Community-maintained fork of fastapi-cache2. The Python import stays
fastapi_cache; intentional safety/behavior changes are described in MIGRATION.md.Latest published stable: 0.5.0 on PyPI (released 2026-11-21) · Release notes · Release evidence · Quickstart · Migration · Compatibility · Security · Support · Changelog. Pre-releases must be independently evaluated before production deployment.
Installation: Python
>=3.8,<3.14. CPython 3.11–3.13 (server-side) has security validation; Python 3.8–3.10 is functional legacy compatibility only. DynamoDB/Memcached remain experimental. Browser/Pyodide/Node are untested. Independent project, not endorsed by the upstream maintainer.
stableis the release branch;maintracks upstream.
Recommended for production: CPython 3.11–3.13, within our tested
FastAPI/Redis/InMemory matrix. Python 3.8–3.10 stays installable and
functionally tested to support migrations from fastapi-cache2, but is
legacy-only: these EOL runtimes and their dependency stacks receive
no security maintenance commitment. Known issues in some Python 3.8/3.9
dependency profiles include AnyIO TLS
and Starlette file serving.
The library does not exercise those specific paths directly, but other
application features might. Use a newer Python and patched compatible
dependencies whenever possible.
A vulnerability found in a package shared by this fork and the original
fastapi-cache2 may affect both if a consumer resolves the same vulnerable
version and invokes the affected code. They do not necessarily install
the same dependency versions. In particular, the published fastapi-cache2==0.2.2 and this fork
both declared redis-py 4.x in the published 0.3.x metadata. The upstream repository's later, unreleased
main branch changed the Redis extra to 5.x. 0.4.x now publishes
Python-conditional redis-py ranges through 8.1.x; older 0.3.x artifacts
remain unchanged.
See compatibility & risk matrix,
security policy and
legacy vulnerability explanation.
| I'm… | Start here | What you'll learn |
|---|---|---|
| Starting from zero | Install in 5 minutes · Visual guide source | Pinned package, InMemory, first HIT/MISS, Redis |
| Already on fastapi-cache2 | Migration walkthrough · Full migration contract | Remove the old distribution, validate privacy, migrate namespaces, rollback |
| Comparing features/backends | Feature matrix · Backend choice | What's actually in 0.5.0 and what remains proposed |
| Checking security | Tenant and request safety · Security policy | Default auth bypass, custom headers, cache-key isolation |
| Looking ahead | Next release plan · #86 tracker | 0.6.0 candidates, not publication promises |
Try it: one-command before/after migration demo (two isolated published environments, no Redis or credentials).
Published: fastapi-cache-stable 0.5.0. The new searchable site is in site_docs/ and has a CI build preview; it is not claimed to be hosted publicly until a separate deployment is approved.
fastapi-cache is a tool to cache FastAPI endpoint and function results, with
backends supporting Redis, Memcached, and Amazon DynamoDB.
The published 0.5.0 line prioritizes safe, low-friction migration over feature count. See the roadmap and its linked issues for maintenance and later opt-in work. Roadmap items are not included in 0.5.0 and are not delivery commitments.
- Same
fastapi_cacheimports and legacy default cache-key algorithm. - Correctness fixes for cache invalidation, error responses and HTTP cache headers.
- Auth/Cookie privacy bypass by default; opt-in typed JSON decoding for Pydantic v2.
- Opt-in
vary_headersfor explicitly declared public header dimensions, with versioned hashed keys. - Redis and InMemory backends; Memcached and DynamoDB remain experimental.
- CI across Python versions, security profiles and published-wheel Redis migration.
- FastAPI with a compatible Python version.
- Redis standalone verified with Python-conditional redis-py ranges: 4.x through 6.1.x on Python 3.8, through 7.0.x on 3.9, and through 8.1.x on 3.10–3.13.
- DynamoDB client extra requires Python 3.10+ and patched aiohttp.
- See COMPATIBILITY.md before using optional backends.
Install the published version you have tested, pinned exactly:
python -m pip uninstall -y fastapi-cache2
python -m pip install "fastapi-cache-stable[redis]==0.5.0"
python -m pip checkThe 0.5.0 release evidence records the published wheel/sdist hashes and workflows. Do not treat the repository version alone as publication evidence.
