Deep Dive: AI Attempt: Stalled worker overwrites database after lock expires
An engineering analysis of why this AI-generated Pull Request passed isolated unit tests, how multi-tenant traffic exposes the architectural flaw, and how interviewers score your response.
The Root Cause Anti-Pattern
In distributed lock algorithms, lease timeouts cannot guarantee safety across network delays or garbage collection pauses without monotonic fencing tokens. Authorizing storage writes by reading a shared mutable instance attribute (like self.current_fence) instead of the worker's individually acquired lease token allows a stalled worker to commit under a newer worker's watermark, silently overwriting fresh data with stale output.
The Interviewer's Scoring Rubric
Candidates are assessed on whether they jump to trial-and-error editing or follow a hypothesis-driven approach. A "Strong Hire" candidate restates the data flow between modules, explains why single-client unit tests passed, crafts an adversarial probe showing cross-tenant leakage, and implements a surgical 1-line patch without breaking existing contracts.
Candidate Strategy for the Code Comprehension Round
- 1Map Module ContractsTrace how the gateway invokes the underlying service. Note what data is passed by reference vs. by value.
- 2Break It With An Adversarial ProbeSimulate concurrent or sequential traffic across multiple identities to force shared state mutation.
- 3Deliver a Surgical PatchMake the minimal required diff to ensure proper per-client state isolation without refactoring unnecessary files.