Dask Delayed#

In this section we parallelize simple for-loop style code with Dask and dask.delayed. Often, this is the only function that you will need to convert functions for use with Dask.

This is a simple way to use dask to parallelize existing codebases or build complex systems. This will also help us to develop an understanding for later sections.

Related Documentation

As we’ll see in the distributed scheduler notebook, Dask has several ways of executing code in parallel. We’ll use the distributed scheduler by creating a dask.distributed.Client. For now, this will provide us with some nice diagnostics. We’ll talk about schedulers in depth later.

from dask.distributed import Client

client = Client(n_workers=4)

Basics#

First let’s make some toy functions, inc and add, that sleep for a while to simulate work. We’ll then time running these functions normally.

In the next section we’ll parallelize this code.

from time import sleep

def inc(x):
    sleep(1)
    return x + 1

def add(x, y):
    sleep(1)
    return x + y

We time the execution of this normal code using the %%time magic, which is a special function of the Jupyter Notebook.

%%time
# This takes three seconds to run because we call each
# function sequentially, one after the other

x = inc(1)
y = inc(2)
z = add(x, y)

Parallelize with the dask.delayed decorator#

Those two increment calls could be called in parallel, because they are totally independent of one-another.

We’ll transform the inc and add functions using the dask.delayed function. When we call the delayed version by passing the arguments, exactly as before, the original function isn’t actually called yet - which is why the cell execution finishes very quickly. Instead, a delayed object is made, which keeps track of the function to call and the arguments to pass to it.

from dask import delayed

%%time
# This runs immediately, all it does is build a graph

x = delayed(inc)(1)
y = delayed(inc)(2)
z = delayed(add)(x, y)

This ran immediately, since nothing has really happened yet.

To get the result, call compute. Notice that this runs faster than the original code.

%%time
# This actually runs our computation using a local thread pool

z.compute()

What just happened?#

The z object is a lazy Delayed object. This object holds everything we need to compute the final result, including references to all of the functions that are required and their inputs and relationship to one-another. We can evaluate the result with .compute() as above or we can visualize the task graph for this value with .visualize().

z

# Look at the task graph for `z`
z.visualize()

Notice that this includes the names of the functions from before, and the logical flow of the outputs of the inc functions to the inputs of add.

Some questions to consider:#

  • Why did we go from 3s to 2s? Why weren’t we able to parallelize down to 1s?

  • What would have happened if the inc and add functions didn’t include the sleep(1)? Would Dask still be able to speed up this code?

  • What if we have multiple outputs or also want to get access to x or y?

Exercise: Parallelize a for loop#

for loops are one of the most common things that we want to parallelize. Use dask.delayed on inc and sum to parallelize the computation below:

data = [1, 2, 3, 4, 5, 6, 7, 8]

%%time
# Sequential code

results = []
for x in data:
    y = inc(x)
    results.append(y)
    
total = sum(results)

total

%%time
results = []

for x in data:
    y = delayed(inc)(x)
    results.append(y)
    
total = delayed(sum)(results)
print("Before computing:", total)  # Let's see what type of thing total is
result = total.compute()
print("After computing :", result)  # After it's computed

How do the graph visualizations compare with the given solution, compared to a version with the sum function used directly rather than wrapped with delayed? Can you explain the latter version? You might find the result of the following expression illuminating

delayed(inc)(1) + delayed(inc)(2)

Exercise: Parallelizing a for-loop code with control flow#

Often we want to delay only some functions, running a few of them immediately. This is especially helpful when those functions are fast and help us to determine what other slower functions we should call. This decision, to delay or not to delay, is usually where we need to be thoughtful when using dask.delayed.

