ApiClient creates a new httpx.AsyncClient per JWKS / OIDC fetch with no timeout and no single-flight, causing "Unknown auth error" storms under load

src/auth0_api_python/utils.py から始め、ApiClient._fetch_jwks と _fetch_oidc_metadata を utils.fetch_jwks、utils.fetch_oidc_metadata、InMemoryCache.get を通じて追跡します。verify_request がこれらのキャッシュミスの経路に到達する仕組みと、httpx.AsyncClient がどのように使われているかを確認します。JWKS と OIDC の再取得で、クライアント、タイムアウト、同時リクエストに関して報告されている失敗動作を回避しつつ、認証成功時のセマンティクスを変更しなければ完了です。

Checklist

• I have looked into the Readme and Examples, and have not found a suitable solution or answer.
• I have looked into the API documentation and have not found a suitable solution or answer.
• I have searched the issues and have not found a suitable solution or answer.
• I have searched the Auth0 Community forums and have not found a suitable solution or answer.
• I agree to the terms within the Auth0 Code of Conduct.

Description

ApiClient._fetch_jwks and _fetch_oidc_metadata (via
utils.fetch_jwks / utils.fetch_oidc_metadata) construct a brand-new
httpx.AsyncClient() on every cache miss with no explicit timeout
configured, and there is no single-flight protection around the
refetch path. Combined, these turn routine cache expiry into a
self-inflicted outage under any non-trivial concurrency.

Reproduction

1. Stand up a FastAPI service that calls ApiClient.verify_request
   on every authenticated route.
2. Drive ≥ ~30 RPS of authenticated requests against it (we hit it
   on a 4 vCPU ECS Fargate task during a k6 load test).
3. Wait for the in-memory JWKS cache to expire — by default Auth0
   returns Cache-Control: max-age=600, so this happens every ~10
   minutes under steady load.
4. Observe a sudden burst of httpx.ConnectTimeout chained out of
   ApiClient._fetch_jwks, surfacing to callers as opaque "Unknown
   auth error" 5xx responses on every authenticated route until the
   herd subsides. Sentry trace shows
   auth0_api_python.errors.UnknownAuth0Exception ←
   httpx.ConnectTimeout.

Additional context

In src/auth0_api_python/utils.py

async def fetch_jwks(jwks_uri, custom_fetch=None):
    ...
    async with httpx.AsyncClient() as client:   # 1
        resp = await client.get(jwks_uri)       # 2
        ...

1. No connection pooling across calls. A fresh client is created
   and torn down per fetch. Every cache miss = a fresh TCP + TLS
   handshake to https://<tenant>.auth0.com/. Under load this
   exhausts ephemeral source-port budget on the host and slows
   everything else on the box.
2. No explicit timeout. httpx.AsyncClient() with no timeout=
   uses httpx's default 5-second connect/read/write/pool budget. On a
   stressed event loop or a slow Auth0 region that 5s budget is
   routinely blown, raising httpx.ConnectTimeout /
   httpx.ReadTimeout. Those bubble up into _fetch_jwks and the
   caller sees ConnectTimeout chained to UnknownAuth0Exception —
   not a 401, not a 503, just an opaque "unknown auth error".
3. No single-flight on refetch. When the in-memory cache expires
   (InMemoryCache.get returns None), every concurrent request
   that reaches _fetch_jwks simultaneously fires its own outbound
   JWKS fetch. N requests in flight at the moment of expiry = N
   concurrent JWKS calls to Auth0. Auth0 throttles some of them, the
   others time out per (1) and (2), and any request that lost the
   race fails auth.

The same three problems apply verbatim to fetch_oidc_metadata and
_fetch_oidc_metadata.

auth0-api-python version

1.0.0b8

Python version

3.11