Benchmarks

Measured via Bench.runAll on kernel hot paths:

Operation	Throughput	p50/op
TokenBucket:take	17.0M ops/s	58ns
Hooks:fire (3-validator chain)	6.6M ops/s	148ns
Bus:publish	5.0M ops/s	198ns
Signal:fire (1 handler)	4.5M ops/s	203ns
Scheduler schedule+step (production)	2.2M ops/s	354ns
Scheduler schedule+step (Studio, MicroProfiler zones)	2.1M ops/s	388ns
Serde schema encode (3 fixed fields, fast engine)	3.5M ops/s	285ns
Serde schema decode (3 fixed fields, fast engine)	2.7M ops/s	367ns
Serde struct delta encode (2 fields)	3.7M ops/s	270ns
Serde struct delta decode (2 fields)	3.4M ops/s	293ns

Reading the numbers

• The full intent security pipeline — rate limit + session check + 3 validators — costs well under a microsecond per packet. Rejecting garbage is effectively free; there is no performance argument for skipping validation.
• The always-on hot-task profiler (per-task timing attributed to script:line) costs ~33ns per task — origins are cached by function identity, and MicroProfiler zones run only in Studio unless a live server opts in via Scheduler.setMicroProfilerZones(true).
• The fast Serde engine vs the wire-engine path, same machine: fixed-schema encode 463ns → 285ns (38% faster), decode 514ns → 367ns, dynamic schemas 22–27% faster — identical bytes either way. See Serde & Base64.

Note

Numbers are from Bench.runAll on kernel hardware and move with Luau VM updates and machine load. Treat them as relative weights, not absolute promises — and re-run the bench suite locally when it matters: Benchmark your own systems.