Fastlane broadcasts

[palkan]

Context

Action Cable decodes and re-encodes every broadcasted message for every client which is unnecessary in most cases. This PR provides a fastlane_broadcasts configuration option to skip double encoding and transmit broadcast messages as is to clients not using custom stream coders or user-defined callbacks.

This is a continuation of the previous work: rails/rails#26999

Benchmarks

The cost of excess encoding heavily depends on the broadcasting payload size and nature. We consider three examples: a small JSON, a larger JSON, and a Turbo Stream "append" action.

Here are the results:

$ be ruby --yjit  benchmarks/broadcasting.rb

Running benchmark with N=10, M=100, adapter=async, fastlane_broadcasts_enabled=false, worker_pool_size=4

ruby 3.4.1 (2024-12-25 revision 48d4efcb85) +YJIT +PRISM [aarch64-linux]
Warming up --------------------------------------
small json (19 Bytes)
                        10.000 i/100ms
large json (80.5 KB)     1.000 i/100ms
turbo stream (2.95 KB)
                         1.000 i/100ms
Calculating -------------------------------------
small json (19 Bytes)
                        101.991 (±19.6%) i/s    (9.80 ms/i) -    980.000 in  10.053077s
large json (80.5 KB)      0.557 (± 0.0%) i/s     (1.80 s/i) -      6.000 in  10.786842s
turbo stream (2.95 KB)
                         16.618 (±24.1%) i/s   (60.18 ms/i) -    156.000 in  10.068446s

$ FASTLANE_BROADCASTS=1 be ruby --yjit  benchmarks/broadcasting.rb

Running benchmark with N=10, M=100, adapter=async, fastlane_broadcasts_enabled=true, worker_pool_size=4

ruby 3.4.1 (2024-12-25 revision 48d4efcb85) +YJIT +PRISM [aarch64-linux]
Warming up --------------------------------------
small json (19 Bytes)
                        38.000 i/100ms
large json (80.5 KB)     1.000 i/100ms
turbo stream (2.95 KB)
                         9.000 i/100ms
Calculating -------------------------------------
small json (19 Bytes)
                        327.354 (±25.0%) i/s    (3.05 ms/i) -      3.116k in  10.076621s
large json (80.5 KB)      3.826 (± 0.0%) i/s  (261.36 ms/i) -     38.000 in  10.019392s
turbo stream (2.95 KB)
                         85.029 (±31.8%) i/s   (11.76 ms/i) -    765.000 in  10.078424s

So, for small JSON payloads the fastlane version is ~3x faster, for larger JSON — ~7x faster, for HTML/Hotwire payloads — ~5x faster.

Here are Vernier profile screenshots:

• Fastlane implementation:

• Current implementation:

Integration benchmarks

These are the results of websocket-bench and k6 benchmarks from actioncable-next-playground:

• websocket-bench:

Clients	Baseline 95p RTT	Fastlane 95p RTT	AnyCable 95p RTT
200	91ms	31ms	91ms
400	124ms	43ms	85ms
600	172ms	59ms	108ms
800	204ms	80ms	123ms
1000	305ms	111ms	143ms
1200	291ms	133ms	191ms
1400	511ms	149ms	187ms
1600	569ms	172ms	182ms
1800	494ms	253ms	225ms
2000	619ms	327ms	208ms

• k6 benchmark.js:

Scenario	broadcast_duration p(95)	rtt p(95)
Baseline	68ms	50ms
Fastlane	21ms	20ms
AnyCable	21ms	18ms