polars.Rmd

---
title: "An Introduction to Polars from R"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{An Introduction to Polars from R}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
options(rmarkdown.html_vignette.check_title = FALSE)
```


## What is Polars?

Polars is [a lightning fast](https://duckdblabs.github.io/db-benchmark/) Data Frame
library. Its embarrassingly parallel execution, cache efficient algorithms and
expressive API makes it perfect for efficient data wrangling, data pipelines,
snappy APIs, and much more besides. Polars also supports "streaming mode" for
out-of-memory operations. This allows users to analyze datasets many times
larger than RAM.

The underlying computation engine is written in Rust and is built on the Apache
Arrow columnar memory format. It can be used in Rust or via Python bindings.
The **polars** R-package provides equivalent bindings from R. To help distinguish
the different language implementations, we typically use a convention of
referring to them with prefixes: rust-polars, py-polars, r-polars,
nodejs-polars, etc. But within each language, the relevant library is always
just called _polars_.

**polars** users can expect orders of magnitude(s) improvement compared to
**dplyr** for simple transformations on datasets >500Mb. The automatic Polars
optimization framework means that that this speed boost can be even greater for
complex queries that chain together many operations. Performance is similar to
that of **data.table**, although **polars** supports additional functionality
via its relationship to the Apache Arrow memory model. For example, it can scan
multiple Parquet files and datasets and selectively import random subsets
without having to read all of the data.

Polars syntax is similar to that of Spark, but the workflow is column-oriented
rather than row-oriented. Since R is itself a column-oriented language, this
should immediately feel familiar to most R users. Like Spark and modern SQL
variants, Polars optimizes queries for memory consumption and speed, so you
don't have to. However, unlike Spark, Polars is natively multithreaded instead
of multinoded. This makes (r)polars much simpler to install and can be used as
one would any other R package.

This R port relies on the excellent [**extendr**](https://github.com/extendr)
package, which is the R equivalent to pyo3+maturin. **extendr** is very
convenient for calling Rust from R, and vice versa, and is what we use to build
the **polars** package. Once built, however, **polars** has no other
dependencies other than R itself. This makes it very fast and lightweight to
install, and so **polars** can immediately be used to tackle your big (or
small!) data wrangling tasks.

## Documentation and tutorials

Users can find detailed documentation for all objects, functions, and methods
on the Reference page of [this website](https://rpolars.github.io/). This documentation
can also be accessed from the R console using the typical `?` syntax. For
example, we will later use the `DataFrame()` constructor function and apply the
`group_by()` method to a `DataFrame` object. The documentation for these can be
accessed by typing these commands:

```r
?DataFrame
?DataFrame_group_by
```

The [Polars book](https://pola-rs.github.io/polars-book/user-guide/) offers a
great introduction to the Polars data frame library, with a very large number
of examples in Python and Rust. The syntax and expressions in the `polars`
package for R are (deliberately) as close to the Python implementation as
possible, so you can always refer to the [polars
book](https://pola-rs.github.io/polars-book/user-guide/) for more ideas. Just
remember to switch out any "." (Python) for a "$" (R) when chaining methods.
For example, here are two equivalent lines of code for some hypothetical
dataset.

```python
# Python
df.group_by("id").mean()
```

```r
# R
df$group_by("id")$mean()
```

## `Series` and `DataFrames`

In `polars` objects of class `Series` are analogous to R vectors. Objects of
class `DataFrame` are analogous to R data frames. To convert R vectors and data
frames to Polars `Series` and `DataFrames`, we load the library and use
constructor functions with the `pl$` prefix. This prefix is very important, as
most of the `polars` functions are made available via `pl$`:

```{r}
library(polars)

ser = pl$Series((1:5) * 5)
ser

dat = pl$DataFrame(mtcars)
dat
```


Both Polars and R are column-orientated. So you can think of `DataFrames`
(data.frames) as being made up of a collection of `Series` (vectors). In fact,
you can create a new Polars `DataFrame` as a mix of `Series` and/or regular R
vectors.

