Skip to content

Scalable Data I/O

Efficient parallel data processing requires data I/O to be parallelized effectively as well. Bodo provides parallel file I/O for many different formats such as Parquet, CSV, JSON, Numpy binaries, HDF5 and SQL databases. This diagram demonstrates how chunks of data are partitioned among parallel execution engines by Bodo.


Bodo reads dataset chunks in parallel

Bodo automatically parallelizes I/O for any number of cores and cluster size without any additional API layers.

Supported formats

Currently, Bodo supports I/O for Parquet, CSV, SQL, JSON, HDF5, and Numpy binaries formats. It can read these formats from multiple filesystems, including S3, HDFS and Azure Data Lake (ADLS) (see File Systems below for more information).

Also see supported pandas Operations for supported arguments of I/O functions.

Parquet

Parquet is a commonly used file format in analytics due to its efficient columnar storage. Bodo supports the standard pandas API for reading Parquet: pd.read_parquet(path), where path can be a parquet file or a directory with multiple parquet files (all are part of the same dataframe):

import pandas as pd
import bodo

@bodo.jit
def write_pq(df):
    df.to_parquet('example.pq')

@bodo.jit
def read_pq():
    df = pd.read_parquet('example.pq')
    return df

to_parquet(name) with distributed data writes to a folder called name. Each process writes one file into the folder, but if the data is not distributed, to_parquet(name) writes to a single file called name:

df = pd.DataFrame({'A': range(10)})

@bodo.jit
def example1_pq(df):
    df.to_parquet('example1.pq')

@bodo.jit(distributed={'df'})
def example2_pq(df):
    df.to_parquet('example2.pq')

if bodo.get_rank() == 0:
    example1_pq(df)
example2_pq(df)

Run the code above with 4 processors:

mpiexec -n 4 python example_pq.py

example1_pq(df) writes 1 single file, and example2_pq(df) writes a folder containing 4 parquet files:

.
├── example1.pq
├── example2.pq
│   ├── part-00.parquet
│   ├── part-01.parquet
│   ├── part-02.parquet
│   └── part-03.parquet

See read_parquet(), to_parquet() for supported arguments.

Filter pushdown

Filter Pushdown

Bodo can detect filters used by the code and optimize the read_parquet call by pushing the filters down to the storage layer, so that only the rows required by the program are read. This can significantly speed up I/O in many cases and will reduce the program's memory footprint, sometimes substantially.

For example, suppose we have a large dataset that spans many years and we only need to read data for a particular year. With pandas, we might perform a query on the year 2021 like this:

@bodo.jit
def query():
    df = pd.read_parquet("s3://my-bucket/data.pq")
    df = df[(df["year"] == 2021)]
    return df.groupby("customer_key")["revenue"].max()

When compiling the above, Bodo detects the df[(df["year"] == 2021)] filter and optimizes the read_parquet call so that it only reads data for year 2021 from S3. Because the data will have already been filtered after reading, Bodo removes the filter operation during compilation. Note that this requires code transformation and optimization and is something that pandas cannot do. Bodo automatically infers which filters can be pushed down.

If your dataset is hive-partitioned and partition columns appear in filter expressions, only the files that contain relevant data are read, and the rest are discarded based on their path. For example, if year is a partition column above and we have a dataset:

.
└── data.pq/
    │   ...
    ├───year=2020/
    │   ├── part-00.parquet
    │   └── part-01.parquet
    └───year=2021/
        ├── part-02.parquet
        └── part-03.parquet

Bodo will only read the files in the year=2021 directory.

For non-partition columns, Bodo may discard files entirely just by looking at their parquet metadata (depending on the filters and statistics contained in the metadata) or filter the rows during read.

Note

Filter pushdown can be a very significant optimization. Please refer to the inlining section to make sure these optimizations are applied in your program.

Exploring Large Data Without Full Read

Exploring Large Data Without Full Read

Exploring large datasets often requires seeing its shape and a sample of the data. Bodo is able to provide this information quickly without loading the full Parquet dataset, which means there is no need for a large cluster with a lot of memory. For example:

@bodo.jit
def head_only_read():
    df = pd.read_parquet("example.pq")
    print(df.shape)
    print(df.head())

In this example, Bodo provides the shape information for the full dataset in df.shape, but only loads the first few rows that are necessary for df.head().

