Add Parquet files in-place to a Snowflake Managed Apache Iceberg™ table¶

When to use this feature¶

Use Zero-Copy Parquet to Managed Iceberg when:

• You’re migrating from a Hive or Spark-based data lake: You have existing Parquet files in cloud storage and want to bring them into a fully managed Snowflake Iceberg table—with DML, optimization, and platform integrations—without reorganizing or copying your data.

• External systems produce Parquet files: Your data pipelines (such as Apache Spark™, Apache Flink, or custom ETL jobs) write Parquet files to cloud storage in an append-only manner, and you want Snowflake to automatically track and query these files.

• You want to minimize ingestion costs: Avoid duplicating large datasets by referencing files in-place instead of scanning and copying them into a table.

• You need interoperable access: Perform read and write operations on Iceberg tables with Snowflake while still making the tables accessible to other engines.

• You want Snowflake to fully manage and optimize storage: You want Snowflake to automatically perform table optimization – compaction and snapshot expiration.

When to use other ingestion methods¶

This feature isn’t the right choice for every scenario. Consider other ingestion methods when:

• Your source files aren’t Parquet: Use the COPY INTO command with LOAD_MODE = FULL_INGEST to load data from CSV, JSON, Avro, ORC, or other formats. See Vue d’ensemble du chargement de données for supported formats.

• You need to transform data during ingestion: Use the COPY INTO command with LOAD_MODE = FULL_INGEST and a SELECT statement to apply transformations as you load.

• Your data sources are streaming: If writers produce data from streaming sources like Apache Kafka instead of writing Parquet files, use Snowpipe Streaming for high-performance, low-latency ingestion into Iceberg tables.

• External engines need to modify or delete data: If external engines need to directly update or delete Parquet files and not just append, consider Parquet Direct tables.

Benefits¶

When you use Zero-Copy Parquet to Managed Iceberg:

• First-class Snowflake experience: Your registered Parquet files become a fully managed Iceberg table with DML support, automatic table optimization (compaction and snapshot expiration), and access to platform features like replication.

• Instant Iceberg interoperability: Turn a bucket of Parquet files into an Iceberg table immediately. Snowflake generates the Iceberg metadata, making your data accessible to any Iceberg-compatible engine through the Horizon Catalog.

• No data movement: Files remain in their original cloud storage location.

• Significantly lower ingestion cost: Skip full ingestion processing.

• No additional storage costs: Avoid duplicating data and associated storage charges.

• Continuous sync: Use Snowpipe to automatically add new files as they arrive.

Requirements¶

When you use this option, the source Parquet files must meet the following conditions:

• The files must be located under the base location of the table that you want to register them to.

• The files must be Iceberg-compatible Apache Parquet files that aren’t registered with an Iceberg catalog.

• The files must be in Apache Parquet format. No other file formats are supported.

Access patterns¶

This feature is designed for the following access patterns:

Prerequisites¶

Before you begin, ensure you have:

• An existing data lake with Parquet files (for example, a Hive-partitioned dataset in Amazon S3).

• Appropriate IAM permissions to create storage integrations and external volumes.

• The Parquet files are Iceberg-compatible (standard Parquet format without Iceberg-specific metadata).

Step 1: Create an external volume¶

Create an external volume that points to the S3 location containing your Parquet files. The external volume must have write access enabled.

CREATE OR REPLACE EXTERNAL VOLUME my_external_volume
  STORAGE_LOCATIONS = (
    (
      NAME = 'my-s3-location'
      STORAGE_PROVIDER = 'S3'
      STORAGE_BASE_URL = 's3://my-bucket/my-data-lake/'
      STORAGE_AWS_ROLE_ARN = 'arn:aws:iam::123456789012:role/my-role'
    )
  )
  ALLOW_WRITES = TRUE;

For detailed instructions, see Configurer un volume externe.

Step 2: Create a storage integration and external stage¶

Create a storage integration and external stage that point to the same S3 location. The stage is used for the COPY INTO command.

