Offload functions to worker threads with shared memory primitives for Node.js.
chore: version 1.2.0
+33
-33
-32
.changeset/stream-atomics.md
-32
.changeset/stream-atomics.md
···
1
-
---
2
-
'moroutine': minor
3
-
---
4
-
5
-
Atomics-based backpressure for streams and channels; new `Int32Atomic#waitAsync` / `#notify`
6
-
7
-
**New public API on `Int32Atomic`**:
8
-
9
-
```ts
10
-
const slot = int32atomic();
11
-
// Park until the slot holds a value other than `expected`, with optional timeout.
12
-
await slot.waitAsync(0); // 'ok' | 'not-equal' | 'timed-out'
13
-
await slot.waitAsync(0, 100 /*ms*/);
14
-
// Wake waiters on this slot.
15
-
slot.notify(); // wake all
16
-
slot.notify(2); // wake at most 2
17
-
```
18
-
19
-
Matches `Atomics.waitAsync` / `Atomics.notify` semantics; returns `'not-equal'` synchronously (no microtask hop) if the slot already holds something different from `expected`.
20
-
21
-
**Streams and channels use atomics for backpressure**: streaming tasks and `channel()` no longer rely on `pause`/`resume` port messages for flow control. Internally, a `SharedArrayBuffer` tracks per-endpoint `inflight` and `state`; the producer parks on `Atomics.waitAsync` when the `highWaterMark` cap is reached and resumes when the consumer drains below it. Behavior-equivalent and much tighter under backpressure (worst-case buffering drops from ~10-20× `highWaterMark` to `highWaterMark + 1-2`).
22
-
23
-
**`highWaterMark` option now honored** for both `run(streamTask, { highWaterMark })` and `channel(src, { highWaterMark })`. Previously read and discarded in the streaming path.
24
-
25
-
**Significant throughput improvements** on streaming and channel fan-out:
26
-
27
-
- Worker → parent stream: ~66K items/s → ~750K items/s (11×)
28
-
- Round-trip stream: ~39K items/s → ~425K items/s (11×)
29
-
- Channel fan-out at 8 consumers: ~412K items/s → ~546K items/s (+33%)
30
-
- Per-task dispatch: ~289K → ~328K (+14%)
31
-
32
-
No changes to user-facing API shape beyond the `Int32Atomic` additions.
+32
CHANGELOG.md
+32
CHANGELOG.md
···
1
1
# moroutine
2
2
3
+
## 1.2.0
4
+
5
+
### Minor Changes
6
+
7
+
- 24741da: Atomics-based backpressure for streams and channels; new `Int32Atomic#waitAsync` / `#notify`
8
+
9
+
**New public API on `Int32Atomic`**:
10
+
11
+
```ts
12
+
const slot = int32atomic();
13
+
// Park until the slot holds a value other than `expected`, with optional timeout.
14
+
await slot.waitAsync(0); // 'ok' | 'not-equal' | 'timed-out'
15
+
await slot.waitAsync(0, 100 /*ms*/);
16
+
// Wake waiters on this slot.
17
+
slot.notify(); // wake all
18
+
slot.notify(2); // wake at most 2
19
+
```
20
+
21
+
Matches `Atomics.waitAsync` / `Atomics.notify` semantics; returns `'not-equal'` synchronously (no microtask hop) if the slot already holds something different from `expected`.
22
+
23
+
**Streams and channels use atomics for backpressure**: streaming tasks and `channel()` no longer rely on `pause`/`resume` port messages for flow control. Internally, a `SharedArrayBuffer` tracks per-endpoint `inflight` and `state`; the producer parks on `Atomics.waitAsync` when the `highWaterMark` cap is reached and resumes when the consumer drains below it. Behavior-equivalent and much tighter under backpressure (worst-case buffering drops from ~10-20× `highWaterMark` to `highWaterMark + 1-2`).
24
+
25
+
**`highWaterMark` option now honored** for both `run(streamTask, { highWaterMark })` and `channel(src, { highWaterMark })`. Previously read and discarded in the streaming path.
26
+
27
+
**Significant throughput improvements** on streaming and channel fan-out:
28
+
- Worker → parent stream: ~66K items/s → ~750K items/s (11×)
29
+
- Round-trip stream: ~39K items/s → ~425K items/s (11×)
30
+
- Channel fan-out at 8 consumers: ~412K items/s → ~546K items/s (+33%)
31
+
- Per-task dispatch: ~289K → ~328K (+14%)
32
+
33
+
No changes to user-facing API shape beyond the `Int32Atomic` additions.
34
+
3
35
## 1.1.0
4
36
5
37
### Minor Changes