OFCA

Ordered concurrent work for Node.js >=22.

Why

Stream processing does not have to mean processing one item at a time. When each item can be handled independently, waiting for one result before starting the next leaves useful work idle.

How

OFCA applies a work step to every input chunk with a bounded amount of concurrency, then combines completed results in their original input order. It works with arrays, sync and async iterables, and Node Readables through async iteration—without being a stream class.

It sits between Array.prototype.map and Transform streams: it keeps map-like work composition while accepting unbounded data from streams.

What

OFCA (Ordered Flow with Concurrent Arrival) is a small standalone ESM library with zero runtime dependencies. It provides:

• ofca(source, mapper, { concurrency }) for ordered asynchronous processing;
• ofcaSync(source, stages) for synchronous iterable processing;
• compose(...stages) for ordered unary stage composition; and
• DROP for intentionally omitting an output.

import { DROP, compose, ofca, ofcaSync } from "@scramjet/ofca";

const doubleThenAdd = compose((value) => value * 2, (value) => value + 1);
console.log(ofcaSync([1, 2], doubleThenAdd)); // [3, 5]

const output = [];
for await (const value of ofca([1, 2, 3], async (value) => (
  value === 2 ? DROP : value * 10
), { concurrency: 2 })) {
  output.push(value);
}
console.log(output); // [10, 30]

See DEVELOPMENT.md for development instructions.

Representative benchmark

Run npm run benchmark to measure the current checkout. Results are configuration-specific (Node version, machine, workload, and concurrency) and are not an intrinsic or universal speedup claim.

What to expect

On Node v22.22.1, a 10,000-item workload with 16 async stages, small deterministic jitter, and OFCA concurrency of 256 completed in:

Method	Time	Throughput
Array.map ×16 + Promise.all	140ms	71,267 items/s
OFCA	971ms	10,301 items/s
16 chained Node Transforms	14.4s	696 items/s

OFCA is slower than unbounded array admission, while processing stream-sized input about 14.8× faster than the equivalent chained Transform pipeline in this run.