Six Things I Learned While Making `tidyfast`

13 Nov 2019

This post highlights six major themes of what I learned while creating the tidyfast R package. This process taught me about the tidyverse, data.table, R, and data science in general.

Before getting into that though, disclaimer: I am not part of the development teams of either the tidyverse or data.table. Instead, I am a big fan of both, and have profound gratitude for what they’ve allowed R to become. As such, my lessons are my opinions and don’t reflect the opinions of either group. And it should be noted that these are things I learned making a package that creates tidy-like functions with data.table. As such, most of the topics discussed will have to do with these packages.

TL;DR

The six lessons I learned:

  1. tidyverse and data.table have many similarities even though their overall frameworks are different
  2. Both the tidyverse and data.table are doing things to make it harder to mess up your data
  3. Data grammar, as explicitly stated in dplyr, is important
  4. Speed and efficiency, as shown by data.table, are important
  5. Both 3 and 4 are not mutually exclusive
  6. Nothing works for every situation, but a general workflow can apply to most situations

These lessons highlight the complimentarity and overlap between the powerful frameworks of the tidyverse and data.table.

Lesson 1: tidyverse and data.table have many similarities even though their overall frameworks are different

With the obvious differences between the two, I’m not sure what I was expecting here. But as I became more familiar with both, similarities became more and more clear. Consider some of those similarities below:

  • Both use an extension of data frames. The tidyverse has tibbles, which come with a cleaner and clearer printing method, is safer (doesn’t change types as easily), and allows informative attributes (e.g. grouped data). Similarly, data.table has a more informative printing method, is safer (again makes accidental changes of types harder), and has informative attributes (e.g. sorted by the key). As an example, the printing approaches are shown below:
library(tidyverse)
library(data.table)
library(nycflights13)

flights2 <- flights[, 1:6]
flights_tbl <- as_tibble(flights2)
flights_dt <- as.data.table(flights2)

flights_tbl
## # A tibble: 336,776 x 6
##     year month   day dep_time sched_dep_time dep_delay
##    <int> <int> <int>    <int>          <int>     <dbl>
##  1  2013     1     1      517            515         2
##  2  2013     1     1      533            529         4
##  3  2013     1     1      542            540         2
##  4  2013     1     1      544            545        -1
##  5  2013     1     1      554            600        -6
##  6  2013     1     1      554            558        -4
##  7  2013     1     1      555            600        -5
##  8  2013     1     1      557            600        -3
##  9  2013     1     1      557            600        -3
## 10  2013     1     1      558            600        -2
## # … with 336,766 more rows

flights_dt
##         year month day dep_time sched_dep_time dep_delay
##      1: 2013     1   1      517            515         2
##      2: 2013     1   1      533            529         4
##      3: 2013     1   1      542            540         2
##      4: 2013     1   1      544            545        -1
##      5: 2013     1   1      554            600        -6
##     ---                                                 
## 336772: 2013     9  30       NA           1455        NA
## 336773: 2013     9  30       NA           2200        NA
## 336774: 2013     9  30       NA           1210        NA
## 336775: 2013     9  30       NA           1159        NA
## 336776: 2013     9  30       NA            840        NA
  • Both rely on non-standard evaluation, making it easier to interact with variables without redundancies. For example, in the tidyverse, interacting with variables happens within functions that don’t require repetitive df$.. code or lots of quotes. data.table does this similarly, but within the data.table square brackes (e.g. dt[var == 1]). The form of non-standard evaluation does differ somewhat between the two (I’ll discuss this later). For now, consider the following examples showing their similar syntax:
# tidyverse
filter(flights_tbl, dep_delay > 100)
# data.table
flights_dt[dep_delay > 100]

  • Both have a way of piping/chaining commands. tidyverse uses pipes (%>%) while data.table has built in functionality with their square brackets ([]). Notably, pipes can be used with data.table but hasn’t seemed to be used a lot yet.

