Frequently Asked Questions about data.table

The first section, Beginner FAQs, is intended to be read in order, from start to finish. It’s just written in a FAQ style to be digested more easily. It isn’t really the most frequently asked questions. A better measure for that is looking on Stack Overflow.

This FAQ is required reading and considered core documentation. Please do not ask questions on Stack Overflow or raise issues on GitHub until you have read it. We can all tell when you ask that you haven’t read it. So if you do ask and haven’t read it, don’t use your real name.

This document has been quickly revised given the changes in v1.9.8 released Nov 2016. Please do submit pull requests to fix mistakes or improvements. If anyone knows why the table of contents comes out so narrow and squashed when displayed by CRAN, please let us know. This document used to be a PDF and we changed it recently to HTML.

Beginner FAQs

Why do `DT[ , 5]` and `DT[2, 5]` return a 1-column data.table rather than vectors like `data.frame`?

For consistency so that when you use data.table in functions that accept varying inputs, you can rely on DT[...] returning a data.table. You don’t have to remember to include drop=FALSE like you do in data.frame. data.table was first released in 2006 and this difference to data.frame has been a feature since the very beginning.

You may have heard that it is generally bad practice to refer to columns by number rather than name, though. If your colleague comes along and reads your code later they may have to hunt around to find out which column is number 5. If you or they change the column ordering higher up in your R program, you may produce wrong results with no warning or error if you forget to change all the places in your code which refer to column number 5. That is your fault not R’s or data.table’s. It’s really really bad. Please don’t do it. It’s the same mantra as professional SQL developers have: never use select *, always explicitly select by column name to at least try to be robust to future changes.

Say column 5 is named "region" and you really must extract that column as a vector not a data.table. It is more robust to use the column name and write DT$region or DT[["region"]]; i.e., the same as base R. Using base R’s $ and [[ on data.table is encouraged. Not when combined with <- to assign (use := instead for that) but just to select a single column by name they are encouraged.

There are some circumstances where referring to a column by number seems like the only way, such as a sequence of columns. In these situations just like data.frame, you can write DT[, 5:10] and DT[,c(1,4,10)]. However, again, it is more robust (to future changes in your data’s number of and ordering of columns) to use a named range such as DT[,columnRed:columnViolet] or name each one DT[,c("columnRed","columnOrange","columnYellow")]. It is harder work up front, but you will probably thank yourself and your colleagues might thank you in the future. At least you can say you tried your best to write robust code if something does go wrong.

However, what we really want you to do is DT[,.(columnRed,columnOrange,columnYellow)]; i.e., use column names as if they are variables directly inside DT[...]. You don’t have to prefix each column with DT$ like you do in data.frame. The .() part is just an alias for list() and you can use list() instead if you prefer. You can place any R expression of column names, using any R package, returning different types of different lengths, right there. We wanted to encourage you to do that so strongly in the past that we deliberately didn’t make DT[,5] work at all. Before v1.9.8 released Nov 2016, DT[,5] used to just return 5. The thinking was that we could more simply teach one fact that the parts inside DT[...] get evaluated within the frame of DT always (they see column names as if they are variables). And 5 evaluates to 5 so that behaviour was consistent with the single rule. We asked you to go through an extra deliberate hurdle DT[,5,with=FALSE] if you really wanted to select a column by name or number. Going forward from Nov 2016, you don’t need to use with=FALSE and we’ll see how greater consistency with data.frame in this regard will help or hinder both new and long-time users. The new users who don’t read this FAQ, not even this very first entry, will hopefully not stumble as soon with data.table as they did before if they had expected it to work like data.frame. Hopefully they will not miss out on understanding our intent and recommendation to place expressions of columns inside DT[i, j, by]. If they use data.table like data.frame they won’t gain any benefits. If you know anyone like that, please give them a friendly nudge to read this document like you are.

Reminder: you can place any R expression inside DT[...] using column names as if they are variables; e.g., try DT[, colA*colB/2]. That does return a vector because you used column names as if they are variables. Wrap with .() to return a data.table; i.e. DT[,.(colA*colB/2)]. Name it: DT[,.(myResult = colA*colB/2)]. And we’ll leave it to you to guess how to return two things from this query. It’s also quite common to do a bunch of things inside an anonymous body: DT[, { x<-colA+10; x*x/2 }] or call another package’s function: DT[ , fitdistr(columnA, "normal")].

Why does `DT[,"region"]` return a 1-column data.table rather than a vector?

See the answer above. Try DT$region instead. Or DT[["region"]].

Why does `DT[, region]` return a vector for the “region” column? I’d like a 1-column data.table.

Try DT[ , .(region)] instead. .() is an alias for list() and ensures a data.table is returned.

Also continue reading and see the FAQ after next. Skim whole documents before getting stuck in one part.

Why does `DT[ , x, y, z]` not work? I wanted the 3 columns `x`,`y` and `z`.

The j expression is the 2nd argument. Try DT[ , c("x","y","z")] or DT[ , .(x,y,z)].

I assigned a variable `mycol="x"` but then `DT[, mycol]` returns an error. How do I get it to look up the column name contained in the `mycol` variable?

The error is that column named "mycol" cannot be found, and this error is correct. data.table’s scoping is different to data.frame in that you can use column names as if they are variables directly inside DT[...] without prefixing each column name with DT$; see FAQ 1.1 above.

To use mycol to select the column x from DT, there are a few options:

DT[, ..mycol]            # .. prefix conveys to look for the mycol one level up in calling scope
DT[, mycol, with=FALSE]  # revert to data.frame behavior
DT[[mycol]]               # treat DT as a list and use [[ from base R

See ?data.table for more details about the .. prefix.

The with argument takes its name from the base function with(). When with=TRUE (default), data.table operates similar to with(), i.e. DT[, mycol] behaves like with(DT, mycol). When with=FALSE, the standard data.frame evaluation rules apply to all variables in j and you can no longer use column names directly.

