Baidu Search Business Unmanned Delivery Practice: Risk-Driven Quality Assurance Exploration

This article introduces Baidu's practice of risk-driven delivery in the context of intelligent testing during the perceptual intelligence phase. The author identifies three key challenges: (1) Not all projects carry risks - over 80% have no associated bugs or online issues; (2) Not all testing tasks can uncover defects - the ratio of effective quality behaviors is quite low; (3) Testers can also make misjudgments, with missed defects being a persistent issue.

The article proposes achieving "unmanned delivery" by addressing three human dependencies: decision-making that relies on humans for determining which tests to run after CI; process dependencies requiring manual communication between developers and testers; and conclusion dependencies where pipelines lack risk analysis capabilities and test results lack binary pass/fail determinations.

Three essential capabilities are required: (1) Comprehensive testing capability - ensuring pipeline test tasks are comprehensive and effective, illustrated through DIFF testing human-like analysis that correlates code changes with field differences using white-box data, models, and business logic; (2) Stable build capability - applying online service stability standards to offline testing, including infrastructure governance, full-link stability assurance (prevention, detection, localization, recovery, closure), and build digitalization; (3) Precise evaluation capability - using quality model and rule-based approaches to automatically assess entry/exit risks by abstracting delivery process data (code white-box data, developer profiles, process data, test results, coverage) as features for model training.

Through these efforts, Baidu's search business has achieved over 90% of requirements completed via developer self-testing, with over 40% achieving unmanned delivery, significantly reducing delivery costs and freeing tester capacity for higher-value work.