7 ms·

Pandas is used in some top 10 banks for analytics. Its performance is abysmal at the scale used there. Nobody wants to invest resources in training analysts to

by hated 6y ago

Pandas is used in some top 10 banks for analytics. Its performance is abysmal at the scale used there. Nobody wants to invest resources in training analysts to write high performance code so here we are. I have never viewed SQL more highly after seeing the mess that analysts make when writing imperative code.

civilized 6y ago
No surprise there - pandas encourages ugly, inefficient code with its bloated, unintuitive API. Once I was a lead on a new project and asked the intern to write some basic ETL code for data in some spreadsheets. I said she could write it in Python if she wanted, because "Python is good for ETL", right? This intern was not dumb by any means, but she wrote code that took 5 minutes to do something that can be done in <1 second with the obvious dplyr approach. Also, if your bank analysts pick up dplyr, they can use dbplyr to write SQL for them :)
- peatmoss 6y ago
  R’s meta programming facilities are head and shoulders above Python’s, which I think explains the brilliance of dplyr and dbplyr. But I feel like with R you have to scrape back a bunch of layers to get to the Schemey parts. I’ve always wondered what Hadley and Co would have done with dplyr and dbplyr had they had something like Racket at their disposal.
  - civilized 6y ago
    I was offended the first time I encountered R's nonstandard evaluation, but it didn't take long to accept it. Now I wonder why anyone would want to write `mytable.column` a million times when it's obvious from context what `column` is referred to, and the computer can reliably figure it out for you with some simple scoping rules. It's a superior notation that facilitates focus on the real underlying problem, and data analysts love that.
  - kgwgk 6y ago
    Unfortunately R success killed xlisp-stat: http://homepage.divms.uiowa.edu/~luke/xls/xlsinfo/ http://homepage.divms.uiowa.edu/~luke/xls/xlsinfo/ Edit: or maybe it's not dead? I just found http://www.user2019.fr/static/pres/t246174.pdf http://www.user2019.fr/static/pres/t246174.pdf
- em500 6y ago
  IMO they should just bite the bullet and learn proper SQL. I say this as a data scientist who learned SQL later than C, Matlab, R, Python/Pandas (though earlier than PySpark).
  - civilized 6y ago
    I agree. SQL is nothing to be afraid of, and there's no happier place to be analyzing huge tabular datasets than in a modern columnar database
    - orhmeh09 6y ago
      R’s data.table package is faster at these things out of the box than any single instance of a database server I’ve encountered. This is frustrating because I’m trying to explain some systemic issues we suffer by not using a relational database, but it’s really hard to make my case when data.table is one install.packages away and a version upgrade from Postgres 9 to something a little faster is gatekept by bureaucracy. I’ve been trying for months!
      - civilized 6y ago
        You need a columnar database for good performance. Try DuckDB to ease them into it, it's a columnar SQLite.
        orhmeh09 6y ago
        Thanks, I’m checking it out, it seems pretty interesting to keep an eye on. Lots of properties that would be useful in our shared computing environment like not requiring root or Docker.
        hated 6y ago
        Might also be worth running a local instance of Postgres 13. Super easy to do on Windows without administrator rights.
smabie 6y ago
Pandas/python is amazingly prevalent at trading firms. And everyday, we bitch about the performance, we bitch about the stupid API, we bitch about the GIL, the lack of expressiveness. The list goes on and on. But for some braindead reason, we never switch to Julia. It's masochistic.
- ivirshup 6y ago
  I do think Julia is a far better language for numerics than python, but compared to DataFrames.jl, pandas can be quite fast. I know, "but it's easier to make it faster in Julia". Last I checked `sort(df, :col)` was significantly slower than `df[sortperm(df[:col])]`. Someone actually has to go through and make these libraries fast. Second issue, in my field (bioinformatics) the script is still a pretty common unit of code. Without cached compilation being a simple flag, Julia often is slower.
  - smabie 6y ago
    Yeah, that's a good point. DataFrames.jl starts to really shine what the cookie cutter pandas functions arent adequate for what you need to do. DataFrames.jl can certainly be slower in some cases, but you should expect a consistent level of performance no matter what you do. This is a farcry from Pandas, which tanks by large factors when you start calling Python code vs C code. In regards to Julia's compilation problem, you can use https://github.com/JuliaLang/PackageCompiler.jl https://github.com/JuliaLang/PackageCompiler.jl to precompile an image, allowing you to avoid paying the JIT performance penalty over and over again.

