Skip to content

research results

Jump to bottom
Caleb Buxton edited this page Jul 5, 2026 · 1 revision

Raw experiment data

All runs on x86_64-darwin24, Cargo release profile, duckling.gemspec's pinned dependency versions (magnus 0.8.2, rb-sys 0.9.128, duckling 0.4.0 per Cargo.lock).

Structural pre-check: rb_fiber_scheduler_blocking_operation_extract availability

for v in 3.2.2 3.3.6 3.3.8 3.4.4 3.4.5 3.4.10 4.0.5; do
  grep -rl "rb_fiber_scheduler_blocking_operation_extract" ~/.rbenv/versions/$v/include/
done
Ruby version Symbol present in headers?
3.2.2 no
3.3.6 no
3.3.8 no
3.4.4 no
3.4.5 no
3.4.10 no
4.0.5 yes (ruby/fiber/scheduler.h)

Live have_func check (mkmf) confirmed: fails on 3.4.10 ("no"), succeeds on 4.0.5 ("yes"). io-event 1.19.1's ext/extconf.rb:

if have_func("rb_fiber_scheduler_blocking_operation_extract")
  if have_header("pthread.h")
    append_cflags(["-DHAVE_IO_EVENT_WORKER_POOL"])
    $srcs << "io/event/worker_pool.c"
  end
end

Async::Scheduler.new.respond_to?(:blocking_operation_wait): false by default, true with ASYNC_SCHEDULER_WORKER_POOL=true set — but only on Ruby 4.0.5 (where WorkerPool compiled at all); on 3.4.10 IO::Event.const_defined?(:WorkerPool) is false regardless of the env var.

Track 2 — Ruby 3.4.10, hand-rolled scheduler (test/spike_77_minimal_scheduler_test.rb)

Scheduler subclass counts blocking_operation_wait invocations directly; ticker Fiber ticks every 1ms for 40 ticks (20 before + 20 after triggering the parse call in a sibling Fiber).

Row (e) — control, flags: 0 (today's rb_thread_call_without_gvl)

Run blocking_operation_wait_calls max_gap (s) parse_duration (s)
seed 48753 0 0.0613 0.0604
seed 1 0 0.0596 0.0590
seed 2 0 0.0556 0.0556
seed 3 0 0.0597 0.0593
seed 4 0 0.0586 0.0585

Hook never fires; max_gap ≈ parse_duration in every run — full reactor stall, matching the known pre-#64 blocking signature.

Row (f) — decisive, RB_NOGVL_OFFLOAD_SAFE set

Run blocking_operation_wait_calls max_gap (s) parse_duration (s) gap as % of parse
seed 48753 1 0.0014 0.0604 2.3%
seed 1 1 0.0014 0.0588 2.4%
seed 2 1 0.0013 0.0533 2.4%
seed 3 1 0.0013 0.0556 2.3%
seed 4 1 0.0014 0.0567 2.5%

Hook fires exactly once per call, every run; max_gap stays flat at ~1.3-1.4ms regardless of parse_duration — the reactor does not stall.

Root-fiber variant (not a formal row — quick manual check)

Called Duckling::Native.parse_nogvl_offload directly from the root fiber (no Fiber.schedule), with a scheduler installed:

before (root fiber, scheduler set, no Fiber.schedule): calls=0
after root-fiber call: calls=0

The hook does not fire outside a fiber-scheduled call stack — consistent with how Duckling.parse would actually be called inside a reactor (always from a scheduled Fiber).

Track 1 — Ruby 4.0.5, real async gem (test/spike_77_falcon_fiber_blocking_test.rb)

ASYNC_SCHEDULER_WORKER_POOL=true set for all rows. Same ticker/parser structure as test/falcon_fiber_blocking_test.rb; allowance = max(TICK_INTERVAL + parse_duration * 0.5, 0.025).

Row SPIKE_NATIVE_METHOD SPIKE_USE_THREAD max_gap (s) parse_duration (s) allowance (s) PASS
(a) baseline parse 1 0.0014 0.1545 0.0783 true
(b) parse 0 0.1303 0.1292 0.0656 false
(c) key, run 1 parse_nogvl_offload 0 0.0012 0.1228 0.0624 true
(c) key, run 2 parse_nogvl_offload 0 0.0013 0.1270 0.0645 true
(c) key, run 3 parse_nogvl_offload 0 0.0013 0.1190 0.0605 true
(c) key, run 4 parse_nogvl_offload 0 0.0013 0.1330 0.0675 true
(d) sanity parse_nogvl_offload 1 0.0012 0.1332 0.0676 true

Row (b) reconfirms #64's existing necessity absent the flag, on Ruby 4.0.5 too. Row (c) — the key row — passes consistently across 4 repeats: the real async/io-event stack auto-offloads the call via the flag alone, no Ruby-level Thread wrapper, once on Ruby 4.0.

Correctness regression check

test/duckling_test.rb + test/duckling_comma_list_test.rb, run with Duckling::Native.parse aliased to call parse_nogvl_offload:

Ruby version Runs Assertions Failures Errors
3.4.10 9 55 0 0
4.0.5 9 55 0 0

No correctness regression from the alternate FFI dispatch on either Ruby version tested.

Gotcha encountered while running this spike

rake compile installs the compiled artifact to a single shared lib/duckling/duckling.bundle path regardless of Ruby version (there is no per-Ruby-version install path in this gem's build, unlike the per-Ruby-version tmp/<platform>/duckling/<ruby-version>/ build directories rb_sys does maintain internally). Switching RBENV_VERSION without re-running rake compile first loads whichever Ruby version's binary was last installed there, which — being a genuine ABI mismatch between Ruby major/minor versions — surfaced as a confusing TypeError: no implicit conversion of Time into Time rather than a load error. Always re-compile immediately after switching RBENV_VERSION, before running anything directly (rake test's task test: :compile prerequisite already does this automatically).

Clone this wiki locally