Python is not a great language for data science
blog.genesmindsmachines.com
blog.genesmindsmachines.com
Most of this is not about Python, it’s about matplotlib. If you want the admittedly very thoughtful design of ggplot in Python, use plotnine
> I would consider the R code to be slightly easier to read (notice how many quotes and brackets the Python code needs)
This isn’t about Python, it’s about the tidyverse. The reason you can use this simpler syntax in R is because it’s non-standard-evaluation allows packages to extend the syntax in a way Python does not expose: http://adv-r.had.co.nz/Computing-on-the-language.html
They are of course now abandoning this idea.
I honestly think that was a coincidence. Perl and Ruby had other disadvantages, Python won despite having bad package management and a bloated standard library, not because of it.
If python had been lean and needed packages to do anything useful, while still having a packaging nightmare, it would have been unusable
Most of the python users are not able nor aware of venv, uv, pip and all of that.
The irony here: We are talking about data science. 98% of "data science" Python projects start by creating a virtual env and adding Pandas and NumPy which have numerous (really: squillions of) dependencies outside the foundation library.
pandas==2.3.3
├── numpy [required: >=1.22.4, installed: 2.2.6]
├── python-dateutil [required: >=2.8.2, installed: 2.9.0.post0]
│ └── six [required: >=1.5, installed: 1.17.0]
├── pytz [required: >=2020.1, installed: 2025.2]
└── tzdata [required: >=2022.7, installed: 2025.2]
e.g.
https://github.com/numpy/numpy/blob/main/.gitmodules (some source code requirements)
https://github.com/numpy/numpy/tree/main/requirements (mostly build/ci/... requirements)
...NumPy will fall back to internal and very slow BLAS and LAPACK implementations if your system does not have a better one, but assuming you're using NumPy for its performance and not just the convenience of adding array programming features to Python, you're really gonna want better ones, and what that is heavily depends on the computer you're using.
This isn't really a Python thing, though. It's a hard problem to solve with any kind of scientific computing. If you insist on using a dynamic interpreted language, which you probably have to do for exploratory interactive analysis, and you still need speed over large datasets, you're gonna need to have a native FFI and link against native libraries. Thanks to standardization, you'll have many choices and which is fastest depends heavily on your hardware setup.
> it’s non-standard-evaluation allows packages to extend the syntax in a way Python does not expose
Well this is a fundamental difference between Python and R.
Oh god no, do people write R like that, pipes at the end? Elixir style pipe-operators at the beginning is the way.
And if you really wanted to "improve" readability by confusing arguments/functions/vars just to omit quotes, python can do that, you'll just need a wrapper object and getattr hacks to get from `my_magic_strings.foo` -> `'foo'`. As for the brackets.. ok that's a legitimate improvement, but again not language related, it's library API design for function sigs.
(-> (gather-some-data)
(map 'Vector #'some-functor)
(filter #'some-predicate)
(reduce #'some-gatherer))
->
gather-some-data
map 'Vector #'some-functor
filter #'some-predicate
reduce #'some-gathererIn my limited experience, Using R feels like to using JavaScript in the browser: it's a platform heavily focused on advanced, feature-rich objects (such as DataFrames and specialized plot objects). but you could also just build almost anything with it.
And I'm still waiting for your examples of "established R packages with bugs caused by R weirdness".
R is not actually competing with those languages. R's design purpose is different. it is a general purpose computational language for scientists. There are FFIs (Foreign Function Interfaces) for all those languages.
R-Kotlin-Jave: https://journal.r-project.org/articles/RJ-2018-066/ R-Rust: https://cran.r-project.org/web/packages/using_rust.html R-C# : https://github.com/Open-Systems-Pharmacology/rsharp/
R is supporting C integration natively anyhow (see Chambers's book.
Regarding VB reference. VB was used in finance a lot to do some advanced maths. just a side remark.
In R it's often that things for which there's a ready made libraries and recipes are easy, but when those don't exist, things become extremely hard. And the usual approach is that if something is not easy with a library recipe, it just is not done.
R: easy things are hard, hard things are easy.
So it actually is about Python vs R.
That said, while this kind of non-standard evaluation is nice when working interactively on the command line, I don't think it's that relevant when writing code for more elaborated analyses. In that context, I'd actually see this as a disadvantage of R because you suddenly have to jump through loops to make trivial things work with that non-standard evaluation.
Yeah, this was so very very painful. I once ended up maintaining a library that basically used all the different NSE approaches, which was not very much fun at all.
If you step back, it's kind of weird that there's no mainstream programming language that has tables as first class citizens. Instead, we're stuck learning multiple APIs (polars, pandas) which are effectively programming languages for tables.
R is perhaps the closest, because it has data.frame as a 'first class citizen', but most people don't seem to use it, and use e.g. tibbles from dplyr instead.
The root cause seems to be that we still haven't figured out the best language to use to manipulate tabular data yet (i.e. the way of expressing this). It feels like there's been some convergence on some common ideas. Polars is kindof similar to dplyr. But no standard, except perhaps SQL.
FWIW, I agree that Python is not great, but I think it's also true R is not great. I don't agree with the specific comparisons in the piece.
(and yes there's special language support for LINQ so it counts as "part of the language" rather than "a library")
SELECT thing1, thing2 FROM things WHERE thing2 != 2;
val thingMap = things.map { it.thing2 to it.thing2 }.filter { it.thing2 !=2 }
Then you've got distinct(), sorting methods, take/drop for limits, count/sumOf/average/minOf/maxOf.
There are set operations, so you can do unions and differences, check for presence, etc.
Joins are the hard part, but map() and some lambda work can pull it off.
What is a paragraph but an array of sentences? What is a sentence but an array of words? What's a word but an array of letters? You can do this all the way down. Eventually you need to assign meaning to things, and when you do, it helps to know what the thing actually is, specifically, because an array of structs can be many things that aren't a table.
Nitpicking aside, a nice library for doing “table stuff” without “the whole ass big table framework” would be nice.
