Ingest data from S3 buckets

When using the S3 connector, RisingWave does not guarantee the order of file reading, nor does it ensure resuming from the same file after a recovery. However, we guarantee that every file will eventually be fully read.

Syntax

CREATE SOURCE [ IF NOT EXISTS ] source_name
schema_definition
[INCLUDE { file | offset | payload } [AS <column_name>]]
WITH (
   connector='s3',
   connector_parameter='value', ...
)
FORMAT data_format ENCODE data_encode (
   without_header = 'true' | 'false',
   delimiter = 'delimiter'
);

For CSV data, specify the delimiter in the delimiter option in ENCODE properties.

schema_definition:

(
   column_name data_type [ PRIMARY KEY ], ...
   [ PRIMARY KEY ( column_name, ... ) ]
)

Parameters

Field	Notes
`connector`	Required. Support the s3 connector only.
`s3.region_name`	Required. The service region.
`s3.bucket_name`	Required. The name of the bucket the data source is stored in.
`s3.credentials.access`	Conditional. Your AWS access key ID. Required if `enable_config_load` is `false` or not set.
`s3.credentials.secret`	Conditional. Your AWS secret access key. Required if `enable_config_load` is `false` or not set.
`enable_config_load`	Optional. Set to `true` to use the default AWS credential provider chain for authentication. This is useful for environments where credentials are provided through IAM roles or environment variables. If this is set to `true`, do not provide `s3.credentials.access` and `s3.credentials.secret`. You must also set the `DISABLE_DEFAULT_CREDENTIAL=false` environment variable on the meta and streaming nodes.
`s3.endpoint_url`	Conditional. The host URL for an S3-compatible object storage server. This allows users to use a different server instead of the standard S3 server.
`compression_format`	Optional. This field specifies the compression format of the file being read. You can define `compression_format` in the CREATE TABLE statement. When set to gzip or gz, the file reader reads all files with the `.gz` suffix. When set to `None` or not defined, the file reader will automatically read and decompress `.gz` and `.gzip` files.
`match_pattern`	Conditional. This field is used to find object keys in `s3.bucket_name` that match the given pattern. Standard Unix-style glob syntax is supported. A typical usage follows the `prefix/.suffix` pattern. For example, `your_directory/.parquet` matches all Parquet files under `your_directory/`. If `match_pattern` does not contain `/`, the scan runs from the container root.
`s3.assume_role`	Optional. Specifies the ARN of an IAM role to assume when accessing S3. It allows temporary, secure access to S3 resources without sharing long-term credentials.
`refresh.interval.sec`	Optional. Configure the time interval between operations of listing files. It determines the delay in discovering new files, with a default value of 60 seconds.

In RisingWave Cloud, the default AWS credential provider chain is disabled. Provide s3.credentials.access and s3.credentials.secret (or use a supported assume-role setup). These credentials cannot be omitted. The enable_config_load option is supported only in self-hosted deployments.

Empty cells in CSV files will be parsed to NULL.

Field	Notes
`_data_format_`	Supported data format: PLAIN.
`_data_encode_`	Supported data encodes: CSV, JSON, PARQUET.
`_without_header_`	This field is only for CSV encode, and it indicates whether the first line is header. Accepted values: `true`, `false`. Default is `true`.
`_delimiter_`	How RisingWave splits contents. For JSON encode, the delimiter is `\n`; for CSV encode, the delimiter can be one of `,`, `;`, `E'\t'`.

Additional columns

Field	Notes
`_file_`	Optional. The column contains the file name where current record comes from.
`_offset_`	Optional. The column contains the corresponding bytes offset (record offset for parquet files) where current message begins

Working with S3 sources

Learn how to work with S3 sources through practical examples and scenarios. To avoid conflicts, double-quote all field names (e.g., "field_name") when working with case-sensitive sources. For more conventions on naming and identifier handling, see Identifiers.

Basic examples

Here are examples of connecting RisingWave to an S3 source to read data from individual streams.

CSV
JSON
PARQUET

CREATE TABLE s(
    id int,
    name varchar,
    age int
)
WITH (
    connector = 's3',
    s3.region_name = 'ap-southeast-2',
    s3.bucket_name = 'example-s3-source',
    s3.credentials.access = 'xxxxx',
    s3.credentials.secret = 'xxxxx'
) FORMAT PLAIN ENCODE CSV (
    without_header = 'true',
    delimiter = ',' -- set delimiter = E'\t' for tab-separated files
);

CREATE TABLE s3(
    id int,
    name TEXT,
    age int,
    mark int,
)
INCLUDE file as file_name
INCLUDE offset -- default column name is `_rw_s3_offset`
WITH (
    connector = 's3',
    match_pattern = '%Ring%*.ndjson',
    s3.region_name = 'ap-southeast-2',
    s3.bucket_name = 'example-s3-source',
    s3.credentials.access = 'xxxxx',
    s3.credentials.secret = 'xxxxx',
    s3.endpoint_url = 'https://s3.us-east-1.amazonaws.com'
) FORMAT PLAIN ENCODE JSON;

