Migrating to 2.2

Tako 2.2.0 adds one field to PerThreadConfig and changes how the per-thread server spreads connections across workers. Handler, router, and transport code is unchanged. See the release notes for the complete release summary.

PerThreadConfig::balance_connections

PerThreadConfig gains balance_connections: bool, which defaults to true. A struct literal that lists every field without ..PerThreadConfig::default() no longer compiles; add the field or fill the rest from the default:

use tako::PerThreadConfig;

let config = PerThreadConfig {
  workers: 4,
  ..PerThreadConfig::default()
};

With the flag set, a Tokio per-thread worker that accepts a connection while it serves more live connections than the least busy worker hands the socket to that worker. This evens out the hash-based SO_REUSEPORT spread on Linux, and on macOS, where the kernel sends every connection to one listener, all workers now serve traffic. Set balance_connections: false to keep kernel-only distribution. The Compio flavour (per-thread-compio) does not balance connections.

Performance changes

These need no code changes:

• The Tokio HTTP/1 servers (plain, TLS, Unix, vsock, and PROXY protocol, plus per-thread workers) reuse one header-read timer per connection instead of registering a new Tokio timer for every request. On the multi-threaded runtime this removes contention on Tokio's time driver under high concurrency. header_read_timeout keeps its meaning.
• Requests no longer update reference counts that every worker thread shares: each connection holds its own router handle, MatchedPath values come from a per-thread copy of the route template (the extension still holds an Arc<str>), and per-thread workers watch a per-worker shutdown token.

See Benchmarks for the measured effect.