```{r}
pl$DataFrame(
  a = pl$Series((1:5) * 5),
  b = pl$Series(letters[1:5]),
  c = pl$Series(c(1, 2, 3, 4, 5)),
  d = c(15, 14, 13, 12, 11),
  c(5, 4, 3, 2, 1),
  1:5
)
```

`Series` and `DataFrame` can be operated on using many standard R functions. For example:

```{r}
# Series
length(ser)

max(ser)

# DataFrame
dat[c(1:3, 12), c("mpg", "hp")]

names(dat)

dim(dat)

head(dat, n = 2)
```

## Methods and pipelines

Although some simple R functions work out of the box on **polars** objects, the
full power of Polars is realized via _methods_. Polars methods are accessed
using the `$` syntax. For example, to convert Polars `Series` and `DataFrames`
back to standard R objects, we use the `$to_vector()` and `$to_data_frame()`
methods:

```{r}
ser$to_vector()
```

There are numerous methods designed to accomplish various tasks:

```{r}
ser$max()

dat$slice(offset = 2, length = 3)
```

One advantage of using methods is that many more operations are possible on
Polars objects using methods than through base R functions. 

A second advantage is _Methods Chaining_, a core part of the Polars workflow.
If you are coming from one of the other popular data wrangling libraries in R,
then you probably already have an innate sense of what this means. For
instance,

- In **dplyr** we use a pipe operator, e.g. `dat |> filter(...) |> select(...)`
- In **data.table** we use its indexing syntax, e.g. `DT[i, j, by][...]`
- Etc.

In **polars** our method chaining syntax takes the form `object$m1()$m2()`,
where `object` is our data object, and `m1()` and `m2()` are appropriate 
methods, like subsetting or aggregation expressions.

This might all seem a little abstract, so let's walk through some quick
examples to help make things concrete. We continue with the `mtcars` dataset that we
coerced to a `DataFrame` in the introduction.^[Similar to how (most) **data.table**
operations are limited to objects of class `data.table`, we can only perform
polars operations on objects that have been converted to an appropriate
**polars** class. Later on, we'll see how to read data from disk directly in Polars format.] 

To start, say we compute the maximum value in each column. We can use the
`max()` method:

```{r}
dat$max()
```

Now, we first use the `$tail` method to select the last 10 rows of the dataset,
and then use the `$max` method to compute the maximums in those 10 rows:

```{r}
dat$tail(10)$max()
```

Finally, we convert the result to a standard R data frame:

```{r}
dat$tail(10)$max()$to_data_frame()
```

Below, we will introduce several other methods, including `$select`, `$filter`,
and `$group_by` which allow us to do powerful data manipulations easily. To give
you a small taste, we now take group-wise means:

```{r}
dat$group_by("cyl")$mean()
```

## Subset

We can now start chaining together various methods (expressions) to manipulate
it in different ways. For example, we can subset the data by rows
([`filter()`](https://rpolars.github.io/reference/DataFrame_filter/))
and also columns
([`select()`](https://rpolars.github.io/reference/DataFrame_select/)):

```{r}
dat$filter(pl$col("cyl") == 6)

dat$filter(pl$col("cyl") == 6 & pl$col("am") == 1)

dat$select(pl$col(c("mpg", "hp")))
```

Of course, we can chain those methods to create a pipeline: 

```{r}
dat$filter(
  pl$col("cyl") == 6
)$select(
  pl$col(c("mpg", "hp", "cyl"))
)
```

## Aggregate and modify

The `select()` method that we introduced above also supports data modification,
so you can simultaneously transform it while you are subsetting. However, the
result will exclude any columns that weren't specified as part of the
expression. To modify or add some columns---whilst preserving all others in the
dataset---it is therefore better to use the
[`with_columns()`](https://rpolars.github.io/reference/DataFrame_with_columns/)
method. This next code chunk is equivalent to `mtcars |>
dplyr::mutate(sum_mpg=sum(mpg), sum_hp=sum(hp), .by = cyl)`.