What are the benefits of being able to use column names as if they are variables inside `DT[...]`?

j doesn’t have to be just column names. You can write any R expression of column names directly in j, e.g., DT[ , mean(x*y/z)]. The same applies to i, e.g., DT[x>1000, sum(y*z)].

This runs the j expression on the set of rows where the i expression is true. You don’t even need to return data, e.g., DT[x>1000, plot(y, z)]. You can do j by group simply by adding by =; e.g., DT[x>1000, sum(y*z), by = w]. This runs j for each group in column w but just over the rows where x>1000. By placing the 3 parts of the query (i=where, j=select and by=group by) inside the square brackets, data.table sees this query as a whole before any part of it is evaluated. Thus it can optimize the combined query for performance. It can do this because the R language uniquely has lazy evaluation (Python and Julia do not). data.table sees the expressions inside DT[...] before they are evaluated and optimizes them before evaluation. For example, if data.table see that you’re only using 2 columns out of 100, it won’t bother to subset the 98 that aren’t needed by your j expression.

OK, I’m starting to see what data.table is about, but why didn’t you just enhance `data.frame` in R? Why does it have to be a new package?

As highlighted above, j in [.data.table is fundamentally different from j in [.data.frame. Even if something as simple as DF[ , 1] was changed in base R to return a data.frame rather than a vector, that would break existing code in many 1000’s of CRAN packages and user code. As soon as we took the step to create a new class that inherited from data.frame, we had the opportunity to change a few things and we did. We want data.table to be slightly different and to work this way for more complicated syntax to work. There are other differences, too (see below ).

Furthermore, data.table inherits from data.frame. It is a data.frame, too. A data.table can be passed to any package that only accepts data.frame and that package can use [.data.frame syntax on the data.table. See this answer for how that is achieved.

We have proposed enhancements to R wherever possible, too. One of these was accepted as a new feature in R 2.12.0 :

unique() and match() are now faster on character vectors where all elements are in the global CHARSXP cache and have unmarked encoding (ASCII). Thanks to Matt Dowle for suggesting improvements to the way the hash code is generated in unique.c.

A second proposal was to use memcpy in duplicate.c, which is much faster than a for loop in C. This would improve the way that R copies data internally (on some measures by 13 times). The thread on r-devel is here.

A third more significant proposal that was accepted is that R now uses data.table’s radix sort code as from R 3.3.0 :

The radix sort algorithm and implementation from data.table (forder) replaces the previous radix (counting) sort and adds a new method for order(). Contributed by Matt Dowle and Arun Srinivasan, the new algorithm supports logical, integer (even with large values), real, and character vectors. It outperforms all other methods, but there are some caveats (see ?sort).

This was big event for us and we celebrated until the cows came home. (Not really.)

Why are the defaults the way they are? Why does it work the way it does?

The simple answer is because the main author originally designed it for his own use. He wanted it that way. He finds it a more natural, faster way to write code, which also executes more quickly.

Isn’t this already done by `with()` and `subset()` in `base`?

Some of the features discussed so far are, yes. The package builds upon base functionality. It does the same sorts of things but with less code required and executes many times faster if used correctly.

Why does `X[Y]` return all the columns from `Y` too? Shouldn’t it return a subset of `X`?

This was changed in v1.5.3 (Feb 2011). Since then X[Y] includes Y’s non-join columns. We refer to this feature as join inherited scope because not only are X columns available to the j expression, so are Y columns. The downside is that X[Y] is less efficient since every item of Y’s non-join columns are duplicated to match the (likely large) number of rows in X that match. We therefore strongly encourage X[Y, j] instead of X[Y]. See next FAQ.

What is the difference between `X[Y]` and `merge(X, Y)`?

X[Y] is a join, looking up X’s rows using Y (or Y’s key if it has one) as an index.

Y[X] is a join, looking up Y’s rows using X (or X’s key if it has one) as an index.

merge(X,Y)¹ does both ways at the same time. The number of rows of X[Y] and Y[X] usually differ, whereas the number of rows returned by merge(X, Y) and merge(Y, X) is the same.

BUT that misses the main point. Most tasks require something to be done on the data after a join or merge. Why merge all the columns of data, only to use a small subset of them afterwards? You may suggest merge(X[ , ColsNeeded1], Y[ , ColsNeeded2]), but that requires the programmer to work out which columns are needed. X[Y, j] in data.table does all that in one step for you. When you write X[Y, sum(foo*bar)], data.table automatically inspects the j expression to see which columns it uses. It will subset those columns only; the others are ignored. Memory is only created for the columns j uses and Y columns enjoy standard R recycling rules within the context of each group. Let’s say foo is in X and bar is in Y (along with 20 other columns in Y). Isn’t X[Y, sum(foo*bar)] quicker to program and quicker to run than a merge of everything wastefully followed by a subset?

Anything else about `X[Y, sum(foo*bar)]`?

This behaviour changed in v1.9.4 (Sep 2014). It now does the X[Y] join and then runs sum(foo*bar) over all the rows; i.e., X[Y][ , sum(foo*bar)]. It used to run j for each group of X that each row of Y matches to. That can still be done as it’s very useful but you now need to be explicit and specify by = .EACHI, i.e., X[Y, sum(foo*bar), by = .EACHI]. We call this grouping by each i.

For example, (further complicating it by using join inherited scope, too):

X = data.table(grp = c("a", "a", "b",
                       "b", "b", "c", "c"), foo = 1:7)
setkey(X, grp)
Y = data.table(c("b", "c"), bar = c(4, 2))
X
# Key: <grp>
#       grp   foo
#    <char> <int>
# 1:      a     1
# 2:      a     2
# 3:      b     3
# 4:      b     4
# 5:      b     5
# 6:      c     6
# 7:      c     7
Y
#        V1   bar
#    <char> <num>
# 1:      b     4
# 2:      c     2
X[Y, sum(foo*bar)]
# [1] 74
X[Y, sum(foo*bar), by = .EACHI]
# Key: <grp>
#       grp    V1
#    <char> <num>
# 1:      b    48
# 2:      c    26

That’s nice. How did you manage to change it given that users depended on the old behaviour?

