User Interfaces

Dask supports several user interfaces:

Each of these user interfaces employs the same underlying parallel computing machinery, and so has the same scaling, diagnostics, resilience, and so on, but each provides a different set of parallel algorithms and programming style.

This document helps you to decide which user interface best suits your needs, and gives some general information that applies to all interfaces. The pages linked above give more information about each interface in greater depth.

High-Level Collections

Many people who start using Dask are explicitly looking for a scalable version of NumPy, Pandas, or Scikit-Learn. For these situations, the starting point within Dask is usually fairly clear. If you want scalable NumPy arrays, then start with Dask array; if you want scalable Pandas DataFrames, then start with Dask DataFrame, and so on.

These high-level interfaces copy the standard interface with slight variations. These interfaces automatically parallelize over larger datasets for you for a large subset of the API from the original project.

# Arrays
import dask.array as da
rng = da.random.default_rng()
x = rng.uniform(low=0, high=10, size=(10000, 10000),  # normal numpy code
                chunks=(1000, 1000))  # break into chunks of size 1000x1000

y = x + x.T - x.mean(axis=0)  # Use normal syntax for high level algorithms

# DataFrames
import dask.dataframe as dd
df = dd.read_csv('2018-*-*.csv', parse_dates='timestamp',  # normal Pandas code
                 blocksize=64000000)  # break text into 64MB chunks

s = df.groupby('name').balance.mean()  # Use normal syntax for high level algorithms

# Bags / lists
import dask.bag as db
b = db.read_text('*.json').map(json.loads)
total = (b.filter(lambda d: d['name'] == 'Alice')
          .map(lambda d: d['balance'])
          .sum())

It is important to remember that, while APIs may be similar, some differences do exist. Additionally, the performance of some algorithms may differ from their in-memory counterparts due to the advantages and disadvantages of parallel programming. Some thought and attention is still required when using Dask.

Low-Level Interfaces

Often when parallelizing existing code bases or building custom algorithms, you run into code that is parallelizable, but isn’t just a big DataFrame or array. Consider the for-loopy code below:

results = []
for a in A:
    for b in B:
        if a < b:
            c = f(a, b)
        else:
            c = g(a, b)
        results.append(c)

There is potential parallelism in this code (the many calls to f and g can be done in parallel), but it’s not clear how to rewrite it into a big array or DataFrame so that it can use a higher-level API. Even if you could rewrite it into one of these paradigms, it’s not clear that this would be a good idea. Much of the meaning would likely be lost in translation, and this process would become much more difficult for more complex systems.

Instead, Dask’s lower-level APIs let you write parallel code one function call at a time within the context of your existing for loops. A common solution here is to use Dask delayed to wrap individual function calls into a lazily constructed task graph:

import dask

lazy_results = []
for a in A:
    for b in B:
        if a < b:
            c = dask.delayed(f)(a, b)  # add lazy task
        else:
            c = dask.delayed(g)(a, b)  # add lazy task
        lazy_results.append(c)

results = dask.compute(*lazy_results)  # compute all in parallel

Combining High- and Low-Level Interfaces

It is common to combine high- and low-level interfaces. For example, you might use Dask array/bag/dataframe to load in data and do initial pre-processing, then switch to Dask delayed for a custom algorithm that is specific to your domain, then switch back to Dask array/dataframe to clean up and store results. Understanding both sets of user interfaces, and how to switch between them, can be a productive combination.

# Convert to a list of delayed Pandas dataframes
delayed_values = df.to_delayed()

# Manipulate delayed values arbitrarily as you like

# Convert many delayed Pandas DataFrames back to a single Dask DataFrame
df = dd.from_delayed(delayed_values)

Laziness and Computing

Most Dask user interfaces are lazy, meaning that they do not evaluate until you explicitly ask for a result using the compute method:

# This array syntax doesn't cause computation
y = x + x.T - x.mean(axis=0)

# Trigger computation by explicitly calling the compute method
y = y.compute()

If you have multiple results that you want to compute at the same time, use the dask.compute function. This can share intermediate results and so be more efficient:

# compute multiple results at the same time with the compute function
min, max = dask.compute(y.min(), y.max())

Note that the compute() function returns in-memory results. It converts Dask DataFrames to Pandas DataFrames, Dask arrays to NumPy arrays, and Dask bags to lists. You should only call compute on results that will fit comfortably in memory. If your result does not fit in memory, then you might consider writing it to disk instead.

