translation-function.Rmd

---
title: "Function translation"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Function translation}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r}
#| include: false
knitr::opts_chunk$set(collapse = TRUE, comment = "#>")
options(tibble.print_min = 6L, tibble.print_max = 6L, digits = 3)
```

There are two parts to dbplyr SQL translation: translating dplyr verbs, and translating expressions within those verbs. This vignette describes how individual expressions (function calls) are translated; `vignette("translation-verb")` describes how entire verbs are translated.

```{r}
library(dbplyr)
library(dplyr, warn.conflicts = FALSE)
```

## Getting started with translations

In this vignette, I'll use `lazy_frame()` to create a toy lazy table that allows us to see the translation without needing to connect to a real database:

```{r}
lf <- lazy_frame(x = 1, y = 2, g = "a")
lf |> mutate(z = (x + y) / 2)
```

The default `lazy_frame()` uses a generic database that generates (approximately) SQL-92 compliant SQL. You can use `simulate_*()` connections to see the translations used by different backends. Different databases generate slightly different SQL; see `vignette("new-backend")` for more details.

```{r}
lf_sqlite <- lazy_frame(x = 1, con = simulate_sqlite())
lf_access <- lazy_frame(x = 1, con = simulate_access())

lf_sqlite |> transmute(z = x^2)
lf_access |> transmute(z = x^2)
```

One key difference between dbplyr-generated SQL and hand-written SQL is that dbplyr always quotes all table and column names. This is verbose but necessary because column names in database tables can be any string, including SQL reserved words like `select` or `if`. Quoting all names ensures that dbplyr-generated SQL always works regardless of the table and column names involved.

In general, perfect translation is not possible because databases don't have all the functions that R does. The goal of dbplyr is to provide a semantic rather than a literal translation: what you mean, rather than precisely what is done. In fact, even for functions that exist both in databases and in R, you shouldn't expect results to be identical; database programmers have different priorities than R core programmers. For example, in R in order to get a higher level of numerical accuracy, `mean()` loops through the data twice. R's `mean()` also provides a `trim` option for computing trimmed means; this is something that databases do not provide. 

If you're interested in how `translate_sql()` is implemented, the basic techniques that underlie the implementation of `translate_sql()` are described in ["Advanced R"](https://adv-r.hadley.nz/translation.html). 

## Basic differences

There are two fundamental differences between R and SQL:

*   `"` and `'` mean different things. 
    R can use either `"` or `'` for strings, but in ANSI SQL, must be `"` used for names and must be `'` used for strings.

    ```{r}
    lf |> filter(x == "x")
    ```

*   R and SQL have different defaults for integers and reals.
    In R, 1 is a real, and 1L is an integer. In SQL, 1 is an integer, and 1.0 is a real.

    ```{r}
    lf |> transmute(z = 1)
    lf |> transmute(z = 1L)
    ```

## Known functions

### Mathematics

* basic math operators: `+`, `-`, `*`, `/`, `^`
* trigonometry: `acos()`, `asin()`, `atan()`, `atan2()`, `cos()`, `cot()`, `tan()`, `sin()`
* hyperbolic: `cosh()`, `coth()`, `sinh()`, `tanh()`
* logarithmic: `log()`, `log10()`, `exp()`
* misc: `abs()`, `ceiling()`, `floor()`, `sqrt()`, `sign()`, `round()`

```{r}
lf |> transmute(x = x / 2, y = x^2 + y^2)

lf |> transmute(x = log(x), y = round(y, 1))
```

### Modulo arithmetic

dbplyr translates `%%` to the SQL equivalents but note that it's not precisely the same: most databases use truncated division where the modulo operator takes the sign of the dividend, where R using the mathematically preferred floored division with the modulo sign taking the sign of the divisor.

