Delta table format#

dlt supports writing Delta tables when using the filesystem destination.

How it works#

dlt uses the deltalake library to write Delta tables. One or multiple Parquet files are prepared during the extract and normalize steps. In the load step, these Parquet files are exposed as an Arrow data structure and fed into deltalake.

Delta dependencies#

You need the deltalake package to use this format:

pip install "dlt[deltalake]"

You also need pyarrow>=17.0.0:

pip install 'pyarrow>=17.0.0'

Set table format#

Set the table_format argument to delta when defining your resource:

@dlt.resource(table_format="delta")
def my_delta_resource():
    ...

or when calling run on your pipeline:

pipeline.run(my_resource, table_format="delta")

Table format partitioning#

Delta tables can be partitioned by specifying one or more partition column hints. This example partitions a Delta table by the foo column:

@dlt.resource(
  table_format="delta",
  columns={"foo": {"partition": True}}
)
def my_delta_resource():
    ...

Table access helper functions#

You can use the get_delta_tables helper function to access native table objects. These are deltalake DeltaTable objects.

from dlt.common.libs.deltalake import get_delta_tables

# get dictionary of DeltaTable objects
delta_tables = get_delta_tables(pipeline)

# execute operations on DeltaTable objects
delta_tables["my_delta_table"].optimize.compact()
delta_tables["another_delta_table"].optimize.z_order(["col_a", "col_b"])
# delta_tables["my_delta_table"].vacuum()
# etc.

Google Cloud Storage authentication#

Note that not all authentication methods are supported when using Delta table format on Google Cloud Storage:

• Service Account - ✅ Supported
• Application Default Credentials - ✅ Supported
• OAuth - ❌ Not supported

Table format merge support#

The upsert and insert-only merge strategies are supported for delta.

@dlt.resource(
    write_disposition={"disposition": "merge", "strategy": "upsert"},
    primary_key="my_primary_key",
    table_format="delta"
)
def my_upsert_resource():
    ...

Known limitations#

• hard_delete hint not supported
• Deleting records from nested tables not supported
  • This means updates to JSON columns that involve element removals are not propagated. For example, if you first load {"key": 1, "nested": [1, 2]} and then load {"key": 1, "nested": [1]}, then the record for element 2 will not be deleted from the nested table.

By default, dlt runs Delta table upserts in streamed mode to reduce memory pressure. To enable the use of source table statistics to derive an early pruning predicate, set:

[destination.filesystem]
deltalake_streamed_exec = false

Delta table format storage options and configuration#

You can pass storage options and configuration by configuring both destination.filesystem.deltalake_storage_options and
destination.filesystem.deltalake_configuration:

[destination.filesystem]
deltalake_configuration = '{"delta.enableChangeDataFeed": "true", "delta.minWriterVersion": "7"}'
deltalake_storage_options = '{"AWS_S3_LOCKING_PROVIDER": "dynamodb", "DELTA_DYNAMO_TABLE_NAME": "custom_table_name"}'

dlt passes these as arguments to the write_deltalake method in the deltalake library (deltalake_configuration maps to configuration and deltalake_storage_options maps to storage_options). Look at their documentation to see which options can be used.

You don't need to specify credentials here. dlt merges the required credentials with the options you provided before passing it as storage_options.

❗When using s3, you need to specify storage options to configure locking behavior.

Delta table format memory usage#