It’s not hard to roll this stuff by hand, but again, a nicer way wouldn’t be bad.
most procs use tables as both input and output, and you better hope the tables have the correct columns.
you want a loop? you either get an implicit loop over rows in a table, write something using syscalls on each row in a table, or you're writing macros (all text).
Simplifying a lot, R is heavily inspired by Scheme, with some lazy evaluation added on top. Julia is another take at the design space first explored by Dylan.
You're forgetting R's data.table, https://cran.r-project.org/web/packages/data.table/vignettes...,
which is amazing. Tibbles only wins because they fought the docs/onboarding battle better, and dplyr ended up getting industry buy-in.
But you can have the best of both worlds with https://dtplyr.tidyverse.org/, using data.table's performance improvements with dplyr syntax.
Relevant to the author's point, Python is pretty poor for this kind of thing. Pandas is a perf mess. Polars, duckdb, dask etc, are fine perhaps for production data pipelines but quite verbose and persnickety for rapid iteration. If you put a gun to my head and told me to find some nuggets of insight in some massive flat files, I would ask for an RStudio cloud instance + data.table hosted on a VM with 256GB+ of RAM.
I suspect that in the fullness of time, mainstream languages will eventually fully incorporate tabular programming in much the same way they have slowly absorbed a variety of idioms traditionally seen as part of functional programming, like map/filter/reduce on collections.
[0] https://en.wikipedia.org/wiki/Q_(programming_language_from_K...
[1] https://ryelang.org/blog/posts/comparing_tables_to_python/
Matlab has them, in fact it has multiple competing concepts of it.
My problem with APL is 1.) the syntax is less amazing at other more mundane stuff, and 2.) the only production worthy versions are all commercial. I'm not creating something that requires me to pay for a development license as well as distribution royalties.
Everyone in R uses data.frame because tibble (and data.table) inherits from data.frame. This means that "first class" (base R) functions work directly on tibble/data.table. It also makes it trivial to convert between tibble, data.table, and data.frames.
The languages devs use are largely Algol derived. Algol is a language that was used to express algorithms, which were largely abstractions over Turing machines, which are based around an infinite 1D tape of memory. This model of 1D memory was built into early computers, and early operating systems and early languages. We call it "mechanical sympathy".
Meanwhile, other languages at the same time were invented that weren't tied so closely to the machine, but were more for the purpose of doing science and math. They didn't care as much about this 1D view of the world. Early languages like Fortran and Matlab had notions of 2D data matrices because math and science had notions of 2D data matrices. Languages like C were happy to support these things by using an array of pointers because that mapped nicely to their data model.
The same thing can be said for 1-based and 0-based indexing -- languages like Matlab, R, and Excel are 1-based because that's how people index tables; whereas languages like C and Java are 0-based because that's how people index memory.
Also, the linear memory model itself is not really only because of Algol/Turing machines/theoretical CS/"early" hardware and mechanical sympathy. DRAM has rows & columns internally, but byte addressability leads to hiding that from HW client systems (unless someone is doing a rowhammer attack or something). More random access than tape rewind/fast forward is indeed a huge deal, but I think the actual popularity of linearity just comes from its simplicity as an interface more than anything else. E.g.s, segmented x86 memory with near/far pointers was considered ugly relative to a big 32-bit address space and disk files and other allocation arenas have internally a large linear address/seek spaces. People just want to defer using >1 number until they really need to. People learn univariate-X before they learn multivariate-X where X could be calculus, statistics, etc., etc.
Soon as you start doing things like joins, it gets complicated but in theory you could do something like an API of an ORM to do most things. With using just operators you quickly run into the fact that you have to overload (abuse) operators or write a new language with different operator semantics:
orders * customers | (customers.id == orders.customer_id | orders.amount > Decimal(‘10.00’)
But it would be nice to have it built in so talking to a database would be a bit more native.
For reference, I think the same is true of Python, so it’s not like I’m a Perl wizard or something.
For one thing, SQL is not really meant to be dynamically constructed in SQL. But we often need to dynamically construct a query (for example customer applied several filters to the product listing). The SQL way to handle that would be to have a general purpose query with a thousand if/elses or stored procedures which I think takes it from “flows beautifully” to “oh god who wrote this?” Or you could just do string concatenation in a language that handles that well, like Python. Then wrap the whole thing in functions and objects and you get an ORM.
I still have not seen a language that incorporates anything like SQL into it that would allow for even basic ORM-like functionality.
This is why key-value stores are so popular, I think. They make you do more but with all local data (that is data in your memory not in the database server). SQL can do a lot but because we almost never represent a user object as just a tuple there is a fundamental impedance mismatch between an environment that only deals with tuples in tables and an environment that deals with objects of some kind. Something that can do both at once would be the ultimate. Maybe the way to look at it isn’t to bring the database into your application but to run the entirety of the application inside a database. Imagine if all your business logic could easily be encoded into stored procedures and all you had to do was expose endpoints to draw a UI for it. That might actually work (and I know there are some systems that try this but none are mainstream enough).
But ultimately data analysis is going beyond Python and R into the realm of Stan and PyMC3, probabilistic programming languages. It’s because we want to do nested integrals and those software ecosystems provide the best way to do it (among other probabilistic programming languages). They allow us to understand complex situations and make good / valuable decisions.
What tools are easily available in a language, by default, shape the pretty path, and by extension, the entire feel of the language. An example that we've largely come around on is key-value stores. Today, they're table stakes for a standard library. Go back to 90's, and the most popular languages at best treated them as second-class citizens, more like imported objects than something fundamental like arrays. Sure, you can implement a hash map in any language, or import some else's implementation, but oftentimes you'll instead end up with nightmarish, hopefully-synchronized arrays, because those are built-in, and ready at hand.
> There's a number of structures that I think are missing in our major programming languages. Tables are one. Matrices are another.
And all of those programmers are either using specialized languages, (suffering problems when they want to turn their program into a shitty web app, for example), or committing crimes against syntax like
rotation_matrix.matmul(vectorized_cat)
You don't even need such construction in most native applications, embedded systems, and OS kernel development.
