> By providing the beta version of FireDucks free of charge and enabling data scientists to actually use it, NEC will work to improve its functionality while verifying its effectiveness, with the aim of commercializing it within FY2024.
I nearly made the mistake of merging Akka into a codebase recently; fortunately I double-checked the license and noticed it was the bullshit BUSL and it would have potentially cost my employer tens of thousands of dollars a year [1]. I ended up switching everything to Vert.x, but I really hate how normalized these ostensibly open source projects are sneaking scary expensive licenses into things now.
[1] Yes I'm aware of Pekko now, and my stuff probably would have worked with Pekko, but I didn't really want to deal with something that by design is 3 years out of date.
> We have an akka app before the BUSL and it is a PITA to maintain
I would imaging the non-Scala use case to be less than ideal.
In Scala land, Pekko - the open source fork of Akka is the way to go if you need compatibility. Personally, I'd avoid new versions of Akka like the plague, and just use more modern alternatives to Pekko/Akka anyway.
I'm not sure what Lightbend's target market is? Maybe they think they have enough critical mass to merit the price tag for companies like Sony/Netflix/Lyft, etc. But they've burnt their bridge right into the water with everyone else, so I see them fading into irrelevance over the next few years.
I actually do have some decision-making power in regards to what tech I use for my job [1] at a mid-size (by tech standards) company, and my initial plan was to use Akka for the thing I was working on, since it more or less fit into the actor model perfectly.
I'm sure that Lightbend feels that their support contract is the bee's knees and worth whatever they charge for it, but it's a complete non-starter for me, and so I look elsewhere.
Vert.x actor-ish model is a bit different, but it's not the that different, and considering that Vert.x tends to perform extremely well in benchmarks, it doesn't really feel like I'm losing a lot by using it instead of Akka, particularly since I'm not using Akka Streams.
[1] Normal disclaimer: I don't hide my employment history, and it's not hard to find, but I politely ask that you do not post it here.
I've found actors (Akka specifically) to be a great model when you have concurrent access to fine grained shared state. It provides such a simple mental model of how to serialize that access. I'm not a fan as a general programming model or even as a general purpose concurrent programming model.
Vert.x has the "Verticle" abstraction, which more or less corresponds to something like an Actor. It's close enough to where I don't feel like I'm missing much by using it instead of Akka.
Yeah, Vert.x actually ended up being pretty great. I feel like it gives me most of the cool features of Akka that I actually care about, but it allows you to gradually move into it; it can be a full-on framework, but it can also just be a decent library to handle concurrency.
Important to upvote this. If there's room for improvement for Polars (which I'm sure there is), go and support the project. But don't fall for a commercial trap when there are competent open source tools available.
No shade to the juggernaut of the open source software movement and everything it has/will enabled, but why the hate for a project that required people’s time and knowledge to create something useful to a segment of users and then expect to charge for using it in the future? Commercial trap seems to imply this is some sort of evil machination but it seems like they are being quite upfront with that language.
While I agree, it's worth noting that this project is a drop-in replacement (they claim that, at least), but Polars has a very different API. I much prefer Polars's API, but it's still a non-trivial cost to switch to it, which is why many people would instead explore Pandas alternatives instead.
I thought I saw on the documentation that it was released under the modified BSD license. I guess they could take future versions closed source, but the current version should be available for folks to use and further develop.
It's a bit sad for me. I find the biggest issue for me with pandas is the API, not the speed.
So many foot guns, poorly thought through functions, 10s of keyword arguments instead of good abstractions, 1d and 2d structures being totally different objects (and no higher-order structures). I'd take 50% of the speed for a better API.
I looked at Polars, which looks neat, but seems made for a different purpose (data pipelines rather than building models semi-interactively).
To be clear, this library might be great, it's just a shame for me that there seems no effort to make a Pandas-like thing with better API. Maybe time to roll up my sleeves...
Yes, every time I write df[df.sth = val], a tiny part of me dies.
For a comparison, dplyr offers a lot of elegant functionality, and the functional approach in Pandas often feels like an afterthought. If R is cleaner than Python, it tells a lot (as a side note: the same story for ggplot2 and matplotlib).
Another surprise for friends coming from non-Python backgrounds is the lack of column-level type enforcement. You write df.loc[:, "col1"] and hope it works, with all checks happening at runtime. It would be amazing if Pandas integrated something like Pydantic out of the box.
I still remember when Pandas first came out—it was fantastic to have a tool that replaced hand-rolled data structures using NumPy arrays and column metadata. But that was quite a while ago, and the ecosystem has evolved rapidly since then, including Python’s gradual shift toward type checking.
> Yes, every time I write df[df.sth = val], a tiny part of me dies.
That's because it's a bad way to use Pandas, even though it is the most popular and often times recommended way. But the thing is, you can just write "safe" immutable Pandas code with method chaining and lambda expressions, resulting in very Polars-like code. For example:
Plus nowadays with the latest Pandas versions supporting Arrow datatypes, Polars performance improvements over Pandas are considerably less impressive.
Column-level name checking would be awesome, but unfortunately no python library supports that, and it will likely never be possible unless some big changes are made in the Python type hint system.
I agree it comes in handy quite often, but that still doesn’t make it great to write compared to what sql or dplyr offers in terms of choosing columns to filter on (`where y > 0.5`, for sql and `filter(y > 0.5)`, for dplyr)
It is superior because you don't need to assign your dataframe to a variable ('df'), then update that variable or create a new one everytime you need to do that operation. Which means it is both safer (you're guaranteed to filter on the current version of the dataframe) and more concise.
For the rest of your comment: it's the best you can do in python. Sure you could write SQL, but then you're mixing text queries with python data manipulation and I would dread that. And SQL-only scripting is really out of question.
Eh, SQL and python can still work together very well where SQL takes the place of pandas. Doing things in waves/batch helps.
Big problem with pandas is that you still have to load the dataframe into memory to work with it. My data's too big for that and postgres makes that problem go away almost entirely.
