Silence Is Not Guilt: Fault-Correlated Missing Telemetry as a Shortcut in Learned Root-Cause Localization

Work on microservice root-cause analysis has shown that models can read silence as health and miss a fault once the failing component stops reporting. This paper is about the opposite error. Some faults silence the very service they break, so in fault-injection data missing telemetry tends to mark the root, and a supervised ranker trained on such data learns to blame whichever component goes quiet. A monitoring gap on a healthy component is then enough to cause a misdiagnosis. On a testbed of 170 incidents in 18 scenario groups, learned rankers placed the root first in only 20 to 22% of incidents when the telemetry of an innocent component was hidden, where a simple anomaly baseline reached 57%. Partial gaps did far less damage. Accuracy fell when an innocent component lost 75% of its metric channels, but not when it lost 50% or less, and not when monitoring stopped partway through an incident. We then trained the same ranker on additional views in which one component is silent yet still a candidate, a variant of an existing augmentation. With one component hidden this raised accuracy by 37.1 points over the uncorrected ranker (95% interval +26.1 to +47.7; pre-registered), and 88% of silent roots were still found. Removing the masked component from the candidate set did not help (descriptive comparison). On RCAEval, with the root silenced in 37% of incidents (real telemetry, simulated silencing), the same correction recovered 24.1 points (+17.9 to +30.4; pre-registered) at a small, unresolved cost on complete telemetry. In a second public benchmark of 1,422 fault injections no root went completely silent: every root kept at least one metric row, and a Java service lost both its logs and its traces in only 18 incidents (exploratory).