The request to change came from users. The feeling was that if a query is doing grouping then an explicit by= should be present for code readability reasons. An option was provided to return the old behaviour: options(datatable.old.bywithoutby), by default FALSE. This enabled upgrading to test the other new features / bug fixes in v1.9.4, with later migration of any by-without-by queries when ready by adding by=.EACHI to them. We retained 47 pre-change tests and added them back as new tests, tested under options(datatable.old.bywithoutby=TRUE). We added a startup message about the change and how to revert to the old behaviour. After 1 year the option was deprecated with warning when used. After 2 years the option to revert to old behaviour was removed.

Of the 66 packages on CRAN or Bioconductor that depended on or import data.table at the time of releasing v1.9.4 (it is now over 300), only one was affected by the change. That could be because many packages don’t have comprehensive tests, or just that grouping by each row in i wasn’t being used much by downstream packages. We always test the new version with all dependent packages before release and coordinate any changes with those maintainers. So this release was quite straightforward in that regard.

Another compelling reason to make the change was that previously, there was no efficient way to achieve what X[Y, sum(foo*bar)] does now. You had to write X[Y][ , sum(foo*bar)]. That was suboptimal because X[Y] joined all the columns and passed them all to the second compound query without knowing that only foo and bar are needed. To solve that efficiency problem, extra programming effort was required: X[Y, list(foo, bar)][ , sum(foo*bar)]. The change to by = .EACHI has simplified this by allowing both queries to be expressed inside a single DT[...] query for efficiency.

General Syntax

How can I avoid writing a really long `j` expression? You’ve said that I should use the column names, but I’ve got a lot of columns.

When grouping, the j expression can use column names as variables, as you know, but it can also use a reserved symbol .SD which refers to the Subset of the Data.table for each group (excluding the grouping columns). So to sum up all your columns it’s just DT[ , lapply(.SD, sum), by = grp]. It might seem tricky, but it’s fast to write and fast to run. Notice you don’t have to create an anonymous function. The .SD object is efficiently implemented internally and more efficient than passing an argument to a function. But if the .SD symbol appears in j then data.table has to populate .SD fully for each group even if j doesn’t use all of it.

So please don’t do, for example, DT[ , sum(.SD[["sales"]]), by = grp]. That works but is inefficient and inelegant. DT[ , sum(sales), by = grp] is what was intended, and it could be 100s of times faster. If you use all of the data in .SD for each group (such as in DT[ , lapply(.SD, sum), by = grp]) then that’s very good usage of .SD. If you’re using several but not all of the columns, you can combine .SD with .SDcols; see ?data.table.

Why is the default for `mult` now `"all"`?

In v1.5.3 the default was changed to "all". When i (or i’s key if it has one) has fewer columns than x’s key, mult was already set to "all" automatically. Changing the default makes this clearer and easier for users as it came up quite often.

In versions up to v1.3, "all" was slower. Internally, "all" was implemented by joining using "first", then again from scratch using "last", after which a diff between them was performed to work out the span of the matches in x for each row in i. Most often we join to single rows, though, where "first","last" and "all" return the same result. We preferred maximum performance for the majority of situations so the default chosen was "first". When working with a non-unique key (generally a single column containing a grouping variable), DT["A"] returned the first row of that group so DT["A", mult = "all"] was needed to return all the rows in that group.

In v1.4 the binary search in C was changed to branch at the deepest level to find first and last. That branch will likely occur within the same final pages of RAM so there should no longer be a speed disadvantage in defaulting mult to "all". We warned that the default might change and made the change in v1.5.3.

A future version of data.table may allow a distinction between a key and a unique key. Internally mult = "all" would perform more like mult = "first" when all x’s key columns were joined to and x’s key was a unique key. data.table would need checks on insert and update to make sure a unique key is maintained. An advantage of specifying a unique key would be that data.table would ensure no duplicates could be inserted, in addition to performance.

I’m using `c()` in `j` and getting strange results.

This is a common source of confusion. In data.frame you are used to, for example:

DF = data.frame(x = 1:3, y = 4:6, z = 7:9)
DF
#   x y z
# 1 1 4 7
# 2 2 5 8
# 3 3 6 9
DF[ , c("y", "z")]
#   y z
# 1 4 7
# 2 5 8
# 3 6 9

which returns the two columns. In data.table you know you can use the column names directly and might try:

DT = data.table(DF)
DT[ , c(y, z)]
# [1] 4 5 6 7 8 9

but this returns one vector. Remember that the j expression is evaluated within the environment of DT and c() returns a vector. If 2 or more columns are required, use list() or .() instead:

DT[ , .(y, z)]
#        y     z
#    <int> <int>
# 1:     4     7
# 2:     5     8
# 3:     6     9

c() can be useful in a data.table too, but its behaviour is different from that in [.data.frame.

I have built up a complex table with many columns. I want to use it as a template for a new table; i.e., create a new table with no rows, but with the column names and types copied from my table. Can I do that easily?

Yes. If your complex table is called DT, try NEWDT = DT[0].

Is a null data.table the same as `DT[0]`?

No. By “null data.table” we mean the result of data.table(NULL) or as.data.table(NULL); i.e.,

data.table(NULL)
# Null data.table (0 rows and 0 cols)
data.frame(NULL)
# data frame with 0 columns and 0 rows
as.data.table(NULL)
# Null data.table (0 rows and 0 cols)
as.data.frame(NULL)
# data frame with 0 columns and 0 rows
is.null(data.table(NULL))
# [1] FALSE
is.null(data.frame(NULL))
# [1] FALSE

The null data.table|frame is NULL with some attributes attached, which means it’s no longer NULL. In R only pure NULL is NULL as tested by is.null(). When referring to the “null data.table” we use lower case null to help distinguish from upper case NULL. To test for the null data.table, use length(DT) == 0 or ncol(DT) == 0 (length is slightly faster as it’s a primitive function).

