The load-bearing distinction is between a test run and a test's health. A run is a point-in-time pass or fail. Health is a claim about every interval the test is accountable for, and Grab's orchestrator treats a test as unhealthy until the underlying issue is fully resolved [8]. The worked example is a completeness check on the T-1 daily partition: it failed two days ago, passed yesterday and today, and remains unhealthy until the older partition has been backfilled and verified [9]. Two green runs in a row do not close the issue, which is inconvenient for dashboards and correct for consumers.
That state machine is forced by the auto-close requirement. Grab designed the lifecycle to run with minimal human in the loop, and a good DPI should close itself when the condition recovers [5]. If closure is keyed to the last run, the stack self-resolves at the next successful schedule and leaves the hole in the partition behind it.
The second decision is where the semantics live. Kinabalu, Grab's incident orchestrator, does not compute freshness or completeness. It asks whether the contract's tests are healthy through Test Health APIs, and one unhealthy test means a breached contract and a new DPI [6][7]. The contract is a versioned producer-consumer agreement covering schema, freshness, completeness and other semantic guarantees, codified as data quality tests in Genchi [13]. So the alert-to-ticket step is a lookup against a named test set, not a heuristic about severity.
Noise control sits in triage. New signals for an already-open breach update that DPI instead of forking a second one, and DPIs with a shared root cause can be grouped [11]. Each promoted issue carries the affected asset, the breached contract, the unhealthy tests, the data interval, and upstream and downstream dependencies [12]. That context list is also the diagnostic input, which is why the detect step has to gather it rather than link to a dashboard.
This only transfers if a few conditions hold. Contracts need to name specific tests rather than describe a table in prose, and the quality layer needs to express health per test across intervals rather than just per run [8]. The ownership graph also needs to name a producing job, a platform incident or a late upstream, since otherwise every breach lands on the asset owner and you pay for the handoff [14]. Grab has the first two through Genchi and the Data Contract Registry [2], which is the part of the certification stack the ticket lifecycle actually reads [6].
The contentious half is assignment, and the published excerpt stops at the sentence that introduces it [17]. Certification, meanwhile, is doing the job it was scoped for: it documents ownership, lineage and controls, and it does not stop a certified Kafka stream from violating a freshness expectation [3].
