Introducing RCA Agent: Trace a Failing Data Check to Its Root Cause
Introducing RCA Agent: Trace a Failing Data Check to Its Root Cause
Introducing RCA Agent: Trace a Failing Data Check to Its Root Cause

Fabiana Ferraz
Fabiana Ferraz
Rédacteur technique chez Soda
Rédacteur technique chez Soda

Tom Baeyens
Tom Baeyens
CTO et co-fondateur chez Soda
CTO et co-fondateur chez Soda
Table des matières
When data quality checks and monitors fail, they can generate many alerts from a single upstream problem. Finding their shared cause means working through pipeline logs, warehouse data and code to understand where things went wrong. Soda’s root cause analysis (RCA) agent automates that investigation in your own Claude Code, examining related check failures together and tracing the evidence upstream.
The agent connects through MCP to the stack you already run: your orchestrator, warehouse, transformation project and code repositories. You choose those connections and configure their access permissions, so you control which systems it can investigate.
It traces the issue upstream to its root cause and reads your code to explain what caused it. After reviewing its findings, you decide which follow-up actions it runs: assess downstream impact, propose a fix, identify owners to notify, or create an incident in Soda Cloud. With a configured Jira or ServiceNow integration, that incident can also open a ticket for the team handling the fix.
That investigation is the part of data observability that follows detection. This post explains how the agent gathers evidence, what its report contains, and how you decide what happens next.
Key takeaways |
|---|
|
In the demo below, I show it in under five minutes: pasting a failing check’s link into Claude Code, letting the RCA agent trace the problem to a code change that divided contract values by 100 twice, and reviewing the report it posts to the Soda Cloud incident, with the evidence trail and the downstream datasets it cleared or flagged.
Why is root cause analysis for data quality so hard?
A failing data quality check tells you which expectation the data violated, but the cause may be several steps upstream. A check on a customer report, for example, might fail because an earlier transformation calculated contract values incorrectly. If other datasets use that same calculation, one mistake can trigger several alerts.
The pipeline’s execution status may offer little help. A transformation can finish successfully while producing incorrect values, so investigating only failed jobs can miss the cause. You need to examine what the code did and how the data changed along the way.
That evidence is spread across tools: pipeline logs show what ran, warehouse queries reveal where values went wrong, and code history shows what changed. Following data lineage upstream helps connect those findings. The work is piecing them together to explain how one problem produced the failures you see downstream.
What is Soda’s RCA agent?
Soda’s root cause analysis (RCA) agent is a Soda plugin for Claude Code that investigates a failing Soda check across the tools you connect and reports the root cause with the evidence behind it. You start it by pasting the link of a failing check from Soda Cloud into Claude Code, or by typing /rca with a description of the problem, such as a dataset that looks stale.

↗ The RCA agent, from a failing check to an incident (demo stack)
It runs next to Soda MCP, which tells it which checks are failing, and the MCP servers you register for the rest of your stack. Soda’s docs point to servers for Airflow 3 and Prefect; Snowflake, Databricks, BigQuery, PostgreSQL and Redshift; and dbt. For code history, the git and gh CLIs already on your path count.
You could point your own Claude Code at the same servers and ask, and for a quick question that may be enough. For the next incident, you’d be writing down the investigation order, the evidence rule, the coverage check and the downstream walk, and keeping that prompt working as your servers change. The agent ships all of it, and its report can land on a Soda Cloud incident with a lead from dataset ownership. Your terminal session ends. The incident page keeps the evidence.
Soda RCA agent sits with the rest of Soda AI. For what each part does and where you review it, see Soda AI, part by part.
How the RCA agent investigates a failing check
The RCA agent follows a fixed order — triage, orchestrator, upstream trace, lineage — as far as your connections reach, so you can check each step against what you’d have done yourself. Every claim in its result rests on evidence it gathered: a log line, a query result, a commit or other tool output.
Triage. Every failing check, with its dataset, type and last evaluation time. Related failures are evidence, so it reads all of them before it settles on a cause.
Orchestrator. Recent runs of the pipeline that loads those datasets, read from the task logs rather than only the task states. That’s how it catches a green run that skipped its work.
Upstream trace. The exact commit the logs say was deployed. Row counts and load timestamps from the warehouse show where fresh data turns stale; the transformation project shows which models were built or skipped.
Lineage. A walk upstream to the first dataset that’s already wrong, with lineage from the first source it can reach: the warehouse’s own, a catalog, the model graph, then the pipeline code. No lineage tool required.
It works with what you connected (it won’t search your filesystem for projects nobody mentioned). If the first wrong dataset is a raw landing table, that’s where bad data entered your stack: the report shows where it came in, not why the source system sent it.
In the demo above, a check on MART_CUSTOMER_360 failed because average contract value per customer had dropped below its threshold. The agent traced the data upstream, verified that the raw SFDC_CONTRACTS data was correct, and located the error in MART_CONTRACTS. Commit 3481ef1 had added a cents-to-dollars conversion without removing the existing division by 100, so contract values were divided by 100 twice.

