Show HN: ML condenses billions of logs into a tiny snapshot your LLM can debug
TL;DR
Rocketgraph is an open-source observability tool that does not dump raw logs into an LLM. It first turns log lines into structural patterns and flags unusual clusters. Its ML pipeline uses Drain3 for templates, Isolation Forest for anomaly scores by service and Half-Space-Trees to score brand-new logs in real time. The README claims a production-burst test where 2M logs became 58 templates and 9 anomalies in 90 seconds on a single container.
Nauti's Take
Smart idea, but the physics of detail decides this one: if the snapshot compresses away the wrong edges, your LLM just hallucinates with better posture. This gets real when Rocketgraph proves which incident classes survive the squeeze and which disappear into pattern soup.
Briefingshow
The interesting part is not the LLM, but the compression step: a debugging agent does not need millions of lines, it needs a small set of robust patterns with rarity signals. If this works, observability moves from dashboard hunting to machine-readable incident snapshots. The catch: a GitHub demo only becomes credible after it survives noisy production data.
Sources