data-quality-gate

Overview

The three engines (dbt-testing, great-expectations, soda-checks) each produce their own per-check output: dbt writes run_results.json (dbt-run-results), Great Expectations returns a JSON result with a success flag and a results list (gx-run-validation-definition), and Soda emits a per-check pass/fail summary in stdout (and to Soda Cloud) (sodacl-overview).

This skill builds a single release-readiness gate that consumes those heterogeneous outputs, applies severity-aware thresholds, and emits one go / no-go decision the CI pipeline can act on. The skill is deliberately engine-agnostic at the surface so a project can mix engines (e.g. dbt for transformation tests + GX for ingestion validation + Soda for cross-team observability) and still ship a single gate.

When to use

• A pipeline runs more than one data-quality engine and needs a unified pass/fail summary (instead of one CI job per engine).
• Some checks are blocking (severity: error) and others should warn but not stop the pipeline (severity: warn) - the gate must distinguish.
• The team wants a structured artifact (JSON / markdown table) for PR comments, dashboards, or post-mortems - not just exit 1.
• A migration is rolling out new checks; the gate needs to honor a per-check "ratchet" (existing failures grandfathered, new failures block).

If a project uses only one engine and does not need severity tiering, prefer that engine's native CI integration directly (see the relevant skill's "CI integration" section) - this gate adds machinery you do not need.

Step 1 - Identify your sources

Enumerate every check-emitting engine the gate must consume. For each:

Persist each engine's raw artifact as a CI build artifact (matching the pattern from each skill's "CI integration" section) so the gate input is reproducible and triageable.

Step 2 - Define the unified check record

Flatten every engine's result into one record shape:

{
  "check_id":     "dbt.test.orders.unique_order_id",
  "engine":       "dbt",
  "subject":      "orders.order_id",
  "status":       "fail",
  "severity":     "error",
  "failures":     12,
  "message":      "12 unique-key violations",
  "ratchet":      false,
  "owner":        "@data-platform"
}

Step 3 - Apply the gate decision rule

Pseudocode:

def gate_decision(records, *, allow_warn_failures=True):
    blockers = [
        r for r in records
        if r["status"] in ("fail", "error")
        and r["severity"] == "error"
        and not r.get("ratchet", False)
    ]
    warnings = [
        r for r in records
        if (r["status"] in ("fail", "error") and r["severity"] == "warn")
        or r["status"] == "warn"
    ]
    return {
        "verdict": "no-go" if blockers else "go",
        "blocker_count": len(blockers),
        "warning_count": len(warnings),
        "blockers": blockers,
        "warnings": warnings,
    }

The default is strict-but-warn-tolerant: any non-ratcheted error-severity failure blocks; warn-severity failures and ratcheted records show in the report but do not block.

For a stricter mode (no warn tolerance), set allow_warn_failures=False and treat warning_count > 0 as a blocker.

Step 4 - Emit the artifact

The gate produces a markdown summary suitable for both $GITHUB_STEP_SUMMARY and Soda Cloud / Slack pipelines:

# Data Quality Gate - verdict: NO-GO

**Blockers: 2**

| Engine | Subject              | Check                  | Failures | Owner            |
|--------|----------------------|------------------------|---------:|------------------|
| dbt    | orders.order_id      | unique                 |       12 | @data-platform   |
| gx     | orders.discount_pct  | ExpectColumnValuesToBeBetween | 4 | @analytics-eng   |

**Warnings: 1**

| Engine | Subject     | Check               | Failures | Owner       |
|--------|-------------|---------------------|---------:|-------------|
| soda   | customers   | row_count > 0       |        0 | @platform   |

Plus a JSON sibling for downstream consumers:

{
  "verdict": "no-go",
  "blocker_count": 2,
  "warning_count": 1,
  "blockers": [...],
  "warnings": [...]
}

A no-go verdict exits non-zero so the CI pipeline halts.

Worked example

A full end-to-end collector (scripts/run_quality_gate.py) that reads each engine's artifact, flattens it into the unified record shape, applies the Step 3 gate decision, and wires the gate into CI is in references/worked-example.md.

References

• dbt-testing/SKILL.md - dbt run_results.json schema and field meanings.
• great-expectations/SKILL.md - GX result object shape and result_format levels.
• soda-checks/SKILL.md - Soda CLI invocation and stdout summary format.
• dbt-run-results - canonical run_results.json schema.
• gx-run-validation-definition - GX validation_definition.run() return shape.
• sodacl-overview - SodaCL check vocabulary.