by hated 6y ago

civilized 6y ago

No surprise there - pandas encourages ugly, inefficient code with its bloated, unintuitive API. Once I was a lead on a new project and asked the intern to write some basic ETL code for data in some spreadsheets. I said she could write it in Python if she wanted, because "Python is good for ETL", right? This intern was not dumb by any means, but she wrote code that took 5 minutes to do something that can be done in <1 second with the obvious dplyr approach. Also, if your bank analysts pick up dplyr, they can use dbplyr to write SQL for them :)

peatmoss 6y ago
R’s meta programming facilities are head and shoulders above Python’s, which I think explains the brilliance of dplyr and dbplyr. But I feel like with R you have to scrape back a bunch of layers to get to the Schemey parts. I’ve always wondered what Hadley and Co would have done with dplyr and dbplyr had they had something like Racket at their disposal.
- civilized 6y ago
  I was offended the first time I encountered R's nonstandard evaluation, but it didn't take long to accept it. Now I wonder why anyone would want to write `mytable.column` a million times when it's obvious from context what `column` is referred to, and the computer can reliably figure it out for you with some simple scoping rules. It's a superior notation that facilitates focus on the real underlying problem, and data analysts love that.
- kgwgk 6y ago
  Unfortunately R success killed xlisp-stat: http://homepage.divms.uiowa.edu/~luke/xls/xlsinfo/ http://homepage.divms.uiowa.edu/~luke/xls/xlsinfo/ Edit: or maybe it's not dead? I just found http://www.user2019.fr/static/pres/t246174.pdf http://www.user2019.fr/static/pres/t246174.pdf
em500 6y ago
IMO they should just bite the bullet and learn proper SQL. I say this as a data scientist who learned SQL later than C, Matlab, R, Python/Pandas (though earlier than PySpark).
- civilized 6y ago
  I agree. SQL is nothing to be afraid of, and there's no happier place to be analyzing huge tabular datasets than in a modern columnar database
  - orhmeh09 6y ago
    R’s data.table package is faster at these things out of the box than any single instance of a database server I’ve encountered. This is frustrating because I’m trying to explain some systemic issues we suffer by not using a relational database, but it’s really hard to make my case when data.table is one install.packages away and a version upgrade from Postgres 9 to something a little faster is gatekept by bureaucracy. I’ve been trying for months!
    - civilized 6y ago
      You need a columnar database for good performance. Try DuckDB to ease them into it, it's a columnar SQLite.
      - orhmeh09 6y ago
        Thanks, I’m checking it out, it seems pretty interesting to keep an eye on. Lots of properties that would be useful in our shared computing environment like not requiring root or Docker.
        hated 6y ago
        Might also be worth running a local instance of Postgres 13. Super easy to do on Windows without administrator rights.

smabie 6y ago

Pandas/python is amazingly prevalent at trading firms. And everyday, we bitch about the performance, we bitch about the stupid API, we bitch about the GIL, the lack of expressiveness. The list goes on and on. But for some braindead reason, we never switch to Julia. It's masochistic.

ivirshup 6y ago
I do think Julia is a far better language for numerics than python, but compared to DataFrames.jl, pandas can be quite fast. I know, "but it's easier to make it faster in Julia". Last I checked `sort(df, :col)` was significantly slower than `df[sortperm(df[:col])]`. Someone actually has to go through and make these libraries fast. Second issue, in my field (bioinformatics) the script is still a pretty common unit of code. Without cached compilation being a simple flag, Julia often is slower.
- smabie 6y ago
  Yeah, that's a good point. DataFrames.jl starts to really shine what the cookie cutter pandas functions arent adequate for what you need to do. DataFrames.jl can certainly be slower in some cases, but you should expect a consistent level of performance no matter what you do. This is a farcry from Pandas, which tanks by large factors when you start calling Python code vs C code. In regards to Julia's compilation problem, you can use https://github.com/JuliaLang/PackageCompiler.jl https://github.com/JuliaLang/PackageCompiler.jl to precompile an image, allowing you to avoid paying the JIT performance penalty over and over again.