And if you do robotics, the chances of encountering a matrix are very high.
Even through UE blueprints (assuming the most high level abstraction here) you will come across the need to perform calculations with matrices. While a lot is abstracted away, you still need to know about coordinate spaces, quirks around order of operations, etc.
Plus, plenty of third party projects have been incorporated into the Python standard library.
Graphs are a good example, as they are a large family of related structures. For example, are the edges undirected, directed, or something more exotic? Do the nodes/edges have identifiers and/or labels? Are all nodes/edges of the same type, or are there multiple types? Can you have duplicate edges between the same nodes? Does that depend on the types of the nodes/edges, or on the labels?
Then there would be more PEG horror stories. In addition, string and indices in regex processing are universal, while a parser is necessarily more framework-like, far more complex and doomed to be mismatched for many applications.
Mainly for programming control systems for robotics and aerospace applications
https://github.com/mech-lang/mech https://www.hytradboi.com/2022/i-tried-rubbing-a-database-on...
Not all features are implemented yet but we are getting there!
They are in q/kdb and it's glorious. Sql expressions are also first class citizens and it makes it very pleasant to write code
The problems that one might encounter in dealing with a 1m row table are quite different to a 1b row table, and a 1b row table is a rounding error compared to the problems that a 1t row table presents. A standard library needs to support these massive variations at least somewhat gracefully and that's not a trivial API surface to design.
Because they were created by before the need for it and maybe before their invention.
Manipulating numeric arrays and matrices in python is a bit clunky because it was not designed as a scientific computing language so they were added as library. It's much more integrated and natural to use in scientific computer languages such as matlab. However the reverse is also true: because matlab wasn't designed to do what python does, it's a bit clunkier to use outside scientific computing
There are clay tablets from ancient Sumeria that represent information using tables.
But on the other hand, that's doesn't mean SQL is ideal - far from it. When using DuckDB with Python, to make things more succinct, reusable and maintainable, I often fall into the pattern of writing Python functions that generate SQL strings.
But that hints at the drawbacks of SQL: it's mostly not composable as a language (compared to general purpose languages with first-class abstractions). DuckDB syntax does improve on this a little, but I think it's mostly fundamental to SQL. All I'm saying is that it feels like something better is possible.
The author's priorities are sensible, and indeed with that set of priorities, it makes sense to end up near R. However, they're not universal among data scientists. I've been a data scientist for eight years, and have found that this kind of plotting and dataframe wrangling is only part of the work. I find there is usually also some file juggling, parsing, and what the author calls "logistics". And R is terrible at logistics. It's also bad at writing maintainable software.
If you care more about logistics and maintenance, your conclusion is pushed towards Python - which still does okay in the dataframes department. If you're ALSO frequently concerned about speed, you're pushed towards Julia.
None of these are wrong priorities. I wish Julia was better at being R, but it isn't, and it's very hard to be both R and useful for general programming.
Edit: Oh, and I should mention: I also teach and supervise students, and I KEEP seeing students use pandas to solve non-table problems, like trying to represent a graph as a dataframe. Apparently some people are heavily drawn to use dataframes for everything - if you're one of those people, reevaluate your tools, but also, R is probably for you.
Except its not. Data science in python pretty much requires you to use numpy. So his example of mean/variance code is a dumb comparison. Numpy has mean and variance functions built in for arrays.
Even when using raw python in his example, some syntax can be condesed quite a bit:
groups = defaultdict(list) [groups[(row['species'], row['island'])].append(row['body_mass_g']) for row in filtered]
It takes the same amount of mental effort to learn python/numpy as it does with R. The difference is, the former allows you to integrate your code into any other applicaiton.
Even outside of Numpy, the stdlib has the statistics packages which provides mean, variance, population/sample standard deviation, and other statistics functions for normal iterables. The attempt to make Python out-of-the-box code look bad was either deliberately constructed to exaggerate the problems complained of, or was the product of a very convenient ignorance of the applicable parts of Python and its stdlib.
I'd say I'm 50/50 Python/R for exactly this reason: I write Python code on HPC or a server to parse many, many files, then I get some kind of MB-scale summary data I analyse locally in R.
R is not good at looping over hundreds of files in the gigabytes, Python is not good at making pretty insights from the summary. A tool for every task.
If you're doing data science all day, you should learn R, even if it's so weird at first (for somebody coming from a C-style language) that it seems way harder; R is made for the way statisticians work and think, not the way computer programmers work and think. If you're doing data science all day, you should start thinking and working like a statistician and working in R, and the fact that it seems to bend your mind is probably at least in part good, because a statistician needs to think differently than a programmer.
I work in python, though, almost all of the time.
groups = {}
for row in filtered:
key = (row['species'], row['island'])
if key not in groups:
groups[key] = []
groups[key].append(row['body_mass_g'])
groups = collections.defaultdict(list)
for row in filtered:
groups[(row['species'], row['island'])].append(row['body_mass_g'])
variance = sum((x - mean) ** 2 for x in values) / (n - 1)
std_dev = math.sqrt(variance)
std_dev = statistics.stddev(values)In the first instance, the original code is readable and tells me exactly what's what. In your example, you're sacrificing readability for being clever.
Clear code(even if verbose) is better than being clever.
defaultdict is ubiquitous in modern python, and is far from a complicated concept to grasp.
The difference between the examples is so trivial I'm not really sure why the parent comment felt compelled to complain.
That said, I'll change my mind here and agree on using std library, but I'd still have separate 'key' assignment here for more clarity.
Once you get used to a language's "quirks" (so long as they're considered idiomatic), they no longer feel quirky, and it's usually pretty quick.
I've definitely written some things that I came back to much later and had to relearn (which is somewhere between embarrassing and humbling).
It's also funny that one would write their own standard deviation function and include Bessel's correction. Usually if I'm manually re-implementing a standard deviation function it's because I'm afraid the implementors blindly applied the correction without considering whether or not it's actually meaningful for the given analysis. At the very least, the correct name for what's implemented there should really be `sample_std_dev`.
