Scalpl provides a lightweight wrapper that helps you to operate
on nested dictionaries seamlessly through the built-in dict
API, by using dot-separated string keys.
It's not a drop-in replacement for your dictionnaries, just syntactic
sugar to avoid this['annoying']['kind']['of']['things']
and
prefer['a.different.approach']
.
No conversion cost, a thin computation overhead: that's Scalpl in a nutshell.
There are a lot of good libraries to operate on nested dictionaries, such as Addict or Box , but if you give Scalpl a try, you will find it:
- 🚀 Powerful as the standard dict API
- ⚡ Lightweight
- 👌 Well tested
Scalpl is a Python3 library that you can install via pip
pip3 install scalpl
Scalpl provides a simple class named Cut that wraps around your dictionary
and handles operations on nested dict
and that can cut accross list
item.
This wrapper strictly follows the standard dict
API, which
means you can operate seamlessly on dict
,
collections.defaultdict
or collections.OrderedDict
by using their methods
with dot-separated keys.
Let's see what it looks like with an example ! 👇
from scalpl import Cut
data = {
'pokemon': [
{
'name': 'Bulbasaur',
'type': ['Grass', 'Poison'],
'category': 'Seed',
'ability': 'Overgrow'
},
{
'name': 'Charmander',
'type': 'Fire',
'category': 'Lizard',
'ability': 'Blaze',
},
{
'name': 'Squirtle',
'type': 'Water',
'category': 'Tiny Turtle',
'ability': 'Torrent',
}
],
'trainers': [
{
'name': 'Ash',
'hometown': 'Pallet Town'
}
]
}
# Just wrap your data, and you're ready to go deeper !
proxy = Cut(data)
You can use the built-in dict
API to access its values.
proxy['pokemon[0].name']
# 'Bulbasaur'
proxy.get('pokemon[1].sex', 'Unknown')
# 'Unknown'
'trainers[0].hometown' in proxy
# True
By default, Scalpl uses dot as a key separator, but you are free to use a different character that better suits your needs.
# You just have to provide one when you wrap your data.
proxy = Cut(data, sep='->')
# Yarrr!
proxy['pokemon[0]->name']
You can also easily create or update any key/value pair.
proxy['pokemon[1].weaknesses'] = ['Ground', 'Rock', 'Water']
proxy['pokemon[1].weaknesses']
# ['Ground', 'Rock', 'Water']
proxy.update({
'trainers[0].region': 'Kanto',
})
Following its purpose in the standard API, the setdefault method allows you to create any missing dictionary when you try to access a nested key.
proxy.setdefault('pokemon[2].moves.Scratch.power', 40)
# 40
And it is still possible to iterate over your data.
proxy.items()
# [('pokemon', [...]), ('trainers', [...])]
proxy.keys()
# ['pokemon', 'trainers']
proxy.values()
# [[...], [...]]
By the way, if you have to operate on a list of dictionaries, the
Cut.all
method is what you are looking for.
# Let's teach these pokemon some sick moves !
for pokemon in proxy.all('pokemon'):
pokemon.setdefault('moves.Scratch.power', 40)
Also, you can remove a specific or an arbitrary key/value pair.
proxy.pop('pokemon[0].category')
# 'Seed'
proxy.popitem()
# ('trainers', [...])
del proxy['pokemon[1].type']
Because Scalpl is only a wrapper around your data, it means you can get it back at will without any conversion cost. If you use an external API that operates on dictionary, it will just work.
import json
json.dumps(proxy.data)
# "{'pokemon': [...]}"
Finally, you can retrieve a shallow copy of the inner dictionary or remove all keys.
shallow_copy = proxy.copy()
proxy.clear()
This humble benchmark is an attempt to give you an overview of the performance
of Scalpl compared to Addict,
Box and the built-in dict
.
It will summarize the number of operations per second that each library is able to perform on a portion of the JSON dump of the Python subreddit main page.
You can run this benchmark on your machine with the following command:
python3 ./benchmarks/performance_comparison.py
Here are the results obtained on an Intel Core i5-7500U CPU (2.50GHz) with Python 3.6.4.
Addict 2.2.1:
instantiate:-------- 271,132 ops per second. get:---------------- 276,090 ops per second. get through list:--- 293,773 ops per second. set:---------------- 300,324 ops per second. set through list:--- 282,149 ops per second.
Box 3.4.2:
instantiate:--------- 4,093,439 ops per second. get:----------------- 957,069 ops per second. get through list:---- 164,013 ops per second. set:----------------- 900,466 ops per second. set through list:---- 165,522 ops per second.
Scalpl latest:
instantiate:-------- 183,879,865 ops per second. get:---------------- 14,941,355 ops per second. get through list:--- 14,175,349 ops per second. set:---------------- 11,320,968 ops per second. set through list:--- 11,956,001 ops per second.
dict:
instantiate:--------- 37,816,714 ops per second. get:----------------- 84,317,032 ops per second. get through list:---- 62,480,474 ops per second. set:----------------- 146,484,375 ops per second. set through list :--- 122,473,974 ops per second.
As a conclusion and despite being an order of magniture slower than the built-in
dict
, Scalpl is faster than Box and Addict by an order of magnitude for any operations.
Besides, the gap increase in favor of Scalpl when wrapping large dictionaries.
Keeping in mind that this benchmark may vary depending on your use-case, it is very unlikely that Scalpl will become a bottleneck of your application.
- What if my keys contain dots ?
If your keys contain a lot of dots, you should use an other key separator when wrapping your data:
proxy = Cut(data, sep='->') proxy['computer->network->127.0.0.1']
Otherwise, split your key in two part:
proxy = Cut(data) proxy['computer.network']['127.0.0.1']
What if my keys contain spaces ?:
proxy = Cut(data) proxy['it works perfectly'] = 'fine'
Contributions are welcomed and anyone can feel free to submit a patch, report a bug or ask for a feature. Please open an issue first in order to encourage and keep tracks of potential discussions ✍️
Scalpl is released into the Public Domain. 🎉
Ps: If we meet some day, and you think this small stuff worths it, you can give me a beer, a coffee or a high-five in return: I would be really happy to share a moment with you ! 🍻