For process-local InMemory caching, omit [redis]. See the
5-minute quickstart for a complete example.
To evaluate a published pre-release candidate:
python -m pip install --pre --upgrade "fastapi-cache-stable[redis]"Other extras: [memcache], [dynamodb] (experimental), or no extra for
InMemory. Before deploying, check PyPI
and pin the exact version you've tested. Never co-install fastapi-cache2;
both distributions provide fastapi_cache files.
The intended low-friction upgrade is fastapi-cache2==0.2.2 →
fastapi-cache-stable, keeping your existing from fastapi_cache ...
imports and usual @cache(...) calls. For GET/JSON with standalone Redis
or InMemory, the default key algorithm is unchanged. Uninstall the old
distribution first: the two packages install the same Python module and
must not coexist. Update the dependency in your application's lockfile.
This is not a promise of identical behavior in every case: authenticated
requests bypass cache by default, second init() calls now raise, and
clear(key=...) is intentionally narrower. Old private/special Redis
entries may require targeted invalidation. See the
quick migration checklist,
verification scenarios and rollback instructions.
Redis 7 / redis-py 4.x example. Initialize once per application lifespan,
then close owned resources and reset the process-global cache configuration
on teardown. Avoid routine use of force=True.
from contextlib import asynccontextmanager
from typing import AsyncIterator
from fastapi import FastAPI
from redis import asyncio as aioredis
from fastapi_cache import FastAPICache
from fastapi_cache.backends.redis import RedisBackend
from fastapi_cache.decorator import cache
@asynccontextmanager
async def lifespan(_: FastAPI) -> AsyncIterator[None]:
redis_client = aioredis.from_url(
"redis://localhost:6379/0", decode_responses=False
)
FastAPICache.init(RedisBackend(redis_client), prefix="fastapi-cache")
try:
yield
finally:
FastAPICache.reset()
await redis_client.close() # redis-py 4.x
app = FastAPI(lifespan=lifespan)
@app.get("/")
@cache(expire=60)
async def index():
return {"hello": "world"}For a process-local test fixture, use
FastAPICache.init(InMemoryBackend(), prefix="fastapi-cache")
after importing InMemoryBackend from
fastapi_cache.backends.inmemory. InMemory data is not shared across workers.
Call FastAPICache.init() once per process at application startup. A
second call raises RuntimeError; test fixtures that re-enter a lifespan
must close owned backend resources and call FastAPICache.reset() during
teardown. force=True permits deliberate replacement but does not
close the old connection pool or isolate multiple FastAPI apps in one process.
If you want to cache a FastAPI response transparently, use the @cache
decorator between the router decorator and the view function.
| Parameter | type | default | description |
|---|---|---|---|
expire |
int |
sets the caching time in seconds | |
namespace |
str |
"" |
namespace to use to store certain cache items |
coder |
Coder |
JsonCoder |
which coder to use, e.g. JsonCoder |
key_builder |
KeyBuilder callable |
default_key_builder |
which key builder to use |
injected_dependency_namespace |
str |
__fastapi_cache |
prefix for injected dependency keywords. |
cache_authenticated |
bool |
None |
per-endpoint override for authenticated requests; requires an identity-aware key_builder. |
infer_response_model |
bool |
False |
opt into FastAPI's response-model inference from a return annotation; can enable output validation and break invalid union/Response annotations. Disabled by default for drop-in compatibility. |
vary_headers |
Sequence[str] |
None |
opt into an isolated, hashed variant cache key for the explicitly listed HTTP request headers; it does not automatically discover tenant/identity/context state. |
The cache_status_header argument belongs to FastAPICache.init(), not
@cache(). It controls the HIT / MISS response header.
You can also use the @cache decorator on regular functions to cache their result.
