resilient-pipeline

Zero dependencies · Node, Bun, Deno, Workers, browsers

Failures are inevitable.
Outages are optional.

Retry, circuit breaker, timeout, hedging, rate limiting and typed errors, composed into one pipeline. A Polly v8 API for TypeScript, from 4 KB.

Get started Open the playground

import { ResiliencePipelineBuilder } from 'resilient-pipeline';

const payments = new ResiliencePipelineBuilder({ name: 'payments' })
  .addFallback({ fallbackAction: () => queueForLater() })
  .addRetry({ maxRetryAttempts: 3, backoffType: 'exponential', useJitter: 'full' })
  .addCircuitBreaker({ failureRatio: 0.5, minimumThroughput: 20 })
  .addTimeout('2s')
  .build();

const receipt = await payments.execute(({ signal }) => charge(order, { signal }));

0dependencies

4 KBretry, min+gzip

14strategies

230+tests

ESM + CJStyped, tree-shakeable

What it does

Every failure mode, one pipeline.

Failures

Retry what is worth retrying. Stop hammering what is down.

Exponential backoff with jitter and Retry-After, a circuit breaker that opens before the outage spreads, and fallbacks that answer anyway. 4xx errors are never retried.

.addRetry({ backoffType: 'exponential', useJitter: 'full' })
.addCircuitBreaker({ consecutiveFailures: 5, breakDuration: '30s' })
.addFallback({ fallbackValue: cached })

Overload

Limits that tune themselves.

Bulkheads with priority queues, token buckets and sliding windows per user, and an adaptive concurrency limit that finds the right number from latency.

.addAdaptiveConcurrencyLimiter({ algorithm: 'gradient2' })
.addPartitionedRateLimiter({ partitionKey: (ctx) => tenant(ctx), … })

Latency

Cut the slow tail.

Hedging races a second attempt against a slow one. Timeouts abort the work, not just the wait, and deadlines flow into every nested call.

.addHedging({ delay: '200ms', maxHedgedAttempts: 2 })
.addTimeout('2s') // nested pipelines inherit the deadline

Typed errors

The compiler knows every way it can fail.

settle() returns a typed outcome. Outcome.match makes you handle each error code, and forgetting one is a compile error.

Outcome.match(await pipeline.settle(load), {
  ok: (user) => user,
  TIMEOUT: () => cached,
  CIRCUIT_OPEN: () => cached,
  UNEXPECTED: (e) => { throw e },
})

Distributed

One decision across every instance.

Share circuit state, rate limits, locks and caches through Redis. Atomic Lua scripts, fencing tokens, Cluster-safe keys.

extensions: [sharedCircuitState({ store: new RedisCircuitStateStore({ client }), key: 'payments' })]

Testing

Test a 30-second outage in milliseconds.

FakeTimeProvider drives every retry delay, timeout and open circuit. No sleeps, no flaky tests.

const time = new FakeTimeProvider();
const done = assert.rejects(pipeline.execute(neverResolves));
await time.runAll(); // 34 virtual seconds, instantly

Only here

What the others don't ship.

HedgingRace a second attempt against the slow tail.
Adaptive concurrencygradient2, vegas and aimd limits, no magic numbers.
Deadline propagationAn outer timeout cancels every nested call.
Typed errors on a Polly APIsettle() and exhaustive Outcome.match.
Retry budgetsRetries capped at a share of traffic, so no retry storms.
Redis-shared stateBreakers, limits, locks and caches across instances.

Full comparison

	resilient-pipeline	cockatiel	opossum	tenace
Retry · breaker · timeout · fallback	✓	✓	✓	✓
Hedging	✓	–	–	–
Adaptive concurrency	✓	–	–	–
Typed error channel	✓	–	–	✓
Shared state across instances	✓	–	–	✓
Fake clock for tests	✓	–	–	✓
Runs in browsers & Workers	✓	✓	✓	✓

The toolbox

Everything a call to a flaky dependency needs.

Failure

addRetryBackoff, jitter, budgets
addCircuitBreakerStop calling what is down
addTimeoutAbort, not just wait
addFallbackAnswer anyway
addHedgingRace the slow tail

Load

addConcurrencyLimiterBulkhead with priorities
addAdaptiveConcurrencyLimiterSelf-tuning limit
addRateLimiterToken bucket
addPartitionedRateLimiterPer user, per tenant
addCache · addCoalesceCache and single-flight

Types & flow

settle()Typed outcome, never throws
Outcome.match()Exhaustive by error code
executeStream()Commit after the first chunk
addLockLeases with fencing tokens
createBatcherDataLoader-style batching

Run & observe

/fetch · /axiosHTTP clients, per host
/http · /fastify · /koaInbound deadlines, shedding
/otel · /prometheusSpans and metrics
FakeTimeProviderTest without waiting
setGlobalChaosGame days on demand

In code

Real problems, a few lines each.

import { createResilientFetch } from 'resilient-pipeline/fetch';

// Per-host breakers, retries that honour Retry-After, idempotency-aware POSTs.
const http = createResilientFetch({ name: 'github', retryOn: [429, 503] });

const res = await http('https://api.github.com/repos/nodejs/node', {
  resilience: { timeout: '3s' },
});

// A timeout that really stops the query on the server.
const db = new ResiliencePipelineBuilder()
  .addRetry({ shouldHandle: ({ outcome }) => isSerializationFailure(outcome) })
  .addTimeout('1s')
  .build();

await db.execute(({ signal }) => transaction(signal, async (tx) => {
  // pg_cancel_backend runs when `signal` aborts
}));

// Retried until the first token arrives, committed after it.
for await (const token of pipeline.executeStream(({ signal }) =>
  openai.stream(prompt, { signal }))) {
  render(token);
}

import { FakeTimeProvider } from 'resilient-pipeline/testing';

const time = new FakeTimeProvider();
const pipeline = new ResiliencePipelineBuilder({ timeProvider: time })
  .addRetry({ delay: '10s', maxRetryAttempts: 3 })
  .addTimeout('1s')
  .build();

const done = assert.rejects(pipeline.execute(() => new Promise(() => {})));
await time.runAll(); // 34 virtual seconds
await done;

Get started

Wrap one call. Grow from there.

Start with a retry and a timeout. Add a breaker when the first incident teaches you why.

Read the quick start Break things in the playground