Ordering and flakes

Two topics that share a root cause: specs that aren't truly independent (ginkgo:overview). Ordering decorators let you declare a dependency on purpose; flake controls let you manage or expose one. Docs: https://onsi.github.io/ginkgo/#serial-specs, https://onsi.github.io/ginkgo/#ordered-containers, https://onsi.github.io/ginkgo/#repeating-spec-runs-and-managing-flaky-specs.

Serial: opt out of parallelism

Serial (container or subject) guarantees a spec never runs in parallel with anything else. Under the hood Ginkgo runs serial specs last, on process #1, after all other processes have exited. Reach for it sparingly. OK for benchmarks, resource hogs, or specs that put an external resource into a known-bad state - but not OK as a band-aid for order-dependence or poorly isolated code/specs.

Describe("Something expensive", Serial, func() {
	It("is a resource hog that can't share the box", func() { ... })
})

→ ginkgo:decorators, ginkgo:parallelism.

Ordered: run specs in definition order

Ordered (container only) guarantees its child specs run sequentially, in written order, on one process — so they may legitimately mutate shared closure state. They can still parallelize against specs in other containers; only the inner order is pinned.

Describe("checking out a book", Ordered, func() {
	var libraryClient *library.Client
	var book *books.Book

	BeforeAll(func() {                         // once, before the first spec
		libraryClient = library.NewClient()
		Expect(libraryClient.Connect()).To(Succeed())
		DeferCleanup(libraryClient.Disconnect) // context-aware: behaves like AfterAll
	})

	It("can fetch a book", func() { book, _ = libraryClient.FetchByTitle("Les Miserables") })
	It("can check it out", func() { Expect(library.CheckOut(book)).To(Succeed()) })
	It("is then out of stock", func() { ... })

	AfterAll(func() { ... })                   // once, after the last spec
})

• BeforeAll/AfterAll are legal only inside an Ordered container (or a container nested within one). That's the whole point of Ordered — once-per-group expensive setup.
• A failing spec skips the rest of the group by default (then AfterAll still runs). Spec independence is gone, so Ginkgo won't pretend the later specs are meaningful.
• ContinueOnFailure (outermost Ordered only — error on a nested container) overrides that: keep running later specs after a failure. Use it when Ordered is just shared setup, not a dependent flow. A failed BeforeAll still skips everything — the setup is presumed broken.
• Nested containers inside an Ordered container are automatically Ordered; there's no way to un-order them.

OncePerOrdered: outer *Each setup that respects groups

A top-level BeforeEach that resets state runs before every spec — which breaks an Ordered group's first-spec-sets-up-rest contract (spec 1 passes, the rest get wiped). Decorate it with OncePerOrdered so it runs once per independent unit: still per-spec for normal specs, but once around each Ordered container (like a BeforeAll).

BeforeEach(OncePerOrdered, func() {
	libraryClient = library.NewClient()
	snapshot := libraryClient.TakeSnapshot()
	DeferCleanup(libraryClient.RestoreSnapshot, snapshot)
})

Applies to BeforeEach/JustBeforeEach/AfterEach/JustAfterEach only — not ReportBeforeEach/ReportAfterEach (reporting is always per-spec).

The Serial-inside-Ordered gotcha

You cannot mark an individual spec inside an Ordered container Serial. To make a whole ordered group run serially, put both decorators on the container: Describe("...", Ordered, Serial, func(){...}).

Ordered vs one big It

Ordered and a single It with By steps (ginkgo:writing-specs) solve the same "dependent flow" problem. Prefer one It + By by default — it stays fully parallelizable and independent. Reach for Ordered when you want each step reported as its own spec, or genuinely need BeforeAll once-per-group setup. Either way, use Ordered only when needed — it gives up internal parallelism.

SpecPriority

SpecPriority(int) (default 0, higher runs earlier, may be negative) is a scheduling hint — e.g. "boulders first, then sand" in big suites. Honored per-spec under --randomize-all, otherwise per outermost container. Anti-pattern: do not use it to impose a full deterministic order — that's what Ordered is for, and it defeats randomization's whole purpose. → ginkgo:decorators.

Flaky specs

The stance: retries are explicit technical debt. Default to none — debug the flake; it's usually telling you something about your architecture. Ginkgo gives you tools to both expose flakes and (reluctantly) tolerate them.

Expose them

ginkgo --until-it-fails -p --randomize-all   # loop forever until something breaks — best flake hunter
ginkgo --repeat=N                            # run 1+N times, stop at first failure (CI-safe)

--until-it-fails compiles once and reruns indefinitely, so it's for local hunting, not CI; --repeat=N is the CI-safe bounded version. Note --repeat=N runs 1+N times — and Ginkgo is not compatible with go test --count; use --repeat. → ginkgo:ci. Granular per-spec version: MustPassRepeatedly(N) (must pass N times in a row).

Tolerate them (judiciously, as a last resort)

ginkgo --flake-attempts=3        # retry any failing spec up to 3x; a later pass = flaky, not failed

Per-spec: FlakeAttempts(N) decorator (CLI overrides it). On a pass-after-retry the suite stays green but reports the spec as flaky.

FlakeAttempts × Ordered: a failure in an It reruns just that It (not BeforeAll/AfterAll); a failure in a BeforeAll runs the AfterAll to clean up, then reruns the BeforeAll.

Related

• Why independence matters in the first place → ginkgo:overview.
• Diagnosing why a spec failed/flaked from reports → ginkgo:debugging-failures.
• The full decorator signatures → ginkgo:decorators.