These days I run some big query on an OLAP database and download the results to parquet stored on the local disk of a cloud notebook VM and then mine it to bits with duckdb reading straight from these parquet files.
The notebooks end up with very clear SQL queries and results (most notebook servers support SQL cells with highlighting and completion etc), and small pockets of python cells for doing those corner case things that an imperative language makes easier.
So when I get to the bottom of the article where it shows the difference between Python and R, I'm screaming "wouldn't that look better in SQL?!" :)
Of course there's a bunch of loops and things; you're exposing what has to happen in both R and Python under the hood of all those packages.
It's pretty clear the post is focused on the context of work being done in an academic research lab. In that context I think most of the points are pretty valid, but most of the real world benefit I've experience from using Python is being able to work more closely with engineering (even on non-Python teams).
I shipped R code to a production environment once over my career and it felt incredibly fragile.
R is great for EDA, but really doesn't work well for iteratively building larger software projects. R is has a great package system, but it's not so great when you need abstraction in between.
BTW AI is not helping and in fact is leading to a generation of scientists who know how to write prompts, but do not understand the code those prompts generate or have the ability to peer review it.
Clojure, unlike lists in traditional Lisps, based on composable, unified abstraction for its collections, they are lazy by default and literal readable data structures, they are far easier to introspect and not so "opaque" compared to anything - not just CL (even Python), they are superb for dealing with heterogeneous data. Clojure's cohesive data manipulation story is where Common Lisp's lists-and-symbols just can't match.
Common Lisp has O[1] vectors, multidimensional arrays, hash-tables (what Clojure calls maps), structs, and objects. It has set operations too but it doesn't enforce membership uniqueness. It also has bignums, several sizes of floats, infinite-precision rationals, and complex numbers. Not to mention characters, strings, and logical operations on individual bits. The main difference from Clojure is that CL data structures are not immutable. But that's an orthogonal issue to the suggestion that CL doesn't contain a rich library of modern data structures.
Common Lisp has never been limited to "List Processing."
You're saying: "hash-tables (what Clojure calls maps)" not only inaccurate, you're hand-waving Clojure's core design philosophy (immutability, structural sharing, lazy sequences) as orthogonal. But those aren't cosmetic differences - they're the reason why Clojure's data structures are fundamentally better for data analysis. I think you're confusing "having equivalent data types" with "solving the same problem the same way"
The early tooling was also pretty dependent on Vim or Emacs. Maybe it's all easier now with VSCode or something like that.
The tooling story is also very solid - I use Emacs, but many of my friends and colleagues use IntelliJ, Vim, Sublime and VSCode, and some of them migrated to it from Atom.
Do you have a habit of referring to yourself in plural, or do you typically like to generalize things based on your personal experiences?
I personally know many Clojurists who never had problems you're describing - hundreds of people. Sure, that could be the case of survivorship bias, perhaps I just don't befriend people who struggled with getting into Clojure specifically in a way you're describing. But like they say: "Those who are willing to make the effort will find the solutions. Those who aren't will find the excuses."
Clojure undeniably had challenges in the past, and still has some today. But not the things you're talking about. This is literally not an exaggeration - it's as easy as installing Calva extention for VSCode - that's all one needs to mess around with Clojure.
I did point out that maybe things had changed a good bit (literally said maybe VSCode made that easier now as it has for other tools) and tried to make it clear that my experience was a bit dated.
As far as excuses go, I don't see how that's relevant. I just pointed out I had issues with a steep learning curve when I was seriously considering it many years ago along with other languages that are hosted on the JVM (Scala, Kotlin) or .NET (F#). Nothing against those languages, but all the tutorials and even many of the books at the time would frequently borrow from the host language in weird ways. Like I'd have to use some random Java library and when it didn't work, had no idea how to troubleshoot why it wasn't there and I didn't want to have to go learn Java first.
I own at least two books on F# and talked with some prominent authors personally and they admitted it was really geared towards intermediate or greater C# users who wanted to move over to functional programming. I could have stuck with it, but decided to stick with other tools.
Clojure certainly is nice and I wanted to take advantage of it...it just ended up not being as ergonomic for my needs as I had hoped.
If you want to use Java you also don't really need to know Java beyond "you create instances of classes and call methods on them". I really don't want to learn a dinosaur like Java, but having access to the universe of Java libs has saved me many times. It's super fun and nice to use and poke around mature Java libs interactively with a REPL :)
All that said I'd have no idea how to write even a helloworld in Java
PS: Agreed on Emacs. I love Emacs.. but it's for turbo nerds. Having to learn Emacs and Clojure in parallel was a crazy barrier. (and no, Emacs is not as easy people make it out to be)
While I agree with you in principal this also leads to what I call the "VB Effect". Back in the day VB was taught at every school as part of the standard curriculum. This made every kid a 'computer wizz'. I have had to fix many a legacy codebase that was started by someone's nephew the whizz kid.
> Python is pretty good for deep learning. There’s a reason PyTorch is the industry standard. When I’m talking about data science here, I’m specifically excluding deep learning.
I've written very little deep learning code over my career, but made very frequent use of the GPU and differentiable programming for non-deep learning specific tasks. In general Python is much easier to write quantitative programs that make use of the hardware, and you have a lot more options when your problem doesn't fit into RAM.
> I have been running a research lab in computational biology for over two decades.
I've been working nearly exclusively in industry for these two decades and a major reason I find Python just better is it's much, much easier to interface with other parts of engineering when you're a using truly general purpose PL. I've actually never worked for a pure Python shop, but it's generally much easier to get production ML/DS solutions into prod when working with Python.
> Data science as I define it here involves a lot of interactive exploration of data and quick one-off analyses or experiments
This re-iterates the previous difference. In my experience I would call this "step one" in all my DS related work. The first step is to understand the problem and de-risk. But the vast majority of code and work is related to delivering a scalable product.
You can say that's not part of "data science", but if you did you'd have a hard time finding a job on most of the teams I've worked on.
