Databases using dplyr
As well as working with local in-memory data stored in data frames,
dplyr also works with remote on-disk data stored in databases. This is particularly useful in two scenarios:
Your data is already in a database.
You have so much data that it does not all fit into memory simultaneously and you need to use some external storage engine.
(If your data fits in memory, there is no advantage to putting it in a database; it will only be slower and more frustrating.)
This vignette focuses on the first scenario because it is the most common. If you are using R to do data analysis inside a company, most of the data you need probably already lives in a database (it’s just a matter of figuring out which one!). However, you will learn how to load data in to a local database in order to demonstrate
dplyr’s database tools. At the end, I’ll also give you a few pointers if you do need to set up your own database.
To use databases with
dplyr, you need to first install
You’ll also need to install a DBI backend package. The
DBI package provides a common interface that allows
dplyr to work with many different databases using the same code.
DBI is automatically installed with
dbplyr, but you need to install a specific backend for the database that you want to connect to.
Five commonly used backends are:
RMySQL connects to MySQL and MariaDB
RPostgreSQL connects to Postgres and Redshift.
RSQLite embeds a SQLite database.
odbc connects to many commercial databases via the open database connectivity protocol.
bigrquery connects to Google’s BigQuery.
If the database you need to connect to is not listed here, you’ll need to do some investigation yourself.
In this vignette, we’re going to use the
RSQLite backend, which is automatically installed when you install
dbplyr. SQLite is a great way to get started with databases because it’s completely embedded inside an R package. Unlike most other systems, you don’t need to set up a separate database server. SQLite is great for demos, but is surprisingly powerful, and with a little practice you can use it to easily work with many gigabytes of data.
Connecting to the database
To work with a database in
dplyr, you must first connect to it, using
DBI::dbConnect(). We’re not going to go into the details of the
DBI package here, but it’s the foundation upon which
dbplyr is built. You’ll need to learn more about if you need to do things to the database that are beyond the scope of
library(dplyr) con <- DBI::dbConnect(RSQLite::SQLite(), path = ":dbname:")
The arguments to
DBI::dbConnect() vary from database to database, but the first argument is always the database backend. It’s
RSQLite::SQLite() for RSQLite,
RMySQL::MySQL() for RMySQL,
RPostgreSQL::PostgreSQL() for RPostgreSQL,
odbc::odbc() for odbc, and
bigrquery::bigquery() for BigQuery. SQLite only needs one other argument: the path to the database. Here we use the special string,
":memory:", which causes SQLite to make a temporary in-memory database.
Most existing databases don’t live in a file, but instead live on another server. In real life that your code will look more like this:
con <- DBI::dbConnect(RMySQL::MySQL(), host = "database.rstudio.com", user = "hadley", password = rstudioapi::askForPassword("Database password") )
(If you’re not using RStudio, you’ll need some other way to securely retrieve your password. You should never record it in your analysis scripts or type it into the console.)
Our temporary database has no data in it, so we’ll start by copying over
nycflights13::flights using the convenient
copy_to() function. This is a quick and dirty way of getting data into a database and is useful primarily for demos and other small jobs.
copy_to(con, nycflights13::flights, "flights", temporary = FALSE, indexes = list( c("year", "month", "day"), "carrier", "tailnum", "dest" ) )
As you can see, the
copy_to() operation has an additional argument that allows you to supply indexes for the table. Here we set up indexes that will allow us to quickly process the data by day, carrier, plane, and destination. Creating the write indices is key to good database performance, but is unfortunately beyond the scope of this article.
Now that we’ve copied the data, we can use
tbl() to take a reference to it:
flights_db <- tbl(con, "flights")
When you print it out, you’ll notice that it mostly looks like a regular tibble:
flights_db #> # Source: table<flights> [?? x 19] #> # Database: sqlite 3.19.3  #> year month day dep_time sched_dep_time dep_delay arr_time #> <int> <int> <int> <int> <int> <dbl> <int> #> 1 2013 1 1 517 515 2. 830 #> 2 2013 1 1 533 529 4. 850 #> 3 2013 1 1 542 540 2. 923 #> 4 2013 1 1 544 545 -1. 1004 #> 5 2013 1 1 554 600 -6. 812 #> 6 2013 1 1 554 558 -4. 740 #> # ... with more rows, and 12 more variables: sched_arr_time <int>, #> # arr_delay <dbl>, carrier <chr>, flight <int>, tailnum <chr>, #> # origin <chr>, dest <chr>, air_time <dbl>, distance <dbl>, hour <dbl>, #> # minute <dbl>, time_hour <dbl>
The main difference is that you can see that it’s a remote source in a SQLite database.
