writeonce/docs/stories/language-runtime-database/refine/11-fibers.md
shoney.arickathil d24c705860 feat: iterations 19 + 17 — Float/Bytes scalars (.wob v5), library kind + internal/
- Float full stack: literals (fraction/exponent; `0..10` still a range), f64
  opcodes 34-41, @table column, WAL bit-exact replay, json fractions in and
  shortest-round-trip out. IEEE-quiet — FDIV never traps where DIV does.
- Bytes: a wo_str with its own class id, so alloc/free/copy are shared but no
  Text builtin accepts one; len/at/slice/eq/concat, base64 both ways, json
  boundary as base64; TEXT_COPY preserves the kind.
- No implicit Int/Float mixing (WO-E201 in the typechecker, not the emitter,
  which picks the opcode from one side and would misread the other).
- One IEEE deviation: float_cmp total order (NaN last, -0.0 == +0.0) for
  indexes and order-by, keys canonicalized to match. `?Float` nil is a
  reserved quiet NaN — the zero word is +0.0, WO_NIL_SCALAR's bits are -2.0.
- Renderer prefers fixed over exponential in 1e-6..1e21: pure shortest makes
  a price of 900.0 read `9e+02`. One renderer for interp/json/float_to_text.
- Fixed en route: lexer double-counted the leading digit; is_scalar_shaped
  took Float/Bytes as Int-shaped; Bytes ownership needed a shared heap-scalar
  predicate or temps never dropped; order-by bit-compared negatives backwards.
- Iteration 17: `kind = "library"` (absent = program; bad value = WO-E109),
  entry-less check mode retiring the `--emit` workaround, Go's `internal/` as
  WO-E108 at the consumer's `use`. Driver-only; VM/.wob/GC untouched.
- Framework reorg: internal/{parse,serve}.wo; http/form.wo split out to keep
  media_type/form_values public (parse.wo had grown public surface).
- Docs: link audit (97 -> 88 broken, conflict markers resolved, 2 duplicate
  stories removed), 00-code-review verified 26/27, iterations re-sequenced.
- Also carries the pre-staged pub(read)/using/#if work from the index.
- Gates: corpus 103/0, test_wal 156/0, web-app 26/0, oop-accept ALL MET.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-20 19:24:15 +02:00

104 lines
5.3 KiB
Markdown

# Iteration 11 — fibers (the 8+11 concurrency arc, part 2)
> Format: `product/story-iteration-template`. Part of
> [Story — one language, one runtime, one database, one binary](../00-story.md).
>
> **REFINED 2026-08-20** (developer decisions, no code): 8 and 11 ship as
> **one arc** — the DB becomes an actor on an owner shard
> ([iteration 8](08-shard-actor-runtime.md)'s decision), so serving
> shards must PARK on cross-shard replies, which is this iteration.
> The spawn-surface question below is SETTLED: one unified actor-address
> surface (`spawn` returns an address, fiber or remote alike; `send`
> moves ownership; same-heap sends take the cheap path). The arc's
> driving workload is [iteration 24: chat](24-chat-websocket-workload.md)
> — fiber-per-WebSocket-connection is the serving model that retires the
> framework's close-when-idle keep-alive policy.
>
> **STAGE-1 SUBSTANCE LANDED 2026-08-20** (branch `concurrency-arc`):
> reduction-budget fibers (back-edge accounting — the livelock lesson is
> in the plan's deviations), `spawn`/`send`/`actor M`, one-message-at-a-
> time delivery, main-return reap, fiber-trap isolation, and parked
> `net`/`time` builtins on the io_uring-first per-shard I/O plane
> (`WO_IO=uring|epoll`, epoll fallback proven). Demonstrated by
> `docs/examples/fibers` (`just fibers` 8/0). The shard-context criteria
> (TID assertions, cross-shard sends, blue-green drain reuse) close with
> the arc's stage 2.
## Goals
- Concurrency **inside** a shard the doctrine way: cooperative fibers
scheduled by the VM — a fiber is the interpreter state wovm already
isolates (register windows + frame stack + pc), made per-fiber. One
kernel task per core stays the rule; thousands of fibers run above it
where the kernel tops out at hundreds of threads (CFS O(log N),
~8 MiB + ~20 µs per thread vs an arena-allocated fiber context).
- **Preemption by reduction budget**, the Erlang shape: the dispatch loop
counts down and parks the fiber at zero — no signals, no safepoints, no
stack copying, deterministic replay.
- **Blocking builtins park instead of block** on server shards: the same
`net`/`time`/`fs` calls that block the thread in program mode hand their
fd to the shard's epoll/io_uring loop and resume on completion. One API,
same source text, no async/await, no function coloring — ever.
- Cross-fiber sends follow the iteration 8 rule — ownership moves — on the
cheaper same-heap path.
## Acceptance Criteria
- What to achieve?
- **Given** a shard running thousands of spawned fibers with one hot
compute loop among them,
- **when** the reduction budget expires,
- **then** the hot fiber parks, every other fiber makes progress
(starvation-free corpus fixture), and output is identical across
runs (deterministic scheduling).
- What to achieve?
- **Given** a fiber blocked in a stdlib builtin on a server shard,
- **when** the completion arrives on the shard's event loop,
- **then** the fiber resumes with the result while the shard served
other fibers in the interim — one OS thread, verified by TID.
- What to achieve?
- **Given** a parked fiber at shard shutdown,
- **when** the shard unwinds it,
- **then** every drop map runs (ASan zero leaks) — parked fibers die
as cleanly as trapped ones. (Iteration 26's blue-green drain reuses
exactly this unwind path.)
- What to achieve?
- **Given** TRACED objects (GC-ness inferred since 7b) referenced only
from a parked fiber's frames,
- **when** the per-shard mark-sweep collector scans,
- **then** parked fibers' frames are roots exactly as the live frame
stack is (`vm_gc_roots` grows fiber awareness) — nothing live is
collected, nothing dead survives. (Originally said `@gc`; restated
2026-08-20 in inference terms.)
## Out Of Scope
- Fiber migration across shards (ownership doctrine forbids it; cross-shard
is a message send, iteration 8).
- Priorities, timers, structured-concurrency policy — recipe-box
capabilities for a later `.wo` library, not runtime policy.
- Program mode: stays single-fiber in v1 (blocking legal, log-watcher
needs nothing more).
- async/await keyword — permanently rejected surface, not deferred.
## Info
- Research note: [`docs/plan/exploration/fibers/00-fibers.md`](../../../plan/exploration/fibers/00-fibers.md)
— kernel's-eye evidence (task_struct costs, CFS collapse at high task
counts) and the precedent survey (BEAM reductions adopted; Go stack
copying and Tokio coloring rejected; Loom's park-under-blocking-API
matches the stdlib posture).
- Vision origin: [blue-green vision §3](../../../plan/exploration/blue-green-vm/00-vision.md);
iteration 8's scheduler is the substrate this extends.
- Open questions to settle in the spec — REDUCED 2026-08-20: the spawn
surface is settled (the unified actor address, iteration 8 decision 2).
Still open for the arc's spec: budget size and check granularity,
parked-fiber drop semantics, run-queue fairness (FIFO v1), and how a
parked fiber's borrow state interacts with the shard's GC safepoints.
## Proposed Solution
- No implementation plan exists yet — this iteration starts with the
brainstorming → spec → writing-plans chain (the superpowers path every
prior iteration followed), then executes that plan. The research note
above is the brainstorm's entry material.