From 7bb3909e4b5ab98d01deef532d02f3e4f7f10e10 Mon Sep 17 00:00:00 2001
From: Peter Steinberger <steipete@gmail.com>
Date: Fri, 14 Aug 2026 23:47:42 -0700
Subject: [PATCH] fix(workers): clean stale bundle install staging (#124057)

Serialize node worker installs per Gateway namespace so a later bundle hash reclaims interrupted staging from older hashes instead of leaking it indefinitely.
---
 src/node-host/node-worker-bundle-installer.test.ts |  5 +++--
 src/node-host/node-worker-bundle-installer.ts      | 10 +++++-----
 2 files changed, 8 insertions(+), 7 deletions(-)

diff --git a/src/node-host/node-worker-bundle-installer.test.ts b/src/node-host/node-worker-bundle-installer.test.ts
index a3de9b9720f8..1fe84a3e57a9 100644
--- a/src/node-host/node-worker-bundle-installer.test.ts
+++ b/src/node-host/node-worker-bundle-installer.test.ts
@@ -95,13 +95,14 @@ describe("node worker bundle installer", () => {
     return { gatewayUrl: `ws://127.0.0.1:${address.port}`, requests };
   }
 
-  it("atomically installs and reuses an exact namespaced bundle", async () => {
+  it("atomically installs, reuses, and cleans prior-hash crash staging", async () => {
     const fixture = await bundleFixture();
+    const staleBundleHash = "f".repeat(64);
     const staleStaging = path.join(
       root,
       fixture.input.gatewayNamespace,
       "bundles",
-      `.staging-${fixture.input.build.bundleHash}-crashed`,
+      `.staging-${staleBundleHash}-crashed`,
     );
     await fs.mkdir(staleStaging, { recursive: true });
     const served = await serve(fixture.archive, fixture.input.archive.token);
diff --git a/src/node-host/node-worker-bundle-installer.ts b/src/node-host/node-worker-bundle-installer.ts
index dd69031a9ee8..ed0826013310 100644
--- a/src/node-host/node-worker-bundle-installer.ts
+++ b/src/node-host/node-worker-bundle-installer.ts
@@ -165,12 +165,11 @@ async function validateInstalledBundle(
   }
 }
 
-async function removeStaleInstallStaging(bundlesRoot: string, bundleHash: string): Promise<void> {
-  const prefix = `.staging-${bundleHash}-`;
+async function removeStaleInstallStaging(bundlesRoot: string): Promise<void> {
   const entries = await fsp.readdir(bundlesRoot, { withFileTypes: true });
   await Promise.all(
     entries.map(async (entry) => {
-      if (entry.name.startsWith(prefix) && entry.isDirectory() && !entry.isSymbolicLink()) {
+      if (entry.name.startsWith(".staging-") && entry.isDirectory() && !entry.isSymbolicLink()) {
         await fsp.rm(path.join(bundlesRoot, entry.name), { recursive: true, force: true });
       }
     }),
@@ -227,7 +226,8 @@ export class NodeWorkerBundleInstaller {
     signal?: AbortSignal;
   }): Promise<WorkerAdmissionHandshake> {
     const { input } = params;
-    const key = `${input.gatewayNamespace}\0${input.build.bundleHash}`;
+    // One namespace owns every staging sibling, so serialize it before sweeping crash residue.
+    const key = input.gatewayNamespace;
     return await this.#operations.enqueue(key, async () => {
       try {
         params.signal?.throwIfAborted();
@@ -237,7 +237,7 @@ export class NodeWorkerBundleInstaller {
           return structuredClone(input.build);
         }
         await fsp.mkdir(bundlesRoot, { recursive: true, mode: 0o700 });
-        await removeStaleInstallStaging(bundlesRoot, input.build.bundleHash);
+        await removeStaleInstallStaging(bundlesRoot);
         const operationRoot = await fsp.mkdtemp(
           path.join(bundlesRoot, `.staging-${input.build.bundleHash}-`),
         );