target-duckdb

Singer target that loads data into DuckDB following the Singer spec.

Install

First, make sure Python 3 is installed on your system or follow these installation instructions for Mac or Ubuntu.

It's recommended to use uv:

uv sync

To run

Like any other target that's following the singer specification:

some-singer-tap | target-duckdb --config [config.json]

It's reading incoming messages from STDIN and using the properties in config.json to load data into DuckDB.

Note: To avoid version conflicts run tap and targets in separate virtual environments.

BATCH message support

In addition to RECORD messages, this target accepts Singer BATCH messages with either encoding:

• encoding.format: "jsonl" (optionally encoding.compression: "gzip") — loaded via DuckDB's native read_json table function.
• encoding.format: "arrow" — loaded via DuckDB's native arrow community extension (read_arrow).

Both paths load each manifest file directly via INSERT INTO ... BY NAME SELECT * FROM read_json(...)/read_arrow(...) — there's no Python-side row-by-row materialization, unlike RECORD-mode's CSV-based loading. The arrow extension is fetched automatically the first time an Arrow BATCH message is processed (INSTALL arrow FROM community), so the environment running this target needs network access at least once; it's cached locally afterwards.

Manifest files are deleted from disk once their data has been durably loaded into the target table — they're treated as consumed, the same way other Singer BATCH consumers handle them. If a multi-file batch fails partway through, none of its files are deleted, so a retry can pick up the full batch again.

Limitations for BATCH-sourced records:

• add_metadata_columns/hard_delete are not populated for BATCH-sourced records (no time_extracted is available per-row the way it is in a RECORD message's envelope) — the _SDC_* columns are left NULL for rows loaded from a BATCH message.
• data_flattening_max_level > 0 is not supported for BATCH-sourced streams whose schema actually has nested object/array properties: the target table gets flattened column names (e.g. c_obj__nested_prop1), but a BATCH file's columns keep their original (unflattened) names, so there's no way to match them by name. The target raises a clear error in this case rather than silently dropping data. Set data_flattening_max_level=0 (the default) for BATCH-mode streams, or don't send BATCH messages for streams that need flattening.

Configuration settings

Running the target connector requires a config.json file. An example with the minimal settings:

{
    "path": "/path/to/local/file.duckdb",
    "default_target_schema": "main"
}

Additional options in config.json:

Property	Type	Required?	Description
path (alias: filepath)	String	Yes	The path to use for the duckdb.connect call; either a local file or a MotherDuck connection uri.
dbname (alias: database)	String		The database name to write to; this will be inferred from the path property if it is not specified.
token	String		For MotherDuck connections, the auth token to use (this may also be set directly via the MOTHERDUCK_TOKEN environment variable.
delimiter	String		(Default: ',') The delimiter to use for the CSV files that are used for record imports.
quotechar	String		(Default: '"') The quote character to use for the CSV files that are used for record imports.
batch_size_rows	Integer		(Default: 100000) Maximum number of rows in each batch. At the end of each batch, the rows in the batch are loaded into DuckDB.
flush_all_streams	Boolean		(Default: False) Flush and load every stream into DuckDB when one batch is full. Warning: This may trigger the COPY command to use files with low number of records.
default_target_schema	String		Name of the schema where the tables will be created. If schema_mapping is not defined then every stream sent by the tap is loaded into this schema.
schema_mapping	Object		Useful if you want to load multiple streams from one tap to multiple DuckDB schemas. If the tap sends the stream_id in <schema_name>-<table_name> format then this option overwrites the default_target_schema value.
add_metadata_columns	Boolean		(Default: False) Metadata columns add extra row level information about data ingestions, (i.e. when was the row read in source, when was inserted or deleted in postgres etc.) Metadata columns are creating automatically by adding extra columns to the tables with a column prefix _SDC_. The column names are following the stitch naming conventions documented at https://www.stitchdata.com/docs/data-structure/integration-schemas#sdc-columns. Enabling metadata columns will flag the deleted rows by setting the _SDC_DELETED_AT metadata column. Without the add_metadata_columns option the deleted rows from singer taps will not be recognisable in DuckDB. Not populated for BATCH-sourced records (see BATCH message support).
hard_delete	Boolean		(Default: False) When hard_delete option is true then DELETE SQL commands will be performed in DuckDB to delete rows in tables. It's achieved by continuously checking the _SDC_DELETED_AT metadata column sent by the singer tap. Due to deleting rows requires metadata columns, hard_delete option automatically enables the add_metadata_columns option as well. Not populated for BATCH-sourced records (see BATCH message support).
data_flattening_max_level	Integer		(Default: 0) Object type RECORD items from taps can be transformed to flattened columns by creating columns automatically. When value is 0 (default) then flattening functionality is turned off. Not supported for BATCH-sourced streams with nested properties when set above 0 (see BATCH message support).
primary_key_required	Boolean		(Default: True) Log based and Incremental replications on tables with no Primary Key cause duplicates when merging UPDATE events. When set to true, stop loading data if no Primary Key is defined.
validate_records	Boolean		(Default: False) Validate every single record message to the corresponding JSON schema. This option is disabled by default and invalid RECORD messages will fail only at load time by DuckDB. Enabling this option will detect invalid records earlier but could cause performance degradation.
temp_dir	String		(Default: platform-dependent) Directory of temporary CSV files with RECORD messages.

Running Tests

1. Install Tox

uv tool install --with tox-uv tox

2. To run unit tests:

tox -e unit

3. To run integration tests:

tox -e integration

Linting

1. Install python dependencies and run python linter

tox -e lint

License

Apache License Version 2.0

See LICENSE to see the full text.