Skip to main content

0005. Missing server: connect_error, immediately, no retry, plus a console.error

Status: Accepted · 2026-07-28 · #65 Governed by: 0000

TL;DR When no server is registered for the origin, smocket fires connect_error on the next tick, does not throw, and does not retry, and it also logs the failure with console.error. Real socket.io retries forever, so this is the one place the two deliberately differ.

Decision​

A connect(url) whose origin has no registered server fails as one decision with four faces.

  • Shape: connect_error. The failure surfaces as a connect_error event on the client, the same event real socket.io uses when a connection cannot be made. smocket does not throw, because throwing would invent a failure shape the real library does not have, and dual-run holds only if both sides take the same failure path.
  • Timing: immediately, no delay. The connect_error fires on the next tick through the same defer as a successful connect, with no artificial wait before it. The shape follows real socket.io's design, but a reconnection delay is not a design choice, it is the physics of a network round trip, which a mock has no source for. Under 0000, an unobservable delay is not invented, so the failure is reported at once.
  • Retry: none. smocket reports the failure once and stops. Real socket.io retries the connection forever, but that retry loop is driven by the same network timing the mock cannot reproduce, so smocket does not simulate it.
  • Diagnostics: a parallel console.error. Alongside the event, smocket writes the failure to the console. Most apps never attach a connect_error handler, so an event-only signal would leave a developer who mistyped a url seeing nothing at all. The real library does not log this, but that is a diagnostics-tool layer over the mock, not a change to the behaviour dual-run compares, so it is not a parity violation.

Taken together, this is the single point where smocket deliberately diverges from real socket.io rather than reproducing it, which is the entry that belongs in differences.md.

Alternatives rejected​

  • Throw synchronously from connect(). A thrown error is a failure shape real socket.io never produces here, so tests would have to branch on the target and dual-run would no longer compare one behaviour.
  • Simulate the real retry loop. Retrying forever would need an invented delay between attempts, exactly the network timing 0000 forbids inventing, and it would hang tests waiting on a server that will never appear.
  • Fire the event only, with no console output. Correct against the wire contract, but it hides the most common real mistake (a wrong or unregistered url) from a developer who attached no handler, so the console line is kept.