It's superior because it is safer. Not because the API (or requirement for using Lambda) looks better. The lambda allows the operation to work on the current state of the dataframe in the chained operation rather than the original dataframe. Alternatively, you could use .query("y > 0.5"). This also works on the current state of the dataframe.
(I'm the first to complain about the many warts in Pandas. Have written multiple books about it. This is annoying, but it is much better than [df.y > 0.5].)
Using `lambda` without care is dangerous because it risks being not vectorized at all. It risks being super slow, operating one row at a time. Is `d` a single row or the entire series or the entire dataframe?
In this case `d` is the entire dataframe. It's just a way of "piping" the object without having to rename it.
You are probably thinking about `df.apply(lambda row: ..., axis=1)` which operates on each row at a time and is indeed very slow since it's not vectorized. Here this is different and vectorized.
All I want is for the IDE and Python to correctly infer types and column names for all of these array objects. 99% of the pain for me is navigating around SQL return values and CSVs as pieces of text instead of code.
I started using Polars for the "rapid iteration" usecase you describe, in notebooks and such, and haven't looked back — there are a few ergonomic wrinkles that I mostly attribute to the newness of the library, but I found that polars forces me to structure my thought process and ask myself "what am I actually trying to do here?".
I find I basically never write myself into a corner with initially expedient but ultimately awkward data structures like I often did with pandas, the expression API makes the semantics a lot clearer, and I don't have to "guess" the API nearly as much.
So even for this usecase, I would recommend trying out polars for anyone reading this and seeing how it feels after the initial learning phase is over.
+1 Seconding this. My limited experience with pandas had a non-trivial number of moments "?? Is it really like this? Nah - I'm mistaken for sure, this can not be, no one would do something insane like that". And yet and yet... Fwiw since I've found that numpy is a must (ofc), but pandas is mostly optional. So I stick to numpy for my writing, and keep pandas read only. (just execute someone else's)
Have you tried polars? It’s a much more regular syntax. The regular syntax fits well with the lazy execution. It’s very composable for programmatically building queries. And then it’s super fast
I found the biggest benefit of polars is ironically the loss of the thing I thought I would miss most, the index; with pandas there are columns, indices, and multi-indices, whereas with polars, everything is a column, it’s all the same so you can delete a lot of conditionals.
However, I still find myself using pandas for the timestamps, timedeltas, and date offsets, and even still, I need a whole extra column just to hold time zones, since polars maps everything to UTC storage zone, you lose the origin / local TZ which screws up heterogeneous time zone datasets. (And I learned you really need to enforce careful manual thoughtful consideration of time zone replacement vs offsetting at the API level)
Had to write a ton of code to deal with this, I wish polars had explicit separation of local vs storage zones on the Datetime data type
I think pandas was so ambitious syntax wise and concept wise. But it got be a bit of a jumble. The index idea in particular is so cool, particular multi-indexes, watching people who really understand it do multi index operations is very cool.
IMO Polars sets a different goal of what's the most pandas like thing that we can build that is fast (and leaves open the possibility for more optimization), and clean.
Polars feels like you are obviously manipulating an advanced query engine. Pandas feels like manipulating this squishy datastructure that should be super useful and friendly, but sometimes it does something dumb and slow
Building on top of Pandas feels like you're only escaping part of the problems. In addition to the API, the datatypes in Pandas are a mess, with multiple confusing (and none of them good) options for e.g. dates/datetimes. Does redframes do anything there?
It does to me. Well, a 1x developer into a .01x dev in my case.
My conclusion was that pandas is not for developers. But for one-offs by managers, data-scientists, scientists, and so on. And maybe for "hackers" who cludge together stuff 'till it works and then hopefully never touch it.
Which made me realize such thoughts can come over as smug, patronizing or belittling. But they do show how software can be optimized for different use-cases.
The danger then lies into not recognizing these use-cases when you pull in smth like pandas. "Maybe using panda's to map and reduce the CSVs that our users upload to insert batches isn't a good idea at all".
This is often worsened by the tools/platforms/lib devs or communities not advertising these sweet spots and limitations. Not in the case of Pandas though: that's really clear about this not being a lib or framework for devs, but a tool(kit) to do data analysis with. Kudo's for that.
Pandas is a commonly known DSL at this point so lots of data scientists know pandas like the back of their hand and thats why a lot of pandas but for X libraries have become popular.
I agree that pandas does not have the best designed api in comparison to say dplyr but it also has a lot of functionality like pivot, melt, unstack that are often not implemented by other libraries. It’s also existed for more than a decade at this point so there’s a plethora of resources and stackoverflow questions.
On top of that, these days I just use ChatGPT to generate some of my pandas tasks. ChatGPT and other coding assistants know pandas really well so it’s super easy.
But I think if you get to know Pandas after a while you just learn all the weird quirks but gain huge benefits from all the things it can do and all the other libraries you can use with it.
I've been living in the shadow of pandas for about a decade now, and the only thing I learned is to avoid using it.
I 100% agree that pandas addresses all the pain points of data analysis in the wild, and this is precisely why it is so popular. My point is, it doesn't address them well. It seems like a conglomerate of special cases, written for a specific problem it's author was facing, with little concern for consistency, generality or other use cases that might arise.
In my usage, any time saved by its (very useful) methods tends to be lost on fixing subtle bugs introduced by strange pandas behaviours.
In my use cases, I reindex the data using pandas and get it to numpy arrays as soon as I can, and work with those, with a small library of utilities I wrote over the years. I'd gladly use a "sane pandas" instead.
Aye, but we've learned it, we've got code bases written in it, many of us are much more data kids than "real devs".
I get it doesn't follow best practices, but it does do what it needs to. Speed has been an issue, and it's exciting seeing that problem being solved.
Interesting to see so many people recently saying "polars looks great, but no way I'll rewrite". This library seems to give a lot of people, myself included, exactly what we want. I look forward to trying it.
Great point that I completely share. I tend to avoid pandas at all costs except for very simple things as I have bitten by many issues related to indexing. For anything complicated I tend to switch to duckdb instead.
Can you explain your use-case and why DuckDB is better?