# Write larger results out to disk rather than store them in memory
my_dask_dataframe.to_parquet('myfile.parquet')
my_dask_array.to_hdf5('myfile.hdf5')
my_dask_bag.to_textfiles('myfile.*.txt')

Persist into Distributed Memory

Alternatively, if you are on a cluster, then you may want to trigger a computation and store the results in distributed memory. In this case you do not want to call compute, which would create a single Pandas, NumPy, or list result. Instead, you want to call persist, which returns a new Dask object that points to actively computing, or already computed results spread around your cluster’s memory.

# Compute returns an in-memory non-Dask object
y = y.compute()

# Persist returns an in-memory Dask object that uses distributed storage if available
y = y.persist()

This is common to see after data loading an preprocessing steps, but before rapid iteration, exploration, or complex algorithms. For example, we might read in a lot of data, filter down to a more manageable subset, and then persist data into memory so that we can iterate quickly.

import dask.dataframe as dd
df = dd.read_parquet('...')
df = df[df.name == 'Alice']  # select important subset of data
df = df.persist()  # trigger computation in the background

# These are all relatively fast now that the relevant data is in memory
df.groupby(df.id).balance.sum().compute()   # explore data quickly
df.groupby(df.id).balance.mean().compute()  # explore data quickly
df.id.nunique()                             # explore data quickly

Lazy vs Immediate

As mentioned above, most Dask workloads are lazy, that is, they don’t start any work until you explicitly trigger them with a call to compute(). However, sometimes you do want to submit work as quickly as possible, track it over time, submit new work or cancel work depending on partial results, and so on. This can be useful when tracking or responding to real-time events, handling streaming data, or when building complex and adaptive algorithms.

For these situations, people typically turn to the futures interface which is a low-level interface like Dask delayed, but operates immediately rather than lazily.

Here is the same example with Dask delayed and Dask futures to illustrate the difference.

Delayed: Lazy

@dask.delayed
def inc(x):
    return x + 1

@dask.delayed
def add(x, y):
    return x + y

a = inc(1)       # no work has happened yet
b = inc(2)       # no work has happened yet
c = add(a, b)    # no work has happened yet

c = c.compute()  # This triggers all of the above computations

Futures: Immediate

from dask.distributed import Client
client = Client()

def inc(x):
    return x + 1

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

a = client.submit(inc, 1)     # work starts immediately
b = client.submit(inc, 2)     # work starts immediately
c = client.submit(add, a, b)  # work starts immediately

c = c.result()                # block until work finishes, then gather result

You can also trigger work with the high-level collections using the persist function. This will cause work to happen in the background when using the distributed scheduler.

Combining Interfaces

There are established ways to combine the interfaces above:

  1. The high-level interfaces (array, bag, dataframe) have a to_delayed method that can convert to a sequence (or grid) of Dask delayed objects

    delayeds = df.to_delayed()
    
  2. The high-level interfaces (array, bag, dataframe) have a from_delayed method that can convert from either Delayed or Future objects

    df = dd.from_delayed(delayeds)
    df = dd.from_delayed(futures)
    
  3. The Client.compute method converts Delayed objects into Futures

    futures = client.compute(delayeds)
    
  4. The dask.distributed.futures_of function gathers futures from persisted collections

    from dask.distributed import futures_of
    
    df = df.persist()  # start computation in the background
    futures = futures_of(df)
    
  5. The Dask.delayed object converts Futures into delayed objects

    delayed_value = dask.delayed(future)
    

The approaches above should suffice to convert any interface into any other. We often see some anti-patterns that do not work as well:

  1. Calling low-level APIs (delayed or futures) on high-level objects (like Dask arrays or DataFrames). This downgrades those objects to their NumPy or Pandas equivalents, which may not be desired. Often people are looking for APIs like dask.array.map_blocks or dask.dataframe.map_partitions instead.

  2. Calling compute() on Future objects. Often people want the .result() method instead.

  3. Calling NumPy/Pandas functions on high-level Dask objects or high-level Dask functions on NumPy/Pandas objects

Conclusion

Most people who use Dask start with only one of the interfaces above but eventually learn how to use a few interfaces together. This helps them leverage the sophisticated algorithms in the high-level interfaces while also working around tricky problems with the low-level interfaces.

For more information, see the documentation for the particular user interfaces below: