[Marketen]

Summary

The Gateway cut every response at five minutes: an absolute server write deadline, plus total-duration deadlines on provider requests. Healthy long answers and streams ended mid-response.

Client side

• The server has no absolute write deadline any more.
• Each read from and write to the client is bounded on its own (CLIENT_IO_TIMEOUT_SECONDS, default 60 s). Transfers are split into chunks of at most 32 KiB and the deadline refreshes on progress, so a transfer that keeps moving is never cut. This also covers encrypted (EHBP) streams.
• Request headers keep their 30 s limit.

Provider side

• Provider requests have no total deadline.
• A request ends only when the provider sends nothing, neither headers nor bytes, for PROVIDER_SILENCE_TIMEOUT_SECONDS (default 900 s). Keepalive bytes count as activity.
• Tinfoil's verified transport is not replaced; only the wait for headers is bounded around it.

Both settings must be positive whole seconds; the Gateway refuses to start otherwise.

Compatibility

• Deploy after feat: keep the credit hold alive while inference runs #10 and dappnode/dappnode-nexus#143. Responses can now last longer than Metering's reservation TTL, and without hold renewal those requests would go unbilled.
• Edge proxies still apply. Proxies in front of the Gateway keep their own timeouts for silent connections.

Tests

• Unit:
  • a stream with steady writes outlives the old deadline;
  • an encrypted stream refreshes its network write deadline;
  • upload reads are bounded per read and cleared after the body;
  • provider silence ends a request and keepalive bytes keep it open;
  • the header wait is bounded without limiting the body;
  • invalid settings are rejected.
• Together with feat: keep the credit hold alive while inference runs #10: the merged tree passes the full suite. End to end with a 2-second silence guard and lease, a 6-second response with keepalives completed and was billed.

🤖 Generated with Claude Code