CSV

CSV is a common text format for data exchange. Bodo supports most of the standard pandas API to read CSV files:

import pandas as pd
import bodo

@bodo.jit
def write_csv(df):
    df.to_csv('example.csv')

@bodo.jit
def read_csv():
    df = pd.read_csv('example.csv')
    return df

Unlike read_csv in regular pandas, Bodo can read a directory that contains multiple partitioned CSV files as well. All files in the folder must have the same number and datatype of columns. They can have different number of rows.

Usage:

@bodo.jit
def read_csv_folder():
    df = pd.read_csv("/path/to/folder/foldername")
    doSomething(df)

Use sep="n" to read text files line by line into a single-column dataframe (without creating separate columns, useful when text data is unstructured or there are too many columns to read efficiently):

@bodo.jit
def read_test():
    df = pd.read_csv("example.csv", sep="n", names=["value"], dtype={"value": "str"})
    return df

Note

Bodo uses nullable integer types of pandas to ensure type stability (see integer NA issue in pandas for more details). Therefore, data types must be specified explicitly for accurate performance comparisons of Bodo and pandas for read_csv.

to_csv(name) has different behaviors for different file systems:

  1. POSIX file systems: always writes to a single file, regardless of the number of processes and whether the data is distributed, but writing is still done in parallel when more than 1 processor is used:

    df = pd.DataFrame({'A': np.arange(n)})
    
    @bodo.jit
    def example1_csv(df):
        df.to_csv('example1.csv')
    
    @bodo.jit(distributed={'df'})
    def example2_csv(df):
        df.to_csv('example2.csv')
    
    if bodo.get_rank() == 0:
        example1_csv(df)
    example2_csv(df)
    

    Run the code above with 4 processors:

    mpiexec -n 4 python example_csv.py
    

    each example1_csv(df) and example2_csv(df) writes to a single file:

    .
    ├── example1.csv
    ├── example2.csv
    
  2. S3 and HDFS: distributed data is written to a folder called name. Each process writes one file into the folder, but if the data is not distributed, to_csv(name) writes to a single file called name:

    df = pd.DataFrame({'A': np.arange(n)})
    
    @bodo.jit
    def example1_csv(df):
        df.to_csv('s3://bucket-name/example1.csv')
    
    @bodo.jit(distributed={'df'})
    def example2_csv(df):
        df.to_csv('s3://bucket-name/example2.csv')
    
    if bodo.get_rank() == 0:
        example1_csv(df)
    example2_csv(df)
    

    Run the code above with 4 processors:

    mpiexec -n 4 python example_csv.py
    

    example1_csv(df) writes 1 single file, and example2_csv(df) writes a folder containing 4 csv files:

    .
    ├── example1.csv
    ├── example2.csv
    │   ├── part-00.csv
    │   ├── part-01.csv
    │   ├── part-02.csv
    │   └── part-03.csv
    

See read_csv(), to_csv() for supported arguments.

JSON

For JSON, the syntax is also the same as pandas.

Usage:

@bodo.jit
def example_write_json(df, fname):
    df.to_json(fname)

@bodo.jit
def example_read_json_lines_format():
    df = pd.read_json('example.json', orient = 'records', lines = True)

@bodo.jit
def example_read_json_multi_lines():
    df = pd.read_json('example_file.json', orient = 'records', lines = False,
        dtype={"A": float, "B": "bool", "C": int})

Note

  • The dtype argument is required when reading a regular multi-line JSON file.
  • Bodo cannot read a directory containing multiple multi-line JSON files

to_json(name) has different behaviors for different file systems:

  1. POSIX file systems: to_json(name) behavior depends on orient and lines arguments.

    1. DataFrame.to_json(name, orient='records', lines=True) (i.e. writing JSON Lines text file format) always writes to a single file, regardless of the number of processes and whether the data is distributed, but writing is still done in parallel when more than 1 processor is used:

      df = pd.DataFrame({'A': np.arange(n)})
      
      @bodo.jit
      def example1_json(df):
          df.to_json('example1.json', orient='records', lines=True)
      
      @bodo.jit(distributed={'df'})
      def example2_json(df):
          df.to_json('example2.json', orient='records', lines=True)
      
      if bodo.get_rank() == 0:
          example1_json(df)
      example2_jsons(df)
      

      Run the code above with 4 processors:

      mpiexec -n 4 python example_json.py
      

      each example1_json(df) and example2_json(df) writes to a single file:

      .
      ├── example1.json
      ├── example2.json
      
    2. All other combinations of values for orient and lines have the same behavior as S3 and HDFS explained below.

  2. S3 and HDFS : distributed data is written to a folder called name. Each process writes one file into the folder, but if the data is not distributed, to_json(name) writes to a file called name:

    df = pd.DataFrame({'A': np.arange(n)})
    
