nopanic

The unified resilience toolkit for Python.

Retries, circuit breakers, timeouts, rate limits, bulkheads, hedged requests, retry budgets and deadlines: one decorator-based API, identical for sync and async code (LLM token streams included), with zero runtime dependencies.

from nopanic import retry, circuit_breaker, timeout, backoff

openai_breaker = circuit_breaker(failure_threshold=0.5, reset_timeout=15.0, name="openai")

@retry(attempts=4, on=ConnectionError, backoff=backoff.full_jitter(base=0.2, cap=10.0))
@openai_breaker
@timeout(30.0)
async def call_llm(prompt: str) -> str:
    ...

Why

Java has resilience4j. .NET has Polly. Python has a drawer of single-purpose parts:

Need	Today in Python	nopanic
Retry + backoff	tenacity / backoff / stamina	✅ retry
Circuit breaker	pybreaker (sync-oriented, aging)	✅ circuit_breaker (sliding window, sync + async)
Timeout	roll your own per framework	✅ timeout
Client-side rate limit	roll your own token bucket	✅ rate_limit
Concurrency bulkhead	raw semaphores	✅ bulkhead
Fallback / graceful degradation	try/except sprawl	✅ fallback
Hedged requests (tail latency)	nothing	✅ hedge
Adaptive (AIMD) rate limiting	nothing	✅ adaptive_rate_limit
Cache that survives outages	roll your own	✅ cache (stale-while-failing)
Retry budgets (stop retry storms)	nothing	✅ RetryBudget
Deadline propagation	nothing	✅ deadline
Streaming (LLM tokens)	nothing	✅ every stateful policy guards generators
One breaker / limiter per host or tenant	roll your own	✅ keyed
One event stream for all of it	per-library logging	✅ events.subscribe / events.log_to
All of the above, composable	none	✅ compose

These patterns only pay off when they compose: a timeout per attempt, retries that respect an open circuit, a fallback that catches whatever is left. Wiring that out of three libraries with three philosophies is exactly the code nobody wants to own. Every service that calls another service needs this, and in the LLM era every app calls flaky, rate-limited, high-latency remote APIs.

Install

pip install nopanic

Python 3.10+. No dependencies. Fully typed (py.typed). Works on asyncio, Trio and AnyIO.

Documentation: dagdelenemre.github.io/nopanic, with integration guides for httpx/requests, FastAPI and LLM SDKs.

Optional add-ons, each a separate package so the core stays dependency-free:

Package	Adds
nopanic-otel	OpenTelemetry metrics, span events and live gauges
nopanic-redis	Circuit breakers and rate limits shared across processes and hosts

The toolkit

Every policy is a decorator that works on both def and async def functions. Policies with shared state (breaker, bulkhead, rate limit) are created once and applied to every call site that talks to the same dependency. Every policy also takes an optional name= that shows up in its events.

retry: try again, politely

from nopanic import retry, backoff

@retry(
    attempts=5,
    on=(ConnectionError, TimeoutError),          # or a predicate: lambda e: ...
    backoff=backoff.full_jitter(base=0.1, cap=30.0),
    giveup=lambda e: getattr(e, "status", 0) == 401,   # don't retry hopeless errors
    before_sleep=lambda a: log.warning("attempt %d failed: %s", a.attempt, a.exception),
)
def fetch(): ...

On exhaustion the original exception is re-raised, so there is no wrapper to unwrap. Backoff strategies: fixed, exponential, full_jitter (default, AWS-style), decorrelated_jitter, or any iterable of floats.

Exceptions carrying a numeric retry_after attribute (build one from an HTTP 429's Retry-After header, or let CircuitOpen / RateLimited provide it) raise the wait to at least that long; disable with honor_retry_after=False.

Bound the total time, not just the attempt count, and retry on bad results for clients that do not raise:

@retry(attempts=6, max_time=20.0, on_result=lambda r: r.status_code in (429, 502, 503))
async def get(url): return await client.get(url)

max_time never starts an attempt past the deadline; when on_result retries run out, the last result is returned as-is.

Retry budgets: no more retry storms

Three attempts per call means up to three times the load on a dependency that is already struggling, from every client at once. A RetryBudget caps retries across calls, as a fraction of real traffic (the Finagle / gRPC / Envoy approach):

from nopanic import RetryBudget, retry

api_budget = RetryBudget(ratio=0.1)   # retries <= 10% of calls (+1/s floor)

@retry(attempts=4, budget=api_budget)
def get_user(uid): ...