  • Both use either C or C++ to improve speed. data.table uses this for nearly all of the functionality, while tidyverse uses it often but not nearly as much.

  • Both have ways to interact with their non-standard evaluation programmatically. They go about it very differently, but both allow this to happen in various ways (e.g. `` operator in dplyr).

  • Both have a team of developers that work together in a synergistic fashion. These teams come with a range of experiences, expertise, and perspectives.

On the other hand, there are important differences in emphasis, style, and framework. These differences make the two, in my opinion, quite complimentary. Where one falls short, the other shines. And vice versa. Some of these differences will be highlighted in these next lessons discussed next. But here I wanted to highlight some notable differences.

  • tidyverse has recently emphasized type safety (see vctrs for example). This matches their overall style of being explicit in all behavior. This, at times, is at the expense of parsimony.

  • data.table emphasizes speed and efficiency in both performance and syntax. This creates flexibility and conciseness, but, at times, can lack explicit syntax. Most of the underlying behavior of data.table relies on their extensive library of C code that almost always creates opportunities to work with data that most other programs could not even begin to work with.

  • tidyverse works with a whole host of packages (possibly somewhat based on the idea of “conscious decoupling”) all designed for specific purposes. This makes it so these have several dependencies (dplyr currently as 11 imports and 25 suggests). data.table, on the other hand, is very self-contained with only 1 import (“methods”, which is always included in R) and 8 suggests. This idea is not as simple as just counting dependencies, however. To better understand the situation, I recommend watching Jim Hester’s talk.

  • Related to dependencies, the tidyverse has many (100’s) functions that are clear about their functionality. data.table has only a few functions total. Depending on your personal preferences, one may work better than the other for you. But keep in mind, and I discuss it more later on, that nearly all functionality in each can be replicated by the other (some of this is shown in dtplyr and tidyfast).

  • data.table uses a special operator that assigns by reference (or in place). This avoids a copy of the data, but with how R does shallow copies (only copies what is absolutely necessary), this has become less necessary. As such, the tidyverse does not use any modifications in place.

These are just some of the differences, and in several ways, it helps meet the needs of a variety of analysts with varying styles. Instead of a weakness in the R world, I see this as a great strength. This strength has been clearly demonstrated even more so with the collaboration between data.table and tidyverse teams on making dtplyr.

Lesson 2: Both the tidyverse and data.table are doing things to make it harder to mess up your data

In vctrs, Hadley and the tidyverse team show some crazy examples of ways R can really mess up. Consider:

c(factor("a"), factor("a"))
## [1] 1 1

If you are like me, your first reaction was a series of emojis: 🤔 😱 😞 😿 🤨. The tidyverse team has been working on ways to make sure things like this don’t happen, particularly with vctrs.

Importantly, though, both data.table and tidyverse have many safety checks to make sure one doesn’t mix data types already integrated into their frameworks. Along with these safety checks, they both have more informative errors, that can help shape how a user can fix any issues they encounter. Compared to base R errors, these are much more approachable.

Personally, I’m grateful for both working hard to make it easier to keep my data consistent and to work through errors.

Lesson 3: Data grammar is important

Being able to communicate about data wrangling across languages is important. It allows individuals to communicate methods, goals, and ideals, without regards to how these are actually done with code. For example, being able to discuss the steps needed to extract, manipulate, and clean data using SQL or R is often beneficial within an organization.

However, even more important is being able to communicate within a language. Recently, this has been a concern for some with the rise of the tidyverse, as it has profoundly changed the way users interact with R. The concern is that the use of R could become split, with some only using tidyverse syntax while others use more base R syntax (or data.table syntax).

