Order of operations

In the “Get Started” vignette, we mostly used eager execution. This is the classic type of execution in R: write some code, link functions with pipes, run the code and it is executed line by line.

Most of the time, this kind of execution is perfectly fine. With datasets of a reasonable size (let’s say up to a few hundreds of thousands of observations and a few dozens columns), we don’t really have to worry about whether our code is optimized to save memory and time.

However, when we start dealing with much larger datasets, the order in which functions are applied is extremely important. Indeed, some functions are much more memory intensive than others.

For instance, let’s say we have a dataset with 1M observations containing country names and a few numeric columns. We would like to keep just a few of these countries and sort them alphabetically. Here, we have two operations: filtering and sorting. Sorting is much harder to do than filtering. To filter data, we simply check whether each row fills some conditions, but to sort data we have to compare rows between them and rearrange them as we go. If we don’t take this into account, sorting data before filtering it can slow down our pipeline significantly.

countries = c(
  "France", "Germany", "United Kingdom", "Japan", "Columbia",
  "South Korea", "Vietnam", "South Africa", "Senegal", "Iran"
)

set.seed(123)
test = data.frame(
  country = sample(countries, 1e6, TRUE),
  x = sample(1:100, 1e6, TRUE),
  y = sample(1:1000, 1e6, TRUE)
)

bench::mark(
  sort_filter = {
    tmp = test[order(test$country), ]
    subset(tmp, country %in% c("United Kingdom", "Japan", "Vietnam"))
  },
  filter_sort = {
    tmp = subset(test, country %in% c("United Kingdom", "Japan", "Vietnam"))
    tmp[order(tmp$country), ]
  }
)

#> Warning: Some expressions had a GC in every iteration; so filtering is
#> disabled.
#> # A tibble: 2 × 6
#>   expression       min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>  <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 sort_filter    10.4s    10.4s    0.0964    87.2MB    0.193
#> 2 filter_sort     2.4s     2.4s    0.417     67.1MB    0.835

How does polars help?

We have seen that the order in which functions are applied matters a lot. But it already takes a long time to deal with large data, we’re not going to spend even more time trying to optimize our pipeline, right?

This is where lazy execution comes into play. The idea is that we write our code as usual, but this time, we won’t apply it directly on a dataset but on a lazy dataset, i.e a dataset that is not loaded in memory yet (in polars terms, these are DataFrames and LazyFrames). Once our code is ready, we call collect() at the end of the pipeline. Before executing our code, polars will internally check whether it can be optimized, for example by reordering some operations.

Let’s re-use the example above but this time with polars syntax and 10M observations. For the purpose of this vignette, we can create a LazyFrame directly in our session, but if the data was stored in a CSV file for instance, we would have to scan it first with pl$scan_csv():

library(polars)

set.seed(123)
df_test = pl$DataFrame(
  country = sample(countries, 1e7, TRUE),
  x = sample(1:100, 1e7, TRUE),
  y = sample(1:1000, 1e7, TRUE)
)

lf_test = df_test$lazy()

Now, we can convert the base R code above to a polars query:

df_test$
  sort(pl$col("country"))$
  filter(
  pl$col("country")$is_in(pl$lit(c("United Kingdom", "Japan", "Vietnam")))
)
#> shape: (3_000_835, 3)
#> ┌─────────┬─────┬─────┐
#> │ country ┆ x   ┆ y   │
#> │ ---     ┆ --- ┆ --- │
#> │ str     ┆ i32 ┆ i32 │
#> ╞═════════╪═════╪═════╡
#> │ Japan   ┆ 97  ┆ 7   │
#> │ Japan   ┆ 96  ┆ 672 │
#> │ Japan   ┆ 17  ┆ 710 │
#> │ Japan   ┆ 68  ┆ 41  │
#> │ Japan   ┆ 100 ┆ 109 │
#> │ …       ┆ …   ┆ …   │
#> │ Vietnam ┆ 89  ┆ 352 │
#> │ Vietnam ┆ 62  ┆ 8   │
#> │ Vietnam ┆ 52  ┆ 988 │
#> │ Vietnam ┆ 85  ┆ 982 │
#> │ Vietnam ┆ 74  ┆ 692 │
#> └─────────┴─────┴─────┘

This works for the DataFrame, that uses eager execution. For the LazyFrame, we can write the same query:

lazy_query = lf_test$
  sort(pl$col("country"))$
  filter(
  pl$col("country")$is_in(pl$lit(c("United Kingdom", "Japan", "Vietnam")))
)

lazy_query
#> polars LazyFrame
#>  $explain(): Show the optimized query plan.
#> 
#> Naive plan:
#> FILTER col("country").is_in([Series]) FROM
#>   SORT BY [col("country")]
#>     DF ["country", "x", "y"]; PROJECT */3 COLUMNS; SELECTION: None

However, this doesn’t do anything to the data until we call collect() at the end. We can now compare the two approaches (in the lazy timing, calling collect() both reads the data and process it, so we include the data loading part in the eager timing as well):

bench::mark(
  eager = df_test$
    sort(pl$col("country"))$
    filter(
      pl$col("country")$is_in(pl$lit(c("United Kingdom", "Japan", "Vietnam")))
    ),
  lazy = lazy_query$collect(),
  iterations = 10
)

#> # A tibble: 2 × 6
#>   expression      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 eager         1.18s    1.28s     0.754    9.95KB        0
#> 2 lazy       526.98ms 577.35ms     1.72       864B        0

On this very simple query, using lazy execution instead of eager execution lead to a 1.7-2.2x decrease in time.

So what happened? Under the hood, polars reorganized the query so that it filters rows while reading the csv into memory, and then sorts the remaining data. This can be seen by comparing the original query with the optimized one:

lazy_query$explain(optimized = FALSE) |>
  cat()
#> FILTER col("country").is_in([Series]) FROM
#>   SORT BY [col("country")]
#>     DF ["country", "x", "y"]; PROJECT */3 COLUMNS; SELECTION: None

lazy_query$explain() |>
  cat()
#> SORT BY [col("country")]
#>   DF ["country", "x", "y"]; PROJECT */3 COLUMNS; SELECTION: col("country").is_in([Series])

Note that queries must be read from bottom to top, i.e the optimized query is “select the dataset where the column ‘country’ matches these values, then sort the data by the values of ‘country’”.