To interact with a database you usually use SQL, the Structured Query Language. SQL is over 40 years old, and is used by pretty much every database in existence. The goal of
dbplyr is to automatically generate SQL for you so that you’re not forced to use it. However, SQL is a very large language, and
dbplyr doesn’t do everything. It focuses on
SELECT statements, the SQL you write most often as an analyst.
Most of the time you don’t need to know anything about SQL, and you can continue to use the
dplyr verbs that you’re already familiar with:
flights_db %>% select(year:day, dep_delay, arr_delay) #> # Source: lazy query [?? x 5] #> # Database: sqlite 3.19.3  #> year month day dep_delay arr_delay #> <int> <int> <int> <dbl> <dbl> #> 1 2013 1 1 2. 11. #> 2 2013 1 1 4. 20. #> 3 2013 1 1 2. 33. #> 4 2013 1 1 -1. -18. #> 5 2013 1 1 -6. -25. #> 6 2013 1 1 -4. 12. #> # ... with more rows flights_db %>% filter(dep_delay > 240) #> # Source: lazy query [?? x 19] #> # Database: sqlite 3.19.3  #> year month day dep_time sched_dep_time dep_delay arr_time #> <int> <int> <int> <int> <int> <dbl> <int> #> 1 2013 1 1 848 1835 853. 1001 #> 2 2013 1 1 1815 1325 290. 2120 #> 3 2013 1 1 1842 1422 260. 1958 #> 4 2013 1 1 2115 1700 255. 2330 #> 5 2013 1 1 2205 1720 285. 46 #> 6 2013 1 1 2343 1724 379. 314 #> # ... with more rows, and 12 more variables: sched_arr_time <int>, #> # arr_delay <dbl>, carrier <chr>, flight <int>, tailnum <chr>, #> # origin <chr>, dest <chr>, air_time <dbl>, distance <dbl>, hour <dbl>, #> # minute <dbl>, time_hour <dbl> flights_db %>% group_by(dest) %>% summarise(delay = mean(dep_time)) #> Warning: Missing values are always removed in SQL. #> Use `AVG(x, na.rm = TRUE)` to silence this warning #> # Source: lazy query [?? x 2] #> # Database: sqlite 3.19.3  #> dest delay #> <chr> <dbl> #> 1 ABQ 2006. #> 2 ACK 1033. #> 3 ALB 1627. #> 4 ANC 1635. #> 5 ATL 1293. #> 6 AUS 1521. #> # ... with more rows
However, in the long run, I highly recommend you at least learn the basics of SQL. It’s a valuable skill for any data scientist, and it will help you debug problems if you run into problems with
dplyr’s automatic translation. If you’re completely new to SQL, you might start with this codeacademy tutorial. If you have some familiarity with SQL and you’d like to learn more, I found how indexes work in SQLite and 10 easy steps to a complete understanding of SQL to be particularly helpful.
The most important difference between ordinary data frames and remote database queries is that your R code is translated into SQL and executed in the database, not in R. When working with databases,
dplyr tries to be as lazy as possible:
It never pulls data into R unless you explicitly ask for it.
It delays doing any work until the last possible moment: it collects together everything you want to do and then sends it to the database in one step.