An empty data.table (DT[0]) has one or more columns, all of which are empty. Those empty columns still have names and types.

DT = data.table(a = 1:3, b = c(4, 5, 6), d = c(7L,8L,9L))
DT[0]
# Empty data.table (0 rows and 3 cols): a,b,d
sapply(DT[0], class)
#         a         b         d 
# "integer" "numeric" "integer"

Why has the `DT()` alias been removed?

DT was introduced originally as a wrapper for a list of jexpressions. Since DT was an alias for data.table, this was a convenient way to take care of silent recycling in cases where each item of the j list evaluated to different lengths. The alias was one reason grouping was slow, though.

As of v1.3, list() or .() should be passed instead to the j argument. These are much faster, especially when there are many groups. Internally, this was a non-trivial change. Vector recycling is now done internally, along with several other speed enhancements for grouping.

But my code uses `j = DT(...)` and it works. The previous FAQ says that `DT()` has been removed.

Then you are using a version prior to 1.5.3. Prior to 1.5.3 [.data.table detected use of DT() in the j and automatically replaced it with a call to list(). This was to help the transition for existing users.

What are the scoping rules for `j` expressions?

Think of the subset as an environment where all the column names are variables. When a variable foo is used in the j of a query such as X[Y, sum(foo)], foo is looked for in the following order :

The scope of X’s subset; i.e., X’s column names.
The scope of each row of Y; i.e., Y’s column names (join inherited scope)
The scope of the calling frame; e.g., the line that appears before the data.table query.
Exercise for reader: does it then ripple up the calling frames, or go straight to globalenv()?
The global environment

This is lexical scoping as explained in R FAQ 3.3.1. The environment in which the function was created is not relevant, though, because there is no function. No anonymous function is passed to j. Instead, an anonymous body is passed to j; for example,

DT = data.table(x = rep(c("a", "b"), c(2, 3)), y = 1:5)
DT
#         x     y
#    <char> <int>
# 1:      a     1
# 2:      a     2
# 3:      b     3
# 4:      b     4
# 5:      b     5
DT[ , {z = sum(y); z + 3}, by = x]
#         x    V1
#    <char> <num>
# 1:      a     6
# 2:      b    15

Some programming languages call this a lambda.

Can I trace the `j` expression as it runs through the groups?

Try something like this:

DT[ , {
  cat("Objects:", paste(objects(), collapse = ","), "\n")
  cat("Trace: x=", as.character(x), " y=", y, "\n")
  sum(y)},
  by = x]
# Objects: -.POSIXt,Cfastmean,print,strptime,x,y 
# Trace: x= a  y= 1 2 
# Objects: -.POSIXt,Cfastmean,print,strptime,x,y 
# Trace: x= b  y= 3 4 5
#         x    V1
#    <char> <int>
# 1:      a     3
# 2:      b    12

Inside each group, why are the group variables length-1?

Above, x is a grouping variable and (as from v1.6.1) has length 1 (if inspected or used in j). It’s for efficiency and convenience. Therefore, there is no difference between the following two statements:

DT[ , .(g = 1, h = 2, i = 3, j = 4, repeatgroupname = x, sum(y)), by = x]
#         x     g     h     i     j repeatgroupname    V6
#    <char> <num> <num> <num> <num>          <char> <int>
# 1:      a     1     2     3     4               a     3
# 2:      b     1     2     3     4               b    12
DT[ , .(g = 1, h = 2, i = 3, j = 4, repeatgroupname = x[1], sum(y)), by = x]
#         x     g     h     i     j repeatgroupname    V6
#    <char> <num> <num> <num> <num>          <char> <int>
# 1:      a     1     2     3     4               a     3
# 2:      b     1     2     3     4               b    12

If you need the size of the current group, use .N rather than calling length() on any column.

Only the first 10 rows are printed, how do I print more?

There are two things happening here. First, if the number of rows in a data.table are large (> 100 by default), then a summary of the data.table is printed to the console by default. Second, the summary of a large data.table is printed by taking the top and bottom n (= 5 by default) rows of the data.table and only printing those. Both of these parameters (when to trigger a summary and how much of a table to use as a summary) are configurable by R’s options mechanism, or by calling the print function directly.

For instance, to enforce the summary of a data.table to only happen when a data.table is greater than 50 rows, you could options(datatable.print.nrows = 50). To disable the summary-by-default completely, you could options(datatable.print.nrows = Inf). You could also call print directly, as in print(your.data.table, nrows = Inf).

If you want to show more than just the top (and bottom) 10 rows of a data.table summary (say you like 20), set options(datatable.print.topn = 20), for example. Again, you could also just call print directly, as in print(your.data.table, topn = 20).

With an `X[Y]` join, what if `X` contains a column called `"Y"`?

When i is a single name such as Y it is evaluated in the calling frame. In all other cases such as calls to .() or other expressions, i is evaluated within the scope of X. This facilitates easy self-joins such as X[J(unique(colA)), mult = "first"].

`X[Z[Y]]` is failing because `X` contains a column `"Y"`. I’d like it to use the table `Y` in calling scope.

The Z[Y] part is not a single name so that is evaluated within the frame of X and the problem occurs. Try tmp = Z[Y]; X[tmp]. This is robust to X containing a column "tmp" because tmp is a single name. If you often encounter conflicts of this type, one simple solution may be to name all tables in uppercase and all column names in lowercase, or some similar scheme.

Can you explain further why data.table is inspired by `A[B]` syntax in `base`?

