Practical Flaky Failure Triage Framework for Distributed Database CI

arXiv·medium signal

Addresses a real DevOps pain point: after a CI failure in distributed databases, operators must decide within milliseconds whether to auto-rerun (likely flaky) or escalate (likely persistent). Existing approaches fail because they rely on post-failure artifacts, produce poorly calibrated scores under telemetry shifts, or require controlled lab conditions. This framework operates under CPU-only millisecond budgets on real production CI workloads. Directly applicable to any team running distributed system CI pipelines.