You havent seen anything until you need to put a 4.2gb gzipped csv into a pandas dataframe, which works without any issues I should note.
I raise you thousands of gzipped files (total > 20GB) combined into one dataframe. Frankly, my work laptop did not like it all that much. But most basic operations still worked fine tho
I really don’t think that’s a lot either. Nowadays we routinely process terabytes of data.
Yeah, it was just a simple example. Although using just pandas (without something like dask) for loading terabytes of data at once into a single dataframe may not be the best idea, even with enough memory.
It’s good to see the occult is still alive and well
Is 600 MB a lot for pandas? Of course, CSV isn’t really optimal but I would’ve sworn pandas happily works with gigabytes of data.
What do you mean not optimal? This is quite literally the most popular format for any serious data handling and exchange. One byte per separator and newline is all you need. It is not compressed so allows you to stream as well. If you don’t need tree structure it is massively better than others
I think portability and easy parsing is the only advantage od CSV. It’s definitely good enough (maybe even the best) for small datasets but if you have a lot of data you need a compressed binary format, something like parquet.
But which separator is it, and which line ending? ASCII, UTF-8, UTF-16 or something else? What about quoting separators and line endings? Yes, there is an RFC, but a million programs were made before the RFC and won’t change their ways now.
Also you can gzip CSV and still stream them.
Have you heard that there are great serialised file formats like .parquet from appache arrow, that can easily be used in typical data science packages like duckdb or polars. Perhaps it even works with pandas (although do not know it that well. I avoid pandas as much as possible as someone who comes from the R tidyverse and try to use polars more when I work in python, because it often feels more intuitive to work with for me.)
I used to export my pandas DataFrames as pickles, but decided to test parquet and it was great. It was like 10x smaller and allowed me to had the the databases on a server directory instead of having to copy everything to the local machine.
If you have a csv bigger than like 500mb you need more than 8gb of ram to open it
Wait till you hear about WSV
It really depends on the machine that is running the code. Pandas will always have the entire thing loaded in memory, and while 600Mb is not a concern for our modern laptops running a single analysis at a time, it can get really messy if the person is not thinking about hardware limitations
Pandas supports lazy loading and can read files in chunks. Hell, even regular ole Python doesn’t need to read the whole file at once with
csvI didn’t know about lazy loading, that’s cool!
Then I guess that the meme doesn’t apply anymore. Though I will state that (from my anedoctal experience) people that can use Panda’s most advanced features* are also comfortable with other data processing frameworks (usually more suitable to large datasets**)
*Anything beyond the standard
groupby-applycan be considered advanced, from the placrs I’ve been**I feel the urge to note that 60Mb isn’ lt a large dataset by any means, but I believe that’s beyond the point
Is 600 MB a lot for pandas?
No, but it’s easy to make a program in Python that doesn’t like it.
Oh, I know, believe me. I have some painful first-hand experience with such code.
I guess it’s more of a critique of how bad CSV is for storing large data than pandas being inefficient
CSV is not optimal, but then someone shows up and gives you 60GB of JSON instead of 600MB of CSV.
Why do you have to personally attack me like that?
Or they dump their entire 6gb SQL database, customer info and all, into a SQL file that you have to load into a mariadb docker container when you just needed a subset that you were going to turn into csv anyway ☺️
Fine! .csv.gz ftw!
It’s more likely you’ll eat up storage when you read a 600mb parquet and try to write it as CSV.
I mean, yeah, that’s the point of compression. I don’t quite get what you mean by that comment.
Ah I was trying to point out that CSV is the inefficient format. Reading a large amount of data from a more efficient format like parquet is more likely to cause trouble because the memory required can be more than the file size. CSV is the opposite where it will almost always use more disk space than is required to represent the data in memory.
Right, true!
600MB? What is this, 2004?
pola.rs enters the chat…
Hell, depending on what you’re doing, reading it into a numpy array instead of a panda dataframe will yeild huge performance gains too.
Been meaning to try polars… Haven’t had a good excuse yet :)
Just read a few at a time…
No, just buy some more RAM. 64Gb is the minimum for a professional data analyst. 128Gb, is the sweet spot.
And there are like 8 software projects dedicated to making pandas wrappers that work with large datasets because this is somehow better than engineers and statisticians learning SQL or some kind of distributed calculations strategy.
Compared to other technical skills, SQL to a level needed by a data analyst has to be the easiest. It’s easier than learning Excel
But how else are the suits going to lick Microsoft’s boots?
I do this daily haha
CSV are a cool concept. Not so much any standard but rather a text doc where values are separated by commas. Sometimes banks use them and its hell to format them for Excel. Sometimes its just a list of readable words and values.
I had to build a Twitch Bot to add banned words in a CSV to a black list a while back, wish they would just let you copy paste like YT does.
“Constipated”