Considering switching from pandas and want to understand what is my best bet. I am just processing feature vectors that are too large for memory, and need an initial simple JOIN to aggregate them.
Look into [Ibis](https://ibis-project.org/). It's a dataframe library built on duckdb. It supports lazy execution, greater than memory datastructures, remote s3 data and is insanely fast. Also works with basically any backend (postgres, mysql, parquet/csv files, etc) though there are some implementation gaps in places.
I previously had a pandas+sklearn transformation stack that would take up to 8 hours. Converted it to ibis and it executes in about 4 minutes now and doesn't fill up RAM.
It's not a perfect apples to apples pandas replacement but really a nice layer on top of sql. after learning it, I'm almost as fast as I was on pandas with expressions.
I made the switch to Ibis a few months ago and have been really enjoying it. It works with all the plotting libraries including seaborn and plotnine. And it makes switching from testing on a CSV to running on a SQL/Spark a one-line change. It's just really handy for analysis (similar to the tidyverse).
I am not necessarily saying duckdb is better. I personally just found it easier, clearer to write a sql query for any complicated set of joins/group by processing than to try to do that in pandas.
Polars is missing a crucial feature for replacing pandas in Finance: first class timeseries handling. Pandas allows me to easily do algebra on timeseries. I can easily resample data with the resample(...) method, I can reason about the index frequency, I can do algebra between timeseries, etc.
You can do the same with Polars, but you have to start messing about with datetimes and convert the simple problem "I want to calculate a monthly sum anchored on the last business day of the month" to SQL-like operations.
Pandas grew a large and obtuse API because it provides specialized functions for 99% of the tasks one needs to do on timeseries. If I want to calculate an exponential weighted covariance between two time series, I can trivially do this with pandas: series1.ewm(...).cov(series2). I welcome people to try and do this with Polars. It'll be a horrible and barely readable contraption.
YC is mostly populated by technologists, and technologists are often completely ignorant about what makes pandas useful and popular. It was built by quants/scientists, for doing (interactive) research. In this respect it is similar to R, which is not a language well liked by technologists, but it is (surprise) deeply loved by many scientists.
I don't know what exponential weighted covariance is, but I've had pretty good luck converting time series-based analyses from pandas to polars (for patient presentations to my emergency department -- patients per hour, per day, per shift, etc.). Resample has a direct (and easier IMO) replacement in polars, and there is group_by_dynamic.
I've had trouble determining whether one timestamp falls between two others across tens of thousands of rows (with the polars team suggesting I use a massive cross product and filter -- which worked but excludes the memory requirement), whereas in pandas I was able to sort the timestamps and thereby only need to compare against the preceding / following few based on the index of the last match.
The other issue I've had with resampling is with polars automatically dropping time periods with zero events, giving me a null instead of zero for the count of events in certain time periods (which then gets dropped from aggregations). This has caught me a few times.
I'm curious how is polars group_by_dynamic easier than resample in pandas. In pandas if I want to resample to a monthly frequency anchored to the last business day of the month, I'd write:
> my_df.resample("BME").apply(...)
Done. I don't think it gets any easier than this. Every time I tried something similar with polars, I got bogged down in calendar treatment hell and large and obscure SQL like contraptions.
Edit: original tone was unintentionally combative - apologies.
"""
Calculate monthly sums anchored to the last business day of each month
Parameters:
df: DataFrame with dates and values
date_column: name of date column
value_column: name of value column to sum
Returns:
DataFrame with sums anchored to last business day
"""
# Ensure date column is datetime
df[date_column] = pd.to_datetime(df[date_column])
# Group by end of business month and sum
monthly_sum = df.groupby(pd.Grouper(
key=date_column,
freq='BME' # Business Month End frequency
))[value_column].sum().reset_index()
return monthly_sum
It's actually much simpler than that. Assuming the index of the dataframe DF is composed of timestamps (which is normal for timeseries):
df.resample("BME").sum()
Done. One line of code and it is quite obvious what it is doing - with perhaps the small exception of BME, but if you want max readability you could do:
It runs on top of Polars so you get those speed gains, but uses the Elixir programming language. This gives the benefit of a simple finctional syntax w/ pipelines & whatnot.
It also benefits from the excellent Livebook (a Jupyter alternative specific to Elixir) ecosystem, which provides all kinds of benefits.
Planning to switch to NumPy 2.0 strings soon. Other than that I feel all the basic operations are fine and solid.
Note for anyone else rolling up their sleeves: You can get quite far with pure Python when building on top of NumPy (or maybe Arrow). The only thing I found needing more performance was group-by-aggregate, where Numba seems to work OK, although a bit difficult as a dependency.
Have you examined siuba at all? It promises to be more similar to the R tidyverse, which IMHO has a much better API. And I personally prefer dplyr/tidyverse to Polars for exploratory analysis.
I have not yet used siuba, but would be interested in others' opinions. The activation energy to learn a new set of tools is so large that I rarely have the time to fully examine this space...
I think the choice of using functions instead of classes + methods doesn't really fit well into Python. Either you need to do a huge amount of imports or use the awful `from siuba import *`. This feels like shoehorning the dplyr syntax into Python when method chaining would be more natural and would still retain the idea.
Also, having (already a while ago) looked at the implementation of the magic `_` object, it seemed like an awful hack that will serve only a part of use cases. Maybe someone can correct me if I'm wrong, but I get the impression you can do e.g. `summarize(x=_.x.mean())` but not `summarize(x=median(_.x))`. I'm guessing you don't get autocompletion in your editor or useful error messages and it can then get painful using this kind of a magic.
Thank you I don't know why people think it's so amazing. I end up sometimes just extracting the numpy arrays from the data frame and doing things like I know how to, because the Panda way is so difficult
i fell on dark days when they changed the multiindex reference level=N, which worked perfectly and was so logical and could be input alongside the axis, was swapped out in favor of a separate call for groupby
In that case I’d recommend dplyr in R. It also integrates with a better plotting library, GGPlot, which not only gives you better API than matplotlib but also prettier plots (unless you really get to work at your matplot code)
> Then came along Polars (written in Rust, btw!) which shook the ground of Python ecosystem due to its speed and efficiency
Polars rocked my world by having a sane API, not by being fast. I can see the value in this approach if, like the author, you have a large amount of pandas code you don't want to rewrite, but personally I'm extremely glad to be leaving the pandas API behind.
