duckdb-transformation

DuckDB Transformation Knowledge

Provide expertise on writing, configuring, and optimizing DuckDB transformations in Keboola Connection.

Overview

DuckDB (keboola.duckdb-transformation) is an in-process analytical database (OLAP) with columnar storage, designed for fast SQL analytics. It is a cost-effective alternative to Snowflake for small-to-medium datasets.

Status: BETA --- breaking changes may occur.

Key advantages:

• In-process execution (no external database server)
• Columnar storage optimized for analytical queries
• Block-based orchestration with automatic dependency analysis
• Parallel execution of independent scripts
• Rich SQL dialect with modern quality-of-life extensions

Configuration

Component ID

keboola.duckdb-transformation

Configuration Structure

{
  "parameters": {
    "blocks": [
      {
        "name": "Block Name",
        "codes": [
          {
            "name": "Script Name",
            "script": [
              "CREATE TABLE \"output\" AS",
              "SELECT * FROM \"input_table\";"
            ]
          }
        ]
      }
    ],
    "threads": 4,
    "max_memory_mb": 2048,
    "image_tag": "parquet",
    "dtypes_infer": false
  },
  "storage": {
    "input": {
      "tables": [
        {
          "source": "in.c-bucket.table",
          "destination": "input_table"
        }
      ]
    },
    "output": {
      "tables": [
        {
          "source": "output",
          "destination": "out.c-bucket.result"
        }
      ]
    }
  }
}

Settings

• Timeout --- maximum execution time (default: 1 hour)
• Backend size --- memory allocation (see Dynamic Backends)
• DuckDB version --- latest (default) or pinned (e.g., 1.5.2, 1.4.4)
• Automatic data types --- auto-assign types to output tables
• Use parquet for input tables --- Parquet instead of CSV (recommended for >1 GB)
• Infer input table data types --- infer types from input tables (useful for non-typed Storage)
• Debug mode --- enable debug logging

Block-Based Orchestration

• Blocks execute sequentially (one after another)
• Scripts within a block execute in parallel (when no dependencies)
• Uses SQLGlot for automatic SQL analysis
• Builds a DAG (Directed Acyclic Graph) of dependencies
• Execution order is automatically optimized

Dynamic Backends

Backend Size	Memory	Recommended For
XSmall	8 GB	Small datasets, testing
Small (default)	16 GB	Most use cases
Medium	32 GB	Large datasets (5 GB+)
Large	113.6 GB	Very large datasets (10 GB+)

Start with Small, scale up as needed. Dynamic backends are not available on Free/Pay As You Go plans.

DuckDB automatically detects available CPU and memory. Manual override via threads and max_memory_mb parameters.

Parquet Format

Enable Parquet for significantly better performance with large datasets:

• Faster processing than CSV
• Lower memory usage
• Columnar storage optimized for analytics

Recommendation: Always use Parquet for datasets > 1 GB.

Configuration: set "image_tag": "parquet" or toggle Use parquet for input tables in UI.

Infer Input Table Data Types

Keboola Storage tables can be non-typed (all columns as VARCHAR). When enabled, DuckDB infers actual types (INTEGER, FLOAT, DATE, etc.) so aggregate functions and type-specific operations work without manual casting.

Without inference: SUM("amount") fails because amount is VARCHAR. With inference: SUM("amount") works because DuckDB detects it as numeric.

Sync Actions

Four debugging actions available without running the full transformation:

1. syntax_check --- validate SQL syntax without execution
2. lineage_visualization --- markdown diagram of data dependencies
3. execution_plan_visualization --- planned execution order with blocks and batches
4. expected_input_tables --- list of expected input tables based on SQL analysis

SQL Writing Rules

Semicolons Between Statements

Each SQL statement MUST be terminated with a semicolon (;). Multiple statements in a single script must be properly separated:

-- Correct: each statement ends with a semicolon
CREATE TABLE "output_a" AS SELECT * FROM "input_a";

CREATE TABLE "output_b" AS SELECT * FROM "input_b";

Missing semicolons cause syntax errors.

Case Sensitivity (CRITICAL)

Table names:

• Unquoted → lowercase (MyTable → mytable)
• Quoted → case-sensitive ("MyTable" stays MyTable)

Column names:

• Always case-sensitive regardless of quoting
• SELECT columnName and SELECT ColumnName refer to different columns

This is different from Snowflake (unquoted → uppercase, quoted columns case-insensitive).

