Refactoring for Scalable Python Code With Pandas | by Charlie Shelbourne | May, 2022

Python design sample for writing scalable data-wrangling pipelines

Photograph by Siora Photography on Unsplash

A number of the beauties of Python are its flexibility and ease. Nonetheless, these talents are a double-edged sword. If you don’t put the work in early on, to design reusable, manageable, and testable code, you’ll run into progress points as your codebase scales.

When utilizing Python’s Pandas module, it’s straightforward to maneuver away from an object-oriented type of coding. A typical pitfall is to write down fast code, that turns into laborious to check and messy to scale.

This put up reveals a design sample for reusable and low upkeep code when data-wrangling with Pandas.

  1. Construct a metrics pipeline with Pandas

2. Refactor pipeline to be simply extendable and testable

🐍 Codebase discovered here.

We’ll be utilizing a free information set from Kaggle containing: “An entire checklist of unicorn firms on this planet.”


“A unicorn firm, or unicorn startup, is a personal firm with a valuation over $1 billion. As of March 2022, there are 1,000 unicorns around the globe. In style former unicorns embody Airbnb, Fb and Google. Variants embody a decacorn, valued at over $10 billion, and a hectocorn, valued at over $100 billion.”

📊 Kaggle Unicorn Corporations dataset discovered here.

We’ll construct 3 tables of Unicorn statistics utilizing Pandas.

  1. Nation-level metrics
  2. Nation-level time collection metrics
  3. Investor metrics

🏆 Successful Mixture of Pandas Strategies

We can be utilizing a mix of Pandas strategies that breeze information manipulation and hold our pipeline wanting clear.

🌍 Nation Degree Stats

Firstly, we are going to calculate the depend of Unicorns per nation and the common Unicorn valuation (in billions) per nation.


📈 Instance Plot utilizing country_stats Output Desk

The plot under is one instance of a plot made with our country_stats desk. We will shortly see the US main the world within the whole variety of Unicorn firms.

⏳ Nation Degree Time Collection

For these metrics, we group by nation and date_joined columns, to depend the variety of Unicorns over time and sum the valuations.

💡 Word I beforehand sorted the dataframe by date_joined.


➕ Cumulative Time Collection

To date we have now solely generated time-series metrics on the cut-off date. Nonetheless, it’s simpler on the attention, to view the cumulative sum over time.

These steps take the generated time_series desk and use an increasing window to calculate a cumulative sum.


📈 Instance Plot Utilizing time_series Output Desk

The time collection plot under is made with our cumulative outcomes, for the variety of Unicorns per nation. We will see since 2020–2021 the US has attain a trajectory for producing Unicorns that would not be matched by China. While, the India and UK could also be simply starting their development phases.

🧑‍💼 Investor Stats

Producing investor metrics is extra complicated. Every firm’s buyers are saved as a coma-separated string.

For instance, “Avant” have select_investors “RRE Ventures, Tiger International, August Capital”.

We need to reuse the identical code format, as with the country-level metrics, to utilize the pandas.DataFrame.groupby methodology. This can assist us refactor afterward.

instance output

🏗 Un-pivoting Buyers

Un-pivoting is vital to this design sample, as we need to make use of the groupby methodology on particular person buyers.

Utilizing pandas.DataFrame.explode, we generate a further column for particular person buyers. Word that we now have a number of rows per firm in our desk.

💡 Right here I’ve used explode to un-pivot. One other methodology to take a look at is melt.


The following step, is to generate easy investor stats, firm depend and valuation of the businesses, in every investor’s portfolio.

📈 Instance Plot utilizing investors_stats

The histogram under reveals the distribution of buyers by the entire variety of Unicorns of their portfolio. We see power-law sort distribution the place most buyers have just one unicorn, while few have invested in lots of. Any such distribution may also be present in populational financial wealth and social networks.

To date we have now a metrics pipeline that look pretty neat, however we’re solely producing a complete of 8 metrics. If we have been to increase this to twenty–30 metics, our script would begin to see loads of repetition.

The format of our code to this point is an easy python script. Thus our code can’t be remoted and unit-tested.

Our solely testing possibility is to run the whole script and assess the output, in an end-to-end type check. This isn’t nice because it may take a very long time to run.

🚪Open-close Precept

Open for extension however closed for modification.”

Refactor our code to comply with the open-close precept as very best.

  1. Transfer metrics features to a Metrics class and enlist the use pandas.DataFrame.apply methodology.
  2. Take away the repeat calls to pandas.DataFrame.groupby, with a generate_metrics perform, utilizing Python’s inbuilt getattr perform.
  3. Create a metrics config file, which is handed to our generate_metrics perform, with the meta-data required to generate our metrics.

💡 Take a look at SOLID design rules here.

🧑‍🎓 Metrics Class

By shifting our metrics to a category, we are able to isolate the metrics, and construct unit assessments for every metric. Through the use of the Pandas.DataFrame.apply methodology, we are able to add personalised metrics, and leverage different python packages that aren’t included in Pandas.

Following the open-close precept, if we needed so as to add metrics we’d create a brand new class that inherits our class Metrics.

📂 Config File

The config file has a listing of metrics for every desk we need to generate. If we need to add or take away metrics or change naming and so forth, we merely change the config file. This fashion we’re not modifying our codebase itself.

⚙️ Generate Metrics Capabilities

This perform takes in our Unicorn information, an occasion of our Metrics class, and our metrics config, and returns a metrics dataframe.


  1. Makes use of getattr to create a pandas.DataFrame.groupby.apply object
  2. Makes use of getattr to create a Metrics class methodology object (e.g. Metrics.depend)
  3. Calls the pandas.DataFrame.groupby.apply object passing the Metrics methodology object

🪄 Refactored Pipeline

Lastly, we arrive at our refactored pipeline. We will simply add metrics to present tables by defining new metrics lessons and including them to our config file.

  • Standardise code in our pipeline to make use of pandas.DataFrame.groupby.apply, to run transformations on our information.
  • Un-pivot information to offer distinctive rows (to be used with pandas.DataFrame.groupby).
  • Host the metrics we want to generate in a category and move them to pandas.DataFrame.groupby.apply.
  • Use Python’s built-in perform getattr, and a metadata dictionary, to loop by means of our metrics, moderately than repeat calls to pandas.DataFrame.groupby.apply.


  • Code is modular and extra manageable for scaling up our metrics.
  • A config file give us extra flexiblility so as to add and takeout metrics with out touching code.
  • Simpler to check performance of our code because it has been remoted.


  • Extra code to keep up for a small pipeline.
  • Readability of our code has been decreased.
  • More durable to debug for a small pipeline.

🐍 Codebase discovered here.

📊 Kaggle Unicorn Corporations dataset discovered here.

Third-party libraries:

Wish to Join?I am simply beginning out my running a blog journey. Let's Join on Twitter!

More Posts