The cache decorator injects dependencies for the Request and Response
objects, so that it can add cache control headers to the outgoing response, and
return a 304 Not Modified response when the incoming request has a matching
If-None-Match header. The ETag is a quoted weak validator derived from a stable SHA-256
digest of the cached representation; If-None-Match supports weak comparison,
comma-separated lists and *. Conditional requests are evaluated on both hits and misses,
and malformed entity-tags are ignored. This only happens if the decorated endpoint doesn't
already list these dependencies already.
The keyword arguments for these extra dependencies are named
__fastapi_cache_request and __fastapi_cache_response to minimize collisions.
Use the injected_dependency_namespace argument to @cache to change the
prefix used if those names would clash anyway.
The default JsonCoder stores any data type that FastAPI can convert to JSON, including
Pydantic models and dataclasses. On a cache hit it returns the decoded JSON value (plain
dicts/lists), not the annotated model; this is the legacy fastapi-cache2 behavior.
Pydantic v2 users can opt into typed restoration with TypedJsonCoder:
from fastapi_cache import FastAPICache
from fastapi_cache.coder import TypedJsonCoder
FastAPICache.init(backend, coder=TypedJsonCoder)or per endpoint:
from .models import SomeModel, create_some_model
@app.get("/foo")
@cache(expire=60, coder=TypedJsonCoder)
async def foo() -> SomeModel:
return create_some_model()TypedJsonCoder uses the endpoint return annotation to restore the annotated type and
propagates validation errors. It requires Pydantic v2; Pydantic v1 and annotations that
cannot be represented by a Pydantic v2 schema keep the legacy decoded behavior.
For broader type support, use fastapi_cache.coder.PickleCoder or implement a custom coder
(see below).
PickleCoder deserializes with Python's pickle and must only be used with a cache backend
whose contents are trusted and integrity-protected. An attacker who can alter cached bytes
may be able to execute code when those values are read. JsonCoder is the default.
By default, @cache preserves the upstream behavior of not exposing
the function's return annotation as a FastAPI response model. Existing
endpoints that return mixed Response and data values therefore keep
registering without new Pydantic validation.
To have FastAPI infer an OpenAPI response model from a valid annotation and validate/serialize the output, opt in explicitly:
@app.get("/item")
@cache(expire=60, infer_response_model=True)
async def item() -> SomeModel:
return create_some_model()Do not opt in with annotations FastAPI cannot use as a response model,
for example Union[Response, dict]. Alternatively, set
response_model=SomeModel on @app.get(...) as with ordinary FastAPI.
This option changes schema and response validation, not stored cache keys.
JsonCoder caches only simple JSONResponse objects: status 200, media type
application/json, no custom headers, and no background task. Other Response subclasses
are returned directly and are not cached: StreamingResponse, FileResponse,
RedirectResponse, plain Response, template responses, and JSONResponse objects with
metadata that would be lost on a cache hit. Custom coders can override
Coder.can_encode() to opt in; PickleCoder caches response objects it can reproduce
faithfully (JSONResponse, plain Response, RedirectResponse) and bypasses streaming,
file, and template responses. Responses with Set-Cookie, Cache-Control: no-store or
Cache-Control: private, or a background task are never cached, regardless of the coder.
If the endpoint or a dependency sets Set-Cookie, Cache-Control: no-store/private, a
background task, a non-200 status, or any header other than the automatic content-length
on the injected response object, the call is not cached. The sole exception is an injected
Vary header whose fields are all covered by explicitly declared vary_headers;
other metadata still bypasses. This covers dict/list returns whose final HTTP metadata
comes from that object. A returned JSONResponse with custom Vary remains uncacheable
under JsonCoder because its headers cannot be faithfully restored on a hit.
User payloads that contain _spec_type are preserved as normal data; the internal marker is
namespaced as __fastapi_cache_spec_type__.
