Go's singleflight dedupes work, but who owns its lifetime?
Coalescing cache-missed S3 requests is the easy part; the trap is that the first caller's context cancels the fetch for everyone. This deep dive from IsleDB's author walks through that footgun and a protocol that lets each waiter stop independently, cancels work when the last waiter leaves, and drains cleanly on shutdown. Concretely demonstrated in Go code, not theory.