```{r}
df <- tibble(
  x = c(10L, 10L, -10L, -10L),
  y = c(3L, -3L, 3L, -3L)
)
db <- copy_to(memdb(), df)

df |> mutate(x %% y)
db |> mutate(x %% y)
```

dbplyr no longer translates `%/%` because there's no robust cross-database translation available.

### Logical comparisons and boolean operations

* logical comparisons: `<`, `<=`, `!=`, `>=`, `>`, `==`, `%in%`, `between()`
* boolean operations: `&`, `&&`, `|`, `||`, `!`, `xor()`

```{r}
lf |> filter(x > 5 | y == 2)

lf |> filter(x %in% c(1, 2, 3))

lf |> filter(between(x, 1, 5))
```

### Bitwise operations

`bitwNot()`, `bitwAnd()`, `bitwOr()`, `bitwXor()`, `bitwShiftL()`, and `bitwShiftR()` are all supported:

```{r}
lf |> transmute(x = bitwAnd(x, 3L), y = bitwShiftL(x, 2L))
```

### Type coercion

Type coercion functions use the corresponding SQL `CAST()` call:

```{r}
lf |> transmute(x = as.integer(y), y = as.character(x))
```

* integer types: `as.integer()`, `as.integer64()`
* floating point: `as.numeric()`, `as.double()`
* character: `as.character()`
* logical: `as.logical()`
* date/time: `as.Date()`, `as.POSIXct()`

For database-specific types not covered by these functions, use `as()`:

```{r}
lf |> transmute(x = as(x, "TIME"), y = as(y, "DECIMAL(10, 2)"))
```

### `NULL`/`NA` handling

* `is.na()`, `is.null()`: test for `NULL`.
* `na_if()`: replace a value with `NULL`.
* `coalesce()`: replace `NULL` with a default value.

```{r}
lf |> filter(!is.na(x))

lf |> transmute(x = coalesce(x, 0L))

lf |> transmute(x = na_if(x, 0L))
```

### Aggregation

All databases provide translation for the basic aggregations: `mean()`, `sum()`, `min()`, `max()`. Databases automatically drop NULLs (their equivalent of missing values) whereas in R you have to ask nicely. The aggregation functions warn you about this important difference:

```{r}
lf |> summarise(z = mean(x))
lf |> summarise(z = mean(x, na.rm = TRUE))
```

Note that aggregation functions used inside `mutate()` or `filter()` generate a window translation:

```{r}
lf |> mutate(z = mean(x, na.rm = TRUE))
lf |> filter(mean(x, na.rm = TRUE) > 0)
```

Most backends also support:

* `sd()`, `var()`, `cor()`, `cov()`
* `median()`, `quantile()`
* `n()`, `n_distinct()`
* `all()`, `any()`
* `str_flatten()`

### Conditional evaluation

`if`, `ifelse()`, and `if_else()` are translated to `CASE WHEN`:

```{r}
lf |> transmute(z = ifelse(x > 5, "big", "small"))
```

`case_when()`, `case_match()`, and `switch()` are also supported:

```{r}
lf |> 
  mutate(z = case_when(
    x > 10 ~ "medium",
    x > 30 ~ "big", 
    .default = "small"
  ))

lf |> mutate(z = switch(g, a = 1L, b = 2L, 3L))
```

### String functions

Base R string functions and their stringr equivalents are widely supported:

* `nchar()`, `str_length()`
* `tolower()`, `toupper()`, `str_to_lower()`, `str_to_upper()`, `str_to_title()`
* `trimws()`, `str_trim()`
* `paste()`, `paste0()`, `str_c()`
* `substr()`, `substring()`, `str_sub()`

```{r}
lf |> transmute(x = paste0(g, " dog"))

lf |> transmute(x = substr(g, 1L, 2L))
```

Many backends also support regular expression functions like `str_detect()`, `str_replace()`, `str_replace_all()`, `str_remove()`, `str_remove_all()`, `str_squish()`, and `str_like()`. Support varies by backend; see the individual backend documentation for details.

### Date/time functions

dbplyr supports many lubridate functions for extracting date components:

* `today()`, `now()`
* `year()`, `month()`, `day()`, `mday()`, `hour()`, `minute()`, `second()`

