Ingest data from S3 buckets

Use the SQL statement below to connect RisingWave to an Amazon S3 source. RisingWave supports CSV, ndjson and Parquet file formats.

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	Required. This field indicates the access key ID of AWS.
s3.credentials.secret	Required. This field indicates the secret access key of AWS.
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.
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.

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

Examples

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

CSV

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
);

JSON

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;

PARQUET

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;

Important considerations

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 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 WITH ( connector = 's3', ... );
CREATE MATERIALIZED VIEW mv AS SELECT * FROM s3_source;

-- Create a table with the S3 connector
CREATE TABLE s3_table ( ... ) WITH ( connector = 's3', ... );

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

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(file_format, storage_type, 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.