Parquet file format

Apache Parquet is a free and open-source column-oriented data storage format in the Apache Hadoop ecosystem. dlt is capable of storing data in this format when configured to do so.

To use this format, you need a pyarrow package. You can get this package as a dlt extra as well:

pip install "dlt[parquet]"

Supported Destinations

Supported by: BigQuery, DuckDB, Snowflake, Filesystem, Athena, Databricks, Synapse

How to configure

There are several ways of configuring dlt to use parquet file format for normalization step and to store your data at the destination:

1. You can set the loader_file_format argument to parquet in the run command:

info = pipeline.run(some_source(), loader_file_format="parquet")

2. You can set the loader_file_format in config.toml or secrets.toml:

[normalize]
loader_file_format="parquet"

3. You can set the loader_file_format via ENV variable:

export NORMALIZE__LOADER_FILE_FORMAT="parquet"

4. You can set the file type directly in the resource decorator.

@dlt.resource(file_format="parquet")
def generate_rows(nr):
    pass

Destination AutoConfig

dlt uses destination capabilities to configure the parquet writer:

• It uses decimal and wei precision to pick the right decimal type and sets precision and scale.
• It uses timestamp precision to pick the right timestamp type resolution (seconds, micro, or nano).

Writer settings

Under the hood, dlt uses the pyarrow parquet writer to create the files. The following options can be used to change the behavior of the writer:

• flavor: Sanitize schema or set other compatibility options to work with various target systems. Defaults to None which is pyarrow default.
• version: Determine which Parquet logical types are available for use, whether the reduced set from the Parquet 1.x.x format or the expanded logical types added in later format versions. Defaults to "2.6".
• data_page_size: Set a target threshold for the approximate encoded size of data pages within a column chunk (in bytes). Defaults to None which is pyarrow default.
• row_group_size: Set the number of rows in a row group. See here how this can optimize parallel processing of queries on your destination over the default setting of pyarrow.
• timestamp_timezone: A string specifying timezone, default is UTC.
• coerce_timestamps: resolution to which coerce timestamps, choose from s, ms, us, ns
• allow_truncated_timestamps - will raise if precision is lost on truncated timestamp.

tip

Default parquet version used by dlt is 2.4. It coerces timestamps to microseconds and truncates nanoseconds silently. Such setting provides best interoperability with database systems, including loading panda frames which have nanosecond resolution by default

Read the pyarrow parquet docs to learn more about these settings.

Example:

[normalize.data_writer]
# the default values
flavor="spark"
version="2.4"
data_page_size=1048576
timestamp_timezone="Europe/Berlin"

Or using environment variables:

NORMALIZE__DATA_WRITER__FLAVOR
NORMALIZE__DATA_WRITER__VERSION
NORMALIZE__DATA_WRITER__DATA_PAGE_SIZE
NORMALIZE__DATA_WRITER__TIMESTAMP_TIMEZONE

Timestamps and timezones

dlt adds timezone (UTC adjustment) to all timestamps regardless of a precision (from seconds to nanoseconds). dlt will also create TZ aware timestamp columns in the destinations. duckdb is an exception here

Disable timezones / utc adjustment flags

You can generate parquet files without timezone adjustment information in two ways:

1. Set the flavor to spark. All timestamps will be generated via deprecated int96 physical data type, without the logical one
2. Set the timestamp_timezone to empty string (ie. DATA_WRITER__TIMESTAMP_TIMEZONE="") to generate logical type without UTC adjustment.

To our best knowledge, arrow will convert your timezone aware DateTime(s) to UTC and store them in parquet without timezone information.

Row group size

The pyarrow parquet writer writes each item, i.e. table or record batch, in a separate row group. This may lead to many small row groups which may not be optimal for certain query engines. For example, duckdb parallelizes on a row group. dlt allows controlling the size of the row group by buffering and concatenating tables and batches before they are written. The concatenation is done as a zero-copy to save memory. You can control the size of the row group by setting the maximum number of rows kept in the buffer.

[extract.data_writer]
buffer_max_items=10e6

Mind that dlt holds the tables in memory. Thus, 1,000,000 rows in the example above may consume a significant amount of RAM.

row_group_size configuration setting has limited utility with pyarrow writer. It may be useful when you write single very large pyarrow tables or when your in memory buffer is really large.