I personally found the polars API much clunkier, especially for rapid prototyping. I use it only for cemented processes where I could do with speed up/memory reduction.
Is there anything specific you prefer moving from the pandas API to polars?
Not OP but the ability to natively implement complex groupby logic is a huge plus for me at least.
Say you want to take an aggergation like "the mean of all values over the 75th percentile" algonside a few other aggregations. In pandas, this means you're gonna be in for a bunch of hoops and messing around with stuff because you can't express it via the api. Polars' api lets you express this directly without having to implement any kind of workaround.
> FireDucks is not a open source library at this moment.
You can get it installed freely using pip and use under BSD-3 license and of course can look into the python part of the source code.
I don't understand what it means. It looks like a contradiction. Does it have a BSD-3 licence or not?
> While the wheel packages are available at https://pypi.org/project/fireducks/#files, and while they do contain Python files, most of the magic happens inside a (BSD-3-licensed) shared object library, for which source code is not provided.
They provide BSD-3-licensed Python files but the interesting bit happens in the shared object library, which is only provided in binary form (but is also BSD-3-licensed it seems, so you can distribute it freely).
Wouldn't it be nice if GitHub was just for source code and you couldn't just slap up a README that's an add for some proprietary shitware with a vague promise of source some day in the glorious future?
> Wouldn't it be nice if GitHub was just for source code
GitHub always been a platform for "We love to host FOSS but we won't be 100% FOSS ourselves", so makes sense they allow that kind of usage for others too.
I think what you want, is something like Codeberg instead, which is explicitly for FOSS and 100% FOSS themselves.
I think the anger comes from the fact that we expect Github repositories to host the actual source code and not be a dead-end with a single README.md file.
High quality code isn't just code that performs well when executed, but also is readable, understandable and maintainable. You can't judge code quality by looking at the compiled result, just because it works well.
> One could also say that quality is related to the functional output.
Right, I said nothing that contradicts that ("High quality code isn't just code that performs well when executed, but also ..."). High quality functional output is a necessary requirement, but it isn't sufficient to determine if code is high quality.
It's not really subjective if you're at all reasonable about it.
Imagine writing a very good program, running it through an obfuscator, and throwing away the original code. Is the obfuscated code "high quality code" now, because the output of the compilation still works as before?
It’s not clear to me why this would be faster than polars, duckdb, vaex or clickhouse. They seem to be taking the same approach of multithreading, optimizing the plan, using arrow, optimizing the core functions like group by.
I’m asking more about what techniques did they use to get the performance improvements in the slides.
They are showing a 20-30% improvement over Polars, Clickhouse and Duckdb. But those 3 tools are SOTA in this area and generally rank near eachother in every benchmark.
So 20-30% improvement over that cluster makes me interested to know what techniques they are using to achieve that over their peers.
In its essence it is a commercial product which has a free trial.
> Future Plans
By providing the beta version of FireDucks free of charge and enabling data scientists to actually use it, NEC will work to improve its functionality while verifying its effectiveness, with the aim of commercializing it within FY2024.
Its freeware under an open source license. Really misleading.
It looks like something you should stay away from unless you need it REALLY badly. Its a proprietary product with unknown pricing and no indication of what their plans are.
Does the fact that the binary is BSD licensed allow reverse-engineering?
I’m sad that R’s tidy syntax is not copied more widely in the python world. Dplyr is incredibly intuitive most don’t ever bother reading the instructions you can look at a handful of examples and you’ve got the gist of it. Polars despite its speed is still verbose and inconsistent while pandas is seemingly a collection of random spells.
Lots of Pandas hate in this thread. However, for folks with lots of lines of Pandas in production, Fireducks can be a lifesaver.
I've had the chance to play with it on some of my code it queries than ran in 8+ minutes come down to 20 seconds.
Re-writing in Polars involves more code changes.
However, with Pandas 2.2+ and arrow, you can use .pipe to move data to Polars, run the slow computation there, and then zero copy back to Pandas. Like so...
Setting aside complaints about the Pandas API, it's frustrating that we might see the community of a popular "standard" tool fragment into two or even three ecosystems (for libraries with slightly incompatible APIs) -- seemingly all with the value proposition of "making it faster". Based on the machine learning experience over the last decade, this kind of churn in tooling is somewhat exhausting.
I wonder how much of this is fundamental to the common approach of writing libraries in Python with the processing-heavy parts delegated to C/C++ -- that the expressive parts cannot be fast and the fast parts cannot be expressive. Also, whether Rust (for polars, and other newer generation of libraries) changes this tradeoff substantially enough.
I think it's a natural path of software life that compatibility often stands in the way of improving the API.
This really does seem like a rare thing that everything speeds up without breaking compatability. If you want a fast revised API for your new project (or to rework your existing one) then you have a solution for that with Polars. If you just want your existing code/workloads to work faster, you have a solution for that now.
It's OK to have a slow, compatible, static codebase to build things on then optimize as-needed.
Trying to "fix" the api would break a ton of existing code, including existing plugins. Orphaning those projects and codebases would be the wrong move, those things take a decade to flesh out.
This really doesn't seem like the worst outcome, and doesn't seem to be creating a huge fragmented mess.
> Based on the machine learning experience over the last decade, this kind of churn in tooling is somewhat exhausting.
Don't come to old web-devs with those complains, every single one of them had to write at least one open source javascript library just to create their linkedin account!
I understand `pandas` is widely used in finance and quantitative trading, but it does not seem to be the best fit especially when you want your research code to be quickly ported to production.
We found `numpy` and `jax` to be a good trade-off between "too high level to optimize" and "too low level to understand". Therefore in our hedge fund we just build data structures and helper functions on top of them. The downside of the above combination is on sparse data, for which we call wrapped c++/rust code in python.