    @bodo.jit
    def example1_json(df):
        df.to_json('s3://bucket-name/example1.json')
    
    @bodo.jit(distributed={'df'})
    def example2_json(df):
        df.to_json('s3://bucket-name/example2.json')
    
    if bodo.get_rank() == 0:
        example1_json(df)
    example2_json(df)
    

    Run the code above with 4 processors:

    mpiexec -n 4 python example_json.py
    

    example1_json(df) writes 1 single file, and example2_json(df) writes a folder containing 4 json files:

    .
    ├── example1.json
    ├── example2.json
    │   ├── part-00.json
    │   ├── part-01.json
    │   ├── part-02.json
    │   └── part-03.json
    

    See read_json()][pandas-f-in], [to_json()` for supported arguments.

SQL

For SQL, the syntax is also the same as pandas. For reading:

@bodo.jit
def example_read_sql():
    df = pd.read_sql('select * from employees', 'mysql+pymysql://admin:server')

See read_sql() for supported arguments.

For writing:

@bodo.jit
def example_write_sql(df):
    df.to_sql('table_name', 'mysql+pymysql://admin:server')

See to_sql() for supported arguments.

Note

sqlalchemy must be installed in order to use pandas.read_sql.

Delta Lake

Reading parquet files from Delta Lake is supported locally, from S3, and from Azure ADLS.

  • The Delta Lake binding python packaged needs to be installed using pip:pip install deltalake.
  • For S3, the AWS_DEFAULT_REGION environment variable should be set to the region of the bucket hosting the Delta Lake table.
  • For ADLS, the AZURE_STORAGE_ACCOUNT and AZURE_STORAGE_KEY environment variables need to be set.

Example code for reading:

@bodo.jit
def example_read_deltalake():
    df = pd.read_parquet('path/to/deltalake')

Note

Writing is currently not supported.

Numpy binaries

Numpy's fromfile and tofile are supported as below:

@bodo.jit
def example_np_io():
    A = np.fromfile("myfile.dat", np.float64)
    ...
    A.tofile("newfile.dat")

Bodo has the same behavior as Numpy for numpy.ndarray.tofile(), where we always write to a single file. However, writing distributed data to POSIX is done in parallel, but writing to S3 & HDFS is done sequentially (due to file system limitations).

HDF5

HDF5 is a common format in scientific computing, especially for multi-dimensional numerical data. HDF5 can be very efficient at scale, since it has native parallel I/O support. For HDF5, the syntax is the same as the h5py package. For example:

@bodo.jit
def example_h5():
    f = h5py.File("data.hdf5", "r")
    X = f['points'][:]
    Y = f['responses'][:]

Filepaths determined at runtime

When reading from a file, Bodo needs to know the types of the resulting dataframe. If the file name is a constant string or function argument, Bodo can look at the file at compile time and infer the types. If the the filepath is not constant, this information must be supplied by the user. For pd.read_csv and pd.read_excel, this information can be supplied through the names and dtypes keyword arguments:

@bodo.jit
def example_csv(fname1, fname2, flag)):
    if flag:
        file_name = fname1
    else:
        file_name = fname2
    return pd.read_csv(file_name, names = ["A", "B", "C"], dtype={"A": int, "B": float, "C": str})

@bodo.jit
def example_excel(fname1, fname2, flag)):
    if flag:
            file_name = fname1
    else:
        file_name = fname2
    return pd.read_excel(
        file_name,
        names=["A", "B", "C", "D", "E"],
        dtype={"A": int, "B": float, "C": str, "D": str, "E": np.bool_},
    )

For the remaining pandas read functions, the existing APIs do not currently allow this information to be supplied. Users can still provide this information by adding type information in the bodo.jit decorator, similar to Numba's typing syntax. For example:

@bodo.jit(locals={'df':{'one': bodo.float64[:],
                  'two': bodo.string_array_type,
                  'three': bodo.bool_[:],
                  'four': bodo.float64[:],
                  'five': bodo.string_array_type,
                  }})
def example_df_schema(fname1, fname2, flag):
    if flag:
        file_name = fname1
    else:
        file_name = fname2
    df = pd.read_parquet(file_name)
    return df


 @bodo.jit(locals={'X': bodo.float64[:,:], 'Y': bodo.float64[:]})
 def example_h5(fname1, fname2, flag):
    if flag:
        file_name = fname1
    else:
        file_name = fname2
     f = h5py.File(file_name, "r")
     X = f['points'][:]
     Y = f['responses'][:]

For the complete list of supported types, please see the pandas dtype section. In the event that the dtypes are improperly specified, Bodo will throw a runtime error.

