An introduction to seplyr

December 14, 2017
By

(This article was first published on Revolutions, and kindly contributed to R-bloggers)

by John Mount, Win-Vector LLC

seplyr is an R package
that supplies improved standard evaluation interfaces for many common data wrangling tasks.

The core of seplyr is a
re-skinning of dplyr's
functionality to seplyr conventions (similar to how stringr
re-skins the implementing package stringi).

Standard Evaluation and Non-Standard Evaluation

"Standard evaluation" is the name we are using for the value oriented calling convention found in many programming languages. The idea is: functions are only allowed to look at the values of their arguments and not how those values arise (i.e., they can not look at source code or variable names). This evaluation principle allows one to transform, optimize, and reason about code.

It is what lets us say the following two snippets of code are equivalent.

  • x <- 4; sqrt(x)
  • x <- 4; sqrt(4)

The mantra is:

"variables can be replaced with their values."

Which is called referential transparency.

"Non-standard evaluation" is the name used for code that more aggressively inspects its environment. It is often used for harmless tasks such as conveniently setting axis labels on plots. For example, notice the following two plots have different y-axis labels (despite plotting identical values).

plot(x = 1:3)

Plot1

plot(x = c(1,2,3))

Plot2

dplyr and seplyr

The dplyr authors appear to strongly prefer a non-standard evaluation interface. Many in the dplyr community have come to think a package such as dplyr requires a non-standard interface. seplyr started as an experiment to show this is not actually the case.

Syntactically the packages are deliberately similar.

We can take a dplyr pipeline:

suppressPackageStartupMessages(library("dplyr"))

starwars %>%
  select(name, height, mass) %>%
  arrange(desc(height)) %>%
  head()

## # A tibble: 6 x 3
##           name height  mass
##             
## 1  Yarael Poof    264    NA
## 2      Tarfful    234   136
## 3      Lama Su    229    88
## 4    Chewbacca    228   112
## 5 Roos Tarpals    224    82
## 6     Grievous    216   159

And re-write it in seplyr notation:

library("seplyr")

starwars %.>%
  select_se(., c("name", "height", "mass")) %.>%
  arrange_se(., "desc(height)") %.>%
  head(.)

## # A tibble: 6 x 3
##           name height  mass
##             
## 1  Yarael Poof    264    NA
## 2      Tarfful    234   136
## 3      Lama Su    229    88
## 4    Chewbacca    228   112
## 5 Roos Tarpals    224    82
## 6     Grievous    216   159

For the common dplyr-verbs (excluding mutate(), which we will discuss next) all the non-standard evaluation is saving us is a few quote marks and array designations (and we have ways of getting rid of the need for quote marks). In exchange for this small benefit the non-standard evaluation is needlessly hard to program over. For instance in the seplyr pipeline it is easy to accept the list of columns from an outside source as a simple array of names.

Until you introduce a substitution system such as rlang or wrapr::let() (which we recommend over rlang and publicly pre-dates the public release of rlang) you have some difficulty writing re-usable programs that use the dplyr verbs over "to be specified later" column names.

We are presumably not the only ones who considered this a limitation:

Github

seplyr is an attempt to make programming a primary concern by
making the value-oriented (standard) interfaces the primary interfaces.

mutate()

The earlier "standard evaluation costs just a few quotes" becomes a bit strained when we talk about the dplyr::mutate() operator. It doesn't seem worth the effort unless you get something more in return. In seplyr 0.5.0 we introduced "the something more": planning over and optimizing dplyr::mutate() sequences.

A seplyr mutate looks like the following:

  select_se(., c("name", "height", "mass")) %.>%
  mutate_se(., c(
    "height" := "height + 1",
    "mass" := "mass + 1",
    "height" := "height + 2",
    "mass" := "mass + 2",
    "height" := "height + 3",
    "mass" := "mass + 3"
  )) %.>%
  arrange_se(., "name") %.>%
  head(.)

## # A tibble: 6 x 3
##                  name height  mass
##                    
## 1              Ackbar    186    89
## 2          Adi Gallia    190    56
## 3    Anakin Skywalker    194    90
## 4        Arvel Crynyd     NA    NA
## 5         Ayla Secura    184    61
## 6 Bail Prestor Organa    197    NA

seplyr::mutate_se() always uses ":=" to denote assignment (dplyr::mutate() prefers "=" for assignment, except in cases where ":=" is required).

The advantage is: once we are go to the trouble to capture the mutate expressions we can treat them as data and apply procedures to them. For example we can re-group and optimize the mutate assignments.

plan <- partition_mutate_se(
  c("name" := "tolower(name)",
    "height" := "height + 0.5",
    "height" := "floor(height)",
    "mass" := "mass + 0.5",
    "mass" := "floor(mass)"))
print(plan)

## $group00001
##            name          height            mass 
## "tolower(name)"  "height + 0.5"    "mass + 0.5" 
## 
## $group00002
##          height            mass 
## "floor(height)"   "floor(mass)"

Notice seplyr::partition_mutate_se() re-ordered and re-grouped the assignments so that:

  • In each group each value used is independent of values produced in other assignments.
  • All dependencies between assignments are respected by the group order.

