timeouts-and-async

Timeouts, interruptible nodes, and async testing

Builds on the failure mental model in ginkgo:overview (failure is a panic Ginkgo recovers; a goroutine that fails needs defer GinkgoRecover()). Docs: https://onsi.github.io/ginkgo/#spec-timeouts-and-interruptible-nodes and https://onsi.github.io/ginkgo/#patterns-for-asynchronous-testing.

Interruptible nodes: take a context, honor cancellation

A setup or subject node becomes interruptible simply by accepting a SpecContext (or plain context.Context) — Ginkgo supplies one automatically:

It("likes to sleep in", func(ctx context.Context) {
  select {
  case <-ctx.Done(): return        // honor cancellation and exit promptly
  case <-time.After(time.Hour): ...
  }
}, NodeTimeout(time.Second))

On timeout or interrupt, Ginkgo cancels the context to tell the node to stop. Pass ctx down into every blocking call (libraryClient.SaveBook(ctx, book), exec.CommandContext(ctx, ...)) so the cancellation actually propagates. Only setup/subject nodes are interruptible — container nodes are not (they run at construction time).

• ctx.Deadline() does NOT report the node's deadline. Ginkgo manages cancellation timing itself (to snapshot a progress report first), so it does not use a WithDeadline context. Trust <-ctx.Done(), not Deadline().
• SpecContext satisfies context.Context; you may wrap it (context.WithValue) and Ginkgo still cancels the result on time.

The timeout decorators (full reference → ginkgo:decorators)

• SpecTimeout can only be more lenient than a node's NodeTimeout — it caps the sum of all nodes (BeforeEach+It+AfterEach); a per-node NodeTimeout inside it is always more stringent.
• Cleanup still runs after a timeout. When SpecTimeout fires, Ginkgo cancels the current node, then runs AfterEach/AfterAll/DeferCleanup (each under its own NodeTimeout/grace). The timeout is a "mark failed" threshold, not a hard kill.
• A node that ignores cancellation is "leaked," not killed. After the grace period Ginkgo gives up waiting and moves on. Leaking is deliberately preferred over hanging forever — but a leaked goroutine keeps running and can call Fail/AddReportEntry and pollute a later spec. Always make blocking code respond to ctx.Done().
• DeferCleanup/DescribeTable entries are interruptible too — give the cleanup/entry func a ctx first arg; don't capture and reuse the parent node's ctx (it's already cancelled by cleanup time).

Interrupting the whole suite

• ginkgo --timeout=DURATION — suite-wide budget across all suites (default 1h).
• Abort("reason") — end the suite immediately from within a spec (programmatic interrupt).
• SIGINT/SIGTERM (^C) — interrupt: cancel the current interruptible node, run its cleanup + reporting nodes, skip the rest, exit failed. Escalation: a 2nd interrupt skips cleanup (still runs reporting); a 3rd bails immediately. To inspect a suite without stopping it, send SIGINFO/SIGUSR1 for a progress report → ginkgo:debugging-failures.

Async assertions: Eventually / Consistently

Use Gomega's Eventually (polls until the matcher passes or it times out) and Consistently (polls and requires the matcher to hold the whole interval — the way to assert something doesn't happen). Three input shapes: bare values (channels, gbytes), functions returning (value[, error]), and functions taking a Gomega.

It("publishes a book", func(ctx SpecContext) {
  buffer := gbytes.NewBuffer()
  c := publisher.Publish(ctx, book, buffer)              // pass ctx so it cancels cleanly
  Eventually(ctx, buffer).Should(gbytes.Say(`Publish complete!`))
  var result publisher.PublishResult
  Eventually(ctx, c).WithTimeout(time.Second).Should(Receive(&result)) // poll, don't <-c
}, SpecTimeout(time.Second*30))

Propagate the spec deadline into the poll with .WithContext(ctx) or the positional Eventually(ctx, ...) — now a node timeout/interrupt makes the Eventually exit immediately instead of running its own clock. Gomega also auto-injects the context and .WithArguments(...) into a polled function whose first params are (ctx), so Eventually(client.Connect).WithContext(ctx).Should(Succeed()) works (pass the method reference, not client.Connect()).

Assertions inside the polled function — use g, never global Expect

Pass a function taking func(g Gomega, ...) and assert with g.Expect so a failed poll retries instead of failing the spec outright:

Eventually(func(g Gomega, ctx SpecContext) {   // g Gomega must be first
  messages, err := gmail.Fetch(ctx, jane.EmailAddress)
  g.Expect(err).NotTo(HaveOccurred())
  g.Expect(messages).To(ContainElement(WithTransform(subjectOf, Equal(want))))
}).WithContext(ctx).Should(Succeed())          // Succeed() = "no failures in the func"

Using the global Expect inside an Eventually defeats the retry — the first failure calls Ginkgo's Fail and the spec dies with no second attempt. The local g lets Eventually catch and re-poll.

Goroutines: the two rules that bite

It("repaginates", func() {
  done := make(chan any)
  go func() {
    defer GinkgoRecover()                 // RULE 1
    Expect(book.SetFontSize(28)).To(Succeed())
    close(done)
  }()
  Eventually(done).Should(BeClosed())     // RULE 2: poll, don't block on <-done
})

1. Any goroutine that may call Fail or a Gomega assertion needs defer GinkgoRecover(). Ginkgo can't recover a panic raised on a goroutine it didn't start — without this, one failed assertion crashes the entire suite. (Ginkgo's crash message reminds you.)
2. Don't block the spec on a channel fed by a failing goroutine. If the goroutine fails before close(done), a bare <-done blocks until the node times out instead of reporting the real failure. Eventually(done).Should(BeClosed()) lets the failure surface immediately.

See also

• Decorator reference (NodeTimeout, SpecTimeout, GracePeriod, PollProgressAfter) → ginkgo:decorators
• Progress reports for a hanging spec, and diagnosing a timeout from the JSON report → ginkgo:debugging-failures
• --timeout and friends in a CI config → ginkgo:ci