to a lower bitrate version so the music keeps playing instead of stopping. Now let's talk about search. If you type blinding lights, Spotify cannot scan millions of songs in a normal database every time. It needs a dedicated search index. That index stores things like track names, artists, albums, and keywords in a structure optimized for very fast lookup. Then there's recommendations. Spotify needs to answer questions like, what should this user listen to next? What belongs in Discover Weekly? What song should autoplay after this one? That requires a completely different system. Spotify can use listening history, skips, likes, playlists, artists, and similarities between users and songs to generate recommendations. Those recommendations are often pre-computed in the background because doing all of that work every time you open the app would be too expensive. Now imagine a new album from a huge artist drops at midnight. Millions of people might request the same songs at once. This is actually a great case for caching and CDNs. Popular tracks can be replicated closer to users, so Spotify doesn't keep fetching the same audio from some central storage system. And finally, Spotify has to keep your state synchronized across devices. You pause a song on your laptop, open Spotify on your phone, and ideally the phone already knows what you were listening to and where you stopped. So Spotify needs fast distributed systems for things like playback state, playlists, likes, and listening history. If you simplify the entire architecture, Spotify has really several systems working together. Object storage holds the audio. CDNs deliver the music. Search indexes find tracks. Recommendation systems decide what you might like. Databases store your account in playlists. And background pipelines process enormous amounts of listening data. The hard part about building Spotify isn't playing an MP3. It's making millions of users feel like every song in the world is sitting right there on their phone, ready to play instantly.

Let's design Spotify. At first, it sounds pretty simple. A user opens the app, searches for a song, presses play, and music starts. But behind that one tap, Sp…