Reliability
Define reliability in the language of the consumer
Freshness, completeness and correctness are separate promises. Give each one an owner and a response.
Name the data product
Identify the consumers, the decisions they make and when they need the data. “The job succeeded” is an execution signal; it does not establish whether the dataset is useful.
Write observable expectations
Specify a freshness deadline, completeness checks and important business invariants. Set thresholds with consumers and document exceptions, including planned source outages and late reporting.
Design an actionable alert
An alert should identify the affected dataset, the failed expectation and the responsible team. Include a runbook and distinguish delayed input from failed transformation.
Practice the incident
Walk through a source schema change or missing partition. Verify that the team can detect the issue, communicate impact and recover without concealing data gaps.