Python数据分析

- PolarsBook中文版: https://www.pythondataanalysis.com/docs/polars_book_cn/
- Polars快速入门: https://www.pythondataanalysis.com/docs/polars_book_cn/quickstart/
- Polars表达式: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/
- Polars表达式: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/expressions/
- Polars上下文: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/contexts/
- Polars分组: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/groupby/
- Polars折叠: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/folds/
- Polars自定义函数: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/custom_functions/
- Polars实例: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/introduction_polars/
- Polars表达式方法: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/api/
- Polars视频介绍: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/video_intro/
- Polars与Numpy交互: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/numpy/
- Polars窗口函数: https://www.pythondataanalysis.com/docs/polars_book_cn/dsl/window_functions/
- Polars索引: https://www.pythondataanalysis.com/docs/polars_book_cn/indexing/
- Polars数据类型: https://www.pythondataanalysis.com/docs/polars_book_cn/datatypes/
- 来自Pandas: https://www.pythondataanalysis.com/docs/polars_book_cn/coming_from_pandas/
- 来自ApacheSpark: https://www.pythondataanalysis.com/docs/polars_book_cn/coming_from_spark/
- Polars性能: https://www.pythondataanalysis.com/docs/polars_book_cn/performance/
- 字符串: https://www.pythondataanalysis.com/docs/polars_book_cn/performance/strings/
- Polars优化: https://www.pythondataanalysis.com/docs/polars_book_cn/optimizations/
- Polars惰性方法: https://www.pythondataanalysis.com/docs/polars_book_cn/optimizations/lazy/
- 谓词下推: https://www.pythondataanalysis.com/docs/polars_book_cn/optimizations/lazy/predicate-pushdown/
- 投影下推: https://www.pythondataanalysis.com/docs/polars_book_cn/optimizations/lazy/projection-pushdown/
- 其它优化: https://www.pythondataanalysis.com/docs/polars_book_cn/optimizations/lazy/other-optimizations/
- Polars参考指南: https://www.pythondataanalysis.com/docs/polars_book_cn/references/
- Polars时间序列: https://www.pythondataanalysis.com/docs/polars_book_cn/timeseries/
- Polars时间序列实例: https://www.pythondataanalysis.com/docs/polars_book_cn/timeseries/time-series/
- Polars使用范围: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/
- IO: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/io/
- Polars操作CSV文件: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/io/csv/
- Polars操作Parquet文件: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/io/parquet/
- Polars处理多个文件: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/io/multiple_files/
- Polars读取数据库: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/io/read_db/
- Polars与AWS交互: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/io/aws/
- Polars与Google BigQuery交互: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/io/google-big-query/
- Polars与Postgres交互: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/io/postgres/
- 互通性: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/interop/
- Arrow: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/interop/arrow/
- Numpy: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/interop/numpy/
- 数据: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/data/
- 字符串: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/data/strings/
- 时间戳: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/data/timestamps/
- 数据帧: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/
- 选中行或列: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/row_col_selection/
- 常用操作: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/common-manipulations/
- 聚合: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/aggregate/
- 分组: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/groupby/
- 过滤: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/filter/
- 连接: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/join/
- 重塑: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/melt/
- 条件应用: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/conditionally-apply/
- 排序: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/sorting/
- 透视: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/df/pivot/
- 应用: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/apply/
- Polars自定义函数: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/apply/udfs/
- Polars窗口函数: https://www.pythondataanalysis.com/docs/polars_book_cn/howcani/apply/window-functions/
- Python数据分析 第二版: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/
- 第 1 章 准备工作: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-01/
- 第 2 章 Python 语法基础: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-02/
- 第 3 章 Python 的数据结构、函数和文件: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-03/
- 第 4 章 NumPy 基础：数组和向量计算: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-04/