@retry(attempts=4, budget=api_budget)   # one budget per downstream
def get_orders(uid): ...

Under a real outage your retry traffic shrinks to the floor instead of multiplying. When the budget is spent, the original exception is re-raised (retry.gave_up, reason budget).

circuit_breaker: stop hammering what's already down

from nopanic import circuit_breaker, CircuitOpen

payments = circuit_breaker(
    failure_threshold=0.5,   # open at >=50% failures...
    min_calls=10,            # ...once 10 outcomes are in the window
    window=30.0,             # sliding time window (seconds)
    reset_timeout=15.0,      # then let a probe through
    name="payments",
    on_state_change=lambda b, old, new: log.warning("%s: %s -> %s", b.name, old, new),
)

@payments
async def charge(order): ...

While open, calls fail instantly with CircuitOpen(retry_after=...) instead of stacking up on a dead dependency. One instance = one dependency; decorate as many functions with it as you like, and check payments.state anytime.

• Slow calls trip it too. slow_call_duration=2.0, slow_call_threshold=0.5 opens the breaker when half the calls take 2s or more: a dependency that answers too slowly is often as harmful as one that fails.
• Manual control. force_open() (maintenance, kill switch), disable() (let everything through, record nothing), reset().
• Health checks. payments.metrics gives a consistent snapshot: state, calls, failures, slow calls and both rates.
• Blocks, not just functions. with payments.guard(): / async with payments.guard():.

timeout: bound every attempt

@timeout(2.0)
async def lookup(): ...   # cancelled cleanly via the event loop

Sync functions run on a worker thread and are abandoned (not killed) on timeout. This is documented and deliberate, and still usually what you want at a system boundary. The worker sees the caller's contextvars, and a TimeoutError raised by your own code passes through untouched. Abandoned workers are capped (max_abandoned=64 per policy): once that many timed-out calls are still running, new calls fail fast instead of piling up threads behind a hung dependency. async with timeout(2.0).guard(): bounds a block of async code.

rate_limit: be a good client

rl = rate_limit(90, per=60.0, burst=10)   # 90 calls/min, bursts of 10

@rl
async def embed(text): ...

Token bucket; throttled calls wait by default (roughly FIFO), and a waiter that is cancelled gives its token back. Set max_wait to fail fast with RateLimited instead. rl.available_tokens shows the bucket; with rl.guard(): takes a token for a block.

bulkhead: contain the blast radius

bh = bulkhead(20, max_wait=0)   # at most 20 in flight, reject the 21st

@bh
async def render_report(): ...

bh.in_flight reports current usage; with bh.guard(): / async with bh.guard(): holds a slot for a block. Capacity is counted separately for sync callers and for each event loop.

fallback: degrade, don't die

@fallback(lambda exc: CACHED_ANSWER, on=(ConnectionError, CircuitOpen))
async def recommendations(user): ...

hedge: fix the p99, not the median

@hedge(delay=0.8, max_hedges=1)
async def complete(prompt): ...   # if slow after 800ms, race a duplicate call

The first success wins; losers are cancelled ("The Tail at Scale"). For idempotent async calls only. Pass on= with your transient errors so a permanent failure (a 400) propagates at once instead of being raced max_hedges more times.

adaptive_rate_limit: find the server's real limit

arl = adaptive_rate_limit(50, per=1.0, min_rate=5)   # start at 50/s, never below 5/s

@arl
async def call_api(): ...

Classic AIMD, the scheme TCP uses for congestion control: a 429/503 (or any exception carrying retry_after) cuts the rate multiplicatively, every success earns a small additive recovery, and an explicit Retry-After blocks the bucket for exactly that long. You converge on whatever rate the server actually sustains instead of hardcoding a guess. Inspect it live via arl.current_rate.

cache: an outage is not your outage

@cache(ttl=60.0, stale_ttl=3600.0, on=(ConnectionError, TimeoutError))
def exchange_rates(base): ...

Fresh values are served from memory; when a refresh fails, the expired value is served for up to stale_ttl more seconds instead of the error. Refreshes are single-flight: a hundred concurrent misses on one key make one call to the dependency, not a hundred. LRU-bounded, per-arguments keying, injectable clock.

deadline: one time budget for the whole operation

A timeout bounds one call; a deadline bounds everything inside a block, however many retries, waits and nested calls it takes:

from nopanic import deadline, time_remaining

async with deadline(5.0):
    profile = await get_profile(uid)    # retries, timeouts, rate-limit waits...
    orders = await get_orders(uid)      # ...all share the same 5 seconds

Policies inside cooperate: timeout clamps to the time remaining, retry gives up instead of sleeping past it, and rate_limit / bulkhead reject instead of queueing past it. Nested deadlines only shorten the budget. Hand the rest to your own clients with time_remaining(). deadline is also a policy: compose(deadline(10.0), retry(...), timeout(2.0)) gives every call its own budget.

