Implement Incremental Extraction in a New Connector

You are an expert in implementing incremental metadata extraction using the Atlan Application SDK. You have deep knowledge of the SDK's single inheritance chain pattern, Temporal workflows, DuckDB file-backed queries, and RocksDB disk-backed state storage.

When to Use This Skill

• Adding incremental extraction to a new SQL-based connector app
• Understanding the SDK's incremental extraction architecture
• Debugging incremental extraction issues (state management, column batching, ancestral merge)
• Extending incremental extraction to support new entity types
• Reviewing PRs that modify incremental extraction logic

When NOT to Use This Skill

• Building a non-SQL connector (REST-based, file-based)
• Working on full extraction only (no incremental support needed)
• Modifying the SDK's core incremental framework (see SDK source directly)

Architecture Overview

Single Inheritance Chain

BaseSQLMetadataExtractionActivities (SDK)
    └── IncrementalSQLMetadataExtractionActivities (SDK)
            └── YourDatabaseActivities (App)

BaseSQLMetadataExtractionWorkflow (SDK)
    └── IncrementalSQLMetadataExtractionWorkflow (SDK)
            └── YourDatabaseWorkflow (App)

SDK vs App Responsibilities

Component	SDK Handles	App Provides
Workflow orchestration	4-phase execution, parallel batching, retry policies	Nothing (inherited)
Marker management	S3 fetch/persist, timestamp normalization, prepone logic	Nothing (inherited)
State management	Current-state read/write, S3 upload/download, ancestral merge	Nothing (inherited)
Table extraction	Switching between full/incremental SQL, placeholder resolution, auto-loading incremental_table_sql from app/sql/	extract_table_incremental.sql file, resolve_database_placeholders() (optional)
Column extraction	Table analysis (DuckDB), backfill detection (DuckDB), batching, parallel execution, auto-loading incremental_column_sql from app/sql/	build_incremental_column_sql() - the SQL building strategy, extract_column_incremental.sql file
SQL execution	Query execution, result counting, output path management	Nothing (inherited)

Workflow 4-Phase Execution

Phase 1: Setup
  get_workflow_args → fetch_incremental_marker → read_current_state → save_state

Phase 2: Base Extraction (inherited from BaseSQLMetadataExtractionWorkflow)
  fetch_databases → fetch_schemas → fetch_tables → fetch_columns (skipped if incremental)
  → fetch_procedures → transform_data → App.upload()

Phase 3: Incremental Column Extraction (if prerequisites met)
  prepare_column_extraction_queries → execute_single_column_batch (parallel) → transform_data

Phase 4: Finalization
  write_current_state (ancestral merge + upload) → update_incremental_marker

Incremental Prerequisites (all must be true)

1. incremental-extraction parameter is "true"
2. marker_timestamp exists (fetched from S3 marker.txt from a previous run)
3. current_state_available is true (previous state snapshot exists in S3)

If any prerequisite is not met, the workflow runs a full extraction instead.

Implementation Checklist

Files You Need to Create/Modify

your-database-app/
├── app/
│   ├── activities/
│   │   └── metadata_extraction/
│   │       └── your_db.py          # Activities class (MAIN FILE)
│   ├── workflows/
│   │   └── metadata_extraction/
│   │       └── your_db.py          # Workflow class (minimal)
│   └── sql/
│       ├── extract_table.sql              # Full table extraction
│       ├── extract_table_incremental.sql  # Incremental table extraction (NEW)
│       ├── extract_column.sql             # Full column extraction
│       └── extract_column_incremental.sql # Incremental column extraction (NEW)
├── tests/
│   └── unit/
│       └── test_column_utils.py    # Tests for build_incremental_column_sql
└── pyproject.toml                  # SDK dependency with [incremental] extra

Step-by-Step Implementation

See the reference files for detailed implementation of each step:

1. references/activities-implementation.md - Activities class with all overrides
2. references/sql-templates.md - SQL template patterns for incremental queries
3. references/workflow-implementation.md - Workflow class setup
4. references/testing-patterns.md - Unit test patterns
5. references/duckdb-patterns.md - DuckDB usage patterns

Quick Start: Minimal Implementation

# connector/app.py

from application_sdk.templates import IncrementalSqlMetadataExtractor
from application_sdk.app import task
from application_sdk.templates.contracts.sql_metadata import (
    FetchDatabasesInput, FetchDatabasesOutput,
    FetchSchemasInput, FetchSchemasOutput,
    FetchTablesInput, FetchTablesOutput,
    TransformInput, TransformOutput,
)

class YourDBExtractor(IncrementalSqlMetadataExtractor):
    sql_client_class = YourDBClient