↗ Tracing a failed check to the model whose checks passed (demo stack)
The difficult part was identifying which model to investigate. All nine dbt models had built successfully, and all four checks on MART_CONTRACTS had passed. The agent found the cause by examining the data and code behind those successful runs.
The report also confirmed that all five evidence sources were reachable and stated that “nothing in this analysis is inferred.” In this investigation, the warehouse queries and code change supported the finding, giving the engineer evidence to review alongside the explanation.
What’s in the RCA report?
The report leads with the finding in a few sentences. Then it lists the failing checks with links to Soda Cloud, the root cause with its evidence trail, and a lineage diagram of the causal path, marked with the offending commit or the timestamps where freshness breaks.

↗ The causal path on the Soda Cloud incident (demo stack)
The report also says what it couldn’t reach: at the start of every investigation, a coverage check finds which systems the session can reach. When one isn’t connected, the report closes with what the agent couldn’t check and what it did instead, so you can weigh “an analysis that leaned on inference rather than a query” (the docs’ words). Read that list as evidence it couldn’t gather, not as every hypothesis it didn’t test.
A missing connection never blocks the run. The agent works with what’s there and offers to help close the gap once per configuration; tell it not to ask again, and it stays quiet until your set of servers changes.
Four next steps, and you choose which ones run
After the report, the agent asks which follow-up actions to run (analyze impact, propose a fix, show whom to notify, create an incident) and runs only the ones you pick.
Analyze impact. A verdict per downstream dataset, decided from the transformation code rather than dataset names, including datasets whose checks are green but whose data is wrong. With dbt exposures, it also names the dashboards and reports that read an affected model. In the demo, it marked the Finance Bookings report as affected and cleared the customer segments model and the GTM coverage dashboard, because segments counts contracts instead of using their value.
Propose a fix. The exact revert or edit as a diff with file paths for a code cause, or a reload, backfill or quarantine for bad data. Either way, it lists what must re-run for the checks to go green.
Show whom to notify. Dataset and data source owners from Soda Cloud, plus the commit’s author, with what each needs to do.
Create an incident. An incident in Soda Cloud, with the failing results linked and the full report and impact assessment on its page.

