All telemetry is stored in one or more ring buffers (data might be split into high importance and low frequency, and low importance and high frequency for example) based on real-time configuration provided by the control plane. Data is drained from the ring buffers, filtered, and sent to bitdrift also based on real-time configuration. Data might be sent only when specific events occur. Data might be sent continuously only for a specific client or set of clients (e.g. alpha users). Any and all policies are possible. The limit is only the imagination of the user!
To illustrate why the ring buffer is so powerful, imagine being a mobile developer at Lyft. The app has thousands of screens, and accordingly a colossal amount of potential telemetry. You're curious about a bug on the ride request screen, but would like to know all of the detailed tracing information that leads up to the bug. By storing a bounded amount of verbose telemetry in the ring buffer and only sending it when the bug is hit, all of that detailed information critical to solving the bug becomes available when it otherwise would have been lost.