By default use JsonCoder, you can write custom coder to encode and decode cache result, just need
inherit fastapi_cache.coder.Coder.
from typing import Any
import orjson
from fastapi.encoders import jsonable_encoder
from fastapi_cache import Coder
class ORJsonCoder(Coder):
@classmethod
def encode(cls, value: Any) -> bytes:
return orjson.dumps(
value,
default=jsonable_encoder,
option=orjson.OPT_NON_STR_KEYS | orjson.OPT_SERIALIZE_NUMPY,
)
@classmethod
def decode(cls, value: bytes) -> Any:
return orjson.loads(value)
@app.get("/")
@cache(expire=60, coder=ORJsonCoder)
async def index():
return dict(hello="world")By default the default_key_builder builtin key builder is used; this creates a
cache key from the function module and name, and the positional and keyword
arguments converted to their repr() representations, encoded as a MD5 hash.
You can provide your own by passing a key builder in to @cache(), or to
FastAPICache.init() to apply globally.
For example, if you wanted to use the request method, URL and query string as a cache key instead of the function identifier you could use:
from starlette.requests import Request
from starlette.responses import Response
def request_key_builder(
func,
namespace: str = "",
*,
request: Request = None,
response: Response = None,
args=(),
kwargs=None,
):
if request is None:
raise ValueError("This builder requires an HTTP Request")
return ":".join([
namespace,
request.method.lower(),
request.url.path,
repr(sorted(request.query_params.items()))
])
@app.get("/")
@cache(expire=60, key_builder=request_key_builder)
async def index():
return dict(hello="world")This request-URL builder is appropriate for public data only. A custom
builder must include every value that affects the response, particularly
validated identity, tenant and permissions when authenticated caching is
explicitly enabled. DB session objects with changing repr() values can
otherwise produce perpetual misses; see the migration checklist.
Versioned API: vary_headers was introduced in published 0.5.0.
It is not part of published 0.4.1. Pin the tested 0.5.0 distribution and
review your application-specific cache-key/identity and namespace-migration
requirements below.
For anonymous/public endpoints that depend on a known header through a
dependency, middleware, or contextvars, opt in to a header-specific key:
from contextvars import ContextVar
from fastapi import Depends, Request
locale_context = ContextVar("locale", default="en")
async def resolve_locale(request: Request) -> None:
locale_context.set(request.headers.get("Accept-Language", "en"))
@app.get("/localized", dependencies=[Depends(resolve_locale)])
@cache(expire=60, namespace="localized-v2",
vary_headers=("Accept-Language",))
async def localized():
return {"locale": locale_context.get()}This is an explicit allowlist, not automatic header detection. Names are
case-normalized and validated; multiple field occurrences, their order, and
missing versus empty fields have distinct keys. Header values are hashed into
a :vary-v1: key scoped under the normal cache namespace, not written as
plaintext. The generated HTTP Vary header is returned on MISS, HIT and 304
for configured cacheable requests. An HTTP Vary response header alone
never partitions the library's internal cache. If the endpoint emits a
Vary field not present in vary_headers, the response bypasses storage.
For a cacheable returned Response, the decorator applies generated
Vary to the actual returned object, not just FastAPI's injected response.
A plain JSONResponse is checked for JsonCoder eligibility before
the header is added; a plain Response with PickleCoder also keeps the
header on MISS and HIT. The generated ETag and cache-status headers follow
the returned object as well, and 304 responses carry Vary. A
JSONResponse that already has custom headers, including Vary,
still bypasses the default JSON coder.
The caller is responsible for listing every representation-defining
header. Application state derived from hidden contextvars, middleware,
non-header sources or arbitrary dependencies cannot be discovered. Using
vary_headers=("Accept-Language",) does not make an endpoint whose
response also depends on X-Tenant safe; include a trustworthy tenant
dimension or use a reviewed, complete custom key_builder. Endpoints
that explicitly declare Request still bypass the default key builder
even with vary_headers; use an identity- and tenant-aware custom builder
there. Only Authorization, Cookie and Proxy-Authorization are
explicitly rejected as credential-related vary_headers names. Names
like X-API-Key, X-Auth-Token and application-defined identity headers
can pass the syntactic name check; that is not a safety endorsement
or an authentication/authorization boundary. The existing
Authorization/Cookie bypass is unchanged and cannot detect every
custom authentication mechanism. Use validated identity/permissions and
a reviewed custom key_builder when needed. A client can choose many
distinct header values and flood the cache with variants: validate and
normalize input, rate-limit, bound cache growth and expiration. Hashing
hides the literal header value from cache keys, but does not authenticate
it, encrypt it or prevent low-entropy guesses.