↗ The RCA agent asks which actions to run next. All four are selected in this demo, awaiting submission.
Selecting Create incident is the go-ahead: a failing check never opens an incident by itself. The lead is the root-cause dataset’s owner, with fallbacks that end at the person whose API key the agent uses. If a teammate already opened an incident for that check result, the agent says which one covers it instead of opening a second.
With a Jira or ServiceNow webhook set up in Soda Cloud, the incident also opens a ticket there through the rule or script you built. The full report stays on the incident page, for anyone who can see all of its datasets.
Want to see the four next steps on a pipeline like yours? Book a demo
What the RCA agent needs, and what it won’t do
The RCA agent needs Claude Code with the Soda plugin and read access to the systems it should investigate, with the warehouse registered read-only. It proposes and you review: it doesn’t commit to a remote, push or open a pull request unless you ask, or create an incident unless you select that step. How far it gets depends on what you connect:
The role behind each server still decides what a trace can query.
The Soda credentials sit in one place in your Claude Code configuration, a single copy to rotate or revoke. Publishing incidents needs the incident RCA feature turned on by Soda, and an API key with Manage incidents on each linked dataset.
It runs on macOS 13+, Ubuntu 20.04+, Debian 10+, RHEL 8+, and Windows 11 23H2+ inside WSL 2. Natively, it’s supported only on Windows 11 23H2+ on ARM64.
Start with your next failing check
First, ask Soda to turn the agent on for your organization. Then add the Soda plugin to Claude Code, connect your data stack by registering servers for your orchestrator, warehouse and transformation project, and paste the link of a failing check.
For a pilot, take the next check that fails and run the agent on it while your on-call engineer investigates the usual way. Then compare: did both reach the same cause, and what did the agent say it couldn’t check?
The agent reads your stack as it is now: what’s failing, the deployed version, recent runs. A failure you already fixed leaves it little to find, so to replay a solved one, re-create it on a staging copy first.
Check the RCA agent documentation for a deeper look at setup.
Frequently Asked Questions
When data quality checks and monitors fail, they can generate many alerts from a single upstream problem. Finding their shared cause means working through pipeline logs, warehouse data and code to understand where things went wrong. Soda’s root cause analysis (RCA) agent automates that investigation in your own Claude Code, examining related check failures together and tracing the evidence upstream.
The agent connects through MCP to the stack you already run: your orchestrator, warehouse, transformation project and code repositories. You choose those connections and configure their access permissions, so you control which systems it can investigate.
It traces the issue upstream to its root cause and reads your code to explain what caused it. After reviewing its findings, you decide which follow-up actions it runs: assess downstream impact, propose a fix, identify owners to notify, or create an incident in Soda Cloud. With a configured Jira or ServiceNow integration, that incident can also open a ticket for the team handling the fix.
That investigation is the part of data observability that follows detection. This post explains how the agent gathers evidence, what its report contains, and how you decide what happens next.
Key takeaways |
|---|
|
In the demo below, I show it in under five minutes: pasting a failing check’s link into Claude Code, letting the RCA agent trace the problem to a code change that divided contract values by 100 twice, and reviewing the report it posts to the Soda Cloud incident, with the evidence trail and the downstream datasets it cleared or flagged.
Why is root cause analysis for data quality so hard?
A failing data quality check tells you which expectation the data violated, but the cause may be several steps upstream. A check on a customer report, for example, might fail because an earlier transformation calculated contract values incorrectly. If other datasets use that same calculation, one mistake can trigger several alerts.
The pipeline’s execution status may offer little help. A transformation can finish successfully while producing incorrect values, so investigating only failed jobs can miss the cause. You need to examine what the code did and how the data changed along the way.
That evidence is spread across tools: pipeline logs show what ran, warehouse queries reveal where values went wrong, and code history shows what changed. Following data lineage upstream helps connect those findings. The work is piecing them together to explain how one problem produced the failures you see downstream.
What is Soda’s RCA agent?
Soda’s root cause analysis (RCA) agent is a Soda plugin for Claude Code that investigates a failing Soda check across the tools you connect and reports the root cause with the evidence behind it. You start it by pasting the link of a failing check from Soda Cloud into Claude Code, or by typing /rca with a description of the problem, such as a dataset that looks stale.

↗ The RCA agent, from a failing check to an incident (demo stack)
It runs next to Soda MCP, which tells it which checks are failing, and the MCP servers you register for the rest of your stack. Soda’s docs point to servers for Airflow 3 and Prefect; Snowflake, Databricks, BigQuery, PostgreSQL and Redshift; and dbt. For code history, the git and gh CLIs already on your path count.
You could point your own Claude Code at the same servers and ask, and for a quick question that may be enough. For the next incident, you’d be writing down the investigation order, the evidence rule, the coverage check and the downstream walk, and keeping that prompt working as your servers change. The agent ships all of it, and its report can land on a Soda Cloud incident with a lead from dataset ownership. Your terminal session ends. The incident page keeps the evidence.
Soda RCA agent sits with the rest of Soda AI. For what each part does and where you review it, see Soda AI, part by part.
How the RCA agent investigates a failing check
The RCA agent follows a fixed order — triage, orchestrator, upstream trace, lineage — as far as your connections reach, so you can check each step against what you’d have done yourself. Every claim in its result rests on evidence it gathered: a log line, a query result, a commit or other tool output.
Triage. Every failing check, with its dataset, type and last evaluation time. Related failures are evidence, so it reads all of them before it settles on a cause.
Orchestrator. Recent runs of the pipeline that loads those datasets, read from the task logs rather than only the task states. That’s how it catches a green run that skipped its work.
Upstream trace. The exact commit the logs say was deployed. Row counts and load timestamps from the warehouse show where fresh data turns stale; the transformation project shows which models were built or skipped.
Lineage. A walk upstream to the first dataset that’s already wrong, with lineage from the first source it can reach: the warehouse’s own, a catalog, the model graph, then the pipeline code. No lineage tool required.
It works with what you connected (it won’t search your filesystem for projects nobody mentioned). If the first wrong dataset is a raw landing table, that’s where bad data entered your stack: the report shows where it came in, not why the source system sent it.
In the demo above, a check on MART_CUSTOMER_360 failed because average contract value per customer had dropped below its threshold. The agent traced the data upstream, verified that the raw SFDC_CONTRACTS data was correct, and located the error in MART_CONTRACTS. Commit 3481ef1 had added a cents-to-dollars conversion without removing the existing division by 100, so contract values were divided by 100 twice.