File Systems

Amazon S3

Reading and writing CSV, Parquet, JSON, and Numpy binary files from and to Amazon S3 is supported.

The fsspec package must be available, and the file path should start with s3://:

@bodo.jit
def example_s3_parquet():
    df = pd.read_parquet('s3://bucket-name/file_name.parquet')

These environment variables are used for File I/O with S3 credentials:

  • AWS_ACCESS_KEY_ID
  • AWS_SECRET_ACCESS_KEY
  • AWS_DEFAULT_REGION: default as us-east-1
  • AWS_S3_ENDPOINT: specify custom host name, default as AWS endpoint(s3.amazonaws.com)

Connecting to S3 endpoints through a proxy is supported. The proxy URI can be provided by setting one of the following environment variables (listed in order of precedence):

  • http_proxy
  • https_proxy
  • HTTP_PROXY
  • HTTPS_PROXY

Bodo uses Apache Arrow internally for read and write of data on S3.

Google Cloud Storage

Reading and writing Parquet files from and to Google Cloud is supported.

The file path should start with gs:// or gcs://:

@bodo.jit
def example_gcs_parquet():
    df = pd.read_parquet('gcs://bucket-name/file_name.parquet')
These environment variables are used for File I/O with GCS credentials:

  • GOOGLE_APPLICATION_CREDENTIALS

Details for GOOGLE_APPLICATION_CREDENTIALS can be seen in the Google docs here.

Bodo uses the fsspec-based gcsfs library internally for read and write of data on GCS.

Hadoop Distributed File System (HDFS) and Azure Data Lake Storage (ADLS) Gen2

Reading and writing CSV, Parquet, JSON, and Numpy binary files from and to Hadoop Distributed File System (HDFS) is supported. Note that Azure Data Lake Storage Gen2 can be accessed through HDFS.

The openjdk version 8 package must be available, and the file path should start with hdfs:// or abfs[s]://:

@bodo.jit
def example_hdfs_parquet():
    df = pd.read_parquet('hdfs://host:port/dir/file_name.pq')

These environment variables are used for File I/O with HDFS:

  • HADOOP_HOME: the root of your installed Hadoop distribution. Often is lib/native/libhdfs.so.
  • ARROW_LIBHDFS_DIR: location of libhdfs. Often is $HADOOP_HOME/lib/native.
  • CLASSPATH: must contain the Hadoop jars. You can set these using:
    export CLASSPATH=`$HADOOP_HOME/bin/hdfs classpath --glob`
    

Bodo uses Apache Arrow internally for read and write of data on HDFS. $HADOOP_HOME/etc/hadoop/hdfs-site.xml provides default behaviors for the HDFS client used by Bodo. Inconsistent configurations (e.g. dfs.replication) could potentially cause errors in Bodo programs.

Databases

Currently, Bodo supports most RDBMS that work with SQLAlchemy, with a corresponding driver.

Snowflake

To read a dataframe from a Snowflake database, users can use pd.read_sql with their Snowflake username and password: pd.read_sql(query,snowflake://<username>:<password>@url).

Prerequisites

In order to be able to query Snowflake from Bodo, you will have to install the Snowflake connector. If you're using Bodo in a conda environment:

conda install -c conda-forge snowflake-connector-python

If you've installed Bodo using pip, then you can install the Snowflake connector using pip as well:

pip install snowflake-connector-python

Usage

Bodo requires the Snowflake connection string to be passed as an argument to the pd.read_sql function. The complete code looks as follows:

import bodo
import pandas as pd

@bodo.jit(distributed=["df"])
def read_snowflake(db_name, table_name):
    df = pd.read_sql(
            f"SELECT * FROM {table_name}",
            f"snowflake://user:password@url/{db_name}/schema?warehouse=warehouse_name",
        )
    return df
df = read_snowflake(db_name, temp_table_name)

We can use the pd.to_sql method to persist a dataframe to a Snowflake table:

df.to_sql('<table_name>',f"snowflake://<username>:<password>@url/<db_name>/public?warehouse=XL_WH",schema="<schema>",if_exists="append",index=False)

Note

  • index=False is required as Snowflake does not support indexes.
  • if_exists=append is needed if the table already exists in snowflake.
Back to top