A loose implementation of deltalake, and the deltalake, written in Python on top of pyarrow, focused on extensibility, customizability, and distributed data.
This is mostly inspired by the deltalake package, and is (much) less battle tested. However, it is more flexible given it's Pythonic design. If you're interested give it a shot and maybe even help make it better.
pip install xdlake
Instantiate a table! This can be a local or remote. For remote, you may need to install the relevant fsspec implementation, for instance s3fs, gcsfs, adlfs for AWS S3, Google Storage, and Azure Storage, respectively.
dt = xdlake.DeltaTable("path/to/my/cool/local/table")
dt = xdlake.DeltaTable("s3://path/to/my/cool/table")
dt = xdlake.DeltaTable("az://path/to/my/cool/table", storage_options=dict_of_azure_creds)
Read the data. For fancy filtering and predicate push down and whatever, use to_pyarrow_dataset and
learn how to filter pyarrow datasets.
ds = dt.to_pyarrow_dataset()
t = dt.to_pyarrow_table()
df = dt.to_pandas()
Instances of DeltaTable are immutable: any method that performs a table operation will return a new DeltaTable.
Write data from memory. Data can be pyarrow tables, datasets, record batches, pandas DataFrames, or iterables of those things.
dt = dt.write(my_cool_pandas_dataframe)
dt = dt.write(my_cool_arrow_table)
dt = dt.write(my_cool_arrow_dataset)
dt = dt.write(my_cool_arrow_record_batches)
dt = dt.write(pyarrow.Table.from_pandas(df))
Import references to foreign data without copying. Data may be heterogeneously located in s3, gs, azure, and local, and cn be partitioned differently than the DeltaTable itself. Go hog wild.
See Credentials if you need different creds for different storage locations.
Import data from various locations in one go. This only works for non-partitioned data.
dt = dt.import_refs(["s3://some/aws/data", "gs://some/gcp/data", "az://some/azure/data" ])
dt = dt.import_refs(my_pyarrow_filesystem_dataset)
Partitioned data needs to be handled a differently. First, you'll need to read up on pyarrow partitioning to do it. Second, you can only import one dataset at a time.
foreign_partitioning = pyarrow.dataset.partitioning(...)
ds = pyarrow.dataset.dataset(
list_of_files,
partitioning=foreign_partitioning,
partition_base_dir,
filesystem=xdlake.storage.get_filesystem(foreign_refs_loc),
)
dt = dt.import_refs(ds, partition_by=my_partition_cols)
Delete rows from a DeltaTable using pyarrow expressions:
import pyarrow.compute as pc
expr = (
(pc.field("cats") == pc.scalar("A"))
|
(pc.field("float64") > pc.scalar(0.9))
)
dt = dt.delete(expr)
I really want to support deletion vectors, but pyarrow can't filter parquet files by row indices (pretty basic if you ask me). If you also would like xdlake to support deletion vectors, let the arrow folks know by chiming in here.
You can clone a deltatable. This is a soft clone (no data is copied, and the new table just references the data). The entire version history is preserved. Writes are written to the new location.
cloned_dt = dt.clone("the/location/of/the/clone")
DeltaTables that reference distributed data may need credentials for various cloud locations.
To register default credentials for s3, gs, etc.
xdlake.storage.register_default_filesystem_for_protocol("s3", s3_creds)
xdlake.storage.register_default_filesystem_for_protocol("gs", gs_creds)
xdlake.storage.register_default_filesystem_for_protocol("az", az_creds)
To register specific credentials for various prefixes:
xdlake.storage.register_filesystem("s3://bucket-doom/foo/bar", s3_creds)
xdlake.storage.register_filesystem("s3://bucket-zoom/biz/baz", other_s3_creds)
xdlake.storage.register_filesystem("az://container-blah/whiz/whaz", az_creds)
Project home page GitHub
The deltalake transaction log protocol
Please report bugs, issues, feature requests, etc. on GitHub.