↗ Tracing a failed check to the model whose checks passed (demo stack)
The difficult part was identifying which model to investigate. All nine dbt models had built successfully, and all four checks on MART_CONTRACTS had passed. The agent found the cause by examining the data and code behind those successful runs.
The report also confirmed that all five evidence sources were reachable and stated that “nothing in this analysis is inferred.” In this investigation, the warehouse queries and code change supported the finding, giving the engineer evidence to review alongside the explanation.
What’s in the RCA report?
The report leads with the finding in a few sentences. Then it lists the failing checks with links to Soda Cloud, the root cause with its evidence trail, and a lineage diagram of the causal path, marked with the offending commit or the timestamps where freshness breaks.

↗ The causal path on the Soda Cloud incident (demo stack)
The report also says what it couldn’t reach: at the start of every investigation, a coverage check finds which systems the session can reach. When one isn’t connected, the report closes with what the agent couldn’t check and what it did instead, so you can weigh “an analysis that leaned on inference rather than a query” (the docs’ words). Read that list as evidence it couldn’t gather, not as every hypothesis it didn’t test.
A missing connection never blocks the run. The agent works with what’s there and offers to help close the gap once per configuration; tell it not to ask again, and it stays quiet until your set of servers changes.
Four next steps, and you choose which ones run
After the report, the agent asks which follow-up actions to run (analyze impact, propose a fix, show whom to notify, create an incident) and runs only the ones you pick.
Analyze impact. A verdict per downstream dataset, decided from the transformation code rather than dataset names, including datasets whose checks are green but whose data is wrong. With dbt exposures, it also names the dashboards and reports that read an affected model. In the demo, it marked the Finance Bookings report as affected and cleared the customer segments model and the GTM coverage dashboard, because segments counts contracts instead of using their value.
Propose a fix. The exact revert or edit as a diff with file paths for a code cause, or a reload, backfill or quarantine for bad data. Either way, it lists what must re-run for the checks to go green.
Show whom to notify. Dataset and data source owners from Soda Cloud, plus the commit’s author, with what each needs to do.
Create an incident. An incident in Soda Cloud, with the failing results linked and the full report and impact assessment on its page.

↗ The RCA agent asks which actions to run next. All four are selected in this demo, awaiting submission.
Selecting Create incident is the go-ahead: a failing check never opens an incident by itself. The lead is the root-cause dataset’s owner, with fallbacks that end at the person whose API key the agent uses. If a teammate already opened an incident for that check result, the agent says which one covers it instead of opening a second.
With a Jira or ServiceNow webhook set up in Soda Cloud, the incident also opens a ticket there through the rule or script you built. The full report stays on the incident page, for anyone who can see all of its datasets.
Want to see the four next steps on a pipeline like yours? Book a demo
What the RCA agent needs, and what it won’t do
The RCA agent needs Claude Code with the Soda plugin and read access to the systems it should investigate, with the warehouse registered read-only. It proposes and you review: it doesn’t commit to a remote, push or open a pull request unless you ask, or create an incident unless you select that step. How far it gets depends on what you connect:
The role behind each server still decides what a trace can query.
The Soda credentials sit in one place in your Claude Code configuration, a single copy to rotate or revoke. Publishing incidents needs the incident RCA feature turned on by Soda, and an API key with Manage incidents on each linked dataset.
It runs on macOS 13+, Ubuntu 20.04+, Debian 10+, RHEL 8+, and Windows 11 23H2+ inside WSL 2. Natively, it’s supported only on Windows 11 23H2+ on ARM64.
Start with your next failing check
First, ask Soda to turn the agent on for your organization. Then add the Soda plugin to Claude Code, connect your data stack by registering servers for your orchestrator, warehouse and transformation project, and paste the link of a failing check.
For a pilot, take the next check that fails and run the agent on it while your on-call engineer investigates the usual way. Then compare: did both reach the same cause, and what did the agent say it couldn’t check?
The agent reads your stack as it is now: what’s failing, the deployed version, recent runs. A failure you already fixed leaves it little to find, so to replay a solved one, re-create it on a staging copy first.
Check the RCA agent documentation for a deeper look at setup.
Frequently Asked Questions
Does the RCA agent fix the problem for me?
No. It proposes: a diff with file paths for a code cause, or a reload, backfill or quarantine for bad data, plus what must re-run. It doesn't commit to a remote, push or open a pull request unless you ask, and you review its findings first.
Do I need a data lineage tool to use it?
No. It takes lineage from the first source it can reach: your warehouse's own lineage, a catalog, your transformation project's model graph, or the pipeline code itself. The report then draws the causal path, from the root cause to the failing checks.
What can it do with only Soda connected?
With only Soda connected, the agent can triage: it sees which checks fail, on which datasets, and when they last ran. It can't query your warehouse or read pipeline logs, so it can't confirm a root cause, and its report says what it couldn't check.
Why not ask Claude Code myself?
You can, and for a quick question that may be enough. For repeat incidents you'd be writing and maintaining the investigation order, the evidence rule, the coverage check and the impact walk yourself. The agent ships them, and can publish its report to a Soda Cloud incident.
Where does the investigation run?
The investigation runs in your Claude Code session, with the MCP servers and credentials you register. What the RCA agent reads goes to the model your Claude Code uses, under your agreement with that model's provider. Creating an incident publishes the report and impact assessment to Soda Cloud.
Trusted by the world’s leading enterprises
Real stories from companies using Soda to keep their data reliable, accurate, and ready for action.
At the end of the day, we don’t want to be in there managing the checks, updating the checks, adding the checks. We just want to go and observe what’s happening, and that’s what Soda is enabling right now.