New opted-in entries cannot read pre-opt-in keys. Clear or rotate affected namespaces (and remove potentially unsafe historical entries) before serving traffic; do not simply assume they will disappear safely. A shared Redis namespace used across deployments must be migrated deliberately. See MIGRATION.md.
By default, requests with a non-empty Authorization or Cookie header bypass the cache:
the endpoint runs normally and nothing is read from or written to the backend. Public
requests keep caching.
This is a conservative default. Session cookies are covered because they use the Cookie
header. Empty headers are treated as unauthenticated. Functions decorated without a
Request parameter have no HTTP identity context and keep using the cache.
To cache authenticated responses, opt in explicitly and provide a key builder that includes the identity:
import hashlib
import hmac
import os
from starlette.requests import Request
from fastapi_cache import FastAPICache
from fastapi_cache.decorator import cache
CACHE_KEY_SECRET = os.environ["CACHE_KEY_SECRET"].encode()
def identity_key_builder(
func,
namespace: str = "",
*,
request: Request = None,
response=None,
args,
kwargs,
):
identity = ""
if request is not None:
identity = "|".join(
[
request.headers.get("Authorization", ""),
request.headers.get("Cookie", ""),
request.headers.get("X-Tenant-Id", ""),
]
)
key_material = f"{identity}|{args}|{kwargs}"
digest = hmac.new(
CACHE_KEY_SECRET,
key_material.encode(),
hashlib.sha256,
).hexdigest()
return f"{namespace}:{func.__module__}:{func.__name__}:{digest}"
FastAPICache.init(backend, cache_authenticated=True, key_builder=identity_key_builder)
@app.get("/me")
@cache(expire=60)
async def me(request: Request):
...The key builder must include every dimension that affects the response: token, session, tenant, permissions, and any other validated principal state. Prefer the principal your app already validated over raw headers. Do not put credentials in plaintext in keys or logs; use HMAC-SHA256 with an application secret.
If cache_authenticated=True is used with the default key builder, an authenticated request
raises ValueError before the cache is accessed. The flag can also be set per endpoint:
@cache(cache_authenticated=True, key_builder=identity_key_builder).
The automatic bypass only detects Authorization and Cookie. If your app uses X-API-Key,
proxy-injected headers, request.state, or internal dependencies, review your cache strategy.
cache_authenticated=True does not make an unsafe key builder safe.
This protects the application cache. It does not configure CDN, reverse-proxy, or browser caching.
Upgrading from the legacy package or early prereleases: old entries written for authenticated requests may remain under the default key. Clear or rotate the cache namespace before deploying, including all instances/workers and persistent Redis caches. The bypass is not retroactive.
The InMemoryBackend stores cache data in memory and only deletes when an
expired key is accessed. This means that if you don't access a function after
data has been cached, the data will not be removed automatically.
expire=None means no expiration. Each backend instance owns its store and lock, so a new
instance does not see entries from another instance. Do not share one instance across event
loops or threads; use one instance per loop/thread. Subclasses that override __init__()
must call super().__init__(). InMemory is process-local; use Redis or Memcached for
cross-process sharing. expire=0 keeps the legacy fallback to the global expire value.
get_with_ttl returns -1 for no-expiration entries; the decorator omits max-age for
those entries and for expire=None.
When using the Redis backend, please make sure you pass in a redis client that does not decode responses (decode_responses must be False, which is the default). Cached data is stored as bytes (binary), decoding these in the Redis client would break caching.
0.4.x publishes Python-conditional Redis client ranges (4.x through 6.1.x
on Python 3.8, through 7.0.x on 3.9 and through 8.1.x on 3.10–3.13). See
COMPATIBILITY.md and audit/redis/REDIS_CLIENT_MATRIX_RESEARCH.md for
real-server evidence. The older 0.3.x artifacts remain immutable.
coverage run -m pytest
coverage html
xdg-open htmlcov/index.htmlThis project is licensed under the Apache-2.0 License.