CREATE OR REPLACE STORAGE INTEGRATION my_storage_integration
  TYPE = EXTERNAL_STAGE
  STORAGE_PROVIDER = 'S3'
  STORAGE_ALLOWED_LOCATIONS = ('s3://my-bucket/my-data-lake/')
  STORAGE_AWS_ROLE_ARN = 'arn:aws:iam::123456789012:role/my-role'
  ENABLED = TRUE;

CREATE OR REPLACE FILE FORMAT my_parquet_format
  TYPE = PARQUET
  USE_VECTORIZED_SCANNER = TRUE;

CREATE OR REPLACE STAGE my_parquet_stage
  STORAGE_INTEGRATION = my_storage_integration
  URL = 's3://my-bucket/my-data-lake/'
  FILE_FORMAT = my_parquet_format;

For detailed instructions, see Option 1: Configure a Snowflake storage integration to access Amazon S3.

Step 3: Create an Iceberg table¶

You can create the Iceberg table with either a declarative schema or an inferred schema.

Option A: Create a table with a declarative schema

If you know the schema of your Parquet files, define it explicitly:

CREATE OR REPLACE ICEBERG TABLE my_table (
  id INTEGER,
  name STRING,
  created_date DATE,
  amount DOUBLE
)
  CATALOG = 'SNOWFLAKE'
  EXTERNAL_VOLUME = 'my_external_volume'
  BASE_LOCATION = 'my-table-1/';

Option B: Create a table with an inferred schema

Use the INFER_SCHEMA function to detect the schema from your Parquet files:

-- First, preview the inferred schema
SELECT *
  FROM TABLE(
    INFER_SCHEMA(
      LOCATION => '@my_parquet_stage/my-table-1/',
      FILE_FORMAT => 'my_parquet_format',
      KIND => 'ICEBERG'
    )
  );

-- Create the table using the inferred schema
CREATE OR REPLACE ICEBERG TABLE my_table
  USING TEMPLATE (
    SELECT ARRAY_AGG(OBJECT_CONSTRUCT(*))
      WITHIN GROUP (ORDER BY order_id)
      FROM TABLE(
        INFER_SCHEMA(
          LOCATION => '@my_parquet_stage/my-table-1/',
          FILE_FORMAT => 'my_parquet_format',
          KIND => 'ICEBERG'
        )
      )
  )
  CATALOG = 'SNOWFLAKE'
  EXTERNAL_VOLUME = 'my_external_volume'
  BASE_LOCATION = 'my-table-1/';

Step 4: Register existing Parquet files¶

Use the COPY INTO command with LOAD_MODE = ADD_FILES_REFERENCE to register your existing Parquet files with the Iceberg table:

COPY INTO my_table
  FROM @my_parquet_stage/my-table-1/
  FILE_FORMAT = (TYPE = PARQUET USE_VECTORIZED_SCANNER = TRUE)
  LOAD_MODE = ADD_FILES_REFERENCE
  MATCH_BY_COLUMN_NAME = CASE_SENSITIVE;

Output:

+-----------------------------------------------+--------+-------------+-------------+
| file                                          | status | rows_parsed | rows_loaded |
|-----------------------------------------------+--------+-------------+-------------|
| my-table-1/year=2024/month=01/data_001.parquet | LOADED |       10000 |       10000 |
| my-table-1/year=2024/month=02/data_002.parquet | LOADED |       15000 |       15000 |
+-----------------------------------------------+--------+-------------+-------------+

Step 5: Continuously add new files with Snowpipe (optional)¶

To automatically register new Parquet files as they arrive, create a Snowpipe with auto-ingest enabled:

CREATE OR REPLACE PIPE my_table_pipe
  AUTO_INGEST = TRUE
  AS
    COPY INTO my_table
      FROM @my_parquet_stage/my-table-1/
      FILE_FORMAT = (TYPE = PARQUET USE_VECTORIZED_SCANNER = TRUE)
      LOAD_MODE = ADD_FILES_REFERENCE
      MATCH_BY_COLUMN_NAME = CASE_SENSITIVE;

For detailed instructions on configuring event notifications for auto-ingest, see Automatisation de Snowpipe pour Amazon S3.

