gcp-pipeline-ref-ingestion 1.0.13

Reference Implementation: GCS-to-BigQuery Ingestion Pipeline using gcp-pipeline-framework

Generic Ingestion

Unit 1 of Generic 3-Unit Deployment

ODP Ingestion Pipeline - reads mainframe extracts from GCS and loads to BigQuery.

---

Flow Diagram

                         Generic INGESTION FLOW
                         ─────────────────

  GCS Landing                  Beam Pipeline                    BigQuery ODP
  ───────────                  ─────────────                    ────────────

  1. .csv splits ──┐
  2. .ok file    ──┼──► Pub/Sub ──► Orchestration ──► 1. Read CSV      ┌──────────────────┐
                   │      Event      (Unit 3)         2. Parse HDR/TRL │ odp_generic.      │
                   │                                  3. Validate      │ - customers       │
                   │                                  4. Add Audit     │ - accounts        │
                   │                                  5. Write to BQ ─►│ - decision        │
                   │                                                   │ - applications    │
                   └──────────────────────────────────┘                └──────────────────┘
                               │
                               ▼
                        ┌─────────────┐
                        │ Archive to  │
                        │ GCS Archive │
                        └─────────────┘

---

Pattern

JOIN + MAP: 4 entities → 4 ODP tables

Entity	ODP Table	Key Fields
Customers	odp_generic.customers	customer_id, ssn, status, created_date
Accounts	odp_generic.accounts	account_id, customer_id, account_type, balance, open_date
Decision	odp_generic.decision	decision_id, customer_id, decision_code, score, decision_date
Applications	odp_generic.applications	application_id, customer_id, loan_amount, interest_rate, term_months, application_date, status, event_type, account_type

---

Components

Directory	Purpose
data_ingestion/pipeline/	Beam pipeline and transforms
data_ingestion/config/	System configuration
data_ingestion/schema/	Python Schemas: Source-of-truth EntitySchema definitions used for validation and PII masking logic.
data_ingestion/schemas/	JSON Schemas: Physical BigQuery table schemas used for manual loads, recovery, and external tool integration.
data_ingestion/validation/	File and record validators

---

Schema Architecture

The ingestion unit maintains two types of schemas to separate logical rules from physical storage contracts:

1. Logical Engine (schema/ Python):
   • Purpose: Used by the Beam pipeline for runtime CSV parsing, record-level validation, and in-flight PII masking.
   • Feature: Can dynamically generate BigQuery schemas during job execution (CREATE_IF_NEEDED).
2. Physical Contract (schemas/ JSON):
   • Purpose: Provides a tool-agnostic representation of the BigQuery tables.
   • Usage: Consumed by Terraform to pre-provision infrastructure (with partitioning/clustering) and used for manual bq load recovery operations.

This dual-schema approach ensures that Infrastructure builds the container, but Ingestion carries the blueprint.

---

Library-Driven Ease of Use

The Generic ingestion pipeline is a Lean Consumer of the library framework. It achieves complex mainframe ingestion with minimal custom code by leveraging:

1. Metadata-Driven Schema: data_ingestion/schema/customers.py simply defines an EntitySchema. The library's SchemaValidator handles all type checking and PII masking automatically.
2. Standardized Parsing: Uses the HDRTRLParser from gcp-pipeline-beam to validate mainframe headers/trailers without regex boilerplate.
3. Audit Integrity: Automatically injects _run_id and _processed_at using the AddAuditColumnsDoFn library transform.

---

How to Replicate this JOIN Ingestion (3-to-3)

To create a new ingestion unit for a multi-entity system, follow the Creating New Deployment Guide.

Key steps for this JOIN pattern:

1. Define Schema: Create a new schema file using gcp_pipeline_core.schema.EntitySchema.
2. Configure Pipeline: Inherit from gcp_pipeline_beam.pipelines.base.BasePipeline.
3. Plug in Transforms: Use the fluent BeamPipelineBuilder to chain read_csv -> validate -> write_to_bigquery.
4. CI/CD Config: Update your project and org identifiers in your CI/CD configuration.

---

Infrastructure & Configurations

Google Cloud Resources

This deployment requires the following GCP infrastructure, provisioned via Terraform:

• Storage: GCS buckets for landing, archive, and error files.
• Messaging: Pub/Sub Topic generic-file-notifications and Subscription generic-file-notifications-sub.
• Processing: Cloud Dataflow (Apache Beam) for running the ingestion pipeline.
• Data Warehouse: BigQuery dataset odp_generic for raw data storage.

For detailed infrastructure definitions, see infrastructure/terraform/systems/generic/ingestion/.

Pipeline Configuration (GenericPipelineOptions)

The ingestion pipeline accepts several command-line arguments to control its behavior:

Argument	Description	Required
--entity	Generic entity to process (customers, accounts, decision, applications)	Yes
--source_file	GCS path to input CSV file	Yes
--output_table	Target BigQuery table (project:dataset.table)	Yes
--error_table	BQ table for failed records	Yes
--run_id	Unique identifier for tracking/auditing	Yes
--extract_date	Extract date in YYYYMMDD format	Yes
--job_control_project	GCP Project for job control table	No

Technology Stack & Documentation

• Google Cloud Dataflow - Managed Apache Beam service
• Apache Beam Python SDK - Programming model for data processing
• Google Cloud Storage - Input file landing zone
• Google BigQuery - Data warehouse target
• Google Cloud Pub/Sub - Event-driven triggers
• Apache Beam BigQuery I/O - Connector for BQ

---

Dependencies

Library	Purpose
gcp-pipeline-core	Audit, logging, error handling
gcp-pipeline-beam	Beam transforms, HDR/TRL parsing

NO Apache Airflow dependency - orchestration is separate unit.

---

Execution & Testing

1. Local Development Setup

Initialize the virtual environment:

./scripts/setup_deployment_venv.sh original-data-to-bigqueryload
source deployments/original-data-to-bigqueryload/venv/bin/activate

2. Unit Testing

Run the ingestion unit tests:

python -m pytest tests/unit/ -v

3. Local Execution (DirectRunner)

Run the Beam pipeline locally to validate parsing and schema:

python -m data_ingestion.pipeline.main \
    --input_file=tests/data/sample_customers.csv \
    --output_table=my-project:odp_generic.customers \
    --runner=DirectRunner \
    --temp_location=/tmp/beam-temp

4. Cloud Execution

The ingestion unit is typically triggered by Airflow. To trigger it manually for testing on GCP:

1. Upload a data file to GCS.
2. Upload a corresponding .ok file.
3. The generic_pubsub_trigger_dag will detect the file and launch a Dataflow job using this ingestion code.