decisions/testing

An invariant checker that applications and adapters run against any store after a scenario: a race, a random walk, a restart.

It reads every occurrence of the audited points through the adapter’s store.Inspection and the points’ desks, and checks:

InvariantMeaning
I1At most one release per occurrence; the gated tool’s effects equal the released occurrences, when the audit counts effects
I2A release follows an acceptance: the policy proceeded, or the required verdicts approved or corrected
I3Nothing changes an outcome after it settles. After settlement only these entries are appended: assessed, effected and verdict_refused on any occurrence, and reviewed, opened, noted and expanded on a sampled one. A recorded refusal names a showing to its validator made before the occurrence settled. One exception: a recommended entry may follow the expired entry of a record not begun whose own deadline passed, and the occurrence stays expired
I4One deciding verdict per validator per output that stands
I5Every wake-up the occurrence needs as it stands (occurrence.wake_times: the end of a hold, the deadline, each later policy answer; for a record not begun, occurrence.pending_wake_times: its own deadline) is pending; a settled occurrence has none
I6A settled occurrence has exactly one settlement notice, an unsettled one none
I7Entries are numbered 1, 2, 3 and so on, and their times never decrease, so the record as known earlier is a prefix of the record now
I8The head equals the head the record implies (occurrence.head), field by field

I4 differs from STRESS.md’s “one deciding verdict per validator per occurrence” on purpose. A correction replaces the output that stands, and four eyes then needs fresh approvals of the new output, possibly from someone who approved the old one; so the rule is one verdict per validator for each output that stands, and a correction starts a new round that its author belongs to.

A record that has not begun (expected, or holding only failed attempts) needs a wake-up at its own deadline when it has one (occurrence.pending_wake_times) and no notice; once a newer event superseded it or its deadline passed, it is settled: no wake-up, exactly one notice, head status Expired. A recommendation may still follow an expiry by deadline: the occurrence then starts expired, and the notice the expiry admitted is its one notice.

import decisions/testing

let audits = [testing.audit(refund_desk, effects: count_refunds)]
let assert Ok(summary) =
  testing.check(store, memory.inspection(), audits)

Target-independent: builds for Erlang and JavaScript.

Types

One point to audit, with its output type erased.

pub opaque type Audit(tx)

What the checker saw when every invariant held.

pub type Summary {
  Summary(occurrences: Int, settled: Int, released: Int)
}

Constructors

  • Summary(occurrences: Int, settled: Int, released: Int)

A broken invariant.

pub type Violation {
  Violation(
    invariant: String,
    occurrence: String,
    detail: String,
  )
}

Constructors

  • Violation(invariant: String, occurrence: String, detail: String)

Values

pub fn audit(
  desk: decisions.Desk(o, tx),
  effects effects: fn(tx) -> Result(Int, String),
) -> Audit(tx)

Audit the occurrences of a desk’s point. effects counts what the gated tool did, for I1.

pub fn audit_record(desk: decisions.Desk(o, tx)) -> Audit(tx)

Audit the occurrences of a desk’s point without counting effects.

pub fn check(
  tx: tx,
  inspect: store.Inspection(tx),
  audits: List(Audit(tx)),
) -> Result(Summary, List(Violation))

Check every invariant over the audited points, in tx.

✨ Search Document