A common concern I see is people being worried that the reaction to having "how much of our user base actually uses these features" answered will be to remove t

A common concern I see is people being worried that the reaction to having "how much of our user base actually uses these features" answered will be to remove the feature entirely, when in practice it is more of "we think implementation A of the feature is no longer in use and we have migrated everyone to implementation B, is that the case?" or "we have implementation A produce user visible effects, while we have implementation B run parallel to it to detect divergences, have any been detected in the wild in the past X months?". In Rust we've had large migrations like that (the "new" borrow checker comes to mind) and we had a really long periods of time where they were tested against the latest crate versions on crates.io until crater came back clean, and even then we had bug reports about regressions in the wild only after it was released on stable. For me personally there's one big blind spot when testing only against published code, no matter how big the corpus is: humans are excellent fuzzers and the malformed code they write and try to compile is hard to replicate. Having visibility into uncommon cases that are only visible on users machines would be incredibly useful. An example of this could be the "botched" 1.52.0 Rust release[1], where invalid incremental compilations were changed from silent to visible Internal Compiler Errors in nightly for several releases to the point where the team felt all of the outstanding incr comp bugs related to them had been addressed, but when turned on in stable immediately hit users in the real world, making it necessary to do an emergency dot-release reverting that change. If we had telemetry on stable compilers, instead of turning on the silent to ICE change we could have added a metric for the silent error and know ahead of time that the feature wasn't ready for prime time. With more telemetry we could have known what was causing this. Snooping over a user's shoulder while they hit the case could make identifying the bug trivial, but of course then that user would then turn around and ask me in unfriendly terms "who are you and how did you get in?" Arguments can be made against even implementing telemetry, and they can be compelling enough to elect against it, but shouting down the conversation from even happening is not helpful. [1]: https://blog.rust-lang.org/2021/05/10/Rust-1.52.1.html https://blog.rust-lang.org/2021/05/10/Rust-1.52.1.html