This may become the case in years to come. I don’t know. But I’m not convinced that it is a problem. The thing that needs to be clear is the grammar of what is happening, regardless of the syntax used to perform it. Whether one is using all base R syntax or using pipes and other tidy functions, it is important that we can communicate what is actually being done with the data in concrete, accepted terminology. Because dplyr and friends already offer a grammar that has been widely accepted in the R community, it is probably smart to build on this and use, at least some, of the verbs offered there. This, whether explicitly used in function names or implicitly through base R or data.table syntax, is less important (although arguably relevant in some contexts). But the adoption of the grammar is what is important—to be able to communicate across modalities and styles.[1]

Lesson 4: Speed and efficiency are important

It is a common misconception that R is super slow, when in fact it can be incredibly fast. A little familiarization with data.table can be convincing pretty quickly that R is often as fast or faster than other data software. Just see the table on data.tables website. It clearly shows how well R stacks up to other major data software (particularly when using data.table).

In many situations, this performance is not just nice, it is necessary. A recent tweet by Ivan Leung highlighted one way an organization sees data work across different packages given they work with much more than 1 million lines of data.

Lesson 5: Grammar and speed/efficiency are not mutually exclusive

It may seem, at times, that you can have an explicit, easy-to-read syntax or speed/efficiency. Turns out, there is no such dichotomy. In fact, both can happen simultaneously. There are several examples:

  • Many functions in dplyr are both relatively fast (sometimes using Rcpp under the hood) and explicit in functionality.
  • tidyfast attempts to do have explicit functionality with other functions (e.g. nesting and unnesting) using data.table behind the scenes.
  • dtplyr makes use of this extensively by making many dplyr functions built on data.table.
  • data.tables fread() function is extremely fast and explicit in functionality.
  • data.tables dt[i, j, by] syntax can be very clear in function while being extremely fast.

Importantly, though, is that familiarity with data.table or base R (or any other syntax) makes many functions that don’t explicitly use the grammar just as clear as if they were. That is, the readability of code depends heavily on who is reading it. But to emphasize lesson 3, we should still be able to communicate what is happening with a shared grammar.

This leads us to our last lesson.

Lesson 6: Nothing works for every situation, but a general workflow can apply to most situations

As much as I’ve wanted a single framework to work well for every data situation that I find myself in, it just doesn’t seem to exist. Some situations lend themselves to one; and another to another. So what can one do? Learn everything?

I don’t think so. Instead, I think learning workflows that allows you to take messy, disconnected data to tidy, connected data is what will help in nearly all situations.

This relates to understanding the data grammar. With such understanding, you can search for the data verb you need to perform and the framework that would work best for your situation. This is ultimately what led me to start working on tidyfast. I wanted to see if a very “tidy” approach, that of nesting and unnesting data into and from list-columns could be done in a “non-tidy” package: data.table. It became clear that it could be done.

The functions found in tidyfast show that a common grammar (using the terms “nesting” and “unnesting” data into list-columns) was transferable and could be replicated with code that looks very different. Even though the functions in tidyfast are not nearly as well tested and used as those found in the tidyverse, it is an option that can be used in situations that call for it without changing the overall workflow or changing the data grammar that is used.

Miscellaneous Notes

Any general workflow should include safe-guards. This is particularly necessary when you are using a syntax you are less familiar with. These safe-guards can be:

  • including simple tests (e.g. assertthat) in your scripts that make sure things are working
  • using data approaches like nesting to keep analyses and cleaning within groups
  • avoiding excessive copying-and-pasting and, instead, rely more on functions and loops
  • read the documentation on any packages you are using

These can help one regardless of workflow preferences.

Conclusions

R is a beautiful mix of various styles of syntax that can provide functionality for nearly any data situation. Both tidyverse and data.table have much overlap in functionality but also offer vast complementarity. This provides great strength to the R community and adds flexibility to handle a wide range of data challenges.

Upcoming Post

In the next few weeks, I will post about a general workflow that can be used with dtplyr and tidyfast together. This will provide a general example that can be used when needed.

[1] Note, the dt[i, j, by] arguably is a grammar itself but is not often appreciated as such.

Related Posts