All that said, my R vs Python experience has boiled down to: If your end result is a PDF report, R is superior. If your end result is shipping a product, then Python is superior. And my experience has been that, outside of university labs, there aren't a lot of jobs out there for DS folks who only want to deliver PDFs.
That’s a bad argument or a naive and obvious one; depending on how you look at it.
Python wasn’t designed for Data Science. It is not a DSL for it. MATLAB was arguably designed for scientific computing, and yet it’s the most disliked language in the StackOverflow liked/disliked index.
Here’s a different way to look at it. A good programming language is like the weather in a city. I would love to live somewhere where it’s 72F/23C all year round. But if it’s in the middle of nowhere and I’ve got no friends to hang out with, would I? I don’t think so.
FWIW, Python is like Sweden or Finland, with shitty weather for 6 months of the year yet thriving against all odds.
PS: I think the article’s topic is a bit click-batey (not a particularly useful discussion) because it’s polarizing and no one will be 100% right about it. It’s perhaps best thought of as an opinion piece.
[1] https://link.springer.com/article/10.1007/s11336-017-9581-x
using TidierData, DataFrames
using PalmerPenguins: load
penguins = load()
@chain penguins begin
DataFrame
@drop_missing(body_mass_g)
@group_by(species, island)
@summarize(
body_weight_mean =
mean(body_mass_g),
body_weight_std =
std(body_mass_g)
)
show(_, allrows=true)
end>The core problems I see with Python as a language for data science are call-by-reference semantics, lack of built-in concepts of missing values, lack of built-in vectorization, and lack of non-standard evaluation.
[Data Preparation] --> [Data Analysis] --> [Result Preparation]
Neither Python or R does a good job at all of these.
The original article seems to focus on challenges in using Python for data preparation/processing, mostly pointing out challenges with Pandas and "raw" Python code for data processing.
This could be solved by switching to something like duckdb and SQL to process data.
As far as data analysis, both Python and R have their own niches, depending on field. Similarly, there are other specialized languages (e.g., SAS, Matlab) that are still used for domain-specific applications.
I personally find result preparation somewhat difficult in both Python and R. Stargazer is ok for exporting regression tables but it's not really that great. Graphing is probably better in R within the ggplot universe (I'm aware of the python port).
Python doesn't need to be the best at any one thing; it just has to be serviceable for a lot of things. You can take someone who has expertise in a completely different domain in software (web dev, devops, sysadmin, etc.) and introduce them to the data science domain without making them learn an entirely new language and toolchain.
It's used in data science because it's used in data science.
Use whatever you want on your one off personal projects but use something more non-data science friendly if you ever want your model to run directly in a production workflow.
Productionizing R models is quite painful. The normal way is to just rewrite it not in R.
If you write it in R and then rewrite it in C (better: rewrite it in English with the R as helpful annotations, then have someone else rewrite it in C), at least there is some chance you've thought about the abstractions and operations that are actually necessary for your problem.
And it got this unprecedented level of support because right from the start it made its focus clear syntax and (perceived) simplicity.
There is also a sort of cumulative effect from being nice for algorithmic work.
Guido's long-term strategy won over numerous other strong candidates for this role.
1. data scientists aren't programmers, so why do they need a programming language? the tools they should be using don't exist. they'd need programmers to make them, and all we have to offer is... more programming languages.
2. the giant problem at the heart of modern software: the most important feature of a modern programming language is being easy to read and write. this feature is conspicuously absent from most important languages.
they're trapped. they can't do what they need without a programming language but there are only a handful they can possibly use. the real reason python ended up with such good library support is they never really had a choice.
And then Python went into a self-reinforcing loop, with scientific community coming up with more and more ways to improve Python support for the kind of interactive work that was required for data analysis. Think ipython -> jupyter -> jupyter forks and other python-centric notebook systems.
So when data analysis evolved into data science and machine learning, gpu-first library vendors already faced a crowd of people knowing python.
It is crazy how right now one can utilize 100s of gpus through these bits of dirty python wrapped in json.
You need to get the data from somewhere. Do you need to scrape that because Python is okay at scraping? Oh, after its scraped, we looked at it and it's in ObtuseBinaryFormat0.0.LOL.Beta and, what do you know, somebody wrote a converter for that for Python. And we need to clean all the broken entries out of that and Python is decent at that. etc.
The trick is that while Python may or may not be anybody's first choice for a particular task, Python is an okay second or third choice for most tasks.
So, you can learn Python. Or you learn <best language> and <something else>. And if <something else> is Python, was <best language> sufficiently better than Python to be worth spending the time learning?
A lot of data science code is already in Python. That’s where it’s going to stay because rewriting code is time consuming. My guess is we will continue to improve Python gradually and keep refactoring the code.
Sounds a lot like "worse is better". Python is the worse option, incomplete and inelegant, but is much more practical due to being there first and receiving the bulk of the attention.
You do not need a DSL.
The Zen of Python is sadly now an absolute lie.
For the strings, just use f-strings and forget all the others. You can even do things like this for debugging:
>>> class User:
... pass
... user = User()
... user.name = "Surac"
...
>>> print(f"{user.name=}")
user.name='Surac'
>>>
VS Code and PyCharm are both free and are great for Python coding. They each have a full debugger, which is invaluable when you are learning a language.
What editor are you using that can't do that? Notepad?
If by data science you mean loading data to memory and running canned routines for regression, classification and other problems, then Python is great and mostly calls C/FORTRAN binaries under the hood, so Python itself has relatively little overhead.
1. Is easy to read
2. Was easy to extend in languages that people who work with scientific data happen to like.
When I did my masters we hacked around in the numpy source and contributed here and there while doing astrophysics.
Stuff existed in Java and R, but we had learned C in the first semester and python was easier to read and contrary to MATLAB numpy did not need a license.
When data science came into the picture, the field was full of physicists that had done similar things. They brought their tools as did others.
It got popular once Linux distributions started relying on a lot of python scripts (e.g. Red Hat and Debian). As a side effect it was present on a lot of Linux and Unix systems early on. Scientists in the early 2000s and late nineties had access to workstations running Linux and Unix. So, Python was simply the approachable thing that was just there already.