Step 6: Query the table¶

Query the table to verify that the data was registered successfully:

SELECT * FROM my_table LIMIT 10;

View file registration history¶

Use the COPY_HISTORY table function to see which files were successfully registered and which failed:

-- Show recent file registration history for a table
SELECT
    file_name,
    status,
    row_count,
    first_error_message,
    last_load_time
  FROM TABLE(INFORMATION_SCHEMA.COPY_HISTORY(
    TABLE_NAME => 'my_table',
    START_TIME => DATEADD(hours, -24, CURRENT_TIMESTAMP())
  ))
  ORDER BY last_load_time DESC;

The status column indicates whether each file was successfully registered:

• Loaded: The file was successfully registered with the table.

• Load failed: The file registration failed. Check first_error_message for details.

• Partially loaded: Some rows were registered, but others failed.

List pipes using ADD_FILES_REFERENCE¶

Query the PIPES Account Usage view to find all pipes configured to use LOAD_MODE = ADD_FILES_REFERENCE:

-- List pipes using ADD_FILES_REFERENCE with their target tables and status
SELECT
    pipe_catalog AS database_name,
    pipe_schema AS schema_name,
    pipe_name,
    definition,
    is_autoingest_enabled,
    notification_channel_name
  FROM SNOWFLAKE.ACCOUNT_USAGE.PIPES
  WHERE UPPER(definition) LIKE '%ADD_FILES_REFERENCE%'
  ORDER BY created;

To check the current status of a specific pipe, use the PIPE_USAGE_HISTORY view or the SYSTEM$PIPE_STATUS function:

-- Check the status of a specific pipe
SELECT SYSTEM$PIPE_STATUS('my_table_pipe');

File format limitations¶

• Parquet only: Only Apache Parquet files are supported. No other file formats (ORC, Avro, CSV, JSON) can be registered using this feature.

• Iceberg-compatible Parquet: The Parquet files must be compatible with Iceberg. Files that contain Iceberg-specific row lineage metadata (_row_id and _last_updated_sequence_number) require the FULL_INGEST option instead.

• Vectorized scanner required: You must use FILE_FORMAT = (TYPE = PARQUET USE_VECTORIZED_SCANNER = TRUE).

Table type limitations¶

• Snowflake-managed Iceberg tables only: This feature works only with Snowflake-managed Iceberg tables (CATALOG = 'SNOWFLAKE'). Externally managed Iceberg tables are not supported.

• Files must be under BASE_LOCATION: All Parquet files must be located under the table’s BASE_LOCATION path in cloud storage.

COPY INTO options not supported¶

The following COPY INTO options are not supported with LOAD_MODE = ADD_FILES_REFERENCE:

• Copying unstaged data by specifying a cloud storage location and a storage integration directly.

• MATCH_BY_COLUMN_NAME = CASE_INSENSITIVE | NONE (must use CASE_SENSITIVE).

• ON_ERROR = CONTINUE | SKIP_FILE_N | SKIP_FILE_X%.

• PURGE = TRUE (source files cannot be purged).

• Transforming or filtering the data before loading. To transform data, use the FULL_INGEST option instead.

Why this matters¶

When you use ADD_FILES_REFERENCE, Snowflake registers your Parquet files in the Iceberg table metadata. If table optimization (compaction) runs, Snowflake might:

• Combine multiple small files into larger files

• Move files from their original location to the Snowflake-managed data/ subdirectory

• Remove the original files after compaction

This behavior is expected for production use but can cause problems during testing if you’re comparing this feature against other approaches using the same source files.

Safe testing approaches¶

Option 1: Copy files to a separate location before testing

Create a copy of your Parquet files in a separate directory before testing:

AWS S3:

aws s3 cp s3://my-bucket/production-data/ s3://my-bucket/test-data/ --recursive

Azure Blob Storage:

az storage blob copy start-batch \
  --source-container production-data \
  --destination-container test-data \
  --account-name mystorageaccount

Google Cloud Storage:

gsutil -m cp -r gs://my-bucket/production-data/* gs://my-bucket/test-data/

Option 2: Disable compaction during testing

If copying files is not practical, disable automatic compaction on the test table:

ALTER ICEBERG TABLE my_test_table
  SET ENABLE_DATA_COMPACTION = FALSE;

Additionally, avoid running any DML operations (INSERT, UPDATE, DELETE, MERGE) during testing, as these can also trigger file modifications.

ALTER ICEBERG TABLE my_test_table
  SET ENABLE_DATA_COMPACTION = TRUE;