Consider A[B] syntax using an example matrix A :

A = matrix(1:12, nrow = 4)
A
#      [,1] [,2] [,3]
# [1,]    1    5    9
# [2,]    2    6   10
# [3,]    3    7   11
# [4,]    4    8   12

To obtain cells (1, 2) = 5 and (3, 3) = 11 many users (we believe) may try this first :

A[c(1, 3), c(2, 3)]
#      [,1] [,2]
# [1,]    5    9
# [2,]    7   11

However, this returns the union of those rows and columns. To reference the cells, a 2-column matrix is required. ?Extract says :

When indexing arrays by [ a single argument i can be a matrix with as many columns as there are dimensions of x; the result is then a vector with elements corresponding to the sets of indices in each row of i.

Let’s try again.

B = cbind(c(1, 3), c(2, 3))
B
#      [,1] [,2]
# [1,]    1    2
# [2,]    3    3
A[B]
# [1]  5 11

A matrix is a 2-dimensional structure with row names and column names. Can we do the same with names?

rownames(A) = letters[1:4]
colnames(A) = LETTERS[1:3]
A
#   A B  C
# a 1 5  9
# b 2 6 10
# c 3 7 11
# d 4 8 12
B = cbind(c("a", "c"), c("B", "C"))
A[B]
# [1]  5 11

So yes, we can. Can we do the same with a data.frame?

A = data.frame(A = 1:4, B = letters[11:14], C = pi*1:4)
rownames(A) = letters[1:4]
A
#   A B         C
# a 1 k  3.141593
# b 2 l  6.283185
# c 3 m  9.424778
# d 4 n 12.566371
B
#      [,1] [,2]
# [1,] "a"  "B" 
# [2,] "c"  "C"
A[B]
# [1] "k"         " 9.424778"

But, notice that the result was coerced to character. R coerced A to matrix first so that the syntax could work, but the result isn’t ideal. Let’s try making B a data.frame.

B = data.frame(c("a", "c"), c("B", "C"))
cat(try(A[B], silent = TRUE))
# Error in `[.default`(A, B) : invalid subscript type 'list'

So we can’t subset a data.frame by a data.frame in base R. What if we want row names and column names that aren’t character but integer or float? What if we want more than 2 dimensions of mixed types? Enter data.table.

Furthermore, matrices, especially sparse matrices, are often stored in a 3-column tuple: (i, j, value). This can be thought of as a key-value pair where i and j form a 2-column key. If we have more than one value, perhaps of different types, it might look like (i, j, val1, val2, val3, ...). This looks very much like a data.frame. Hence data.table extends data.frame so that a data.frame X can be subset by a data.frame Y, leading to the X[Y] syntax.

Can base be changed to do this then, rather than a new package?

data.frame is used everywhere and so it is very difficult to make any changes to it. data.table inherits from data.frame. It is a data.frame, too. A data.table can be passed to any package that only accepts data.frame. When that package uses [.data.frame syntax on the data.table, it works. It works because [.data.table looks to see where it was called from. If it was called from such a package, [.data.table diverts to [.data.frame.

I’ve heard that data.table syntax is analogous to SQL.

Yes :

i $\Leftrightarrow$ where
j $\Leftrightarrow$ select
:= $\Leftrightarrow$ update
by $\Leftrightarrow$ group by
i $\Leftrightarrow$ order by (in compound syntax)
i $\Leftrightarrow$ having (in compound syntax)
nomatch = NA $\Leftrightarrow$ outer join
nomatch = NULL $\Leftrightarrow$ inner join
mult = "first"|"last" $\Leftrightarrow$ N/A because SQL is inherently unordered
roll = TRUE $\Leftrightarrow$ N/A because SQL is inherently unordered

The general form is :

DT[where, select|update, group by][order by][...] ... [...]

A key advantage of column vectors in R is that they are ordered, unlike SQL². We can use ordered functions in data.table queries such as diff() and we can use any R function from any package, not just the functions that are defined in SQL. A disadvantage is that R objects must fit in memory, but with several R packages such as ff, bigmemory, mmap and indexing, this is changing.

What are the smaller syntax differences between `data.frame` and data.table

DT[3] refers to the 3rd row, but DF[3] refers to the 3rd column
DT[3, ] == DT[3], but DF[ , 3] == DF[3] (somewhat confusingly in data.frame, whereas data.table is consistent)
For this reason we say the comma is optional in DT, but not optional in DF
DT[[3]] == DF[, 3] == DF[[3]]
DT[i, ], where i is a single integer, returns a single row, just like DF[i, ], but unlike a matrix single-row subset which returns a vector.
DT[ , j] where j is a single integer returns a one-column data.table, unlike DF[, j] which returns a vector by default
DT[ , "colA"][[1]] == DF[ , "colA"].
DT[ , colA] == DF[ , "colA"] (currently in data.table v1.9.8 but is about to change, see release notes)
DT[ , list(colA)] == DF[ , "colA", drop = FALSE]
DT[NA] returns 1 row of NA, but DF[NA] returns an entire copy of DF containing NA throughout. The symbol NA is type logical in R and is therefore recycled by [.data.frame. The user’s intention was probably DF[NA_integer_]. [.data.table diverts to this probable intention automatically, for convenience.
DT[c(TRUE, NA, FALSE)] treats the NA as FALSE, but DF[c(TRUE, NA, FALSE)] returns NA rows for each NA
DT[ColA == ColB] is simpler than DF[!is.na(ColA) & !is.na(ColB) & ColA == ColB, ]
data.frame(list(1:2, "k", 1:4)) creates 3 columns, data.table creates one list column.
check.names is by default TRUE in data.frame but FALSE in data.table, for convenience.
data.table has always set stringsAsFactors=FALSE by default. In R 4.0.0 (Apr 2020), data.frame’s default was changed from TRUE to FALSE and there is no longer a difference in this regard; see stringsAsFactors, Kurt Hornik, Feb 2020.
Atomic vectors in list columns are collapsed when printed using ", " in data.frame, but "," in data.table with a trailing comma after the 6th item to avoid accidental printing of large embedded objects.

