virgoMachineLabs/toph-eval

Toph Eval · Pipeline Failure Diagnosis Benchmark

Version: 0.1
Maintainer: Virgo Machine Labs
License: Apache 2.0
Taxonomy: github.com/vaishsagar-cfo/toph-eval

---

Overview

Toph Eval is a structured evaluation benchmark for automated root cause analysis (RCA) of data pipeline failures in enterprise health technology environments.

Standard LLM benchmarks address single-step tasks (question answering, multiple choice, text completion). Pipeline RCA is a multi-step reasoning task: a system must read logs across multiple heterogeneous tools simultaneously, construct a causal graph of pipeline execution, and distinguish the origin of a failure from the component at which the error surfaces. No existing general-purpose benchmark addresses this task category.

This dataset provides evaluation scenarios for systems performing automated pipeline RCA. Each scenario includes realistic simulated logs from a 4-tool enterprise ETL pipeline and a structured ground truth answer key scored across four evaluation dimensions.

---

Pipeline Architecture

All scenarios are built on the Optum pharmacy claims nightly extract pipeline:

Phase	Tool	Description
Orchestrate	Jenkins	Triggers pipeline, monitors stage execution
Extract	Azure Data Factory	Pulls from source systems, lands raw files in ADLS Gen2
Transform	Databricks	Normalizes clinical codes, applies business rules
Load	Azure Synapse Analytics	Writes clean records to data warehouse

Log formats reflect real production tool output: Jenkins produces plain text logs; ADF, Databricks, and Synapse produce structured JSON (as exported to Azure Monitor).

---

Dataset Structure

data/
  {FAILURE_CODE}_{pipeline}_{scenario_number}/
    success_run/
      jenkins.log       # Plain text orchestration log (baseline)
      adf.json          # Azure Data Factory structured log
      databricks.json   # Databricks job event log
      synapse.json      # Azure Synapse Analytics pipeline log
    failure_run/
      jenkins.log
      adf.json
      databricks.json
      synapse.json
    ground_truth.json   # Answer key in toph-eval scoring format

Each scenario includes a success run (normal operation baseline) and a failure run (the scenario under evaluation). The success run enables duration-based anomaly detection and establishes expected row counts.

---

Scenarios (v0.1)

Scenario ID	Failure Code	Category	Description
DEPEND_002_optum_pharmacy_001	DEPEND_002	Upstream Dependency	Silent zero-row extract — all statuses Succeeded
PERM_001_optum_pharmacy_001	PERM_001	Access & Permissions	INSERT permission denied on Synapse target table
AUTH_003_optum_pharmacy_001	AUTH_003	Authentication	Expired client secret in Key Vault — Databricks auth failure
SCHEMA_002_optum_pharmacy_001	SCHEMA_002	Schema & Data Contract	Column removed from upstream source schema
VOLUME_001_optum_pharmacy_001	VOLUME_001	Data Volume & Quality	Zero rows from direct extract, HTTP 200, no error raised
CONN_006_optum_pharmacy_001	CONN_006	Connectivity	EHR vendor API rate limit hit mid-extract
RESOURCE_001_optum_pharmacy_001	RESOURCE_001	Resource Exhaustion	Spark executor OOM on month-end batch volume
ORCH_001_optum_pharmacy_001	ORCH_001	Orchestration	DST timezone shift triggers duplicate pipeline run
FILE_003_optum_pharmacy_001	FILE_003	File & Format	Upstream payer changed EDI file naming convention
HEALTH_001_optum_pharmacy_001	HEALTH_001	Healthcare-Specific	ICD-10 FY2025 codes reference table not updated before October 1

---

Evaluation Dimensions

Each scenario's ground_truth.json scores across four dimensions, assessed independently:

Dimension	Definition	Weight
System identification	Correctly identifies the infrastructure component where failure manifested	Standard
Error classification	Correctly assigns the failure type code from the toph-eval taxonomy	Standard
Causal chain accuracy	Correctly traces failure to its origin, not just where error surfaced	Highest
Fix correctness	Recommended remediation correctly addresses the identified root cause	Standard

Causal chain accuracy is weighted highest. It is the dimension that distinguishes root cause analysis from surface-level error detection, and the dimension on which general-purpose language models without domain-specific tooling perform worst.

---

Ground Truth Format

{
  "scenario_id": "DEPEND_002_optum_pharmacy_001",
  "taxonomy_version": "0.1",
  "failure_code": "DEPEND_002",
  "failure_category": "Upstream Dependency Failure",
  "description": "Upstream job completed with success status but wrote zero rows",
  "evaluation_dimensions": {
    "system_identification": { "correct_answer": "..." },
    "error_classification": { "correct_answer": "DEPEND_002" },
    "causal_chain_accuracy": {
      "correct_answer": ["step 1", "step 2", "..."],
      "failure_originates_at": "...",
      "error_surfaces_at": "..."
    },
    "fix_correctness": {
      "correct_answer": "...",
      "preventive_fix": "..."
    }
  },
  "diagnostic_signals": [...],
  "common_misdiagnoses": [...]
}

---

Why Healthcare-Specific Scenarios Matter

The HEALTH_001 scenario (ICD-10 code set update) represents a failure mode with no analog in general-purpose pipeline monitoring literature. Healthcare pipelines operate under annual CMS code set update cycles (ICD-10 effective October 1, CPT effective January 1, NDC updated quarterly). A pipeline that ran correctly on September 30 will fail on October 1 if reference tables are not updated before the effective date. General-purpose LLMs without domain-specific tooling consistently misclassify this as a schema failure or data quality issue rather than a healthcare code set versioning failure.

---

Taxonomy

The full taxonomy of 63 failure types across 10 categories is maintained at:
github.com/vaishsagar-cfo/toph-eval/blob/main/TAXONOMY.md

This dataset covers one scenario per category. Contributions adding additional scenarios are welcomed via the GitHub repository.

---

Citation

Sagar, V. (2026). Pipeline Failure Taxonomy and Evaluation Benchmark
for Enterprise Health Technology Data Pipelines (Version 0.1).
Virgo Machine Labs.
https://github.com/vaishsagar-cfo/toph-eval

---

Virgo Machine Labs · virgomachinelabs.com · Built in Minnesota