Many of the complaints about Pandas here (and around the internet) are about the weird API. However, if you follow a few best practices, you never run into the issue folks are complaining about.
I wrote a nice article about chaining for Ponder. (Sadly, it looks like the Snowflake acquisition has removed that. My book, Effective Pandas 2, goes deep into my best practices.)
I don't quite agree, but if this was true, what would you tell a junior colleague in a code review? You can't use this function/argument/convention/etc you found in the official API documentation because...I don't like it? I think any team-maintained Pandas codebase will unavoidably drift into the inconsistent and bad. If you're always working alone, then it can of course be a bit better.
I have strong opinions about Pandas. I've used it since it came out and have coalesced on patterns that make it easy to use.
(Disclaimer: I'm a corporate trainer and feed my family teaching folks how to work with their data using Pandas.)
When I teach about "readable" code, I caveat that it should be "readable for a specific audience". I hold that if you are a professional, that audience is other professionals. You should write code for professionals and not for newbies. Newbies should be trained up to write professional code. YMMV, but that is my bias based on experience seeing this work at some of the biggest companies in the world.
If they could just make a dplyr for py it would be so awesome. But sadly I don’t think the python language semantics will support such a tool. It all comes down to managing the namespace I guess
Any explanation what makes it faster than pandas and polars would be nice (at least something more concrete than "leverage the C engine").
My easy guess is that compared to pandas, it's multi-threaded by default, which makes for an easy perf win. But even then, 130-200x feels extreme for a simple sum/mean benchmark. I see they are also doing lazy evaluation and some MLIR/LLVM based JIT work, which is probably enough to get an edge over polars; though its wins over DuckDB _and_ Clickhouse are also surprising out of nowhere.
Also, I thought one of the reasons for Polars's API was that Pandas API is way harder to retrofit lazy evaluation to, so I'm curious how they did that.
Lots of people have mentioned Polars' sane API as the main reason to favor it, but the other crucial reason for us is that it's based on Apache Arrow. That allows us to use it where it's the best tool and then switch to whatever else we need when it isn't.
The killer app for Polars in my day-to-day work is its direct Parquet export. It's become indispensable for cleaning up stuff that goes into Spark or similar engines.
The promise of a 100x speedup with 0 changes to your codebase is pretty huge, but even a few correctness / incompatibility issues would probably make it a no-go for a bunch of potential users.
Just because I haven't jumped into the data ecosystem for a while - is Polars basically the same as Pandas but accelerated? Is Wes still involved in either?
I have never heard of FireDucks! I'm curious if anyone else here has used it. Polars is nice, but it's not totally compatible. It would be interesting how much faster it is for more complex calculations
Looks very cool, BUT: it's closed source? That's an immediate deal breaker for me as a quant. I'm happy to pay for my tools, but not being able to look and modify the source code of a crucial library like this makes it a non-starter.
Why is that crazy? (I think the crazy thing is that they are faster at all. Taking an existing api and making it fast is harder than creating the api from scratch with performance in mind)
Sure but single node performance. This makes it not very useful IMO since quite a few data science folks work with Hadoop clusters or Snowflake clusters or DataBricks where data is distributed and querying is handled by Spark executors.
The comparison is to pandas, so single node performance is understood in the scope.
This is for people running small tasks that may only take a couple days on a single node with a 32 core CPU or something, not tasks that take 3 months using thousands of cores.
My understanding for the latter is that pyspark is a decent option, while ballista is the better option for which to look forward. Perhaps using bastion-rs as a backend can be useful for an upcoming system as well. Databricks et al are cloud trash IMO, as is anything that isn't meant to be run on a local single node system and a local HPC cluster with zero code change and a single line of config change.
While for most of my jobs I ended up being able to evade the use of HPC by simply being smarter and discovering better algorithms to process information, I recall like pyspark decently, but preferring the simplicity of ballista over pyspark due to the simpler installation of Rust over managing Java and JVM junk.
The constant problems caused by anything using JVM backend and the environment config with it was terrible to add to a new system every time I ran a new program.
In this regard, ballista is a enormous improvement. Anything that is a one-line install via pip on any new system, runs local-first without any cloud or telemetry, and requires no change in code to run on a laptop vs HPC is the only option worth even beginning to look into and use.
Hadoop hasn't been relevant for a long time, which is telling.
Unless I had thousands of files to work with, I would be loathe to use cluster computing. There's so much overhead, cost, waiting for nodes to spin up, and cloud architecture nonsense.
My "single node" computer is a refurbished tower server with 256GB RAM and 50 threads.
Most of these distributed computing solutions arose before data processing tools started taking multi-threading seriously.
> By providing the beta version of FireDucks free of charge and enabling data scientists to actually use it, NEC will work to improve its functionality while verifying its effectiveness, with the aim of commercializing it within FY2024.
In other words, it's free only to trap you.
I nearly made the mistake of merging Akka into a codebase recently; fortunately I double-checked the license and noticed it was the bullshit BUSL and it would have potentially cost my employer tens of thousands of dollars a year [1]. I ended up switching everything to Vert.x, but I really hate how normalized these ostensibly open source projects are sneaking scary expensive licenses into things now.
[1] Yes I'm aware of Pekko now, and my stuff probably would have worked with Pekko, but I didn't really want to deal with something that by design is 3 years out of date.
Vert.x and other frameworks are far better and easier for most devs to grok.
I would imaging the non-Scala use case to be less than ideal.
In Scala land, Pekko - the open source fork of Akka is the way to go if you need compatibility. Personally, I'd avoid new versions of Akka like the plague, and just use more modern alternatives to Pekko/Akka anyway.
I'm not sure what Lightbend's target market is? Maybe they think they have enough critical mass to merit the price tag for companies like Sony/Netflix/Lyft, etc. But they've burnt their bridge right into the water with everyone else, so I see them fading into irrelevance over the next few years.
I'm sure that Lightbend feels that their support contract is the bee's knees and worth whatever they charge for it, but it's a complete non-starter for me, and so I look elsewhere.