In [.data.frame we very often set drop = FALSE. When we forget, bugs can arise in edge cases where single columns are selected and all of a sudden a vector is returned rather than a single column data.frame. In [.data.table we took the opportunity to make it consistent and dropped drop.

When a data.table is passed to a data.table-unaware package, that package is not concerned with any of these differences; it just works.

I’m using `j` for its side effect only, but I’m still getting data returned. How do I stop that?

In this case j can be wrapped with invisible(); e.g., DT[ , invisible(hist(colB)), by = colA]³

Why does `[.data.table` now have a `drop` argument from v1.5?

So that data.table can inherit from data.frame without using .... If we used ... then invalid argument names would not be caught.

The drop argument is never used by [.data.table. It is a placeholder for non-data.table-aware packages when they use the [.data.frame syntax directly on a data.table.

Rolling joins are cool and very fast! Was that hard to program?

The prevailing row on or before the i row is the final row the binary search tests anyway. So roll = TRUE is essentially just a switch in the binary search C code to return that row.

Why does `DT[i, col := value]` return the whole of `DT`? I expected either no visible value (consistent with `<-`), or a message or return value containing how many rows were updated. It isn’t obvious that the data has indeed been updated by reference.

This has changed in v1.8.3 to meet your expectations. Please upgrade.

The whole of DT is returned (now invisibly) so that compound syntax can work; e.g., DT[i, done := TRUE][ , sum(done)]. The number of rows updated is returned when verbose is TRUE, either on a per-query basis or globally using options(datatable.verbose = TRUE).

OK, thanks. What was so difficult about the result of `DT[i, col := value]` being returned invisibly?

R internally forces visibility on for [. The value of FunTab’s eval column (see src/main/names.c) for [ is 0 meaning “force R_Visible on” (see R-Internals section 1.6 ). Therefore, when we tried invisible() or setting R_Visible to 0 directly ourselves, eval in src/main/eval.c would force it on again.

To solve this problem, the key was to stop trying to stop the print method running after a :=. Instead, inside := we now (from v1.8.3) set a global flag which the print method uses to know whether to actually print or not.

Why do I have to type `DT` sometimes twice after using `:=` to print the result to console?

This is an unfortunate downside to get #869 to work. If a := is used inside a function with no DT[] before the end of the function, then the next time DT is typed at the prompt, nothing will be printed. A repeated DT will print. To avoid this: include a DT[] after the last := in your function. If that is not possible (e.g., it’s not a function you can change) then print(DT) and DT[] at the prompt are guaranteed to print. As before, adding an extra [] on the end of := query is a recommended idiom to update and then print; e.g.> DT[,foo:=3L][].

I’ve noticed that `base::cbind.data.frame` (and `base::rbind.data.frame`) appear to be changed by data.table. How is this possible? Why?

It was a temporary, last resort solution before rbind and cbind S3 method dispatch was fixed in R >= 4.0.0. Essentially, the issue was that data.table inherits from data.frame, and base::cbind and base::rbind (uniquely) do their own S3 dispatch internally as documented by ?cbind. The data.table workaround was adding one for loop to the start of each function directly in base. That modification was made dynamically, i.e., the base definition of cbind.data.frame was fetched, the for loop added to the beginning, and then assigned back to base. This solution was designed to be robust to different definitions of base::cbind.data.frame in different versions of R, including unknown future changes. It worked well. The competing requirements were:

cbind(DT, DF) needs to work. Defining cbind.data.table didn’t work because base::cbind does its own S3 dispatch and required (before R 4.0.0) that the first cbind method for each object it is passed is identical. This is not true in cbind(DT, DF) because the first method for DT is cbind.data.table but the first method for DF is cbind.data.frame. base::cbind then fell through to its internal bind code which appears to treat DT as a regular list and returns very odd looking and unusable matrix output. See below. We cannot just advise users not to call cbind(DT, DF) because packages such as ggplot2 make such a call (test 167.2).
This naturally led to trying to mask cbind.data.frame instead. Since a data.table is a data.frame, cbind would find the same method for both DT and DF. However, this didn’t work either because base::cbind appears to find methods in base first; i.e., base::cbind.data.frame isn’t maskable.
Finally, we tried masking cbind itself (v1.6.5 and v1.6.6). This allowed cbind(DT, DF) to work, but introduced compatibility issues with package IRanges, since IRanges also masks cbind. It worked if IRanges was lower on the search() path than data.table, but if IRanges was higher then data.table’s, cbind would never be called and the strange-looking matrix output occurs again (see below).

Many thanks to the R core team for fixing the issue in Sep 2019. data.table v1.12.6+ no longer applies the workaround in R >= 4.0.0.

I’ve read about method dispatch (e.g. `merge` may or may not dispatch to `merge.data.table`) but how does R know how to dispatch? Are dots significant or special? How on earth does R know which function to dispatch and when?

This comes up quite a lot but it’s really earth-shatteringly simple. A function such as merge is generic if it consists of a call to UseMethod. When you see people talking about whether or not functions are generic functions they are merely typing the function without () afterwards, looking at the program code inside it and if they see a call to UseMethod then it is generic. What does UseMethod do? It literally slaps the function name together with the class of the first argument, separated by period (.) and then calls that function, passing along the same arguments. It’s that simple. For example, merge(X, Y) contains a UseMethod call which means it then dispatches (i.e. calls) paste("merge", class(X), sep = "."). Functions with dots in their name may or may not be methods. The dot is irrelevant really, other than dot being the separator that UseMethod uses. Knowing this background should now highlight why, for example, it is obvious to R folk that as.data.table.data.frame is the data.frame method for the as.data.table generic function. Further, it may help to elucidate that, yes, you are correct, it is not obvious from its name alone that ls.fit is not the fit method of the ls generic function. You only know that by typing ls (not ls()) and observing it isn’t a single call to UseMethod.

