fix(http): raise keep-alive timeout above proxy idle timeouts (#473)

The HTTP transport kept Node's default keepAliveTimeout of 5s. Proxies
that pool upstream connections (AWS ALB, API Gateway VPC links, nginx)
hold them idle for 60s, so a request written into a socket Node was
closing surfaced as a proxy 502/503 that never reached DBHub's logs.

Set keepAliveTimeout to 65s and headersTimeout to 66s so the proxy is
always the side that closes an idle connection. Hard-coded on purpose;
no new flag.

Fixes #472


Claude-Session: https://claude.ai/code/session_013piqks7dS3HsaAKC6YVWLN

Co-authored-by: Claude <[email protected]>

Author: tianzhou

src/server.ts

@@ -18,6 +18,7 @@ import { getToolsForSource } from "./utils/tool-metadata.js";
 import { startConfigWatcher } from "./utils/config-watcher.js";
 import { validateOrigin, buildAllowedHosts, getSelfHosts, ALLOW_ANY_HOST } from "./utils/cross-origin.js";
 import { validateAuthToken } from "./utils/auth-token.js";
+import { configureHttpKeepAlive } from "./utils/http-keep-alive.js";
 
 // Create __dirname equivalent for ES modules
 const __filename = fileURLToPath(import.meta.url);
@@ -299,6 +300,12 @@ See documentation for more details on configuring database connections.
       // is registered. Matches the pattern used in utils/ssh-tunnel.ts.
       const httpServer = http.createServer(app);
 
+      // Outlive proxy idle timeouts (ALB/API Gateway/nginx default to 60s) so
+      // the proxy, not DBHub, closes idle connections. Otherwise a request
+      // written into a socket Node is closing surfaces as a proxy 502/503
+      // that never reaches our logs (issue #472).
+      configureHttpKeepAlive(httpServer);
+
       httpServer.on('error', (err) => {
         const displayHost = host!.includes(':') ? `[${host!}]` : host!;
         console.error(`Failed to bind HTTP server to ${displayHost}:${port}: ${err.message}`);

src/utils/__tests__/http-keep-alive.test.ts

@@ -0,0 +1,25 @@
+import { describe, it, expect } from "vitest";
+import http from "http";
+import {
+  configureHttpKeepAlive,
+  HTTP_KEEP_ALIVE_TIMEOUT_MS,
+  HTTP_HEADERS_TIMEOUT_MS,
+} from "../http-keep-alive.js";
+
+describe("configureHttpKeepAlive", () => {
+  it("raises keep-alive above a 60s proxy idle timeout (issue #472)", () => {
+    const server = http.createServer();
+    expect(server.keepAliveTimeout).toBe(5_000); // Node default, the bug
+
+    configureHttpKeepAlive(server);
+
+    expect(server.keepAliveTimeout).toBe(HTTP_KEEP_ALIVE_TIMEOUT_MS);
+    expect(server.headersTimeout).toBe(HTTP_HEADERS_TIMEOUT_MS);
+    expect(server.keepAliveTimeout).toBeGreaterThan(60_000);
+    server.close();
+  });
+
+  it("keeps headersTimeout strictly above keepAliveTimeout, as Node requires", () => {
+    expect(HTTP_HEADERS_TIMEOUT_MS).toBeGreaterThan(HTTP_KEEP_ALIVE_TIMEOUT_MS);
+  });
+});

src/utils/http-keep-alive.ts

@@ -0,0 +1,34 @@
+import type http from "http";
+
+/**
+ * Idle keep-alive timeout for the HTTP transport, in milliseconds.
+ *
+ * Node's default is 5s. Reverse proxies and load balancers that pool
+ * upstream connections hold them idle for longer than that (AWS ALB and
+ * API Gateway VPC links default to 60s, as does nginx `keepalive`). When the
+ * proxy writes a request into a socket that Node is closing at the same
+ * moment, the proxy answers the client with its own 502/503 and the request
+ * never reaches DBHub (issue #472).
+ *
+ * The backend's idle timeout must be longer than the proxy's so the proxy is
+ * always the side that closes an idle connection. 65s is the widely used
+ * value for 60s proxies.
+ */
+export const HTTP_KEEP_ALIVE_TIMEOUT_MS = 65_000;
+
+/**
+ * Headers timeout for the HTTP transport, in milliseconds.
+ *
+ * Node requires this to be strictly greater than `keepAliveTimeout`;
+ * otherwise a reused socket can be closed while a request's headers are
+ * still arriving.
+ */
+export const HTTP_HEADERS_TIMEOUT_MS = HTTP_KEEP_ALIVE_TIMEOUT_MS + 1_000;
+
+/**
+ * Apply keep-alive timeouts that survive a 60s proxy idle timeout.
+ */
+export function configureHttpKeepAlive(server: http.Server): void {
+  server.keepAliveTimeout = HTTP_KEEP_ALIVE_TIMEOUT_MS;
+  server.headersTimeout = HTTP_HEADERS_TIMEOUT_MS;
+}