Dask DataFrames

Parallel Programming with Dask in Python

James Fulton

Climate Informatics Researcher

pandas DataFrames vs. Dask DataFrames

import pandas as pd

# Read a single csv file
pandas_df = pd.read_csv(
    "dataset/chunk1.csv"
)

This reads a single CSV file immediately.

import dask.dataframe as dd

# Lazily read all csv files
dask_df = dd.read_csv(
    "dataset/*.csv"
)

This reads all the CSV files in the dataset folder lazily.

Dask DataFrames

print(dask_df)

Dask DataFrame Structure:
              ID     col1     col2    col3     col4     ...
npartitions=3                               
           int64   object   object   int64  float64     ...
             ...      ...      ...     ...      ...     ...
             ...      ...      ...     ...      ...     ...
             ...      ...      ...     ...      ...     ...
Dask Name: getitem, 3 tasks

Dask DataFrame task graph

dask.visualize(dask_df)

The task graph shows 3 loading operations.

Controlling the size of blocks

# Set the maximum memory of a chunk
dask_df = dd.read_csv("dataset/*.csv", blocksize="10MB")

print(dask_df)

Dask DataFrame Structure:
              ID     col1     col2    col3     col4     ...
npartitions=7                               
           int64   object   object   int64  float64     ...
             ...      ...      ...     ...      ...     ...
             ...      ...      ...     ...      ...     ...
             ...      ...      ...     ...      ...     ...
Dask Name: getitem, 7 tasks

Explaining partitions

# Set the maximum memory of a chunk
dask_df = dd.read_csv("dataset/*.csv", blocksize="10MB")

Why 7 partitions?

size      file
  9M      dataset/chunk1.csv
 18M      dataset/chunk2.csv
 32M      dataset/chunk3.csv

Explaining partitions

# Set the maximum memory of a chunk
dask_df = dd.read_csv("dataset/*.csv", blocksize="10MB")

Why 7 partitions?

size      file
  9M      dataset/chunk1.csv    # becomes 1 partition
 18M      dataset/chunk2.csv    # becomes 2 partitions
 32M      dataset/chunk3.csv    # becomes 4 partitions

Analysis with Dask DataFames

Select column
```
col1 = dask_df['col1']
```
Assigning columns
```
dask_df['double_col1'] = 2 * col1
```
Mathematical operations, e.g.
```
dask_df.std()
dask_df.min()
```

Groupby
```
dask_df.groupby(col1).mean()
```
Even fuctions you used earlier
```
dask_df.nlargest(n=3, columns='col1')
```

Datetimes and other pandas functionality

import pandas as pd

# Converting string to datetime format
pd.to_datetime(pandas_df['start_date'])


# Accessing datetime attributes
pandas_df['start_date'].dt.year
pandas_df['start_date'].dt.day
pandas_df['start_date'].dt.hour
pandas_df['start_date'].dt.minute

import dask.dataframe as dd

# Converting string to datetime format
dd.to_datetime(dask_df['start_date'])


# Accessing datetime attributes
dask_df['start_date'].dt.year
dask_df['start_date'].dt.day
dask_df['start_date'].dt.hour
dask_df['start_date'].dt.minute

Making results non-lazy

# Show 5 rows
print(dask_df.head())

        ID     double_col1     col1    col2     col3     ...
0   543795              20       10     436        0     ...
1   874535              24       12     268        0     ...
2   781326              62       31     211        0     ...
3   112457              18        9     898        1     ...
4   103256             142       71     663        0     ...

# Convert lazy Dask DataFrame to in-memory pandas DataFrame
results_df = df.compute()

Sending answer straight to file

# 7 partitions (chunks) so 7 output files
dask_df.to_csv('answer/part-*.csv')

part-0.csv
part-1.csv
part-2.csv
part-3.csv
part-4.csv
part-5.csv
part-6.csv

Faster file formats - Parquet

# Read from parquet
dask_df = dd.read_parquet('dataset_parquet')

# Save to parquet
dask_df.to_parquet('answer_parquet')

Parquet format is multiple times faster to read data from than CSV
Can be faster to write to

Let's practice!

Parallel Programming with Dask in Python