knitr::opts_chunk$set(collapse = TRUE, comment = "#>") options(tibble.print_min = 4L, tibble.print_max = 4L)
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("translate-verb") describes how entire verbs are translated.
dbplyr::translate_sql() powers translation of individual function calls, and I'll use it extensively in this vignette to show what's happening. You shouldn't need to use it ordinary code as dbplyr takes care of the translation automatically.
translate_sql((x + y) / 2)
translate_sql() takes an optional
con parameter. If not supplied, this causes dplyr to generate (approximately) SQL-92 compliant SQL. If supplied, dplyr uses
sql_translate_env() to look up a custom environment which makes it possible for different databases to generate slightly different SQL: see
vignette("new-backend") for more details. You can use the various simulate helpers to see the translations used by different backends:
translate_sql(x ^ 2L) translate_sql(x ^ 2L, con = simulate_sqlite()) translate_sql(x ^ 2L, con = simulate_access())
Perfect translation is not possible because databases don't have all the functions that R does. The goal of dplyr 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 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".
The following examples work through some of the basic differences between R and SQL.
'mean different things
# In SQLite variable names are escaped by double quotes: translate_sql(x) # And strings are escaped by single quotes translate_sql("x")
And some functions have different argument orders:
translate_sql(substr(x, 5, 10)) translate_sql(log(x, 10))
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
- basic math operators:
%/% to their SQL equivalents but note that they are 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.
df <- tibble( x = c(10L, 10L, -10L, -10L), y = c(3L, -3L, 3L, -3L) ) mf <- src_memdb() %>% copy_to(df, overwrite = TRUE) df %>% mutate(x %% y, x %/% y) mf %>% mutate(x %% y, x %/% y)
- logical comparisons:
- boolean operations:
All database provide translation for the basic aggregations:
var(). 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:
translate_sql(mean(x)) translate_sql(mean(x, na.rm = TRUE))
Note that, by default,
translate() assumes that the call is inside a
filter() and generates a window translation. If you want to see the equivalent
summarise()/aggregation translation, use
window = FALSE:
translate_sql(mean(x, na.rm = TRUE), window = FALSE)
switch() are translate to
translate_sql(if (x > 5) "big" else "small") translate_sql(switch(x, a = 1L, b = 2L, 3L))
- string functions:
- coerce types:
Any function that dplyr doesn't know how to convert is left as is. This means that database functions that are not covered by dplyr can be used directly via
translate_sql(). Here a couple of examples that will work with SQLite:
translate_sql(glob(x, y)) translate_sql(x %like% "ab%")
vignette("sql") for more details.
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 the ranking functions,
The partition clause specifies how the window function is broken down over groups. It plays an analogous role to
GROUP BYfor 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 preceeding 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.
knitr::include_graphics("windows.png", dpi = 300)
Of the many possible specifications, there are only three that commonly used. They select between aggregation variants:
BETWEEN UNBOUND PRECEEDING AND UNBOUND FOLLOWING
BETWEEN UNBOUND PRECEEDING AND CURRENT ROW
BETWEEN 2 PRECEEDING AND 2 FOLLOWING
dplyr generates the frame clause based on whether your using a recycled aggregate or a cumulative aggregate.
To see how individual window functions are translated to SQL, we can again use
translate_sql(mean(G)) translate_sql(rank(G)) translate_sql(ntile(G, 2)) translate_sql(lag(G))
If the tbl has been grouped or arranged previously in the pipeline, then dplyr will use that information to set the "partition by" and "order by" clauses. For interactive exploration, you can achieve the same effect by setting the
vars_order arguments to
translate_sql(cummean(G), vars_order = "year") translate_sql(rank(), vars_group = "ID")
There are some challenges when translating window functions between R and SQL, because dplyr 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:
ntile(y, 2). If omitted or
NULL, will use the default ordering associated with the tbl (as set by
Accumulating aggregates only take a single argument (the vector to aggregate). To control ordering, use
Aggregates implemented in dplyr (
last_value) have an
order_byargument. Supply it to override the default ordering.
The three options are illustrated in the snippet below:
mutate(players, min_rank(yearID), order_by(yearID, cumsum(G)), lead(G, order_by = yearID) )
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.