[wrangler] Remote Hyperdrive bindings under getPlatformProxy() never re-seed edge credentials, so a stale dev session never recovers

[mack-erel]

What versions & operating system are you using?

Wrangler 4.129.0 (built from main plus #14900, commit 29410697), Node v26.4.0, macOS 26.5.2.

Please provide a link to a minimal reproduction

No public repository — the reproduction needs a real Hyperdrive configuration on the account and several hours of dev-server idle time to surface. The setup is small enough to inline; see below.

Describe the Bug

A remote: true Hyperdrive binding under getPlatformProxy() (e.g. vite dev via @sveltejs/adapter-cloudflare's platformProxy option) works correctly for a while, then every connection attempt starts failing — and never recovers on its own. Only restarting the dev process fixes it.

Reproduction. A Worker with one or more remote: true Hyperdrive bindings and no localConnectionString, run under vite dev (or any other getPlatformProxy() consumer) for an extended period — in the sessions that surfaced this, failure appeared after roughly 5–8 hours of the process sitting idle (no requests, no file edits). Once it starts, every subsequent request to every Hyperdrive binding fails identically until the process is restarted.

What is established.

• The failure happens at connection-establishment time, not during query execution. Instrumenting the failure's full cause chain shows the throw originates from the Node driver's createConnection(...) call (not from a query already in flight), with a MySQL protocol-level error: errno=1002 code=ER_NO message=Internal error. This is consistent with the edge rejecting the connection attempt itself, not a query timing out.
• This is not a stale/reused socket: this driver setup opens a brand-new connection per request (no pooling), so a new connection attempt fails the same way an old one would have. Ruling that out points at whatever credential or session state is used to authenticate a fresh connection attempt, not at connection reuse.
• seedRemoteHyperdriveBindings() (in packages/remote-bindings/src/seed-hyperdrive-bindings.ts) fetches the edge session's per-session Hyperdrive credentials and is called exactly once, inside getPlatformProxy() (packages/wrangler/src/api/integrations/platform/index.ts). Its result is baked into the one Miniflare instance built for the life of the process. Nothing calls it again — not on a dev reload (which doesn't re-invoke getPlatformProxy()), and not on any kind of periodic or on-error basis.
• maybeStartOrUpdateRemoteProxySession() (packages/remote-bindings/src/maybe-start-or-update-session.ts) reuses the existing remote proxy session indefinitely as long as auth and the set of remote bindings haven't changed — it never checks whether the underlying session (or, in this case, the credentials seeded from it) is still valid.

Put together: if the edge-minted per-session Hyperdrive credentials have any expiry, there is no code path anywhere that would notice and re-seed them. The dev process keeps presenting the same now-invalid credentials to every new connection attempt, forever, until it is restarted and getPlatformProxy() runs again from scratch.

I have not verified whether this also affects other remote binding types that go through the same shared session-reuse layer (KV, R2, D1, etc.) — only Hyperdrive was reproduced and traced in this investigation.

Please provide any relevant error logs

Full cause chain for one failed connection attempt, captured with a temporary diagnostic that walks error.cause recursively (not present in the shipped code):

[HD-DIAG] cause[0] name=Error message=Failed query: select version() as v
[HD-DIAG] cause[1] name=Error message=Internal error. code=ER_NO errno=1002
[HD-DIAG] cause[1] stack=Error: Internal error.
    at eval (.../src/lib/server/db/driver/workers.ts:118:89)
    at open (.../src/lib/server/db/driver/workers.ts:84:15)
    at Object.query (.../src/lib/server/db/driver/workers.ts:92:29)
    at .../drizzle-orm/mysql2/session.js:53:29
    ...

open() at that line is the driver's lazy-connect path (createConnection({...}).then(...)) — i.e. this is a fresh connection attempt failing, not a query on an existing one.

[workers-devprod]

Automated triage

Issue assessment

• Actionability: ✅ Actionable.
• Type: Bug (high confidence).
• Reproduction: Automatic attempt requested.
• Labels:
  • package:wrangler (high)
  • package:remote-bindings (high)
  • product:hyperdrive (high)

Actionability and missing information

Bug report: getPlatformProxy seeds remote Hyperdrive edge credentials once per process with no re-seed or validity-recheck path, so long-idle dev sessions fail every new connection until restart; named code paths make it actionable.

Reproduction reason: Describes a relevant current behavior defect with an inlined setup and concrete failure signature, so a reproduction attempt is warranted.

Missing information:

• affected scope
• minimal reproduction

Reproduction

• Result: ⚠️ Blocked.
• Summary: Built the issue's code state (main + PR [miniflare] Support remote Hyperdrive bindings in local dev #14900, commit 2941069) and a getPlatformProxy() fixture with a remote: true Hyperdrive binding + fresh mysql2 connections. Blocked: the sandbox has no Cloudflare credentials, so the edge session and its Hyperdrive credentials can never be minted.

Evidence and attempted steps

Steps attempted:

1. Cloned workers-sdk; confirmed PR [miniflare] Support remote Hyperdrive bindings in local dev #14900 ('[miniflare] Support remote Hyperdrive bindings in local dev') is OPEN/unmerged with head exactly 2941069 (the commit cited in the issue); checked it out; verified seed-hyperdrive-bindings.ts exists there.
   • Command: git clone --depth 1 https://github.com/cloudflare/workers-sdk.git && git fetch --depth 1 origin pull/14900/head && git checkout -b pr-14900 FETCH_HEAD
2. Installed the monorepo and built wrangler 4.129.0 plus workspace deps from the PR branch; packed the built wrangler and miniflare 5.20260903.0-alpha (containing the remote Hyperdrive TCP bridge) into tarballs.
   • Command: npm install -g pnpm@10.33.0 && pnpm install --frozen-lockfile && pnpm --filter wrangler... build
3. Fixture: wrangler.jsonc with one remote: true Hyperdrive binding (no localConnectionString), src/index.ts, and repro.mjs calling getPlatformProxy() then opening a brand-new mysql2 connection per attempt (select version() as v), logging the cause chain; deps pinned to vendored PR tarballs.
   • Command: pnpm install && node repro.mjs (via `pnpm reproduce`)
4. Ran the attempt: exit 1 inside getPlatformProxy(), via maybeStartOrUpdateRemoteProxySession() then requireAuth: "In a non-interactive environment, it is necessary to set a CLOUDFLARE_API_TOKEN environment variable" (source RemoteRuntimeController); blocked before any edge session could exist.
   • Command: pnpm reproduce 2>&1 | tee evidence/attempt-output.log

Blocker: The sandbox contains no Cloudflare credentials: no wrangler login cache (~/.wrangler and XDG .wrangler/config absent) and no CLOUDFLARE_API_TOKEN/CLOUDFLARE_ACCOUNT_ID environment variables (the CLOUDFLARE_* variables present are infra metadata such as region/node id; values never read). Without an authenticated edge session, getPlatformProxy() aborts before Hyperdrive edge credentials are ever minted. Additionally, the reported failure surfaces only after roughly 5-8 hours of dev-server idle time against a real Hyperdrive configuration on a Cloudflare account, which no feasible observation window in this sandbox can cover.

Artifact: Download reproduction archive (Cloudflare Access required).

Treat archive contents as untrusted. Inspect them before running anything.

---

This summary was generated by an LLM and may contain errors. Please verify its findings.