Good read. This is an interesting take. Although, I was on the edge reading about the "segment upload vs rotate race". Those timing heuristics work but might result in profound bugs.

Additionally, it feels like the sidecar is doing a race with the native broker in almost all formats and functions (dir watch, assignment updates, rotation gotchas, forcing consumers to set high rotate delays, consumer modes etc). How does Pinterest plan to evolve and maintain the broker-decoupled approach going forward?

Additionally, What were the cost savings with the tiered approach and was it worth it?