flonb
lets you develop data pipelines (a.k.a task graphs) with very rapid "modify -> run -> inspect" iteration loops.
It lets you:
- Label each run of the pipeline with a human-readable list of options. Easy to see what you did to the data, easy to reproduce.
- Cache results automagically based on the labels. Change an option, add a new option -
flonb
will manage the cache without extra boilerplate! - Parallel execution of your pipeline.
flonb
pipelines are interoperable with dask schedulers.
So you can spend more time focussing on your logic, and less time on error-prone boilerplate code.
$ pip install flonb
Note this also installs dask.core
. If you require dask extensions (e.g. dask.bag, dask.dataframe, dask.delayed, etc...) then make sure to install them yourself. It can be a bit unclear with dask
what extensions you have installed! https://docs.dask.org/en/latest/install.html
Here is a simple pipeline that counts the number of times a word appears in a block of text.
Define a Task Graph using:
flonb.task_func
decorator for each processing step.flonb.Dep
as the default parameter value to tellflonb
which tasks to use as inputs.
import flonb
@flonb.task_func
def parse_text(normalise):
text = (
"Badger badger badger Mushroom mushroom "
"Badger badger badger Mushroom mushroom "
"Badger badger badger SNAKE its a SNAKE "
)
words = text.split(" ")
if normalise:
words = [w.lower() for w in words]
return words
@flonb.task_func
def word_count(word, text=flonb.Dep(parse_text)):
return sum([w == word for w in text])
Run the Task Graph using the .compute
method that @flonb.task_func
adds to your function.
Note that we supply all the options that specify the entire Task Graph. Even though the function word_count
does not explicitly require the option normalise
, its dependency parse_text
does.
word_count.compute(normalise=False, word="badger")
6
You can still run your pipeline by manually calling the functions the usual way. This can be handy for debugging:
text = parse_text(normalise=False)
word_count(word="badger", text=text)
6
Change the options, and everything works like you'd expect.
word_count.compute(normalise=True, word="badger")
9
Importantly, you can use a single option in multiple places in the same pipeline:
@flonb.task_func
def summary_str(word, normalise, count=flonb.Dep(word_count)):
return f"word={word}, normalise={normalise} -> count={count}"
print(summary_str.compute(word="mushroom", normalise=True))
word=mushroom, normalise=True -> count=4
A typical workflow using flonb
is:
- Run pipeline with different options to see the impacts on your data.
- Add new options and/or steps.
- Repeat.
flonb
is especially useful because you can add new options early in the Task Graph without having to manually propagate the option through every function in the pipeline!
We can visualize the graph using dask[complete]
+ graphviz
. You can install these with conda:
$ conda install dask python-graphviz
See https://docs.dask.org/en/latest/graphviz.html for other installation options.
import dask
graph, key = word_count.graph_and_key(normalise=False, word="badger")
dask.visualize(graph)
Oftentimes you will want to combine multiple computations with different versions of an option.
You can do this by supplying a list
to flonb.Dep
and using the .partial(**options)
method.
@flonb.task_func
def count_many_words(
counts=flonb.Dep(
[word_count.partial(word="badger"), word_count.partial(word="mushroom")]
)
):
return counts # so we can see how flonb substitutes the counts flonb.Dep parameter
count_many_words.compute(normalise=False)
[6, 2]
Lists can be nested, and non-flonb.Task
objects are ignored:
@flonb.task_func
def count_many_words_heterogenous_nested_list(
counts=flonb.Dep(
[
[word, word_count.partial(word=word)]
for word in ["badger", "mushroom", "snake", "cow"]
]
)
):
return counts # so we can see how flonb substitutes the counts flonb.Dep parameter
count_many_words_heterogenous_nested_list.compute(normalise=True)
[['badger', 9], ['mushroom', 4], ['snake', 2], ['cow', 0]]
flonb
uses the options supplied to a task to uniquely identify it, and can automagically cache on disk the results using this identifier.
If a task has been cached, flonb
will fetch the cached result instead of computing not only the task but also any of its dependencies.
Warning: If you invalidate your cache by changing what an option means (e.g. fixing a bug in your code), make sure to delete your cache!
flonb.set_cache_dir("/tmp")
@flonb.task_func(cache_disk=True)
def power(x, y):
print(f"computing {x}**{y}") # side effect - doesn't happen when cached result called
return x**y
# only prints once
power.compute(x=3, y=3)
power.compute(x=3, y=3)
computing 3**3
# each unique set of options labels the cache
power.compute(x=3, y=2)
power.compute(x=2, y=3)
computing 3**2
computing 2**3
Sometimes you want to know the value of an option before you resolve the dependencies. flonb.DynamicDep
has your back here.
@flonb.task_func
def add(x, y):
return x + y
@flonb.task_func
def multiply(a, b):
return a * b
@flonb.task_func
def multiply_or_add(
result=flonb.DynamicDep(lambda mode: {"add": add, "multiply": multiply}[mode])
):
return result
Here the option 'mode'
determines what the task graph looks like, and even what options are available:
multiply_or_add.compute(mode="multiply", a=4, b=3)
12
multiply_or_add.compute(mode="add", x=4, y=3)
7
flonb
implements the dask
Task Graph specification.
- call
.graph_and_key()
on your task. - feed into
dask
as required.
import dask.multiprocessing
import time
@flonb.task_func
def add_one_in_one_second(x):
time.sleep(1) # artifically slow down the task
result = x + 1
return result
@flonb.task_func
def do_sums(data=flonb.Dep([add_one_in_one_second.partial(x=x) for x in range(10)])):
return data
graph, key = do_sums.graph_and_key()
# single process
tic = time.time()
dask.get(graph, key)
print(f"Single-process execution took {time.time() - tic:.2f} seconds.")
print("------------")
# multiple process
tic = time.time()
dask.multiprocessing.get(graph, key, num_workers=4)
print(f"Multiproccessing execution took {time.time() - tic:.2f} seconds.")
Single-process execution took 10.01 seconds.
------------
Multiproccessing execution took 3.30 seconds.
There are many high quality frameworks that let you build and run task graphs. flonb
is lightweight and easy to experiment with, but make sure to check out others if you want to delve further into your options. Here are some suggestions:
dask
luigi
airflow
A big thanks to Solcast (https://solcast.com) for supporting the development of flonb
.