And because it's so easy, there are lots of people getting into Python. So it got its own dynamic of generations of researchers in all sorts of fields knowing about Python being the goto thing to reach for. It never really was the best at anything it does. That wasn't even a goal. It's a bit slow. A bit verbose/clumsy compared to some of the alternatives that some data scientists prefer. It lacks a lot of features other languages have. Etc. This doesn't matter because it is simple and easy. The type of users that are new to programming are looking for something simple that they can understand. Not the platonic ideal of a language that mathematicians or computer scientists might prefer.
Python is the modern equivalent of BASIC which had this role before python was created. It wasn't that amazing. But early home computers had it as part of their OS. E.g. the Commodore 64 that was my first computer had an interactive Basic shell with the ability to load games from a tape as the main OS experience.
If your data is already in a table, and you’re using Python, you’re doing it because you want to learn Python for your next job. Not because it’s the best tool for your current job. The one thing Python has on all those other options is $$$. You will be far more employable than if you stick to R.
And the reason for that is because Python is one of the best languages for data and ML engineering, which is about 80% of what a data science job actually entails.
I'd say dplyr/tidyverse is a lot more a separate programming language to R than pandas is to Python.
No it isn't.
But the article says that very exotic syntax is more readable. I think this is mostly about the libraries, where honestly I equally don’t like matplotlib and R’s ggplot. But I would not think it’s language problem.
I was hoping to find some performance benchmarks or something more than feelings about certain block of code. Don’t get me wrong I am also not a die hard fan of Python although I have written a lot or production code in it. Mentioning bloated, boilerplate code…I am afraid author should look on Java or any modern JavaScript project.
As annoying as it is to admit it, python is a great language for data science almost strictly because it has so many people doing data science with it. The popularity is, itself, a benefit.
There's Julia -- it has serious drawbacks, like slow cold start if you launch a Julia script from the shell, which makes it unsuitable for CLI workflows.
Otherwise you have to switch to compiled languages, with their tradeoffs.
Have you tried Polars? It really discourages the inefficient creation of intermediate boolean arrays such as in the code that you are showing.
> There's Julia -- it has serious drawbacks, like slow cold start if you launch a Julia script from the shell, which makes it unsuitable for CLI workflows.
Julia has gotten significantly better over time with regard to startup, especially with regard to plotting. There is definitely a preference for REPL or notebook based development to spread the costs of compilation over many executions. Compilation is increasingly modular with package based precompilation as well as ahead-of-time compilation modes. I do appreciate that typical compilation is an implicit step making the workflow much more similar to a scripting language than a traditionally compiled language.
I also do appreciate that traditional ahead-of-time static compilation to binary executable is also available now for deployment.
After a day of development in R or Python, I usually start regretting that I am not using Julia because I know yesterday's code could be executing much faster if I did. The question really becomes do I want to pay with time today or over the lifetime of the project.
The problem is not usually inefficiency, but syntactic noise. Polars does remove that in some cases, but in general gets even more verbose (apparently by design), which gets annoying fast when doing explorative data analysis.
If I want to wrangle, explore, or visualise data I’ll always reach for R.
If I want to build ML/DL models or work with LLM’s I will usually reach for Python.
Often in the same document - nowadays this is very easy with Quarto.
Julia allows embedding both R and Python code, and has some very nice tools for drilling down into datasets:
It is the first language I've seen in decades that reduces entire paradigms into single character syntax, often outperforming both C and Numpy in many cases. =3
Not a "smear", but rather a well known limitation of the language. Perhaps your environment context works differently than mine.
It is bizarre people get emotionally invested in something so trivial and mundane. Julia is at v1.12.2 so YMMV, but Queryverse is a lot of fun =3
https://github.com/python/cpython/issues
Griefers ranting about years old _closed_ tickets on v1.0.5 versions on a blog as some sort of proof of lameness... is a poorly structured argument. Julia includes regression testing features built into even its plotting library output, and thus issues usually stay resolved due to pedantic reproducibility. Also, running sanity-checks in any llvm language code is usually wise.
Best of luck =3
Languages like Julia trivially handle conditional parallelism much more cleanly with the broadcast operator, and transparent remote host process instancing over ssh (still needs a lot of work to reach OTP like cluster functionality.)
Much like Go, library resources ported into the native language quietly moves devs away from the same polyglot issues that hit Python.
Best of luck. =3
While plotting may be clunky, I just don’t see r as much better. Plus in 2025 I can just provide a sample of data and what plot I want in an LLM and I get zero shot code of the plot I want.
Author sounds very academic to me.
This is completely aside, but I wouldn't hold this against the students or Python. The students may be following an age-old rule of office politics: "Never troubleshoot in front of an audience." And why this is more prevalent among the students who use Python, well... sample size of 30.
Just a simple one that can get you, R is 1-indexed. Yet if you have a vector, accessing myvec[0] is not an error. Alternatively, if you had say, a vector length of 3 and do myvec[10] that gets NA (an otherwise legal value). Or you could make an assignment past the end of the vector myvec[15] <- 3.14 , which will silently extend the array, inserting NAs
My take (and my own experience) is that python won because the rest of the team knows it. I prefer R but our web developers don't know it, and it's way better for me to write code that the rest of our team can review, extend, and maintain.
- [1] https://scicloj.github.io
I think munging the input into a clean enough data set that you can work on is another place Python excels compared to analysis specific tools like R.
The other thing is that a lot of R’s strengths are really the tidyverse’s. Some of that is to R’s credit as an extensible language that enables a skilled API designer to really shine of course, but I think there’s no reason Python the language couldn’t have similar libraries. In fact it has, in plotnine. (I haven’t tried Polars yet but it does at least seem to have a more consistent API.)