```{r}
# Add the grouped sums of some selected columns.
dat$with_columns(
  pl$col("mpg")$sum()$over("cyl")$alias("sum_mpg"),
  pl$col("hp")$sum()$over("cyl")$alias("sum_hp")
)
```

For what it's worth, the previous query could have been written more concisely as:

```{r}
dat$with_columns(
  pl$col(c("mpg", "hp"))$sum()$over("cyl")$name$prefix("sum_")
)
```

Similarly, here's how we could have aggregated (i.e., collapsed) the dataset
by groups instead of modifying them. We need simply invoke the `group_by()` and
[`agg()`](https://rpolars.github.io/reference/Expr_agg_groups/) methods.

```{r}
dat$group_by(
  "cyl",
  maintain_order = TRUE
)$agg(
  pl$col(c("mpg", "hp"))$sum()
)
```

(arg `maintain_order = TRUE` is optional, since **polars** doesn't sort the results
of grouped operations by default. This is similar to what **data.table** does 
and is also true for newer versions of **dplyr**.)

The same principles of method chaining can be combined very flexibly to group by
multiple variables and aggregation types.

```{r}
dat$group_by(
  "cyl",
  pl$col("am")$cast(pl$Boolean)$alias("manual")
)$agg(
  pl$col("mpg")$mean()$alias("mean_mpg"),
  pl$col("hp")$median()$alias("med_hp")
)
```

Note that we used the `cast` method to convert the data type of the `am`
column. See the section below for more details on data types.

## Reshape

Polars supports data reshaping, going from both long to wide ("pivot") and wide
to long ("melt"). Let's switch to the `Indometh` dataset to demonstrate some
basic examples. Note that the data are currently in long format.

```{r}
indo = pl$DataFrame(Indometh)
```

To go from long to wide, we use the
[`pivot`](https://rpolars.github.io/reference/DataFrame_pivot/) method.
Here we pivot the data so that every subject takes its own column.

```{r}
indo_wide = indo$pivot(values = "conc", index = "time", columns = "Subject")
```

To go from wide to long, we use the
[melt](https://rpolars.github.io/reference/DataFrame_melt/) method.

```{r}
# indo_wide$melt(id_vars = "time") # default column names are "variable" and "value"
indo_wide$melt(id_vars = "time", variable_name = "subject", value_name = "conc")
```

Basic functionality aside, it should be noted that `pivot` can perform
aggregations as part of the reshaping operation. This is useful when you have
multiple observations per ID variable that need to be collapsed into unique
values. The aggregating functions can be arbitrarily complex, but let's
consider a relatively simple example using our `dat` ("mtcars") DataFrame from
earlier: what is the median MPG value (`mpg`) across cylinders (`cyl`), cut by
different combinations of transmission type (`am`) and engine shape (`vs`)?

```{r}
dat$pivot(
  values = "mpg",
  index = c("am", "vs"),
  columns = "cyl",
  aggregate_fun = "median" # aggregating function
)
```

Here, `"median"` is a convenience string that is equivalent to the more verbose
`pl$element()$median()`. Other convenience strings include "first", "last",
"min", "max", "mean", "sum", and "count".

## Join

As a final example of how **polars** can be used for standard data wrangling
tasks, let's implement a (left) join. For this example, we'll borrow some
datasets from the **nycflights13** package.

```{r}
data("flights", "planes", package = "nycflights13")
flights = pl$DataFrame(flights)
planes = pl$DataFrame(planes)

flights$join(
  planes,
  on = "tailnum",
  how = "left"
)
```

More information on the **polars** joining method can be found in the
[reference manual](https://rpolars.github.io/reference/DataFrame_join/).

The package supports many other data manipulation operations, which we won't 
cover here. Hopefully, you will already have a sense of the key syntax features.
We now turn to another core idea of the Polars ecosystem: _lazy execution_.

## Lazy execution

While the "eager" execution engine of **polars** works perfectly well---as
evidenced by all of the previous examples---to get the most out of the package
you need to go _lazy_.
Lazy execution enables several benefits, but the most important is that it improves
performance. Delaying execution until the last possible moment allows Polars to
apply automatic optimization to every query. Let's take a quick look.

To create a so-called
"[LazyFrame](https://rpolars.github.io/reference/LazyFrame_class/)" from an
existing object in memory, we can invoke the `lazy()` constructor.

