I was reading the draft Go 1.28 release notes and noticed a change in how the HTTP/2 server handles its connection-serving goroutines.

I benchmarked the change with 1,000 and 5,000 idle HTTP/2 connections, comparing the same Go development binary with parking enabled and disabled.

With 5,000 connections, parking reduced RSS by 85-93 MiB across three runs. The server also had 5,000 fewer goroutines.

These results are from a development revision, not a final release, and the benchmark focuses on mostly idle connections.

Full methodology, charts, and reproducible benchmark:
https://abdolirad.me/articles/go-1-28-http2-parking-memory/