Load and Save Data with Dask DataFrames

You can create a Dask DataFrame from various data storage formats like CSV, HDF, Apache Parquet, and others. For most formats, this data can live on various storage systems including local disk, network file systems (NFS), the Hadoop Distributed File System (HDFS), Google Cloud Storage, and Amazon S3 (excepting HDF, which is only available on POSIX like file systems).

See the DataFrame overview page for more on dask.dataframe scope, use, and limitations and DataFrame Best Practices for more tips and solutions to common problems.


The following functions provide access to convert between Dask DataFrames, file formats, and other Dask or Python collections.

File Formats:

read_csv(urlpath[, blocksize, ...])

Read CSV files into a Dask.DataFrame

read_parquet(path[, columns, filters, ...])

Read a Parquet file into a Dask DataFrame

read_hdf(pattern, key[, start, stop, ...])

Read HDF files into a Dask DataFrame

read_orc(path[, engine, columns, index, ...])

Read dataframe from ORC file(s)

read_json(url_path[, orient, lines, ...])

Create a dataframe from a set of JSON files

read_sql_table(table_name, con, index_col[, ...])

Read SQL database table into a DataFrame.

read_sql_query(sql, con, index_col[, ...])

Read SQL query into a DataFrame.

read_sql(sql, con, index_col, **kwargs)

Read SQL query or database table into a DataFrame.

read_table(urlpath[, blocksize, ...])

Read delimited files into a Dask.DataFrame

read_fwf(urlpath[, blocksize, ...])

Read fixed-width files into a Dask.DataFrame

from_array(x[, chunksize, columns, meta])

Read any sliceable array into a Dask Dataframe

to_csv(df, filename[, single_file, ...])

Store Dask DataFrame to CSV files

to_parquet(df, path[, engine, compression, ...])

Store Dask.dataframe to Parquet files

to_hdf(df, path, key[, mode, append, ...])

Store Dask Dataframe to Hierarchical Data Format (HDF) files

to_sql(df, name, uri[, schema, if_exists, ...])

Store Dask Dataframe to a SQL table

Dask Collections:

from_delayed(dfs[, meta, divisions, prefix, ...])

Create Dask DataFrame from many Dask Delayed objects

from_dask_array(x[, columns, index, meta])

Create a Dask DataFrame from a Dask Array.

from_map(func, *iterables[, args, meta, ...])

Create a DataFrame collection from a custom function map

dask.bag.core.Bag.to_dataframe([meta, ...])

Create Dask Dataframe from a Dask Bag.


Convert into a list of dask.delayed objects, one per partition.


Create Dask Array from a Dask Dataframe

to_bag(df[, index, format])

Create Dask Bag from a Dask DataFrame



Construct a Dask DataFrame from a Pandas DataFrame

DataFrame.from_dict(data, *, npartitions[, ...])

Construct a Dask DataFrame from a Python Dictionary


Read from CSV

You can use read_csv() to read one or more CSV files into a Dask DataFrame. It supports loading multiple files at once using globstrings:

>>> df = dd.read_csv('myfiles.*.csv')

You can break up a single large file with the blocksize parameter:

>>> df = dd.read_csv('largefile.csv', blocksize=25e6)  # 25MB chunks

Changing the blocksize parameter will change the number of partitions (see the explanation on partitions). A good rule of thumb when working with Dask DataFrames is to keep your partitions under 100MB in size.

Read from Parquet

Similarly, you can use read_parquet() for reading one or more Parquet files. You can read in a single Parquet file:

>>> df = dd.read_parquet("path/to/mydata.parquet")

Or a directory of local Parquet files:

>>> df = dd.read_parquet("path/to/my/parquet/")

For more details on working with Parquet files, including tips and best practices, see the documentation on Dask Dataframe and Parquet.

Read from cloud storage

Dask can read data from a variety of data stores including cloud object stores. You can do this by prepending a protocol like s3:// to paths used in common data access functions like dd.read_csv:

>>> df = dd.read_csv('s3://bucket/path/to/data-*.csv')
>>> df = dd.read_parquet('gcs://bucket/path/to/data-*.parq')

For remote systems like Amazon S3 or Google Cloud Storage, you may need to provide credentials. These are usually stored in a configuration file, but in some cases you may want to pass storage-specific options through to the storage backend. You can do this with the storage_options parameter:

>>> df = dd.read_csv('s3://bucket-name/my-data-*.csv',
...                  storage_options={'anon': True})
>>> df = dd.read_parquet('gs://dask-nyc-taxi/yellowtrip.parquet',
...                      storage_options={'token': 'anon'})

See the documentation on connecting to Amazon S3 or Google Cloud Storage.

Mapping from a function

For cases that are not covered by the functions above, but can be captured by a simple map operation, from_map() is likely to be the most convenient means for DataFrame creation. For example, this API can be used to convert an arbitrary PyArrow Dataset object into a DataFrame collection by mapping fragments to DataFrame partitions:

>>> import pyarrow.dataset as ds
>>> dataset = ds.dataset("hive_data_path", format="orc", partitioning="hive")
>>> fragments = dataset.get_fragments()
>>> func = lambda frag: frag.to_table().to_pandas()
>>> df = dd.from_map(func, fragments)

Dask Delayed

Dask delayed is particularly useful when simple map operations aren’t sufficient to capture the complexity of your data layout. It lets you construct Dask DataFrames out of arbitrary Python function calls, which can be helpful to handle custom data formats or bake in particular logic around loading data. See the documentation on using dask.delayed with collections.


Writing files locally

You can save files locally, assuming each worker can access the same file system. The workers could be located on the same machine, or a network file system can be mounted and referenced at the same path location for every worker node. See the documentation on accessing data locally.

Writing to remote locations

Dask can write to a variety of data stores including cloud object stores. For example, you can write a dask.dataframe to an Azure storage blob as:

>>> d = {'col1': [1, 2, 3, 4], 'col2': [5, 6, 7, 8]}
>>> df = dd.from_pandas(pd.DataFrame(data=d), npartitions=2)
>>> dd.to_parquet(df=df,
...               path='abfs://CONTAINER/FILE.parquet'
...               storage_options={'account_name': 'ACCOUNT_NAME',
...                                'account_key': 'ACCOUNT_KEY'}

See the how-to guide on connecting to remote data.