Vert.x actor-ish model is a bit different, but it's not the that different, and considering that Vert.x tends to perform extremely well in benchmarks, it doesn't really feel like I'm losing a lot by using it instead of Akka, particularly since I'm not using Akka Streams.
[1] Normal disclaimer: I don't hide my employment history, and it's not hard to find, but I politely ask that you do not post it here.
Plus the license isn't stupid.
Basically it's a debate about how many dark patterns can you squeeze next to that "upfront language" before "marketing" slides into "bait-n-switch."
In fact, what's stopping the pandas library from incorporating fireducks code into the mainline branch? pandas itself is BSD.
So many foot guns, poorly thought through functions, 10s of keyword arguments instead of good abstractions, 1d and 2d structures being totally different objects (and no higher-order structures). I'd take 50% of the speed for a better API.
I looked at Polars, which looks neat, but seems made for a different purpose (data pipelines rather than building models semi-interactively).
To be clear, this library might be great, it's just a shame for me that there seems no effort to make a Pandas-like thing with better API. Maybe time to roll up my sleeves...
For a comparison, dplyr offers a lot of elegant functionality, and the functional approach in Pandas often feels like an afterthought. If R is cleaner than Python, it tells a lot (as a side note: the same story for ggplot2 and matplotlib).
Another surprise for friends coming from non-Python backgrounds is the lack of column-level type enforcement. You write df.loc[:, "col1"] and hope it works, with all checks happening at runtime. It would be amazing if Pandas integrated something like Pydantic out of the box.
I still remember when Pandas first came out—it was fantastic to have a tool that replaced hand-rolled data structures using NumPy arrays and column metadata. But that was quite a while ago, and the ecosystem has evolved rapidly since then, including Python’s gradual shift toward type checking.
That's because it's a bad way to use Pandas, even though it is the most popular and often times recommended way. But the thing is, you can just write "safe" immutable Pandas code with method chaining and lambda expressions, resulting in very Polars-like code. For example:
Plus nowadays with the latest Pandas versions supporting Arrow datatypes, Polars performance improvements over Pandas are considerably less impressive.Column-level name checking would be awesome, but unfortunately no python library supports that, and it will likely never be possible unless some big changes are made in the Python type hint system.
For the rest of your comment: it's the best you can do in python. Sure you could write SQL, but then you're mixing text queries with python data manipulation and I would dread that. And SQL-only scripting is really out of question.
Big problem with pandas is that you still have to load the dataframe into memory to work with it. My data's too big for that and postgres makes that problem go away almost entirely.
(I'm the first to complain about the many warts in Pandas. Have written multiple books about it. This is annoying, but it is much better than [df.y > 0.5].)
You are probably thinking about `df.apply(lambda row: ..., axis=1)` which operates on each row at a time and is indeed very slow since it's not vectorized. Here this is different and vectorized.
I find I basically never write myself into a corner with initially expedient but ultimately awkward data structures like I often did with pandas, the expression API makes the semantics a lot clearer, and I don't have to "guess" the API nearly as much.
So even for this usecase, I would recommend trying out polars for anyone reading this and seeing how it feels after the initial learning phase is over.
However, I still find myself using pandas for the timestamps, timedeltas, and date offsets, and even still, I need a whole extra column just to hold time zones, since polars maps everything to UTC storage zone, you lose the origin / local TZ which screws up heterogeneous time zone datasets. (And I learned you really need to enforce careful manual thoughtful consideration of time zone replacement vs offsetting at the API level)
Had to write a ton of code to deal with this, I wish polars had explicit separation of local vs storage zones on the Datetime data type
IMO Polars sets a different goal of what's the most pandas like thing that we can build that is fast (and leaves open the possibility for more optimization), and clean.
Polars feels like you are obviously manipulating an advanced query engine. Pandas feels like manipulating this squishy datastructure that should be super useful and friendly, but sometimes it does something dumb and slow
[1]: https://github.com/maxhumber/redframes
My conclusion was that pandas is not for developers. But for one-offs by managers, data-scientists, scientists, and so on. And maybe for "hackers" who cludge together stuff 'till it works and then hopefully never touch it.
Which made me realize such thoughts can come over as smug, patronizing or belittling. But they do show how software can be optimized for different use-cases.
The danger then lies into not recognizing these use-cases when you pull in smth like pandas. "Maybe using panda's to map and reduce the CSVs that our users upload to insert batches isn't a good idea at all".
This is often worsened by the tools/platforms/lib devs or communities not advertising these sweet spots and limitations. Not in the case of Pandas though: that's really clear about this not being a lib or framework for devs, but a tool(kit) to do data analysis with. Kudo's for that.
I agree that pandas does not have the best designed api in comparison to say dplyr but it also has a lot of functionality like pivot, melt, unstack that are often not implemented by other libraries. It’s also existed for more than a decade at this point so there’s a plethora of resources and stackoverflow questions.
On top of that, these days I just use ChatGPT to generate some of my pandas tasks. ChatGPT and other coding assistants know pandas really well so it’s super easy.
But I think if you get to know Pandas after a while you just learn all the weird quirks but gain huge benefits from all the things it can do and all the other libraries you can use with it.
I 100% agree that pandas addresses all the pain points of data analysis in the wild, and this is precisely why it is so popular. My point is, it doesn't address them well. It seems like a conglomerate of special cases, written for a specific problem it's author was facing, with little concern for consistency, generality or other use cases that might arise.
In my usage, any time saved by its (very useful) methods tends to be lost on fixing subtle bugs introduced by strange pandas behaviours.
In my use cases, I reindex the data using pandas and get it to numpy arrays as soon as I can, and work with those, with a small library of utilities I wrote over the years. I'd gladly use a "sane pandas" instead.
I get it doesn't follow best practices, but it does do what it needs to. Speed has been an issue, and it's exciting seeing that problem being solved.
Interesting to see so many people recently saying "polars looks great, but no way I'll rewrite". This library seems to give a lot of people, myself included, exactly what we want. I look forward to trying it.
Considering switching from pandas and want to understand what is my best bet. I am just processing feature vectors that are too large for memory, and need an initial simple JOIN to aggregate them.