A better stdlib-only version would be:
from palmerpenguins import load_penguins
import math
from itertools import groupby
from statistics import fmean, stdev
penguins = load_penguins()
# Convert DataFrame to list of dictionaries
penguins_list = penguins.to_dict('records')
# create key function for grouping/sorting by species/island
def key_func(x):
return x['species'], x['island']
# Filter out rows where body_mass_g is missing and sort by species and island
filtered = sorted((row for row in penguins_list if not math.isnan(row['body_mass_g'])), key=key_func)
# Group by species and island
groups = groupby(filtered, key=key_func)
# Calculate mean and standard deviation for each group
results = []
for (species, island), group in groups:
values = [row['body_mass_g'] for row in group]
mean_value = fmean(values)
sd_value = stdev(values, xbar=mean_value)
results.append({
'species': species,
'island': island,
'body_weight_mean': mean_value,
'body_weight_sd': sd_value
})It also helps that in R any function can completely change how its arguments are evaluated, allowing the tidyverse packages to do things like evaluate arguments in the context of a data frame or add a pipe operator as a new language feature. This is a very dangerous feature to put in the hands of statisticians, but it allows more syntactic innovation than is possible in Python.
Julia and Nim [1] are dynamic and static approaches (respectively) to 1 language systems. They both have both user-defined operators and macros. Personally, I find the surface syntax of Julia rather distasteful and I also don't live in PLang REPLs / emacs all day long. Of course, neither Julia nor Nim are impractical enough to make calling C/Fortran all that hard, but the communities do tend to implement in the new language without much prompting.
Python, the language itself, might not be a great language for data science. BUT the author can use Pandas or Polars or another data-science-related library/framework in Python to get the job done that s/he was trying to write in R. I could read both her R and Pandas code snippets and understand them equally.
This article reads just like, "Hey, I'm cooking everything by making all ingredients from scratch and see how difficult it is!".
But in my personal projects, my favorite language to use it Dart.
Personally I've found polars has solved most of the "ugly" problems that I had with pandas. It's way faster, has an ergonomic API, seamless pandas interop and amazing support for custom extensions. We have to keep in mind Pandas is almost 20 years old now.
I will agree that Shiny is an amazing package, but I would argue it's less important now that LLMs will write most of your code.
Now, is Python a SUCCESSFUL language? Very.
That the author avoided saying Python was a bad language outright speaks a great deal of its suitability. Well, that, and the majority data science in practice.
I feel like I'm using Python more and more in a way that is just not native to it. Strict typing is simply necessary at some point in time, but I don't get any of the performance benefits or compile-time warnings that other languages provide. But more than Pandas/Polars/DuckDB, I think it is the plotting ecosystem that keeps me in the Python universe. Seaborn, altair, plotnine all take ggplot's GoG and bring it to Python and I'm really grateful for that.
I don't see an alternative to Python for someone who needs to: 1) Work with data that fits in memory and is mostly tabular (thus Pandas/Python), 2) needs to visualize this data often, 3) does exploratory work (Jupyter notebooks).
> Either way, I’ll not discuss it further here. I’ll also not consider proprietary languages such as Matlab or Mathematica, or fairly obscure languages lacking a wide ecosystem of useful packages, such as Octave.
I feel, to most programming folks R is in the same category. R is to them what Octave is to the author. R is nice nice, but do they really want to learn a "niche" language, even if it has better some features than Python? Is holding a whole new paradigm, syntax, library ecosystem in your head worth it?
> Contrast this with equivalent code that is full of logistics, where I’m using only basic Python language features and no special data wrangling package:
n = len(values)
# Calculate mean
mean = sum(values) / n
# Calculate standard deviation
variance = sum((x - mean) \* 2 for x in values) / (n - 1)
std_dev = math.sqrt(variance)
Work experience says that 90% of work is gathering, cleaning and transforming data from different sources. In this capacity Python has more options available.
While I am not a python cheerleader, but a user because the reality is that it is a pretty good glue language, the above is a bit of a problem.
Duckdb, pandas, numpy etc.. is what makes python nice.
About a decade ago I worked at a major BI software company and ran into another silly problem when trying to evangelize R, wikis kbs and search engines don’t like single letter search terms.
So it didn’t matter how much better R was at the time, people found learning it more difficult than it should have been.
At the same time it is an absolute necessity to know if you are doing numerics. What this shows, at least to me, is that it is "good enough" and that the million integrations, examples and pieces of documentation matter more than whether the peculiarities of the language work in favor of its given use case, as long as the shortcomings can be mostly addressed.
Recently I am seeing that Python is heavily pushed for all data science related things. Sometimes objectively Python may not be the best option especially for stats. It is hard to change something after it becomes the "norm" regardless of its usability.
[0] https://kotlinlang.org/docs/kotlin-notebook-overview.html
It lives in a sterile, idealized world.
Python is a great language for data science in practice because it turns out data science is also:
- gluing a lot of data sources
- cleaning up a ton of terribly shaped data
- validation and error handling
- I/O, networking, and format conversion
- emboarding non-programmers into programming
- wrapping a lot of compiled languages' libs or plugging system
- prototyping stuff and exposing that prototype to some people
- turning prototypes into more permanent projects
There are other languages excellent at some of those, or some of the other things, but rarely good at most. And because humanity is vast, diverse, and constantly renewing, being the second best at those is eventually always winning.
Because whoever you are, you will be annoyed at not having the best experience at task X. But you would be mortified if you had the worst experience at doing task Y and Z. And task X, Y, and Z change depending on who you ask.
And you want to get things done, while days have 24 hours.
As usual, to understand the Python phenomenon, you have to see the whole picture. Not your little corner of the bubble. Not the ideal world in your head either. Life is not a maths problem with a clearly laid out premise and an elegant answer.
That's the same debate about why PHP won the web in 2000 no matter the size of the spaghetti plate, why Windows stayed used for so long despite it being terrible, why people keep using iphones after all the abuses, etc. There is more to it than the use case you have every day. People have needs you don't haven't thought about.
So it's not "let the language war begin". It's, "dude, get more experience, go work with accountants, ngos, govs and logistic chains, go work in china, africa and south america, go from a startup to schools to corporate, satisfy the geeks, the artists and the business people, than we'll talk".