Best practice: Use consistent lowercase naming or always quote identifiers.

Type Casting Patterns

For non-typed Storage tables, use safe casting patterns:

-- Safe: handles empty strings without errors
TRY_CAST(NULLIF("column", '') AS DATE)
TRY_CAST(NULLIF("column", '') AS DOUBLE)
TRY_CAST(NULLIF("column", '') AS INT)

-- Shorthand type cast
"column"::BOOLEAN
"column"::INTEGER

-- Regular cast
CAST("column" AS DECIMAL)

Pattern: TRY_CAST(NULLIF("col", '') AS TYPE) --- converts empty strings to NULL before casting, avoiding errors on empty values.

DuckDB SQL Extensions

Quality-of-life extensions not available in all SQL dialects:

-- GROUP BY ALL: auto-groups by all non-aggregated columns
SELECT product, category, SUM(sales) FROM orders GROUP BY ALL;

-- EXCLUDE: select all columns except specific ones
SELECT * EXCLUDE (password, ssn) FROM users;

-- ASOF JOIN: time-series data with non-matching timestamps
SELECT s.player_id, s.score, w.temperature
FROM scores s ASOF JOIN weather w ON s.score_time >= w.timestamp;

-- SUMMARIZE: quick data profiling
SUMMARIZE SELECT * FROM my_table;

Memory Management for Large Datasets

For datasets > 10 GB, configure on-disk processing:

PRAGMA memory_limit='8GB';
PRAGMA temp_directory='/tmp/duckdb_temp';
PRAGMA threads=4;
PRAGMA enable_object_cache;

Snowflake to DuckDB Migration

Data Type Mapping

Snowflake	DuckDB	Notes
VARIANT	JSON or VARCHAR	Semi-structured data
ARRAY	LIST	Native arrays
OBJECT	STRUCT or JSON	Nested objects
INTEGER	INTEGER	Same
VARCHAR	VARCHAR	Same
TIMESTAMP	TIMESTAMP	Same

SQL Function Differences

Category	Snowflake	DuckDB
Window functions	QUALIFY ROW_NUMBER() OVER (...) = 1	Subquery with WHERE rn = 1
Null handling	NVL(a, b)	COALESCE(a, b)
Conditionals	IFF(cond, a, b)	CASE WHEN cond THEN a ELSE b END
Date add	DATEADD(day, 7, date)	date + INTERVAL '7 days'
Date diff	DATEDIFF(day, d1, d2)	date_diff('day', d1, d2)
String agg	LISTAGG(...) WITHIN GROUP (ORDER BY ...)	STRING_AGG(...)
Transient tables	CREATE TRANSIENT TABLE	Not supported (use regular or TEMP)

Case Sensitivity Migration

Snowflake:

• Unquoted → UPPERCASE (MyTable → MYTABLE)
• Quoted → case-sensitive ("MyTable" ≠ "MYTABLE")

DuckDB:

• Unquoted → lowercase (MyTable → mytable)
• Quoted tables → case-sensitive
• Columns → always case-sensitive regardless of quoting

Migration tip: Use consistent lowercase naming: CREATE TABLE mytable ...

Window Functions Migration

-- Snowflake
SELECT * FROM table QUALIFY ROW_NUMBER() OVER (PARTITION BY id ORDER BY date DESC) = 1;

-- DuckDB
SELECT * FROM (
    SELECT *, ROW_NUMBER() OVER (PARTITION BY id ORDER BY date DESC) AS rn
    FROM table
) WHERE rn = 1;

Temporary Table Differences

• Snowflake: Temp tables can be in any schema
• DuckDB: Temp tables ALWAYS in temp.main schema (cannot specify another)
• Both: Auto-drop after session end
• DuckDB name collision: Temp table has priority over regular table with same name. Access regular table via memory.main.table_name

Automatic Migration

SQLGlot can automatically convert ~85% of Snowflake SQL to DuckDB syntax. However, complex queries, Snowflake-specific functions, and edge cases may not convert correctly.

Transformation Migration Component (keboola.app-transformation-migration):

• Experimental component that automates Snowflake → DuckDB migration
• Handles SQL dialect translation via SQLGlot
• Preserves input/output table mappings and runtime settings
• If translation fails for a code block, original SQL is preserved as comments

Important: The automatic migration is not 100% accurate. Roughly 25% of migrated transformations will require manual adjustments --- typically:

• Case sensitivity issues
• Unsupported Snowflake-specific functions
• Data type differences
• Column alias handling in GROUP BY

Always review and test migrated configurations before using in production.

