apache/gluten/ansi-analysis

ANSI Test Analysis Skill

Step 0 — MUST READ FIRST: shared analysis prompt

Before doing anything else, read the shared prompt that defines the analysis output format and reference source locations:

.github/skills/ansi-analysis/shared.md

This file is the single source of truth — the same content is consumed by the CI Python pipeline (.github/skills/ansi-analysis/analyze-ansi.py --ai-analysis). Your output structure, reference source locations, and self-investigation steps MUST follow it. If the file is missing, STOP and tell the user the repo is in a broken state.

Step 1 — Decide entry point

Ask the user (or infer from request):

• Run new tests? → Step 2
• Re-analyze existing JSON in target/ansi-offload/? → Step 3
• Diagnose a single test failure? → Step 4

Step 2 — Run the verification script

./dev/verify-ansi-expressions.sh <category> <spark41|spark40|all> [--clean]

Categories: cast | arithmetic | collection | datetime | math | decimal | string | aggregate | errors | all

Logs: /tmp/ansi-matrix/latest/ (bash logs). JSON: target/ansi-offload/*.json (written by GlutenExpressionOffloadTracker.scala, this is the structured input for analysis).

Notes from prior runs:

• Use all mode in single JVM (~28 min) when full coverage is needed
• After rebase / branch switch, run ./dev/builddep-veloxbe-inc.sh first to refresh libvelox.so / libgluten.so

Step 3 — Analyze JSON results

Two options:

3a. Local AI orchestration (this skill, recommended for interactive review)

1. Read .github/skills/ansi-analysis/shared.md (Step 0)
2. List target/ansi-offload/*.json
3. Read each JSON; extract: suite name, total/passed/failed/ignored counts, per-test failCause
4. Apply the analysis template from shared.md verbatim (sections, tables, constraints)
5. For each failure: extract Velox file:line from failCause, read those C++ files, verify root cause
6. Always grep isAnsiSupported in ep/build-velox/build/velox_ep/velox/functions/sparksql/specialforms/SparkCastExpr.cpp when the failure involves Cast — most NO_EXCEPTION/Cast failures stem from the small whitelist there
7. Output the markdown report

3b. Python script (CI / batch)

python3 .github/skills/ansi-analysis/analyze-ansi.py \
  --json-dir target/ansi-offload/ \
  --ai-analysis \
  --output ansi-report.md

The script loads the same shared prompt and calls the GitHub Models API.

Step 4 — Single-failure diagnosis

When the user pastes one failing test:

1. Locate its JSON entry under target/ansi-offload/
2. Apply the self-investigation steps from shared.md (extract Velox file:line, check isAnsiSupported, cross-check withAnsiEvalMode in ExpressionUtils)
3. Output: Symptom / Root Cause / Fix Point / Representative Tests / Estimated Impact

Step 5 — Optional PR comment

If the user wants the report posted to a PR:

gh pr comment <pr-number> --body-file ansi-report.md

(or use the GitHub MCP server tool when available)

Environment requirements

For Step 2 (running tests):

• SPARK_ANSI_SQL_MODE=true
• SPARK_TESTING=true
• SPARK_SCALA_VERSION=2.13
• JVM: -Dspark.gluten.sql.ansiFallback.enabled=false
• Maven profile: include -Pdelta

What NOT to do

• Do NOT invent reference paths or line numbers — always grep / verify
• Do NOT skip Step 0 — drift between shared.md and your output is the failure mode this skill is designed to prevent
• Do NOT bypass the shared prompt by writing your own analysis structure