The "safe block" assignments can then be used in a pipeline:

starwars %.>%
  select_se(., c("name", "height", "mass")) %.>%
  mutate_seb(., plan) %.>%
  arrange_se(., "name") %.>%
  head(.)

## # A tibble: 6 x 3
##                  name height  mass
##                    
## 1              ackbar    180    83
## 2          adi gallia    184    50
## 3    anakin skywalker    188    84
## 4        arvel crynyd     NA    NA
## 5         ayla secura    178    55
## 6 bail prestor organa    191    NA

This may not seem like much. However, when using dplyr with a SQL database (such as PostgreSQL or even Sparklyr) keeping the number of dependencies in a block low is critical for correct calculation (which is why I recommend keeping dependencies low). Furthermore, on Sparklyr sequences of mutates are simulated by nesting of SQL statements, so you must also keep the number of mutates at a moderate level (i.e., you want a minimal number of blocks or groups).

Machine Generated Code

Because we are representing mutate assignments as user manipulable data we can also enjoy the benefit of machine generated code. seplyr 0.5.* uses this opportunity to introduce a simple function named if_else_device(). This device uses R's ifelse() statement (which conditionally chooses values in a vectorized form) to implement a more powerful block-if/else statement (which conditionally simultaneously controls blocks of values and assignments; SAS has such a feature).

For example: suppose we want to NA-out one of height or mass for each row of the starwars data uniformly at random. This can be written naturally using the if_else_device.

if_else_device(
  testexpr = "runif(n())>=0.5",
  thenexprs = "height" := "NA",
  elseexprs = "mass" := "NA")

##                           ifebtest_30etsitqqutk 
##                               "runif(n())>=0.5" 
##                                          height 
##    "ifelse( ifebtest_30etsitqqutk, NA, height)" 
##                                            mass 
## "ifelse( !( ifebtest_30etsitqqutk ), NA, mass)"

Notice the if_else_device translates the user code into a sequence of dplyr::mutate() expressions (using only the weaker operator ifelse()). Obviously the user could perform this translation, but if_else_device automates the record keeping and can even be nested. Also many such steps can be chained together and broken into a minimal sequence of blocks by partition_mutate_se() (not forcing a new dplyr::mutate() step for each if-block encountered).

When we combine the device with the partitioned we get performant database-safe code where the number of blocks is only the level of variable dependence (and not the possibly much larger number of initial value uses that a straightforward non-reordering split would give; note: seplyr::mutate_se() 0.5.1 and later incorporate the partition_mutate_se() in mutate_se()).

starwars %.>%
  select_se(., c("name", "height", "mass")) %.>%
  mutate_se(., if_else_device(
    testexpr = "runif(n())>=0.5",
    thenexprs = "height" := "NA",
    elseexprs = "mass" := "NA")) %.>%
  arrange_se(., "name") %.>%
  head(.)

## # A tibble: 6 x 4
##                  name height  mass ifebtest_wwr9k0bq4v04
##                                     
## 1              Ackbar     NA    83                  TRUE
## 2          Adi Gallia    184    NA                 FALSE
## 3    Anakin Skywalker     NA    84                  TRUE
## 4        Arvel Crynyd     NA    NA                  TRUE
## 5         Ayla Secura    178    NA                 FALSE
## 6 Bail Prestor Organa    191    NA                 FALSE

Conclusion

The value oriented notation is a bit clunkier, but this is offset by it's greater
flexibility in terms of composition and working parametrically.

Our group has been using seplyr::if_else_device() and seplyr::partition_mutate_se() to greatly simplify porting powerful SAS procedures to R/Sparklyr/Apache Spark clusters.

This is new code, but we are striving to supply sufficient initial documentation and examples.

To leave a comment for the author, please follow the link and comment on their blog: Revolutions.

R-bloggers.com offers daily e-mail updates about R news and tutorials on topics such as: Data science, Big Data, R jobs, visualization (ggplot2, Boxplots, maps, animation), programming (RStudio, Sweave, LaTeX, SQL, Eclipse, git, hadoop, Web Scraping) statistics (regression, PCA, time series, trading) and more...



If you got this far, why not subscribe for updates from the site? Choose your flavor: e-mail, twitter, RSS, or facebook...

Comments are closed.

Search R-bloggers


Sponsors

Never miss an update!
Subscribe to R-bloggers to receive
e-mails with the latest R posts.
(You will not see this message again.)

Click here to close (This popup will not appear again)