Troubleshooting dbt Job Errors

Systematically diagnose and resolve dbt Cloud job failures using available MCP tools, CLI commands, and data investigation.

When to Use

• •dbt Cloud / dbt platform job failed and you need to find the root cause
• •Intermittent job failures that are hard to reproduce
• •Error messages that don't clearly indicate the problem
• •Post-merge failures where a recent change may have caused the issue

Not for: Local dbt development errors - use the skill using-dbt-for-analytics-engineering instead

The Iron Rule

Never modify a test to make it pass without understanding why it's failing.

A failing test is evidence of a problem. Changing the test to pass hides the problem. Investigate the root cause first.

Rationalizations That Mean STOP

Workflow

flowchart TD
    A[Job failure reported] --> B{MCP Admin API available?}
    B -->|yes| C[Use list_jobs_runs to get history]
    B -->|no| D[Ask user for logs and run_results.json]
    C --> E[Use get_job_run_error for details]
    D --> F[Classify error type]
    E --> F
    F --> G{Error type?}
    G -->|Infrastructure| H[Check warehouse, connections, timeouts]
    G -->|Code/Compilation| I[Check git history for recent changes]
    G -->|Data/Test Failure| J[Use discovering-data skill to investigate]
    H --> K{Root cause found?}
    I --> K
    J --> K
    K -->|yes| L[Create branch, implement fix]
    K -->|no| M[Create findings document]
    L --> N[Add test - prefer unit test]
    N --> O[Create PR with explanation]
    M --> P[Document what was checked and next steps]

Step 1: Gather Job Run Information

If dbt MCP Server Admin API Available

Use these tools first - they provide the most comprehensive data:

# Example: Get recent runs for job 12345
list_jobs_runs(job_id=12345, limit=10)

# Example: Get error details for specific run
get_job_run_error(run_id=67890)

Without MCP Admin API

Ask the user to provide these artifacts:

1. •Job run logs from dbt Cloud UI (Debug logs preferred)
2. •run_results.json - contains execution status for each node

To get the run_results.json, generate the artifact URL for the user:

https://<DBT_ENDPOINT>/api/v2/accounts/<ACCOUNT_ID>/runs/<RUN_ID>/artifacts/run_results.json?step=<STEP_NUMBER>

Where:

• •<DBT_ENDPOINT> - The dbt Cloud endpoint. e.g
  • •cloud.getdbt.com for the US multi-tenant platform (there are other endpoints for other regions)
  • •ACCOUNT_PREFIX.us1.dbt.com for the cell-based platforms (there are different cell endpoints for different regions and cloud providers)
• •<ACCOUNT_ID> - The dbt Cloud account ID
• •<RUN_ID> - The failed job run ID
• •<STEP_NUMBER> - The step that failed (e.g., if step 4 failed, use ?step=4)

Example request:

"I don't have access to the dbt MCP server. Could you provide:

1. •The debug logs from dbt Cloud (Job Run → Logs → Download)
2. •The run_results.json - open this URL and copy/paste or upload the contents: https://cloud.getdbt.com/api/v2/accounts/12345/runs/67890/artifacts/run_results.json?step=4

Step 2: Classify the Error

Step 3: Investigate Root Cause

For Infrastructure Errors

1. •Check job configuration (timeout settings, execution steps, etc.)
2. •Look for concurrent jobs competing for resources
3. •Check if failures correlate with time of day or data volume

For Code/Compilation Errors

1. •

   Check git history for recent changes:

   If you're not in the dbt project directory, use the dbt MCP server to find the repository:

   code

   # Get project details including repository URL and project subdirectory
   get_project_details(project_id=<project_id>)
   

   The response includes:

   • •repository - The git repository URL
   • •dbt_project_subdirectory - Optional subfolder where the dbt project lives (e.g., dbt/, transform/analytics/)

   Then either:

   • •Query the repository directly using gh CLI if it's on GitHub
   • •Clone to a temporary folder: git clone <repo_url> /tmp/dbt-investigation