In the example below we iterate through a list of inputs. If that input is even then we want to call inc. If the input is odd then we want to call double. This is_even decision to call inc or double has to be made immediately (not lazily) in order for our graph-building Python code to proceed.

def double(x):
    sleep(1)
    return 2 * x

def is_even(x):
    return not x % 2

data = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]

%%time
# Sequential code

results = []
for x in data:
    if is_even(x):
        y = double(x)
    else:
        y = inc(x)
    results.append(y)
    
total = sum(results)
print(total)

%%time
results = []
for x in data:
    if is_even(x):  # even
        y = delayed(double)(x)
    else:          # odd
        y = delayed(inc)(x)
    results.append(y)
    
total = delayed(sum)(results)

%time total.compute()

total.visualize()

Some questions to consider:#

  • What are other examples of control flow where we can’t use delayed?

  • What would have happened if we had delayed the evaluation of is_even(x) in the example above?

  • What are your thoughts on delaying sum? This function is both computational but also fast to run.

Exercise: Parallelizing a Pandas Groupby Reduction#

In this exercise we read several CSV files and perform a groupby operation in parallel. We are given sequential code to do this and parallelize it with dask.delayed.

The computation we will parallelize is to compute the mean departure delay per airport from some historical flight data. We will do this by using dask.delayed together with pandas. In a future section we will do this same exercise with dask.dataframe.

Inspect data#

import os
sorted(os.listdir(os.path.join('/work/classes/ds_1300/data', 'nycflights')))

Read one file with pandas.read_csv and compute mean departure delay#

import pandas as pd
df = pd.read_csv(os.path.join('/work/classes/ds_1300/data', 'nycflights', '1990.csv'))
df.head()

# What is the schema?
df.dtypes

# What originating airports are in the data?
df.Origin.unique()

# Mean departure delay per-airport for one year
df.groupby('Origin').DepDelay.mean()

Sequential code: Mean Departure Delay Per Airport#

The above cell computes the mean departure delay per-airport for one year. Here we expand that to all years using a sequential for loop.

from glob import glob
filenames = sorted(glob(os.path.join('/work/classes/ds_1300/data', 'nycflights', '*.csv')))

%%time

sums = []
counts = []
for fn in filenames:
    # Read in file
    df = pd.read_csv(fn)
    
    # Groupby origin airport
    by_origin = df.groupby('Origin')
    
    # Sum of all departure delays by origin
    total = by_origin.DepDelay.sum()
    
    # Number of flights by origin
    count = by_origin.DepDelay.count()
    
    # Save the intermediates
    sums.append(total)
    counts.append(count)

# Combine intermediates to get total mean-delay-per-origin
total_delays = sum(sums)
n_flights = sum(counts)
mean = total_delays / n_flights

mean

Parallelize the code above#

from dask import compute

%%time

sums = []
counts = []
for fn in filenames:
    # Read in file
    df = delayed(pd.read_csv)(fn)

    # Groupby origin airport
    by_origin = df.groupby('Origin')

    # Sum of all departure delays by origin
    total = by_origin.DepDelay.sum()

    # Number of flights by origin
    count = by_origin.DepDelay.count()

    # Save the intermediates
    sums.append(total)
    counts.append(count)

# Compute the intermediates
sums, counts = compute(sums, counts)

# Combine intermediates to get total mean-delay-per-origin
total_delays = sum(sums)
n_flights = sum(counts)
mean = total_delays / n_flights

# ensure the results still match
mean

Some questions to consider:#

  • How much speedup did you get? Is this how much speedup you’d expect?

  • Experiment with where to call compute. What happens when you call it on sums and counts? What happens if you wait and call it on mean?

  • Experiment with delaying the call to sum. What does the graph look like if sum is delayed? What does the graph look like if it isn’t?

  • Can you think of any reason why you’d want to do the reduction one way over the other?

Learn More#

Visit the Delayed documentation. In particular, this delayed screencast will reinforce the concepts you learned here and the delayed best practices document collects advice on using dask.delayed well.

Close the Client#

Before moving on to the next exercise, make sure to close your client or stop this kernel.

client.close()

Adapted from Dask Tutorial.