I previously had a pandas+sklearn transformation stack that would take up to 8 hours. Converted it to ibis and it executes in about 4 minutes now and doesn't fill up RAM.
It's not a perfect apples to apples pandas replacement but really a nice layer on top of sql. after learning it, I'm almost as fast as I was on pandas with expressions.
You can do the same with Polars, but you have to start messing about with datetimes and convert the simple problem "I want to calculate a monthly sum anchored on the last business day of the month" to SQL-like operations.
Pandas grew a large and obtuse API because it provides specialized functions for 99% of the tasks one needs to do on timeseries. If I want to calculate an exponential weighted covariance between two time series, I can trivially do this with pandas: series1.ewm(...).cov(series2). I welcome people to try and do this with Polars. It'll be a horrible and barely readable contraption.
YC is mostly populated by technologists, and technologists are often completely ignorant about what makes pandas useful and popular. It was built by quants/scientists, for doing (interactive) research. In this respect it is similar to R, which is not a language well liked by technologists, but it is (surprise) deeply loved by many scientists.
I've had trouble determining whether one timestamp falls between two others across tens of thousands of rows (with the polars team suggesting I use a massive cross product and filter -- which worked but excludes the memory requirement), whereas in pandas I was able to sort the timestamps and thereby only need to compare against the preceding / following few based on the index of the last match.
The other issue I've had with resampling is with polars automatically dropping time periods with zero events, giving me a null instead of zero for the count of events in certain time periods (which then gets dropped from aggregations). This has caught me a few times.
But other than that I've had good luck.
`.join_where()`[1] was also added recently.
[1]: https://docs.pola.rs/api/python/stable/reference/dataframe/a...
> my_df.resample("BME").apply(...)
Done. I don't think it gets any easier than this. Every time I tried something similar with polars, I got bogged down in calendar treatment hell and large and obscure SQL like contraptions.
Edit: original tone was unintentionally combative - apologies.
But I'm guessing it's something like this:
import pandas as pd
def calculate_monthly_business_sum(df, date_column, value_column):
# Example usage:df = pd.DataFrame({ 'date': ['2024-01-01', '2024-01-31', '2024-02-29'], 'amount': [100, 200, 300] })
result = calculate_monthly_business_sum(df, 'date', 'amount')
print(result)
Which you can run here => https://python-fiddle.com/examples/pandas?checkpoint=1732114...
df.resample("BME").sum()
Done. One line of code and it is quite obvious what it is doing - with perhaps the small exception of BME, but if you want max readability you could do:
df.resample(pd.offsets.BusinessMonthEnd()).sum()
This is why people use pandas.
> df.resample("BME").sum()
Assuming `df` is a dataframe (ie table) indexed by a timestamp index, which is usual for timeseries analysis.
"BME" stands for BusinessMonthEnd, which you can type out if you want the code to be easier to read by someone not familiar with pandas.
It so easy for my analyst team because of daily uses but my developers probavly will never thought/know BME and decided to implement the code again.
https://hexdocs.pm/explorer/exploring_explorer.html
It runs on top of Polars so you get those speed gains, but uses the Elixir programming language. This gives the benefit of a simple finctional syntax w/ pipelines & whatnot.
It also benefits from the excellent Livebook (a Jupyter alternative specific to Elixir) ecosystem, which provides all kinds of benefits.
Here's my alternative: https://github.com/otsaloma/dataiter https://dataiter.readthedocs.io/en/latest/_static/comparison...
Planning to switch to NumPy 2.0 strings soon. Other than that I feel all the basic operations are fine and solid.
Note for anyone else rolling up their sleeves: You can get quite far with pure Python when building on top of NumPy (or maybe Arrow). The only thing I found needing more performance was group-by-aggregate, where Numba seems to work OK, although a bit difficult as a dependency.
https://siuba.org
I have not yet used siuba, but would be interested in others' opinions. The activation energy to learn a new set of tools is so large that I rarely have the time to fully examine this space...
Also, having (already a while ago) looked at the implementation of the magic `_` object, it seemed like an awful hack that will serve only a part of use cases. Maybe someone can correct me if I'm wrong, but I get the impression you can do e.g. `summarize(x=_.x.mean())` but not `summarize(x=median(_.x))`. I'm guessing you don't get autocompletion in your editor or useful error messages and it can then get painful using this kind of a magic.
A pity their compares don’t have tidyverse or R’s data.table. I think R would look simpler but now it remains unclear.
Yeah, Pandas has that early PHP feel to it, probably out of being a successful first mover.
Writing pandas code is a bit redundant. So what?
Who is to say that fireducks won't make their own API?
Polars rocked my world by having a sane API, not by being fast. I can see the value in this approach if, like the author, you have a large amount of pandas code you don't want to rewrite, but personally I'm extremely glad to be leaving the pandas API behind.
Is there anything specific you prefer moving from the pandas API to polars?
Say you want to take an aggergation like "the mean of all values over the 75th percentile" algonside a few other aggregations. In pandas, this means you're gonna be in for a bunch of hoops and messing around with stuff because you can't express it via the api. Polars' api lets you express this directly without having to implement any kind of workaround.
Nice article on it here: https://labs.quansight.org/blog/dataframe-group-by
I don't understand what it means. It looks like a contradiction. Does it have a BSD-3 licence or not?
> While the wheel packages are available at https://pypi.org/project/fireducks/#files, and while they do contain Python files, most of the magic happens inside a (BSD-3-licensed) shared object library, for which source code is not provided.
Edit: To use, redistribute, and modify, and distribute modified versions.
https://opensource.org/license/bsd-3-clause
Imagine being like "the project is GPL - just the compiled machine code".
GitHub always been a platform for "We love to host FOSS but we won't be 100% FOSS ourselves", so makes sense they allow that kind of usage for others too.
I think what you want, is something like Codeberg instead, which is explicitly for FOSS and 100% FOSS themselves.
Is this shitware? It seems to be very high quality code
How would you define "quality" in this context?
One could also say that quality is related to the functional output.