- 第 5 章 Pandas 入门: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-05/
- 第 6 章 数据加载、存储与文件格式: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-06/
- 第 7 章 数据清洗和准备: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-07/
- 第 10 章 数据聚合与分组运算: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-10/
- 第 11 章 时间序列: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-11/
- 第 12 章 pandas 高级应用: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-12/
- 第 13 章 Python 建模库介绍: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-13/
-  第 14 章 数据分析案例: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-14/
-  附录 A NumPy 高级应用: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Appendix-A/
-  附录 B 更多关于 IPython 的内容: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Appendix-B/
- 第 8 章 数据规整：聚合、合并和重塑: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-08/
- 第 9 章 绘图和可视化: https://www.pythondataanalysis.com/docs/Python_Data_Analysis_2nd_Editon/Chapter-09/
- Polars用户指南: https://www.pythondataanalysis.com/docs/Polars_user_guide/
- Polars入门: https://www.pythondataanalysis.com/docs/Polars_user_guide/polars_getting_started/
- 安装Polars: https://www.pythondataanalysis.com/docs/Polars_user_guide/polars_installation/
- Polars核心概念: https://www.pythondataanalysis.com/docs/Polars_user_guide/concepts/
- Polars数据类型和结构: https://www.pythondataanalysis.com/docs/Polars_user_guide/concepts/data-types-and-structures/
- Polars表达式和上下文: https://www.pythondataanalysis.com/docs/Polars_user_guide/concepts/expressions-and-contexts/
- Polars延迟API: https://www.pythondataanalysis.com/docs/Polars_user_guide/concepts/lazy-api/
- Streaming: https://www.pythondataanalysis.com/docs/Polars_user_guide/concepts/_streaming/
- Polars表达式: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/
- Polars基本操作: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/basic-operations/
- Aggregation: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/aggregation/
- Casting: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/casting/
- Categorical Data and Enums: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/categorical-data-and-enums/
- Expression Expansion: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/expression-expansion/
- Folds: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/folds/
- Lists and Arrays: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/lists-and-arrays/
- Missing Data: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/missing-data/
- Numpy Functions: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/numpy-functions/
- Strings: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/strings/
- Structs: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/structs/
- User Defined Python Functions: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/user-defined-python-functions/
- Window Functions: https://www.pythondataanalysis.com/docs/Polars_user_guide/expressions/window-functions/
- Reference: https://www.pythondataanalysis.com/docs/Polars_user_guide/api/reference/
- Index: https://www.pythondataanalysis.com/docs/Polars_user_guide/development/contributing/
- Versioning: https://www.pythondataanalysis.com/docs/Polars_user_guide/development/versioning/
- Index: https://www.pythondataanalysis.com/docs/Polars_user_guide/polars-cloud/
- Ecosystem: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/ecosystem/
- Gpu Support: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/gpu-support/
- Index: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/io/
- Index: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/lazy/
- Pandas: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/migration/pandas/
- Spark: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/migration/spark/
- Arrow: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/misc/arrow/
- Comparison: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/misc/comparison/
- Multiprocessing: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/misc/multiprocessing/
- Polars Llms: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/misc/polars_llms/
- Styling: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/misc/styling/
- Visualization: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/misc/visualization/
- Index: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/plugins/
- Create: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/sql/create/
- Cte: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/sql/cte/
- Intro: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/sql/intro/
- Select: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/sql/select/
- Show: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/sql/show/
- Index: https://www.pythondataanalysis.com/docs/Polars_user_guide/user-guide/transformations/