```{r}
lf_dt <- lazy_frame(dt = Sys.time())

lf_dt |> transmute(
  year = year(dt),
  month = month(dt),
  day = day(dt)
)
```

Some backends also support additional lubridate functions including `yday()`, `wday()`, `week()`, `isoweek()`, `quarter()`, `isoyear()`, `floor_date()`, and period functions like `seconds()`, `minutes()`, `hours()`, `days()`, `weeks()`, `months()`, `years()`.

Several backends (including PostgreSQL, Snowflake, SQL Server, Redshift, and Spark SQL) support [clock](https://clock.r-lib.org) functions for date arithmetic. 

* `add_days()`, `add_years()`
* `date_build()`
* `get_year()`, `get_month()`, `get_day()`
* `date_count_between()`
* `difftime()`

clock functions tend to be easier to translate than lubridate functions because they are more specific.

### Other functions

* `pmin()`, `pmax()` for parallel min/max
* `desc()` for descending order
* `cut()` for binning numeric values into categories

## Unknown functions

Any function that dbplyr doesn't know how to convert is left as is. This means that database functions that are not covered by dbplyr can often be used directly.

### Prefix functions

Any function that dbplyr doesn't know about will be left as is:

```{r}
lf |> mutate(z = foofify(x, y))
```

But to make it clear that you're deliberately calling a SQL function, we recommend using the `.sql` pronoun:

```{r}
lf |> transmute(z = .sql$foofify(x, y))
```

If you're working inside a package, this also makes it easier to avoid `R CMD CHECK` notes. Just import `.sql` from dbplyr using a roxygen2 tag like `@importFrom dbplyr .sql`

### Infix functions

As well as prefix functions (where the name of the function comes before the arguments), dbplyr also translates infix functions. That allows you to use expressions like `LIKE`, which does a limited form of pattern matching:

```{r}
lf |> filter(x %LIKE% "%foo%")
```

You can also use `str_like()` for this common case:

```{r}
lf |> filter(str_like(x, "%foo%"))
```

You could use `%||%` for string concatenation, but in most cases it's more R-like to use `paste()` or `paste0()`:

```{r}
lf |> transmute(z = x %||% y)
lf |> transmute(z = paste0(x, y))
lf |> transmute(z = paste(x, y))
```

### Special forms

SQL functions tend to have a greater variety of syntax than R. That means there are a number of expressions that can't be translated directly from R code. To insert these in your own queries, you can use literal SQL inside `sql()`:

```{r}
lf |> transmute(z = sql("x!"))
lf |> transmute(z = x == sql("ANY VALUES(1, 2, 3)"))
```

This gives you a lot of freedom to generate the SQL you need:

```{r}
lf |> transmute(factorial = sql("x!"))
lf |> transmute(factorial = sql("CAST(x AS FLOAT)"))
```

### Error for unknown translations

If needed, you can also use the `dplyr.strict_sql` option to force dbplyr to error if it doesn't know how to translate a function:

