ECONNRESET, connection reset by peer or socket hang up means an established connection was forcibly closed. Unlike connection refused, a connection existed; unlike a clean completion, one side ended it unexpectedly.

Locate the reset

Correlate client, gateway, application and model-server logs using a request ID. Record whether the reset happened:

• before response headers;
• after streaming began;
• during an upload or tool call;
• at a consistent idle or total duration;
• during deployment, scaling or node termination.

The reporting process may be the victim rather than the component that initiated the reset.

Common AI causes

• model server crashed or was OOM-killed;
• load balancer idle timeout ended a quiet stream;
• rolling deployment terminated an in-flight request;
• provider closed an overloaded or long-running call;
• client cancellation was surfaced as a reset upstream;
• proxy and upstream disagreed about HTTP or streaming behavior;
• keep-alive connection was reused after an intermediary expired it.

Inspect restarts, memory, saturation and rollout events. Link model memory failures to AI out-of-memory troubleshooting.

Handle interrupted streams

Never parse a truncated model stream as complete output. Keep an explicit completion state, discard or label incomplete structured values and propagate cancellation upstream.

If an agent already invoked a tool, inspect durable workflow state before retrying. A new connection does not mean the previous side effect did not occur.

Safe retry pattern

Retry only transient resets, with bounded exponential backoff and a total deadline. Separate read-only generation from state-changing workflows. Use idempotency keys or resumable job state when repetition could duplicate work.

Do not retry authentication, invalid input or deterministic schema failures. Do not hide a rising reset rate behind client retries; that increases provider load and cost.

Infrastructure fixes

• align keep-alive and idle timeouts across load balancer, gateway and upstream;
• drain connections during deployments;
• provide readiness before routing and graceful termination before shutdown;
• bound context, concurrency and memory pressure;
• verify proxy streaming and buffering behavior;
• monitor upstream resets separately from client cancellations.

Test resets before headers and mid-stream through AI Testing & Evaluation. Connect remediation to AI Operations, AI Deployment & Hosting and AI Application Architecture.

Related: broken AI streams, AI API timeouts and 502 gateway failures.