Metron FAQ
What Metron is, how it works with your warehouse, and what it will and won't claim about a change in your metrics.
Updated · Zello Labs
Metron is self-hosted root-cause analysis for business metrics, built by Zello Labs and in private beta. It checks a metric in your warehouse against its own normal, finds the segment that holds most of a change, lines it up with nearby events and writes up the evidence. See how Metron works, step by step.
Questions and answers
What is Metron?
Metron is self-hosted root-cause analysis for business metrics, built by Zello Labs. It checks a metric in your data warehouse against its own normal, finds the segment that holds most of a change, lines the change up with deploys and logged events, and writes up the evidence in plain language for your team.
Is Metron self-hosted?
Yes. Metron runs in your own environment and queries your warehouse in place with a read-only role. There is no replication pipeline and no second copy of your data. Raw rows stay where they are. The language model you connect sees only the engine's findings, and the write-up goes wherever you send it.
Which warehouses does Metron support?
PostgreSQL and Google BigQuery are live today, so you can define metrics and run detection on either. Snowflake is coming soon, and Amazon Redshift and Azure Synapse are planned. Metron needs SQL and a timestamp column, so if you run something else, name it in the beta form and it moves up the list.
Does Metron send my data to a language model?
No raw data. Metron's engine does the statistics inside your environment, then hands the language model only its findings, such as the metric, the segment, the size of the change and the events near it. You supply the model endpoint, so you decide where that text goes. The model words the result and never reads your tables.
Does Metron tell you what caused a change?
It tells you where the change sits and what happened right before it, and it scores each of those claims on its own. It does not claim cause, because warehouse data can't prove one. You get the segment, the timing and the nearby events, and your team makes the call.
How is Metron different from data observability tools?
Data observability tools watch the health of your data, such as freshness, schema changes and broken pipelines. Metron works on the business numbers themselves and asks why one of them moved. It explains a change in revenue, conversion or any other KPI, and leaves pipeline failures to the tools built to catch them.
How much history does a metric need?
Metron reads a metric's history straight from your warehouse, so a metric built on existing data can start right away. It needs about two weeks of history before it flags anything. Until there is enough history to separate the weekly pattern and the trend from a real change, it uses a simpler model.
How does setup work?
Metron runs in your own environment. You connect your warehouse with a read-only role and point Metron at the language model endpoint you want it to use. Then you define each metric once, as a SQL aggregate, a timestamp column and the columns to slice it by. During the private beta, we set up each team by hand.
What do I need to start?
You need a PostgreSQL or BigQuery warehouse, a role that can SELECT the tables behind your metric, and a language model endpoint you supply. A GitHub webhook is optional and brings in deploys and releases. To apply for the beta, you only need to tell us your warehouse and one metric you want Metron to watch.
How does Metron deliver alerts?
By webhook or email. You give Metron a webhook URL, an email address, or both. When a run flags a new change, Metron sends the write-up there, such as to your team's channel or on-call inbox. Each change is sent once, so running Metron again doesn't repeat the alert.
Does marking an alert useful or wrong change how Metron detects?
No. You can mark an investigation as useful, a false positive, a wrong attribution or expected behavior, and Metron stores that verdict with the evidence behind it. Your verdicts do not retrain or tune detection. Over time they build a record of which alerts held up and which wasted your morning.
What does a write-up contain?
A few plain sentences with the numbers in them. It says how far the metric is from expected, which segment holds most of the change, whether the rate or the mix moved for a ratio metric, and which events landed near the start. Each number is checked against the engine's findings, and cause is marked as not established.
Can I ask follow-up questions?
Yes. After a write-up, you can ask about a segment, such as a region or a plan. Metron keeps the same metric and the same change, reruns the breakdown inside that segment, and links the new answer to the first one. It answers from the same evidence. Follow-ups stay focused on segments Metron already knows about.
What does private beta mean for getting access?
Metron is not generally available yet. We are setting up a small number of teams by hand. Request access with your work email, the warehouse you run and one metric your team can't explain, and we will get back to you. We use your email only to set you up for the beta.