.EACHI in data.table?

I cannot seem to find any documentation on what exactly .EACHI does in data.table. I see a brief mention of it in the documentation:

Aggregation for a subset of known groups is particularly efficient when passing those groups in i and setting by=.EACHI. When i is a data.table, DT[i,j,by=.EACHI] evaluates j for the groups of DT that each row in i joins to. We call this grouping by each i.

But what does "groups" in the context of DT mean? Is a group determined by the key that is set on DT? Is the group every distinct row that uses all the columns as the key? I fully understand how to run something like DT[i,j,by=my_grouping_variable] but am confused as to how .EACHI would work. Could someone explain please?

Solution 1:

I've added this to the list here. And hopefully we'll be able to deliver as planned.

The reason is most likely that by=.EACHI is a recent feature (since 1.9.4), but what it does isn't. Let me explain with an example. Suppose we have two data.tables X and Y:

X = data.table(x = c(1,1,1,2,2,5,6), y = 1:7, key = "x")
Y = data.table(x = c(2,6), z = letters[2:1], key = "x")

We know that we can join by doing X[Y]. this is similar to a subset operation, but using data.tables (instead of integers / row names or logical values). For each row in Y, taking Y's key columns, it finds and returns corresponding matching rows in X's key columns (+ columns in Y) .

X[Y]
#    x y z
# 1: 2 4 b
# 2: 2 5 b
# 3: 6 7 a

Now let's say we'd like to, for each row from Y's key columns (here only one key column), we'd like to get the count of matches in X. In versions of data.table < 1.9.4, we can do this by simply specifying .N in j as follows:

# < 1.9.4
X[Y, .N]
#    x N
# 1: 2 2
# 2: 6 1

What this implicitly does is, in the presence of j, evaluate the j-expression on each matched result of X (corresponding to the row in Y). This was called by-without-by or implicit-by, because it's as if there's a hidden by.

The issue was that this'll always perform a by operation. So, if we wanted to know the number of rows after a join, then we'd have to do: X[Y][ .N] (or simply nrow(X[Y]) in this case). That is, we can't have the j expression in the same call if we don't want a by-without-by. As a result, when we did for example X[Y, list(z)], it evaluated list(z) using by-without-by and was therefore slightly slower.

Additionally data.table users requested this to be explicit - see this and this for more context.

Hence by=.EACHI was added. Now, when we do:

X[Y, .N]
# [1] 3

it does what it's meant to do (avoids confusion). It returns the number of rows resulting from the join.

And,

X[Y, .N, by=.EACHI]

evaluates j-expression on the matching rows for each row in Y (corresponding to value from Y's key columns here). It'd be easier to see this by using which=TRUE.

X[.(2), which=TRUE] # [1] 4 5
X[.(6), which=TRUE] # [1] 7

If we run .N for each, then we should get 2,1.

X[Y, .N, by=.EACHI]
#    x N
# 1: 2 2
# 2: 6 1

So we now have both functionalities. Hope this helps.