You might now ask: where is this documented in R? Answer: it’s quite clear, but, you need to first know to look in ?UseMethod and that help file contains :

When a function calling UseMethod('fun') is applied to an object with class attribute c('first', 'second'), the system searches for a function called fun.first and, if it finds it, applies it to the object. If no such function is found a function called fun.second is tried. If no class name produces a suitable function, the function fun.default is used, if it exists, or an error results.

Happily, an internet search for “How does R method dispatch work” (at the time of this writing) returns the ?UseMethod help page in the top few links. Admittedly, other links rapidly descend into the intricacies of S3 vs S4, internal generics and so on.

However, features like basic S3 dispatch (pasting the function name together with the class name) is why some R folk love R. It’s so simple. No complicated registration or signature is required. There isn’t much needed to learn. To create the merge method for data.table all that was required, literally, was to merely create a function called merge.data.table.

Questions relating to compute time

I have 20 columns and a large number of rows. Why is an expression of one column so quick?

Several reasons:

Only that column is grouped, the other 19 are ignored because data.table inspects the j expression and realises it doesn’t use the other columns.
One memory allocation is made for the largest group only, then that memory is re-used for the other groups. There is very little garbage to collect.
R is an in-memory column store; i.e., the columns are contiguous in RAM. Page fetches from RAM into L2 cache are minimised.

I don’t have a `key` on a large table, but grouping is still really quick. Why is that?

data.table uses radix sorting. This is significantly faster than other sort algorithms. See our presentations for more information, in particular from useR!2015 Denmark.

This is also one reason why setkey() is quick.

When no key is set, or we group in a different order from that of the key, we call it an ad hoc by.

Why is grouping by columns in the key faster than an ad hoc `by`?

Because each group is contiguous in RAM, thereby minimising page fetches and memory can be copied in bulk (memcpy in C) rather than looping in C.

What are primary and secondary indexes in data.table?

Manual: ?setkey S.O. : What is the purpose of setting a key in data.table?

setkey(DT, col1, col2) orders the rows by column col1 then within each group of col1 it orders by col2. This is a primary index. The row order is changed by reference in RAM. Subsequent joins and groups on those key columns then take advantage of the sort order for efficiency. (Imagine how difficult looking for a phone number in a printed telephone directory would be if it wasn’t sorted by surname then forename. That’s literally all setkey does. It sorts the rows by the columns you specify.) The index doesn’t use any RAM. It simply changes the row order in RAM and marks the key columns. Analogous to a clustered index in SQL.

However, you can only have one primary key because data can only be physically sorted in RAM in one way at a time. Choose the primary index to be the one you use most often (e.g. [id,date]). Sometimes there isn’t an obvious choice for the primary key or you need to join and group many different columns in different orders. Enter a secondary index. This does use memory (4*nrow bytes regardless of the number of columns in the index) to store the order of the rows by the columns you specify, but doesn’t actually reorder the rows in RAM. Subsequent joins and groups take advantage of the secondary key’s order but need to hop via that index so aren’t as efficient as primary indexes. But still, a lot faster than a full vector scan. There is no limit to the number of secondary indexes since each one is just a different ordering vector. Typically you don’t need to create secondary indexes. They are created automatically and used for you automatically by using data.table normally; e.g. DT[someCol == someVal, ] and DT[someCol %in% someVals, ] will create, attach and then use the secondary index. This is faster in data.table than a vector scan so automatic indexing is on by default since there is no up-front penalty. There is an option to turn off automatic indexing; e.g., if somehow many indexes are being created and even the relatively small amount of extra memory becomes too large.

We use the words index and key interchangeably.

Error messages

“Could not find function `DT`”

See above here and here.

“unused argument(s) (`MySum = sum(v)`)”

This error is generated by DT[ , MySum = sum(v)]. DT[ , .(MySum = sum(v))] was intended, or DT[ , j = .(MySum = sum(v))].

“`translateCharUTF8` must be called on a `CHARSXP`”

This error (and similar, e.g., “getCharCE must be called on a CHARSXP”) may be nothing do with character data or locale. Instead, this can be a symptom of an earlier memory corruption. To date these have been reproducible and fixed (quickly). Please report it to our issues tracker.

`cbind(DT, DF)` returns a strange format, e.g. `Integer,5`

This occurs prior to v1.6.5, for rbind(DT, DF) too. Please upgrade to v1.6.7 or later.

“cannot change value of locked binding for `.SD`”

.SD is locked by design. See ?data.table. If you’d like to manipulate .SD before using it, or returning it, and don’t wish to modify DT using :=, then take a copy first (see ?copy), e.g.,

DT = data.table(a = rep(1:3, 1:3), b = 1:6, c = 7:12)
DT
#        a     b     c
#    <int> <int> <int>
# 1:     1     1     7
# 2:     2     2     8
# 3:     2     3     9
# 4:     3     4    10
# 5:     3     5    11
# 6:     3     6    12
DT[ , { mySD = copy(.SD)
      mySD[1, b := 99L]
      mySD},
    by = a]
#        a     b     c
#    <int> <int> <int>
# 1:     1    99     7
# 2:     2    99     8
# 3:     2     3     9
# 4:     3    99    10
# 5:     3     5    11
# 6:     3     6    12

“cannot change value of locked binding for `.N`”

Please upgrade to v1.8.1 or later. From this version, if .N is returned by j it is renamed to N to avoid any ambiguity in any subsequent grouping between the .N special variable and a column called ".N".

The old behaviour can be reproduced by forcing .N to be called .N, like this :

DT = data.table(a = c(1,1,2,2,2), b = c(1,2,2,2,1))
DT
#        a     b
#    <num> <num>
# 1:     1     1
# 2:     1     2
# 3:     2     2
# 4:     2     2
# 5:     2     1
DT[ , list(.N = .N), list(a, b)]   # show intermediate result for exposition
#        a     b    .N
#    <num> <num> <int>
# 1:     1     1     1
# 2:     1     2     1
# 3:     2     2     2
# 4:     2     1     1
cat(try(
    DT[ , list(.N = .N), by = list(a, b)][ , unique(.N), by = a]   # compound query more typical
, silent = TRUE))
# Error : The column '.N' can't be grouped because it conflicts with the special .N variable. Try setnames(DT,'.N','N') first.

