SS
Back to posts

Notes on Sporadic 500s in Nuxt 3 x ECS Fargate x ALB

Why a Nuxt 3 app behind ALB returned sporadic 500 (timeout) errors and how aligning Nitro keep-alive settings in the build output fixed it.

Oct 27, 20252 min read
Nuxt3
AWS
ECS
Fargate
Troubleshooting

TL;DR

  • ALB -> ECS Fargate (Node 20) occasionally returned 500s with timeout in the logs.
  • ALB Idle timeout = 60s, while Nuxt (Nitro) generated server keepAliveTimeout was 5s.
  • Patching keepAliveTimeout / headersTimeout before server.listen in .output/server/*.mjs fixed it.
  • Nuxt updates can change output paths, so always rg "server." -n .output/server after build.

Suspected Cause

AWS metrics showed low TargetResponseTime and plenty of ECS CPU/memory headroom, yet ELB 504 appeared. That pointed to keep-alive mismatch between ALB and the Node server.

Because of the default mismatch between ALB and Node/Nitro:

  1. ALB waits
  2. Node server closes the connection earlier
  3. ALB sees a stalled response and returns 504/500

Force-Patch with sed

Since patching the build output is enough, I injected this during CI/CD after build and before final image creation.

sed -i -e "/server\.listen/i server.keepAliveTimeout = 62 * 1000\nserver.headersTimeout = 65 * 1000" \
  .output/server/chunks/nitro/nitro.mjs

Insert 62s / 65s right before server.listen so it is longer than ALB. Depending on the build, .output/server/index.mjs might be the main entry, so I patched both.

sed -i -e "/server\.listen/i server.keepAliveTimeout = 62 * 1000\nserver.headersTimeout = 65 * 1000" \
  .output/server/index.mjs

Points to note

  • Insert newlines with \n in sed.
  • server.headersTimeout must be slightly longer than keepAliveTimeout to avoid Node warnings.
  • Emit a CI log like echo '[patch] applied keepAliveTimeout tweak' to confirm the patch ran.

Watch Out When Updating Nuxt

Even minor Nuxt 3 updates can change Nitro output code. Run rg "createServer" -n .output/server or rg "server." -n .output/server to confirm the server.listen location.

If there is no chunks/nitro/*.mjs and only server/index.mjs exists, patch only that. For Lambda presets, the path may look like server/chunks/app/server.mjs, so always locate it with:

find .output/server -name '*.mjs' | xargs rg -n "server\.listen"

Aftermath

  • No alerts for a week after the patch, and ALB 5XX stayed at zero.
  • Longer keep-alives increase concurrent connections, so set maxConcurrency and ECS task counts with headroom.
  • I would love a Nitro option like NITRO_PRESET_KEEP_ALIVE, but for now the build patch is a practical workaround.

When in doubt, just grep for server. in the output and patch the right spot.