```{r}
ldat = dat$lazy()
ldat
```

Now consider what happens when we run our subsetting query from earlier on this
LazyFrame.

```{r}
subset_query = ldat$filter(
  pl$col("cyl") == 6
)$select(
  pl$col(c("mpg", "hp", "cyl"))
)

subset_query
```

Right now we only have a tree of instructions. But underneath the hood, 
Polars has already worked out a more optimized version of the query. We can
view this optimized plan this by requesting it.

```{r}
subset_query$describe_optimized_plan()
```

Here we see a simple, but surprisingly effective component in query
optimization: _projection_. Changing the order in which our subsetting
operations occurs---in this case, subsetting on columns first---reduces the
memory overhead of the overall query and leads to a downstream speedup. Of
course, you would hardly notice a difference for this small dataset. But the
same principles carry over to much bigger datasets and more complex queries.

To actually execute the plan, we just need to invoke the `collect()` method.
This should feel very familiar if you have previously used other lazy execution
engines like those provided by **arrow** or **dbplyr**.

```{r}
subset_query$collect()
```

## Data import

**polars** supports data import of both CSV and Parquet files formats. Here we
demonstrate using the `airquality` dataset that also comes bundled with base R.

```{r}
write.csv(airquality, "airquality.csv")

pl$read_csv("airquality.csv")
```

Again, however, the package works best if we take the lazy approach. 

```{r}
pl$scan_csv("airquality.csv")
```

We could obviously append a set of query operators to the above LazyFrame and
then collect the results. However, this workflow is even better suited to 
Parquet files, since we can leverage their efficient storage format 
on disk. Let's see an example.

```{r}
library(arrow)

write_parquet(airquality, "airquality.parquet")

# aq = read_parquet("airquality.parquet) # eager version (okay)
aq = pl$scan_parquet("airquality.parquet") # lazy version (better)

aq$filter(
  pl$col("Month") <= 6
)$group_by(
  "Month"
)$agg(
  pl$col(c("Ozone", "Temp"))$mean()
)$collect()
```

Finally, can read/scan multiple files in the same directory through pattern
globbing. However, please note that partition-aware scanning is not yet 
supported out of the box (e.g., Hive-style partitioned datasets). Follow
[this issue](https://github.com/pola-rs/polars/issues/4347) for more details
about when this will be resolved.

```{r}
dir.create("airquality-ds")
write_dataset(airquality, "airquality-ds", partitioning = "Month")

# Use pattern globbing to scan all parquet files in the folder
aq2 = pl$scan_parquet("airquality-ds/*/*.parquet")

# Just print the first two rows. But note that the Month column
# (which we used for partitioning) is missing.
aq2$limit(2)$collect()
```

Before continuing, don't forget to clean up by removing the newly created temp files
and directory on disk.

```{r}
file.remove(c("airquality.csv", "airquality.parquet"))
unlink("airquality-ds", recursive = TRUE)
```

## Execute R functions within a Polars query

It is possible to mix R code with Polars by passing R functions to **polars**.
This can unlock a lot of flexibility, but note that it can inhibit performance.
R functions are typically slower, so we recommend using native Polars functions
and expressions wherever possible.

```{r}
pl$DataFrame(iris)$select(
  pl$col("Sepal.Length")$map(\(s) { # map with a R function
    x = s$to_vector() # convert from Polars Series to a native R vector
    x[x >= 5] = 10
    x[1:10] # if return is R vector, it will automatically be converted to Polars Series again
  })
)$to_data_frame()
```

## Data types

Polars is
[strongly typed](https://en.wikipedia.org/wiki/Strong_and_weak_typing) and new
types can be created with the `dtypes` constructor. For example:

```{r}
pl$dtypes$Float64
```

The full list of valid Polars types can be found by typing `pl$dtypes` 
into your R console. These include _Boolean_, _Float32(64)_, _Int32(64)_,
_Utf8_, _Categorical_, _Date_, etc. Note that some type names differ from what
they are called in R (e.g., _Boolean_ in Polars is equivalent to `logical()` in
R). This might occasionally require you to look up a specific type. But the good
news is that **polars** generally does a good job of inferring types
automatically.