If you are already running v1.8.1 or later then the error message is now more helpful than the “cannot change value of locked binding” error, as you can see above, since this vignette was produced using v1.8.1 or later.

The more natural syntax now works :

if (packageVersion("data.table") >= "1.8.1") {
    DT[ , .N, by = list(a, b)][ , unique(N), by = a]
  }
#        a    V1
#    <num> <int>
# 1:     1     1
# 2:     2     2
# 3:     2     1
if (packageVersion("data.table") >= "1.9.3") {
    DT[ , .N, by = .(a, b)][ , unique(N), by = a]   # same
}
#        a    V1
#    <num> <int>
# 1:     1     1
# 2:     2     2
# 3:     2     1

Warning messages

“The following object(s) are masked from `package:base`: `cbind`, `rbind`”

This warning was present in v1.6.5 and v.1.6.6 only, when loading the package. The motivation was to allow cbind(DT, DF) to work, but as it transpired, this broke (full) compatibility with package IRanges. Please upgrade to v1.6.7 or later.

“Coerced numeric RHS to integer to match the column’s type”

Hopefully, this is self explanatory. The full message is:

Coerced numeric RHS to integer to match the column’s type; may have truncated precision. Either change the column to numeric first by creating a new numeric vector length 5 (nrows of entire table) yourself and assigning that (i.e. ‘replace’ column), or coerce RHS to integer yourself (e.g. 1L or as.integer) to make your intent clear (and for speed). Or, set the column type correctly up front when you create the table and stick to it, please.

To generate it, try :

DT = data.table(a = 1:5, b = 1:5)
suppressWarnings(
DT[2, b := 6]         # works (slower) with warning
)
class(6)              # numeric not integer
# [1] "numeric"
DT[2, b := 7L]        # works (faster) without warning
class(7L)             # L makes it an integer
# [1] "integer"
DT[ , b := rnorm(5)]  # 'replace' integer column with a numeric column

Reading data.table from RDS or RData file

*.RDS and *.RData are file types which can store in-memory R objects on disk efficiently. However, storing data.table into the binary file loses its column over-allocation. This isn’t a big deal – your data.table will be copied in memory on the next by reference operation and throw a warning. Therefore it is recommended to call setalloccol() on each data.table loaded with readRDS() or load() calls.

General questions about the package

v1.3 appears to be missing from the CRAN archive?

That is correct. v1.3 was available on R-Forge only. There were several large changes internally and these took some time to test in development.

Is data.table compatible with S-plus?

Not currently.

A few core parts of the package are written in C and use internal R functions and R structures.
The package uses lexical scoping which is one of the differences between R and S-plus explained by R FAQ 3.3.1

Is it available for Linux, Mac and Windows?

Yes, for both 32-bit and 64-bit on all platforms. Thanks to CRAN. There are no special or OS-specific libraries used.

I think it’s great. What can I do?

Please file suggestions, bug reports and enhancement requests on our issues tracker. This helps make the package better.

Please do star the package on GitHub. This helps encourage the developers and helps other R users find the package.

You can submit pull requests to change the code and/or documentation yourself; see our Contribution Guidelines.

I think it’s not great. How do I warn others about my experience?

We add all articles we know about (whether positive or negative) to the Articles page. All pages in the project’s wiki on GitHub are open-access with no modify restrictions. Feel free to write an article, link to a negative one someone else wrote that you found, or add a new page to our wiki to collect your criticisms. Please make it constructive so we have a chance to improve.

I have a question. I know the r-help posting guide tells me to contact the maintainer (not r-help), but is there a larger group of people I can ask?

Please see the support guide on the project’s homepage which contains up-to-date links.

Where are the datatable-help archives?

The homepage contains links to the archives in several formats.

I’d prefer not to post on the Issues page, can I mail just one or two people privately?

Sure. You’re more likely to get a faster answer from the Issues page or Stack Overflow, though. Further, asking publicly in those places helps build the general knowledge base.

I have created a package that uses data.table. How do I ensure my package is data.table-aware so that inheritance from `data.frame` works?

Please see this answer.

Here we mean either the merge method for data.table or the merge method for data.frame since both methods work in the same way in this respect. See ?merge.data.table and below for more information about method dispatch.↩
It may be a surprise to learn that select top 10 * from ... does not reliably return the same rows over time in SQL. You do need to include an order by clause, or use a clustered index to guarantee row order; i.e., SQL is inherently unordered.↩
e.g., hist() returns the breakpoints in addition to plotting to the graphics device.↩

2023-12-01

Beginner FAQs

Why do DT[ , 5] and DT[2, 5] return a 1-column data.table rather than vectors like data.frame?

Why does DT[,"region"] return a 1-column data.table rather than a vector?

Why does DT[, region] return a vector for the “region” column? I’d like a 1-column data.table.

Why does DT[ , x, y, z] not work? I wanted the 3 columns x,y and z.

I assigned a variable mycol="x" but then DT[, mycol] returns an error. How do I get it to look up the column name contained in the mycol variable?

What are the benefits of being able to use column names as if they are variables inside DT[...]?

OK, I’m starting to see what data.table is about, but why didn’t you just enhance data.frame in R? Why does it have to be a new package?

Why are the defaults the way they are? Why does it work the way it does?

Isn’t this already done by with() and subset() in base?

Why does X[Y] return all the columns from Y too? Shouldn’t it return a subset of X?

What is the difference between X[Y] and merge(X, Y)?

Anything else about X[Y, sum(foo*bar)]?

That’s nice. How did you manage to change it given that users depended on the old behaviour?

General Syntax

How can I avoid writing a really long j expression? You’ve said that I should use the column names, but I’ve got a lot of columns.

Why is the default for mult now "all"?

I’m using c() in j and getting strange results.