# Multiprocessing

TLDR: if you find that using Python's built-in `multiprocessing` module together with Polars results
in a Polars error about multiprocessing methods, you should make sure you are using `spawn`, not
`fork`, as the starting method:

{{code_block('user-guide/misc/multiprocess','recommendation',[])}}

## When not to use multiprocessing

Before we dive into the details, it is important to emphasize that Polars has been built from the
start to use all your CPU cores. It does this by executing computations which can be done in
parallel in separate threads. For example, requesting two expressions in a `select` statement can be
done in parallel, with the results only being combined at the end. Another example is aggregating a
value within groups using `group_by().agg(<expr>)`, each group can be evaluated separately. It is
very unlikely that the `multiprocessing` module can improve your code performance in these cases. If
you're using the GPU Engine with Polars you should also avoid manual multiprocessing. When used
simultaneously, they can compete for system memory and processing power, leading to reduced
performance.

See [the optimizations section](../lazy/optimizations.md) for more optimizations.

## When to use multiprocessing

Although Polars is multithreaded, other libraries may be single-threaded. When the other library is
the bottleneck, and the problem at hand is parallelizable, it makes sense to use multiprocessing to
gain a speed up.

## The problem with the default multiprocessing config

### Summary

The [Python multiprocessing documentation](https://docs.python.org/3/library/multiprocessing.html)
lists the three methods to create a process pool:

1. spawn
1. fork
1. forkserver

The description of fork is (as of 2022-10-15):

> The parent process uses os.fork() to fork the Python interpreter. The child process, when it
> begins, is effectively identical to the parent process. All resources of the parent are inherited
> by the child process. Note that safely forking a multithreaded process is problematic.

> Available on Unix only. The default on Unix.

The short summary is: Polars is multithreaded as to provide strong performance out-of-the-box. Thus,
it cannot be combined with `fork`. If you are on Unix (Linux, BSD, etc), you are using `fork`,
unless you explicitly override it.

The reason you may not have encountered this before is that pure Python code, and most Python
libraries, are (mostly) single threaded. Alternatively, you are on Windows or MacOS, on which `fork`
is not even available as a method (for MacOS it was up to Python 3.7).

Thus one should use `spawn`, or `forkserver`, instead. `spawn` is available on all platforms and the
safest choice, and hence the recommended method.

### Example

The problem with `fork` is in the copying of the parent's process. Consider the example below, which
is a slightly modified example posted on the
[Polars issue tracker](https://github.com/pola-rs/polars/issues/3144):

{{code_block('user-guide/misc/multiprocess','example1',[])}}

Using `fork` as the method, instead of `spawn`, will cause a dead lock.

The fork method is equivalent to calling `os.fork()`, which is a system call as defined in
[the POSIX standard](https://pubs.opengroup.org/onlinepubs/9699919799/functions/fork.html):

> A process shall be created with a single thread. If a multi-threaded process calls fork(), the new
> process shall contain a replica of the calling thread and its entire address space, possibly
> including the states of mutexes and other resources. Consequently, to avoid errors, the child
> process may only execute async-signal-safe operations until such time as one of the exec functions
> is called.

In contrast, `spawn` will create a completely new fresh Python interpreter, and not inherit the
state of mutexes.

So what happens in the code example? For reading the file with `pl.read_parquet` the file has to be
locked. Then `os.fork()` is called, copying the state of the parent process, including mutexes. Thus
all child processes will copy the file lock in an acquired state, leaving them hanging indefinitely
waiting for the file lock to be released, which never happens.

What makes debugging these issues tricky is that `fork` can work. Change the example to not having
the call to `pl.read_parquet`:

{{code_block('user-guide/misc/multiprocess','example2',[])}}

This works fine. Therefore debugging these issues in larger code bases, i.e. not the small toy
examples here, can be a real pain, as a seemingly unrelated change can break your multiprocessing
code. In general, one should therefore never use the `fork` start method with multithreaded
libraries unless there are very specific requirements that cannot be met otherwise.

### Pro's and cons of fork

Based on the example, you may think, why is `fork` available in Python to start with?

First, probably because of historical reasons: `spawn` was added to Python in version 3.4, whilst
`fork` has been part of Python from the 2.x series.

Second, there are several limitations for `spawn` and `forkserver` that do not apply to `fork`, in
particular all arguments should be pickable. See the
[Python multiprocessing docs](https://docs.python.org/3/library/multiprocessing.html#the-spawn-and-forkserver-start-methods)
for more information.

Third, because it is faster to create new processes compared to `spawn`, as `spawn` is effectively
`fork` + creating a brand new Python process without the locks by calling
[execv](https://pubs.opengroup.org/onlinepubs/9699919799/functions/exec.html). Hence the warning in
the Python docs that it is slower: there is more overhead to `spawn`. However, in almost all cases,
one would like to use multiple processes to speed up computations that take multiple minutes or even
hours, meaning the overhead is negligible in the grand scheme of things. And more importantly, it
actually works in combination with multithreaded libraries.

Fourth, `spawn` starts a new process, and therefore it requires code to be importable, in contrast
to `fork`. In particular, this means that when using `spawn` the relevant code should not be in the
global scope, such as in Jupyter notebooks or in plain scripts. Hence in the examples above, we
define functions where we spawn within, and run those functions from a `__main__` clause. This is
not an issue for typical projects, but during quick experimentation in notebooks it could fail.

## References

1. https://docs.python.org/3/library/multiprocessing.html

1. https://pythonspeed.com/articles/python-multiprocessing/

1. https://pubs.opengroup.org/onlinepubs/9699919799/functions/fork.html

1. https://bnikolic.co.uk/blog/python/parallelism/2019/11/13/python-forkserver-preload.html