```{r}
#| error = TRUE
options(dplyr.strict_sql = TRUE)
lf |> mutate(z = glob(x, y))
```

## Window functions

Things get a little trickier with window functions, because SQL's window functions are considerably more expressive than the specific variants provided by base R or dplyr. They have the form `[expression] OVER ([partition clause] [order clause] [frame_clause])`:

* The __expression__ is a combination of variable names and window functions.
  Support for window functions varies from database to database, but most
  support:

  - ranking: `row_number()`, `min_rank()`, `rank()`, `dense_rank()`,
    `percent_rank()`, `cume_dist()`, `ntile()`;
  - offsets: `lead()`, `lag()`, `first()`, `last()`, `nth()`;
  - aggregates: `mean()`, `sum()`, `min()`, `max()`, `n()`, `n_distinct()`;
  - cumulative: `cummean()`, `cumsum()`, `cummin()`, `cummax()`.

* The __partition clause__ specifies how the window function is broken down
  over groups. It plays an analogous role to `GROUP BY` for aggregate functions,
  and `group_by()` in dplyr. It is possible for different window functions to 
  be partitioned into different groups, but not all databases support it, and
  neither does dplyr.
  
*   The __order clause__ controls the ordering (when it makes a difference).
    This is important for the ranking functions since it specifies which 
    variables to rank by, but it's also needed for cumulative functions and lead.
    Whenever you're thinking about before and after in SQL, you must always tell 
    it which variable defines the order. If the order clause is missing when 
    needed, some databases fail with an error message while others return 
    non-deterministic results.
  
*   The __frame clause__ defines which rows, or __frame__, that are passed 
    to the window function, describing which rows (relative to the current row)
    should be included. The frame clause provides two offsets which determine
    the start and end of frame. There are three special values: -Inf means
    to include all preceding rows (in SQL, "unbounded preceding"), 0 means the
    current row ("current row"), and Inf means all following rows ("unbounded
    following"). The complete set of options is comprehensive, but fairly 
    confusing, and is summarised visually below.

    ```{r}
    #| echo: false
    #| out-width: 100%
    #| fig-alt: >
    #|   A visual summary of the frame clause using the real line labelled
    #|   with negative infinity, -3, -2, -1, 0, 1, 2, 3, infinity. The most
    #|   important clauses are rolling, cumulative, and recycling. 
    #|   Rolling, e.g. between 1 preceding and 1, following, run from 
    #|   -1 to -1. Cumulative, between unbounded preceding and 
    #|   current row, runs from negative infinity to 0. Recycled, 
    #|   between unbound preceeding and unbound following, runs from 
    #|   negative infinity to positive infinity.
    knitr::include_graphics("windows.png", dpi = 300)
    ```

    Of the many possible specifications, only three are commonly
    used. They select between aggregation variants:

    * Recycled: `BETWEEN UNBOUND PRECEDING AND UNBOUND FOLLOWING`
    
    * Cumulative: `BETWEEN UNBOUND PRECEDING AND CURRENT ROW`
    
    * Rolling: `BETWEEN 2 PRECEDING AND 2 FOLLOWING`
    
    dbplyr generates the frame clause based on whether you're using a recycled
    aggregate or a cumulative aggregate.
    
To see how individual window functions are translated to SQL, we can use `transmute()`:

```{r}
lf <- lazy_frame(g = 1, year = 2020, id = 3, con = simulate_dbi())

lf |> transmute(
  mean = mean(g), 
  rank = min_rank(g), 
  cumsum = cumsum(g),
  lag = lag(g)
)
```

If the lazy frame has been grouped or arranged previously in the pipeline, then dbplyr will use that information to set the "partition by" and "order by" clauses:

```{r}
lf |> arrange(year) |> mutate(z = cummean(g))
lf |> group_by(id) |> mutate(z = rank())
```

There are some challenges when translating window functions between R and SQL, because dbplyr tries to keep the window functions as similar as possible to both the existing R analogues and to the SQL functions. This means that there are three ways to control the order clause depending on which window function you're using:

* For ranking functions, the ordering variable is the first argument: `rank(x)`,
  `ntile(y, 2)`. If omitted or `NULL`, will use the default ordering associated 
  with the tbl (as set by `arrange()`).

* Accumulating aggregates only take a single argument (the vector to aggregate).
  To control ordering, use `order_by()`.
  
* Aggregates implemented in dplyr (`lead()`, `lag()`, `nth()`, `first()`,
  `last()`) have an `order_by` argument. Supply it to override the
  default ordering.

The three options are illustrated in the snippet below:

```{r}
lf |> transmute(
  x1 = min_rank(g),
  x2 = order_by(year, cumsum(g)),
  x3 = lead(g, order_by = year)
)
```

Currently there is no way to order by multiple variables, except by setting the default ordering with `arrange()`. This will be added in a future release.