For example, take the following code:
tailnum_delay_db <- flights_db %>% group_by(tailnum) %>% summarise( delay = mean(arr_delay), n = n() ) %>% arrange(desc(delay)) %>% filter(n > 100)
Surprisingly, this sequence of operations never touches the database. It’s not until you ask for the data (e.g., by printing
dplyr generates the SQL and requests the results from the database. Even then it tries to do as little work as possible and only pulls down a few rows.
tailnum_delay_db #> Warning: Missing values are always removed in SQL. #> Use `AVG(x, na.rm = TRUE)` to silence this warning #> # Source: lazy query [?? x 3] #> # Database: sqlite 3.19.3  #> # Ordered by: desc(delay) #> tailnum delay n #> <chr> <dbl> <int> #> 1 N11119 30.3 148 #> 2 N16919 29.9 251 #> 3 N14998 27.9 230 #> 4 N15910 27.6 280 #> 5 N13123 26.0 121 #> 6 N11192 25.9 154 #> # ... with more rows
Behind the scenes,
dplyr is translating your R code into SQL. You can see the SQL it’s generating with
tailnum_delay_db %>% show_query() #> Warning: Missing values are always removed in SQL. #> Use `AVG(x, na.rm = TRUE)` to silence this warning #> <SQL> #> SELECT * #> FROM (SELECT * #> FROM (SELECT `tailnum`, AVG(`arr_delay`) AS `delay`, COUNT() AS `n` #> FROM `flights` #> GROUP BY `tailnum`) #> ORDER BY `delay` DESC) #> WHERE (`n` > 100.0)
If you’re familiar with SQL, this probably isn’t exactly what you’d write by hand, but it does the job. You can learn more about the SQL translation in
Typically, you’ll iterate a few times before you figure out what data you need from the database. Once you’ve figured it out, use
collect() to pull all the data down into a local tibble:
tailnum_delay <- tailnum_delay_db %>% collect() #> Warning: Missing values are always removed in SQL. #> Use `AVG(x, na.rm = TRUE)` to silence this warning tailnum_delay #> # A tibble: 1,201 x 3 #> tailnum delay n #> <chr> <dbl> <int> #> 1 N11119 30.3 148 #> 2 N16919 29.9 251 #> 3 N14998 27.9 230 #> 4 N15910 27.6 280 #> 5 N13123 26.0 121 #> 6 N11192 25.9 154 #> # ... with 1,195 more rows
collect() requires that database does some work, so it may take a long time to complete. Otherwise,
dplyr tries to prevent you from accidentally performing expensive query operations:
Because there’s generally no way to determine how many rows a query will return unless you actually run it,
Because you can’t find the last few rows without executing the whole query, you can’t use
nrow(tailnum_delay_db) #>  NA tail(tailnum_delay_db) #> Error: tail() is not supported by sql sources
You can also ask the database how it plans to execute the query with
explain(). The output is database-dependent and can be esoteric, but learning a bit about it can be very useful because it helps you understand if the database can execute the query efficiently, or if you need to create new indices.
Creating your own database
If you don’t already have a database, here’s some advice from my experiences setting up and running all of them. SQLite is by far the easiest to get started with, but the lack of window functions makes it limited for data analysis. PostgreSQL is not too much harder to use and has a wide range of built-in functions. In my opinion, you shouldn’t bother with MySQL/MariaDB; it’s a pain to set up, the documentation is sub par, and it’s less feature-rich than Postgres. Google BigQuery might be a good fit if you have very large data, or if you’re willing to pay (a small amount of) money to someone who’ll look after your database.
All of these databases follow a client-server model - a computer that connects to the database and the computer that is running the database (the two may be one and the same, but usually aren’t). Getting one of these databases up and running is beyond the scope of this article, but there are plenty of tutorials available on the web.
In terms of functionality, MySQL lies somewhere between SQLite and PostgreSQL. It provides a wider range of built-in functions, but it does not support window functions (so you can’t do grouped mutates and filters).
PostgreSQL is a considerably more powerful database than SQLite. It has:
BigQuery is a hosted database server provided by Google. To connect, you need to provide your
dataset and optionally a project for
billing (if billing for
project isn’t enabled).
It provides a similar set of functions to Postgres and is designed specifically for analytic workflows. Because it’s a hosted solution, there’s no setup involved, but if you have a lot of data, getting it to Google can be an ordeal (especially because upload support from R is not great currently). (If you have lots of data, you can ship hard drives!)