Right, I said nothing that contradicts that ("High quality code isn't just code that performs well when executed, but also ..."). High quality functional output is a necessary requirement, but it isn't sufficient to determine if code is high quality.
My point was that it's all subjective in the end.
Imagine writing a very good program, running it through an obfuscator, and throwing away the original code. Is the obfuscated code "high quality code" now, because the output of the compilation still works as before?
https://fireducks-dev.github.io/files/20241003_PyConZA.pdf
The main reasons are
* multithreading
* rewriting base pandas functions like dropna in c++
* in-built compiler to remove unused code
Pretty impressive especially given you import fireducks.pandas as pd instead of import pandas as pd, and you are good to go
However I think if you are using a pandas function that wasn't rewritten, you might not see the speedups
They are showing a 20-30% improvement over Polars, Clickhouse and Duckdb. But those 3 tools are SOTA in this area and generally rank near eachother in every benchmark.
So 20-30% improvement over that cluster makes me interested to know what techniques they are using to achieve that over their peers.
> Future Plans By providing the beta version of FireDucks free of charge and enabling data scientists to actually use it, NEC will work to improve its functionality while verifying its effectiveness, with the aim of commercializing it within FY2024.
Its freeware under an open source license. Really misleading.
It looks like something you should stay away from unless you need it REALLY badly. Its a proprietary product with unknown pricing and no indication of what their plans are.
Does the fact that the binary is BSD licensed allow reverse-engineering?
Reversing and re-compiling should count as modification?
I've had the chance to play with it on some of my code it queries than ran in 8+ minutes come down to 20 seconds.
Re-writing in Polars involves more code changes.
However, with Pandas 2.2+ and arrow, you can use .pipe to move data to Polars, run the slow computation there, and then zero copy back to Pandas. Like so...
to:Where can I find the code? I don't see it on GitHub.
> contact@fireducks.jp.nec.com
So it's from NEC (a major Japanese computer company), presumably a research artifact?
> https://fireducks-dev.github.io/docs/about-us/ Looks like so.
I wonder how much of this is fundamental to the common approach of writing libraries in Python with the processing-heavy parts delegated to C/C++ -- that the expressive parts cannot be fast and the fast parts cannot be expressive. Also, whether Rust (for polars, and other newer generation of libraries) changes this tradeoff substantially enough.
This really does seem like a rare thing that everything speeds up without breaking compatability. If you want a fast revised API for your new project (or to rework your existing one) then you have a solution for that with Polars. If you just want your existing code/workloads to work faster, you have a solution for that now.
It's OK to have a slow, compatible, static codebase to build things on then optimize as-needed.
Trying to "fix" the api would break a ton of existing code, including existing plugins. Orphaning those projects and codebases would be the wrong move, those things take a decade to flesh out.
This really doesn't seem like the worst outcome, and doesn't seem to be creating a huge fragmented mess.
Don't come to old web-devs with those complains, every single one of them had to write at least one open source javascript library just to create their linkedin account!
Is it actually? Do people see that level of compatibility in practice?
It should be pretty close, though.
We found `numpy` and `jax` to be a good trade-off between "too high level to optimize" and "too low level to understand". Therefore in our hedge fund we just build data structures and helper functions on top of them. The downside of the above combination is on sparse data, for which we call wrapped c++/rust code in python.
I wrote a nice article about chaining for Ponder. (Sadly, it looks like the Snowflake acquisition has removed that. My book, Effective Pandas 2, goes deep into my best practices.)
(Disclaimer: I'm a corporate trainer and feed my family teaching folks how to work with their data using Pandas.)
When I teach about "readable" code, I caveat that it should be "readable for a specific audience". I hold that if you are a professional, that audience is other professionals. You should write code for professionals and not for newbies. Newbies should be trained up to write professional code. YMMV, but that is my bias based on experience seeing this work at some of the biggest companies in the world.
My easy guess is that compared to pandas, it's multi-threaded by default, which makes for an easy perf win. But even then, 130-200x feels extreme for a simple sum/mean benchmark. I see they are also doing lazy evaluation and some MLIR/LLVM based JIT work, which is probably enough to get an edge over polars; though its wins over DuckDB _and_ Clickhouse are also surprising out of nowhere.
Also, I thought one of the reasons for Polars's API was that Pandas API is way harder to retrofit lazy evaluation to, so I'm curious how they did that.
```
>>> df['year'].dtype == np.dtype('int32')
True
```
EDIT: I've found some benchmarks https://fireducks-dev.github.io/docs/benchmarks/
Would be nice to know what are internals of FireDucks
The promise of a 100x speedup with 0 changes to your codebase is pretty huge, but even a few correctness / incompatibility issues would probably make it a no-go for a bunch of potential users.
I haven’t seen that in other system like Polars, but maybe I’m wrong.
edit: I know pandas uses numpy under the hood, but "raw" numpy is typically faster (and more flexible), so curious as to why it's not mentioned
At some point I think it's more honest to say "the python ecosystem keeps getting more awesome".
Q: Why do ducks have big flat feet?
A: So they can stomp out forest fires.
Q: Why do elephants have big flat feet?
A: So they can stomp out flaming ducks.
While for most of my jobs I ended up being able to evade the use of HPC by simply being smarter and discovering better algorithms to process information, I recall like pyspark decently, but preferring the simplicity of ballista over pyspark due to the simpler installation of Rust over managing Java and JVM junk. The constant problems caused by anything using JVM backend and the environment config with it was terrible to add to a new system every time I ran a new program.
In this regard, ballista is a enormous improvement. Anything that is a one-line install via pip on any new system, runs local-first without any cloud or telemetry, and requires no change in code to run on a laptop vs HPC is the only option worth even beginning to look into and use.
Unless I had thousands of files to work with, I would be loathe to use cluster computing. There's so much overhead, cost, waiting for nodes to spin up, and cloud architecture nonsense.
My "single node" computer is a refurbished tower server with 256GB RAM and 50 threads.
Most of these distributed computing solutions arose before data processing tools started taking multi-threading seriously.