   Important: If the project is in a subfolder, navigate to it after cloning:

   bash

   cd /tmp/dbt-investigation/<project_subdirectory>
   

   Once in the project directory:

   bash

   git log --oneline -20
   git diff HEAD~5..HEAD -- models/ macros/
   

2. •

   Use the CLI and LSP tools from the dbt MCP server or use the dbt CLI to check for errors:

   If the dbt MCP server is available, use its tools:

   code

   # CLI tools
   mcp__dbt_parse()                              # Check for parsing errors
   mcp__dbt_list_models()                        # With selectos and `+` for finding models dependencies
   mcp__dbt_compile(models="failing_model")      # Check compilation
   
   # LSP tools
   mcp__dbt_get_column_lineage()                 # Check column lineage
   

   Otherwise, use the dbt CLI directly:

   bash

   dbt parse          # Check for parsing errors
   dbt list --select +failing_model          # Check for models upstream of the failing model
   dbt compile --select failing_model  # Check compilation
   

3. •

   Search for the error pattern:

   • •Find where the undefined macro/model should be defined
   • •Check if a file was deleted or renamed

For Data/Test Failures

Use the discovering-data skill to investigate the actual data.

1. •

   Get the test SQL

   bash

   dbt compile --select project_name.folder1.folder2.test_unique_name --output json
   

   the full path for the test can be found with a dbt ls --resource-type test command

2. •

   Query the failing test's underlying data:

   bash

   dbt show --inline "<query_from_the_test_SQL>" --output json
   

3. •

   Compare to recent git changes:

   • •Did a transformation change introduce new values?
   • •Did upstream source data change?

Step 4: Resolution

If Root Cause Is Found

1. •

   Create a new branch:

   bash

   git checkout -b fix/job-failure-<description>
   

2. •

   Implement the fix addressing the actual root cause

3. •

   Add a test to prevent recurrence:

   • •Prefer unit tests for logic issues
   • •Use data tests for data quality issues
   • •Example unit test for transformation logic:
   yaml

   unit_tests:
     - name: test_status_mapping
       model: orders
       given:
         - input: ref('stg_orders')
           rows:
             - {status_code: 1, expected_status: 'pending'}
             - {status_code: 2, expected_status: 'shipped'}
       expect:
         rows:
           - {status: 'pending'}
           - {status: 'shipped'}
   

4. •

   Create a PR with:

   • •Description of the issue
   • •Root cause analysis
   • •How the fix resolves it
   • •Test coverage added

If Root Cause Is NOT Found

Do not guess. Create a findings document.

Create docs/investigations/job-failure-<date>.md:

# Job Failure Investigation: <Job Name>

**Date:** YYYY-MM-DD
**Job ID:** <id>
**Status:** Unresolved

## Summary
Brief description of the failure and symptoms.

## What Was Checked

### Tools Used
- [ ] list_jobs_runs - findings
- [ ] get_job_run_error - findings
- [ ] git history - findings
- [ ] Data investigation - findings

### Hypotheses Tested
| Hypothesis | Evidence | Result |
|------------|----------|--------|
| Recent code change | No changes to affected models in 7 days | Ruled out |

## Patterns Observed
- Failures occur between 2-4 AM (peak load time?)
- Always fails on model X

## Suggested Next Steps
1. [ ] Check the data ingestion process to see if new data was added
2. [ ] Check if a new version of dbt or of the dbt adapter was released

## Related Resources
- Link to job run logs
- Link to relevant documentation

Commit this document to the repository so findings aren't lost.

Quick Reference

Common Mistakes

Modifying tests to pass without investigation

• •A failing test is a signal, not an obstacle. Understand WHY before changing anything.

Skipping git history review

• •Most failures correlate with recent changes. Always check what changed.

Not documenting when unresolved

• •"I couldn't figure it out" leaves no trail. Document what was checked and what remains.

Making best-guess fixes under pressure

• •A wrong fix creates more problems. Take time to diagnose properly.

Ignoring data investigation for test failures

• •Test failures often reveal data issues. Query the actual data before assuming code is wrong.