EPIPE or Broken pipe means an application wrote to a pipe or socket after the reader had closed it. In an AI application, the reader may be a browser that cancelled an SSE stream, a gateway that timed out, or an upstream model connection that ended unexpectedly.

Find which side closed first

Trace one request through:

browser/client → CDN or proxy → application → model provider/inference server

Correlate request IDs, timestamps, cancellation signals and bytes sent. The component reporting EPIPE is often the last writer, not the original cause.

Classify the event:

• normal client cancellation: navigation, stop button or closed tab;
• idle timeout: proxy or load balancer ended a quiet stream;
• upstream failure: provider reset, crashed model server or interrupted container;
• buffering: intermediary waited instead of forwarding events incrementally;
• application leak: producer kept generating after downstream cancellation.

SSE and streaming checks

For Server-Sent Events, verify the response uses the expected event-stream content type, events are framed correctly and intermediaries do not buffer the stream. If infrastructure has an idle timeout, use protocol-appropriate heartbeat comments only when they are operationally justified.

Do not treat a partial stream as a complete structured response. Record an explicit completion event or state, and keep partial output visibly incomplete.

Propagate cancellation

When the client disconnects:

1. stop writing to the response;
2. abort the upstream model request where supported;
3. release database, tool and queue resources;
4. mark the run cancelled rather than failed or completed;
5. prevent background actions from continuing silently.

Handle the stream’s error and close events. Do not crash the process for an expected client disconnect, but do not suppress an infrastructure reset as harmless either.

Retry safety

Never restart a partially executed agent workflow blindly. Before retrying, determine whether the request was read-only, whether a tool already changed state and whether the client can resume from an event ID or job state.

Use idempotency keys for retryable state-changing operations and bounded backoff for transient upstream failures. The AI Application Architecture foundation covers cancellation, idempotency and background jobs.

Proxies and observability

Align connect, first-byte, idle-stream and total-workflow timeouts across client, gateway and upstream. Verify buffering and streaming behavior with a real end-to-end request, not only a direct call to the model server.

Monitor disconnect reason, stream duration, last event, upstream status, retry count and resource cleanup. Add interrupted-stream scenarios to AI Testing & Evaluation and operate them through AI Operations.

Related foundations: 502 errors in AI gateways, AI API timeouts and Nginx for AI gateways.