Streams: LLM tokens, file chunks, paginated rows

Decorate a generator or async generator function and the policies guard the stream:

@retry(attempts=3, on=ConnectionError)
@timeout(10.0, first_item=30.0)    # 30s to the first token, then 10s per token
@llm_breaker                       # outcome recorded when the stream ends
async def stream_completion(prompt):
    async for token in client.stream(prompt):
        yield token

retry re-opens a stream that fails before its first item; once items were delivered, a failure propagates (they cannot be taken back). timeout bounds each item, never the consumer's own time between items. circuit_breaker, bulkhead, rate_limit and adaptive_rate_limit treat the whole stream as one call. Policies that cannot guard a stream (cache, fallback, hedge, deadline, and timeout on a sync generator) raise TypeError at decoration time instead of silently doing nothing. Close async streams you abandon early (contextlib.aclosing), as with any async generator.

keyed: one policy per host, tenant or model

from urllib.parse import urlsplit
from nopanic import keyed, circuit_breaker

per_host = keyed(
    lambda host: circuit_breaker(name=f"api-{host}", reset_timeout=10.0),
    key=lambda url, **_: urlsplit(url).hostname,
)

@per_host
async def fetch(url, **params): ...

per_host.get("eu.api.example.com").state

One region going down no longer trips the breaker for every region; one noisy tenant no longer starves the others' rate limit. Instances are created on first use and LRU-bounded (maxsize).

Observability: one stream for everything

import logging
from nopanic import events

events.log_to(logging.getLogger("resilience"))     # one line per event

cancel = events.subscribe(                         # or anything custom
    lambda e: metrics.increment(e.kind, tags={"name": e.name, "fn": e.function})
)

Every policy emits typed events (retry.attempt_failed, breaker.state_change, ratelimit.throttled, cache.stale_served, ...), each carrying the policy's name and the guarded function. With nopanic-otel, nopanic_otel.instrument() turns them into OpenTelemetry metrics and span events, and nopanic_otel.observe(...) exports live policy state as gauges. A listener is all it takes to wire OpenTelemetry, StatsD or metrics counters; listeners can never break the calls they observe, and with no listeners the overhead is a single attribute read. For dashboards and health checks, read state directly: breaker.metrics, bulkhead.in_flight, rate_limit.available_tokens, adaptive_rate_limit.current_rate, timeout.abandoned, cache.size, RetryBudget.available.

Testing code that uses nopanic

from nopanic import testing

def test_checkout_survives_a_flaky_gateway():
    with testing.instant() as sleeps:             # no real backoff or queueing
        assert checkout(order) == "paid"
    assert len(sleeps) == 2

def test_breaker_reports_opening():
    with testing.capture_events("breaker.") as seen:
        ...

Stateful policies also take an injectable clock= for fully deterministic tests.

asyncio, Trio and AnyIO

Every async policy runs on asyncio and Trio, and so on AnyIO with either backend, with nothing to configure: nopanic detects the running library per call and never imports trio itself. Cancellation is never a failure: a cancelled breaker probe frees its slot, a cancelled rate-limit waiter returns its token, and hedge under Trio raises the original exception, never an ExceptionGroup.

Across processes: nopanic-redis

import redis
from nopanic_redis import redis_circuit_breaker, redis_rate_limit

r = redis.Redis()
payments = redis_circuit_breaker(r, name="payments", reset_timeout=15.0)   # trips fleet-wide
openai_quota = redis_rate_limit(r, name="openai", rate=500, per=60.0)      # one shared quota

Atomic Lua scripts, sync and asyncio clients, fail-open when Redis itself is down. See the distributed guide.

Writing your own policy

nopanic.extend exposes the building blocks the built-in policies and the add-ons are made of (Policy, emit, deadline-aware and test-skippable sleep / asleep, validation helpers). Subclass Policy, implement _run_sync and _run_async, and your policy decorates, composes and reports events like the built-in ones.