Best part is, write a --help, and you can load them into LLMs as tools to help the LLMs figure it out for you.
Fight me.
I use mlr, sqlite, rye, souffle, and goawk in the shell scripts, and visidata to interactively review the intermediate files.
More terse, more efficient, less error prone, hopefully more numerically accurate, as if Python had an ecosystem of well designed libraries on par with R.
Worked quite well, but the TS/JS langgraph version is way behind. React agents are just a few lines of code, compared to 50 odd lines for the same thing in JS/TS.
Better to use a different language, even one i'm not familiar with, to be able to maintain a few lines of code vs 50 lines.
I can't help to conclude that Python is as good as R because I still have the choice of using Pandas when I need it. What did I get wrong?
also, we didn't define "good".
it was easy to think about the structures (iterators) it was easy to extend. it had a good community.
And for that, people start extending it via libraries.
There are plenty more alternatives now.
I agree that Python is not great at anything specifically, but it is good at almost everything, and that's what makes it great.
float(len(str)) str.len().float()
There should be one-- and preferably only one --obvious way to do it.
So, the answer is to pick the obvious choice, which in this case will depend on the style guide of the project you are working on, and whether they prefer the lisp style syntax or the OO property of style.
Python is not a great language
First, the white space requirements are a bad flashback to 1970s fortran.
Second, it is the language that is least compatible with itself.
You will say that not everything that is successful is great, and you will be right, but the success of python came organically, and not because of advertisement, de facto monopoly, politics, money, or first-arrived-advantage.
Although there is one cause that isn't intrinsic to python but from the people who built numpy. The fact there is a single numerical library, extremely easy to use, fast and extensive in the whole ecosystem was very very huge
mediocre!
Also add to the fact that TS based exploratory code can potentially plot SVG via d3 and maybe even exported to a webpage.
But the language has many rough edges
1. non standard eval is very weird, rlang fixes these shortcomings 2. unintuitive names or functions not belonging to packages, base has a mix of functions 3. S3 mixes with naming, no problem personally with S3 and S7 is even better, but mixing S3 names with ordinary names is unintuitive, keep snake case 4. data.frames are unintuitive, tidyverse fixes this 5. f(a=) seriously? or working with unintuitive functions in body for discrete ranges of function arguments? 6. no imports per file in packages, I can live with this .. still ... 7. AST functions are unintuitive
R has some excellent parts:
non-standard evaluation, AST in the base language, lazy evaluation
but it is being killed by the bad parts
I think all the external fixes and sanity in names should go into base
but it will take a lot of time if it ever happens due to legacy.
Julia fixes many of these not as elegantly as R but it's pragmatic approach is too attractive.
I was all hyped up, ready to see the amazing examples and arguments that would convince me to pick up R, and it gave me absolutely nothing (except quotes and brackets..).
Disappointing.
Often they'd be doing very simplistic querying and then manipulating via DATAstep prior to running whatever modeling and/or reporting PROCs later, rather than pushing it upstream into a far faster native database SQL pull via pass-through.
Back in 2008/2009, I saved 30h+ runtime on a regular report by refactoring everything in SQL via pass-through as opposed to the data scientists' original code that simply pulled the data down from the external source and manipulated it in DATAstep. Moving from 30h to 3m (Oracle backend) freed up an entire FTE to do more than babysit a long-running job 3x a week to multiple times per day.
The focus of SAS and R were primarily limited to data science-related fields; however, Python is a far more generic programming language, thus the number of folks exposed to it is wider and thus the hiring pool of those who come in exposed to Python is FAR LARGER than SAS/R ever were, even when SAS was actively taught/utilized in undergraduate/graduate programs.
As a hiring leader in the Data Science and Engineering space, I have extensive experience with all of these + SQL, among others. Hiring has become much easier to go cross-field/post-secondary experience and find capable folks who can hit the ground running.
In the first place data science is more a label someone put on bag full of cats, rather than a vast field covered by similarly sized boxes.
- A General programming language like Python is good enough for data science but isn't specifically designed for it.
- A language that is specifically designed for Data Science like R is better at Data Science.
Who would have thought?
It makes it look like perl, on a bad day, or worse autogenerated javascript.
Why on earth is it so many levels deep in objects?
R is kind of a super-specialized language. Python is much more general purpose.
R failed to evolve, let's be honest. Python won via jupyter - I see this used ALL the time in universities. R is used too, but mostly for statistics related courses only, give or take.
Perhaps R is better for its niche, but Python has more momentum and in thus, dominates over R. That's simply the reality of the situation. It is like the bulldozer moving forward, at a fast speed.
> I say “This is great, but could you quickly plot the data in this other way?”
Ok so ... he would have to adjust R code too, right? And finding good info on that is simply harder. He says he has experience with universities. Well, I do too, and my experience is that people are WAY better with python than with R. You simply see that more students will drop out from R than from python. That's also simply the reality of the situation.
> They appear to be sufficiently cumbersome or confusing that requests that I think should be trivial frequently are not.
I am sure the reverse also applies. Pick some python library, do something awesome, then tell the R students to do the same. I bet he will have the same problems.
> So many times, I felt that things that would be just a few lines of simple R code turned out to be quite a bit longer and fairly convoluted.
Ok, so here he is trolling. Flat out - I said it.
I wrote a LOT of python and quite a bit of R. There is no way in life that the R code is more succinct than the python code for about 90% of the use cases out there. Sorry, that's simply not the case. R is more verbose.
> Here is the relevant code in R, using the tidyverse approach:
penguins |>
filter(!is.na(body_mass_g)) |>
group_by(species, island) |>
summarize(
This professor just hasn't realised that he is slowly becoming a fossil himself, by being unable to see that x is better than y.
Ju = Julia Pyt = Python Er = R
R is not only supported in Jupyter, it was there from the start. I’ve never written a single line of R. It is bizarre how little people know about their tools.
Of course, if your DS is mixed with ML & modern AI you just:
pip install rpy2
But then, why choose? No need to be dogmatic, if R is nice for you:
install.packages("reticulate")