Sid Srivastava
Director of Data Governance, Quality and MLOps
Investing in data quality is key for cross-functional teams to make accurate, complete decisions with fewer risks and greater returns, using initiatives such as product thinking, data governance, and self-service platforms.

Mario Konschake
Director of Product-Data Platform
Soda has integrated seamlessly into our technology stack and given us the confidence to find, analyze, implement, and resolve data issues through a simple self-serve capability.

Sutaraj Dutta
Data Engineering Manager
Our goal was to deliver high-quality datasets in near real-time, ensuring dashboards reflect live data as it flows in. But beyond solving technical challenges, we wanted to spark a cultural shift - empowering the entire organization to make decisions grounded in accurate, timely data.

Gu Xie
Head of Data Engineering
4,4 sur 5
Commencez à faire confiance à vos données. Aujourd'hui.
Trouvez, comprenez et corrigez tout problème de qualité des données en quelques secondes.
Du niveau de la table au niveau des enregistrements.
Adopté par
Product
Blogue





Trusted by the world’s leading enterprises
Real stories from companies using Soda to keep their data reliable, accurate, and ready for action.
At the end of the day, we don’t want to be in there managing the checks, updating the checks, adding the checks. We just want to go and observe what’s happening, and that’s what Soda is enabling right now.

Sid Srivastava
Director of Data Governance, Quality and MLOps
Investing in data quality is key for cross-functional teams to make accurate, complete decisions with fewer risks and greater returns, using initiatives such as product thinking, data governance, and self-service platforms.

Mario Konschake
Director of Product-Data Platform
Soda has integrated seamlessly into our technology stack and given us the confidence to find, analyze, implement, and resolve data issues through a simple self-serve capability.

Sutaraj Dutta
Data Engineering Manager
Our goal was to deliver high-quality datasets in near real-time, ensuring dashboards reflect live data as it flows in. But beyond solving technical challenges, we wanted to spark a cultural shift - empowering the entire organization to make decisions grounded in accurate, timely data.

Gu Xie
Head of Data Engineering
4,4 sur 5
Commencez à faire confiance à vos données. Aujourd'hui.
Trouvez, comprenez et corrigez tout problème de qualité des données en quelques secondes.
Du niveau de la table au niveau des enregistrements.
Adopté par
Product
Solutions
Blogue





Trusted by the world’s leading enterprises
Real stories from companies using Soda to keep their data reliable, accurate, and ready for action.
At the end of the day, we don’t want to be in there managing the checks, updating the checks, adding the checks. We just want to go and observe what’s happening, and that’s what Soda is enabling right now.

Sid Srivastava
Director of Data Governance, Quality and MLOps
Investing in data quality is key for cross-functional teams to make accurate, complete decisions with fewer risks and greater returns, using initiatives such as product thinking, data governance, and self-service platforms.

Mario Konschake
Director of Product-Data Platform
Soda has integrated seamlessly into our technology stack and given us the confidence to find, analyze, implement, and resolve data issues through a simple self-serve capability.

Sutaraj Dutta
Data Engineering Manager
Our goal was to deliver high-quality datasets in near real-time, ensuring dashboards reflect live data as it flows in. But beyond solving technical challenges, we wanted to spark a cultural shift - empowering the entire organization to make decisions grounded in accurate, timely data.

Gu Xie
Head of Data Engineering
4,4 sur 5
Commencez à faire confiance à vos données. Aujourd'hui.
Trouvez, comprenez et corrigez tout problème de qualité des données en quelques secondes.
Du niveau de la table au niveau des enregistrements.
Adopté par
Product
Solutions
Company
Blogue