    # All SQL queries are auto-loaded from app/sql/ by the SDK:
    #   fetch_database_sql          ← extract_database.sql
    #   fetch_schema_sql            ← extract_schema.sql
    #   fetch_table_sql             ← extract_table.sql
    #   fetch_column_sql            ← extract_column.sql
    #   incremental_table_sql       ← extract_table_incremental.sql
    #   incremental_column_sql      ← extract_column_incremental.sql
    #
    # No need to set these manually — just place the SQL files in app/sql/.

    @task(timeout_seconds=3600)
    async def fetch_databases(self, input: FetchDatabasesInput) -> FetchDatabasesOutput:
        ...

    @task(timeout_seconds=3600)
    async def fetch_schemas(self, input: FetchSchemasInput) -> FetchSchemasOutput:
        ...

    @task(timeout_seconds=3600)
    async def fetch_tables(self, input: FetchTablesInput) -> FetchTablesOutput:
        ...

    @task(timeout_seconds=3600)
    async def transform_data(self, input: TransformInput) -> TransformOutput:
        ...

    def build_incremental_column_sql(self, table_ids, workflow_args):
        """Build SQL for incremental column extraction."""
        # Your database-specific SQL building logic here
        ...

Key Patterns and Best Practices

1. SQL Template Placeholders

The SDK handles {marker_timestamp} automatically. Your app only needs to handle database-specific placeholders via resolve_database_placeholders():

# SDK resolves automatically:
#   {marker_timestamp} → "2024-01-15T00:00:00Z"

# App resolves via resolve_database_placeholders():
#   {system_schema} → "SYS" (Oracle)
#   Any other database-specific placeholders

2. Column Extraction SQL Strategies

Each database has a different way to pass table IDs to the column query:

Database	Strategy	Reason
Oracle	FROM dual CTE with UNION ALL	1000-element IN clause limit
ClickHouse	WHERE ... IN (...) clause	No element limit
PostgreSQL	ANY(ARRAY[...])	PostgreSQL array syntax

3. State Mutation Prevention

Temporal reuses activity instances across workflow runs. Never permanently modify class attributes with resolved SQL:

# BAD - mutates class attribute permanently
self.fetch_table_sql = resolved_sql  # Breaks on next run!

# GOOD - SDK saves originals internally via _original_fetch_table_sql
# The SDK's fetch_tables() and fetch_columns() handle this automatically

4. Incremental Table SQL Labeling

Your extract_table_incremental.sql must include an incremental_state column:

SELECT ...,
  CASE
    WHEN created_time > '{marker_timestamp}' THEN 'CREATED'
    WHEN modified_time > '{marker_timestamp}' THEN 'UPDATED'
    ELSE 'NO CHANGE'
  END AS incremental_state
FROM ...

5. Incremental Column SQL Template

Your extract_column_incremental.sql must include a placeholder for table IDs that your build_incremental_column_sql() method will replace:

-- Oracle pattern: --TABLE_FILTER_CTE-- placeholder
--TABLE_FILTER_CTE--
SELECT ... FROM ... JOIN table_filter ...

-- ClickHouse pattern: {table_ids_in_clause} placeholder
SELECT ... FROM ... WHERE ... IN ({table_ids_in_clause})

6. pyproject.toml Configuration

[project]
dependencies = [
    "atlan-application-sdk[incremental,iam-auth,tests,workflows]==X.Y.Z",
    "rocksdict>=0.3.0",
]

The [incremental] extra brings in DuckDB, pyarrow, pandas, and sqlalchemy. rocksdict is required for disk-backed table state storage.

Common Pitfalls

1. Missing incremental_table_sql: If not set, incremental mode falls back to full table extraction
2. Not escaping quotes in table IDs: Table names with special characters (e.g., O'Brien) must be escaped in SQL
3. Empty table_ids list: build_incremental_column_sql should raise ValueError for empty lists
4. Forgetting resolve_database_placeholders: If your SQL has custom placeholders, they won't be replaced
5. Testing with wrong method name: Tests must call build_incremental_column_sql (not a private method name)
6. Overriding execute_column_batch: Don't override it - it's concrete in the SDK. Only implement build_incremental_column_sql