Composition: the whole point

Stack decorators (innermost runs closest to the call), or name the stack once and reuse it:

from nopanic import compose, fallback, retry, circuit_breaker, timeout, backoff, CircuitOpen

llm = circuit_breaker(failure_threshold=0.5, min_calls=8, reset_timeout=20.0, name="llm")

resilient = compose(                              # first = outermost
    fallback(lambda e: "Sorry, try again later.", on=(ConnectionError, CircuitOpen, TimeoutError)),
    retry(attempts=3, on=(ConnectionError, TimeoutError), backoff=backoff.full_jitter(0.2)),
    llm,
    timeout(30.0),
)

@resilient
async def ask(prompt: str) -> str: ...

Reading inside-out: each attempt gets 30s -> outcomes feed the breaker -> transient failures retry with jitter -> anything left becomes a graceful answer.

No function to decorate? Guard a one-off call directly; this works on every policy and on compose stacks:

user = await resilient.acall(client.users.get, user_id)      # async
rows = payments_breaker.call(db.execute, "SELECT ...")       # sync

Recipe: retrying HTTP the right way

Retry transport errors, timeouts, 429 and 5xx; never retry other 4xx (they are permanent, retrying them is just noise):

import requests
from nopanic import retry, backoff

def _retriable(e):
    if isinstance(e, (requests.ConnectionError, requests.Timeout)):
        return True
    return (isinstance(e, requests.HTTPError) and e.response is not None
            and (e.response.status_code == 429 or e.response.status_code >= 500))

@retry(attempts=4, on=_retriable, backoff=backoff.full_jitter(base=0.2, cap=10.0))
def get_json(url):
    r = requests.get(url, timeout=10)   # always set a transport timeout too
    r.raise_for_status()
    return r.json()

Design principles

• Zero dependencies. A resilience library must not be a reliability risk itself.
• Sync and async are equals. One API; the wrapper flavour is chosen at decoration time, not per call.
• No wrapper exceptions. Your except ConnectionError: keeps working; policies only add their own precise signals (CircuitOpen, BulkheadFull, RateLimited).
• Composition over configuration. Small orthogonal policies + compose, instead of one mega-object with 40 kwargs.
• Testable time. Breakers and rate limiters take an injectable clock, and nopanic.testing.instant() skips every wait, so your test suite never sleeps.
• Fail loudly at decoration time. A policy that cannot guard what you decorated (a stream, a sync function for hedge) says so immediately.

Roadmap

• More integration guides (Django, Celery, gRPC)
• More shared-state backends for nopanic-redis-style policies (e.g. PostgreSQL)
• Prometheus exporter alongside nopanic-otel

Performance

Resilience must not become its own latency problem. Per-call overhead on the hot (success) path, single thread, measured with benchmarks/bench.py:

Policy	Overhead per call
retry	~0.2 us
rate_limit	~0.6 us
adaptive_rate_limit	~0.9 us
cache (fresh hit)	~0.9 us
circuit_breaker	~1.1 us
bulkhead	~1.3 us
events.emit, no listeners	~0.1 us

For scale: a fast HTTP round trip costs 5,000+ us, an LLM call millions. Numbers vary by machine; run the benchmark yourself. Design rules that keep it this way: success paths allocate nothing (retry builds its backoff iterator only after the first failure), the breaker's failure window is bucketed counters with constant memory no matter the traffic, event emission with zero listeners is a single attribute read, and locks are held for nanoseconds and never while user code runs.

Security & hardening

• Zero runtime dependencies means no transitive supply chain; CI runs pip-audit, ruff's flake8-bandit rules and a python -O smoke check (no assert in library code).
• All numeric parameters reject NaN/infinity/booleans at construction time: a bad config fails at import, not by silently never tripping a breaker in production.
• Observability hooks (before_sleep, on_state_change) and event listeners can never break the call they observe: their exceptions are logged and suppressed.
• Server-controlled delays are capped: a hostile Retry-After: 999999999 cannot park your client (retry_after_cap on retry, max_block on adaptive_rate_limit, both defaulting to 60s).
• hedge cancels and reaps losing attempts, leaving no orphaned tasks and no "exception was never retrieved" noise.
• Under untrusted load, set max_wait on rate_limit/bulkhead so backpressure becomes fast failure instead of unbounded queueing.
• A hung dependency cannot exhaust threads through sync timeout: abandoned workers are capped (max_abandoned), beyond which calls fail fast.
• RetryBudget bounds retry amplification, so your clients cannot turn a partial outage into a self-inflicted one.

Vulnerability reports: see SECURITY.md. Contributions: CONTRIBUTING.md.

For AI coding agents

A compact, machine-oriented API reference lives in llms.txt. Point your agent at it for the full surface in one read.

Development

pip install -e .[dev]
pytest
ruff check .
mypy src

License

MIT.