# Best Practices¶

It is easy to get started with Dask delayed, but using it well does require some experience. This page contains suggestions for best practices, and includes solutions to common problems.

## Call delayed on the function, not the result¶

Dask delayed operates on functions like dask.delayed(f)(x, y), not on their results like dask.delayed(f(x, y)). When you do the latter, Python first calculates f(x, y) before Dask has a chance to step in.

 Don’t Do # This executes immediately dask.delayed(f(x, y))  # This makes a delayed function, acting lazily dask.delayed(f)(x, y) 

## Compute on lots of computation at once¶

To improve parallelism, you want to include lots of computation in each compute call. Ideally, you want to make many dask.delayed calls to define your computation and then call dask.compute only at the end. It is ok to call dask.compute in the middle of your computation as well, but everything will stop there as Dask computes those results before moving forward with your code.

 Don’t Do # Avoid calling compute repeatedly results = [] for x in L: y = dask.delayed(f)(x) results.append(y.compute()) results  # Collect many calls for one compute results = [] for x in L: y = dask.delayed(f)(x) results.append(y) results = dask.compute(*results) 

Calling y.compute() within the loop would await the result of the computation every time, and so inhibit parallelism.

## Don’t mutate inputs¶

Your functions should not change the inputs directly.

 Don’t Do # Mutate inputs in functions @dask.delayed def f(x): x += 1 return x  # Return new values or copies @dask.delayed def f(x): x = x + 1 return x 

If you need to use a mutable operation, then make a copy within your function first:

@dask.delayed
def f(x):
x = copy(x)
x += 1
return x


## Avoid global state¶

Ideally, your operations shouldn’t rely on global state. Using global state might work if you only use threads, but when you move to multiprocessing or distributed computing then you will likely encounter confusing errors.

 Don’t L = [] # This references global variable L @dask.delayed def f(x): L.append(x) 

## Don’t rely on side effects¶

Delayed functions only do something if they are computed. You will always need to pass the output to something that eventually calls compute.

 Don’t Do # Forget to call compute dask.delayed(f)(1, 2, 3) ...  # Ensure delayed tasks are computed x = dask.delayed(f)(1, 2, 3) ... dask.compute(x, ...) 

In the first case here, nothing happens, because compute() is never called.

## Break up computations into many pieces¶

Every dask.delayed function call is a single operation from Dask’s perspective. You achieve parallelism by having many delayed calls, not by using only a single one: Dask will not look inside a function decorated with @dask.delayed and parallelize that code internally. To accomplish that, it needs your help to find good places to break up a computation.

 Don’t Do # One giant task def load(filename): ... def process(filename): ... def save(filename): ... @dask.delayed def f(filenames): results = [] for filename in filenames: data = load(filename) data = process(data) result = save(data) return results dask.compute(f(filenames))  # Break up into many tasks @dask.delayed def load(filename): ... @dask.delayed def process(filename): ... @dask.delayed def save(filename): ... def f(filenames): results = [] for filename in filenames: data = load(filename) data = process(data) result = save(data) return results dask.compute(f(filenames)) 

The first version only has one delayed task, and so cannot parallelize.

Every delayed task has an overhead of a few hundred microseconds. Usually this is ok, but it can become a problem if you apply dask.delayed too finely. In this case, it’s often best to break up your many tasks into batches or use one of the Dask collections to help you.

 Don’t Do # Too many tasks results = [] for x in range(10000000): y = dask.delayed(f)(x) results.append(y)  # Use collections import dask.bag as db b = db.from_sequence(range(10000000), npartitions=1000) b = b.map(f) ... 

Here we use dask.bag to automatically batch applying our function. We could also have constructed our own batching as follows

def batch(seq):
sub_results = []
for x in seq:
sub_results.append(f(x))
return sub_results

batches = []
for i in range(0, 10000000, 10000):
result_batch = dask.delayed(batch)(range(i, i + 10000))
batches.append(result_batch)


Here we construct batches where each delayed function call computes for many data points from the original input.

## Avoid calling delayed within delayed functions¶

Often, if you are new to using Dask delayed, you place dask.delayed calls everywhere and hope for the best. While this may actually work, it’s usually slow and results in hard-to-understand solutions.

Usually you never call dask.delayed within dask.delayed functions.

 Don’t Do # Delayed function calls delayed @dask.delayed def process_all(L): result = [] for x in L: y = dask.delayed(f)(x) result.append(y) return result  # Normal function calls delayed def process_all(L): result = [] for x in L: y = dask.delayed(f)(x) result.append(y) return result 

Because the normal function only does delayed work it is very fast and so there is no reason to delay it.

When you place a Dask array or Dask DataFrame into a delayed call, that function will receive the NumPy or Pandas equivalent. Beware that if your array is large, then this might crash your workers.

Instead, it’s more common to use methods like da.map_blocks

 Don’t Do # Call delayed functions on Dask collections import dask.dataframe as dd df = dd.read_csv('/path/to/*.csv') dask.delayed(train)(df)  # Use mapping methods if applicable import dask.dataframe as dd df = dd.read_csv('/path/to/*.csv') df.map_partitions(train) 

Alternatively, if the procedure doesn’t fit into a mapping, you can always turn your arrays or dataframes into many delayed objects, for example

partitions = df.to_delayed()
for part in partitions]


However, if you don’t mind turning your Dask array/DataFrame into a single chunk, then this is ok.

dask.delayed(train)(..., y=df.sum())


## Avoid repeatedly putting large inputs into delayed calls¶

Every time you pass a concrete result (anything that isn’t delayed) Dask will hash it by default to give it a name. This is fairly fast (around 500 MB/s) but can be slow if you do it over and over again. Instead, it is better to delay your data as well.

This is especially important when using a distributed cluster to avoid sending your data separately for each function call.

 Don’t Do x = np.array(...) # some large array results = [dask.delayed(train)(x, i) for i in range(1000)]  x = np.array(...) # some large array x = dask.delayed(x) # delay the data once results = [dask.delayed(train)(x, i) for i in range(1000)] 

Every call to dask.delayed(train)(x, ...) has to hash the NumPy array x, which slows things down.