Functions That Work the Same

• SUBSTRING(str, start, len)
• POSITION('x' IN str)
• CONCAT(a, b) (both handle NULL gracefully)
• a || b (both return NULL if any input is NULL)
• LIMIT n
• FETCH FIRST n ROWS
• CREATE TEMP TABLE / CREATE TEMPORARY TABLE
• CREATE OR REPLACE TABLE
• CREATE TABLE IF NOT EXISTS
• CREATE TABLE AS SELECT (CTAS)

Real-World Example: CRM Data Transformation

Typical transformation processing CRM data (e.g., HubSpot):

/* companies */
CREATE TABLE "out_companies" AS
SELECT
  "companyId",
  "name",
  "website",
  TRY_CAST(NULLIF("createdate", '') AS DATE) AS "createdate",
  "isDeleted"::BOOLEAN AS "isDeleted"
FROM "companies";

/* contacts */
CREATE TABLE "out_contacts" AS
SELECT
  "canonical_vid",
  "firstname",
  "lastname",
  "email",
  TRY_CAST(NULLIF("createdate", '') AS DATE) AS "createdate",
  "hs_analytics_source" AS "email_source",
  "associatedcompanyid",
  "lifecyclestage"
FROM "contacts";

/* deals */
CREATE TABLE "out_deals" AS
SELECT
  "dealId",
  "isDeleted"::BOOLEAN AS "isDeleted",
  "dealname",
  TRY_CAST(NULLIF("createdate", '') AS DATE) AS "createdate",
  TRY_CAST(NULLIF("closedate", '') AS DATE) AS "closedate",
  "dealtype",
  TRY_CAST(NULLIF("amount", '') AS DOUBLE) AS "amount",
  "pipeline",
  "dealstage",
  "hubspot_owner_id",
  "hs_analytics_source"
FROM "deals";

/* pipeline stages */
CREATE TABLE "out_stages" AS
SELECT
  "stageId",
  "label",
  TRY_CAST(NULLIF("displayOrder", '') AS INT) AS "displayOrder",
  TRY_CAST(NULLIF("probability", '') AS DOUBLE) AS "probability",
  "closedWon"::BOOLEAN AS "closedWon"
FROM "pipeline_stages";

/* deals-contacts association */
CREATE TABLE "out_deals_contacts" AS
SELECT
  "contact_vid",
  "dealId"
FROM "deals_contacts_list";

/* activities */
CREATE TABLE "out_activities" AS
SELECT
  "engagement_id",
  "metadata_subject",
  TRY_CAST(NULLIF("engagement_createdAt", '') AS DATE) AS "engagement_createdAt",
  "metadata_durationMilliseconds",
  "associations_contactIds",
  "associations_dealIds",
  "associations_ownerIds"
FROM "activities";

Best Practices

1. Start with Small backend, scale up as needed
2. Use Parquet for datasets > 1 GB
3. Terminate every statement with ; when multiple statements in a script
4. Use consistent casing for identifiers (prefer lowercase)
5. Quote identifiers with mixed case or special characters
6. Filter and project early --- apply WHERE close to source, select only needed columns
7. Use EXPLAIN to analyze execution plans for expensive queries
8. Split complex transformations into smaller steps, each producing one output table
9. Use consistent naming for output tables (e.g., stg_, fact_, dim_ prefixes)
10. Use sync actions for debugging (syntax_check, lineage_visualization, execution_plan)
11. Test migrations carefully when moving from Snowflake
12. Use TRY_CAST(NULLIF(...) AS TYPE) pattern for safe type conversion of potentially empty values

When to Use DuckDB vs. Snowflake

Choose DuckDB for:

• Ad-hoc analysis and small-to-medium datasets (up to a few TB)
• Rapid prototyping of transformations
• Projects with limited budgets
• Development and testing

Choose Snowflake for:

• Very large datasets (TB+)
• Complex enterprise workloads
• Sharing warehouses across multiple processes
• Maximum scalability
• Advanced Snowflake-specific features

DuckDB is NOT for transactional workloads. It is an OLAP database optimized for SELECT statements. Avoid frequent INSERT/UPDATE operations.

Reference Links

• DuckDB Official
• SQLGlot --- SQL transpilation
• DuckDB Snowflake Extension --- for local development
• Component README
• Migration Component README
• Keboola DuckDB Documentation