Use the payload keyword to ingest JSON data when you are unsure of the exact schema beforehand. Instead of defining specific column names and types at the very beginning, you can load all JSON data first and then prune and filter the data during runtime. Check the example below:

CREATE TABLE table_include_payload (v1 int, v2 varchar)
INCLUDE payload
WITH (
    connector = 's3',
    topic = 's3_1_partition_topic',
    properties.bootstrap.server = 'message_queue:29092',
    scan.startup.mode = 'earliest'
) FORMAT PLAIN ENCODE JSON;

CREATE TABLE s(
    id int,
    name varchar,
    age int
)
WITH (
    connector = 's3_v2',
    match_pattern = '*.parquet',
    s3.region_name = 'ap-southeast-2',
    s3.bucket_name = 'example-s3-source',
    s3.credentials.access = 'xxxxx',
    s3.credentials.secret = 'xxxxx'
) FORMAT PLAIN ENCODE PARQUET;

Object filtering in S3 buckets

RisingWave has a prefix argument designed for filtering objects in the S3 bucket. It relies on Apache Opendal whose prefix filter implementation is expected to be released soon.

Handle new files in the bucket

RisingWave periodically scans for new files matching the pattern (default interval: 60 seconds). To customize the discovery interval, set refresh.interval.sec (in seconds). RisingWave automatically ingests new files added to the bucket. However, it does not detect updates to a file if a file is deleted and a new file with the same name is added simultaneously. Additionally, RisingWave will ignore file deletions.

Read data from the source

You need to create a materialized view from the source or create a table with the S3 connector to read the data. Here are some examples:

-- Create a materialized view from the source
CREATE SOURCE s3_source (
    product_id INT,
    sales_date DATE,
    quantity INT,
    revenue DOUBLE
) WITH (
    connector = 's3',
    ...
);
CREATE MATERIALIZED VIEW mv AS SELECT * FROM s3_source;

-- Create a table with the S3 connector
CREATE TABLE s3_table (
    product_id INT,
    sales_date DATE,
    quantity INT,
    revenue DOUBLE
) WITH (
    connector = 's3',
    ...
);

-- Select from the source directly
SELECT count(*) from s3_source;

Column names and types are not inferred automatically from the source. You must explicitly define them when creating the table or materialized view.

Read Parquet files from S3

You can use the table function file_scan() to read Parquet files from S3, either a single file or a directory of Parquet files.

Function signature

file_scan('parquet', 's3', s3_region, s3_access_key, s3_secret_key, file_location_or_directory)

When reading a directory of Parquet files, the schema will be based on the first Parquet file listed. Please ensure that all Parquet files in the directory have the same schema.

For example, assume you have a Parquet file named sales_data.parquet that stores a company’s sales data, containing the following fields:

product_id: Product ID
sales_date: Sales date
quantity: Sales quantity
revenue: Sales revenue

You can use the following SQL statement to read this Parquet file: Read a single Parquet file

SELECT
  product_id,
  sales_date,
  quantity,
  revenue
FROM file_scan(
  'parquet',
  's3',
  'ap-southeast-2',
  'xxxxxxxxxx',
  'yyyyyyyy',
  's3://your-bucket/path/to/sales_data.parquet'
);

----RESULT
product_id |  sales_date  | quantity | revenue
------------+-------------+----------+----------
         12 | 2023-04-01   |       50 |   1000.00
         12 | 2023-04-02   |       30 |    600.00
         15 | 2023-04-01   |       20 |    400.00
(3 rows)

If you have several such Parquet files, you can also read by their file directory: Read a directory of Parquet files

SELECT
  product_id,
  sales_date,
  quantity,
  revenue
FROM file_scan(
  'parquet',
  's3',
  'ap-southeast-2',
  'xxxxxxxxxx',
  'yyyyyyyy',
  's3://your-bucket/path/to/sales_data_file_directory/'
);

----RESULT
product_id |  sales_date  | quantity | revenue
------------+-------------+----------+----------
         12 | 2023-04-01   |       50 |   1000.00
         12 | 2023-04-02   |       30 |    600.00
         15 | 2023-04-01   |       20 |    400.00
         15 | 2023-04-03   |       40 |    800.00
         18 | 2023-04-02   |       25 |    500.00
         18 | 2023-04-04   |       35 |    700.00
(6 rows)

Handle unexpected file types or poorly formatted files

RisingWave will attempt to interpret and parse files, regardless of their type, as CSV or ndjson, based on the specified rules. Warnings will be reported for parts of the file that cannot be parsed, but the source part will not fail. Poorly formatted parts of a file will be discarded.

​Syntax

​Parameters

​Additional columns

​Working with S3 sources

​Basic examples

​Object filtering in S3 buckets

​Handle new files in the bucket

​Read data from the source

​Read Parquet files from S3

​Handle unexpected file types or poorly formatted files

Syntax

Parameters

Additional columns

Working with S3 sources

Basic examples

Object filtering in S3 buckets

Handle new files in the bucket

Read data from the source

Read Parquet files from S3

Handle unexpected file types or poorly formatted files