The Missing Modality

Microservice incident management uses multiple data modalities — logs, metrics, traces, alerts — to detect, diagnose, and resolve incidents. Recent frameworks fuse these modalities for joint optimization: the log says what happened, the metric says when the anomaly started, the trace says which service propagated the fault. Together, they provide a richer signal than any modality alone.

The unrealistic assumption: all modalities are always available. In practice, network fluctuations drop metric streams, agent failures silence log collectors, and trace sampling misses the relevant request. The missing data isn't random — it's correlated with the incident. The same network partition that causes the incident also prevents the monitoring data from reaching the analysis system. The modalities most likely to be missing are the ones most relevant to the current failure.

Missing-aware fusion handles incomplete data by design rather than by imputation. Instead of filling in missing modalities with synthetic data (which can hallucinate patterns that don't exist), the system adapts its fusion strategy based on which modalities are available. When all modalities are present, full fusion. When logs are missing, the system relies on metrics and traces with adjusted confidence. When only metrics are available, the system falls back to a metrics-only model with wider uncertainty bounds.

The adaptation is learned, not hardcoded: the system trains on deliberately masked modalities, learning how to degrade gracefully across all possible absence patterns. The key insight is that the fusion weights should change when modalities disappear — not just by zeroing the missing modality's weight but by re-weighting the remaining modalities to compensate for the lost information.

The through-claim: multimodal incident management systems fail at the worst possible time — during incidents — because the same failure that creates the incident also disrupts the monitoring. Robustness to missing modalities isn't a nice-to-have; it's a requirement for the system to work when it matters most.