Secondary indices and auto indexing


This vignette assumes that the reader is familiar with data.table’s [i, j, by] syntax, and how to perform fast key based subsets. If you’re not familiar with these concepts, please read the “Introduction to data.table”, “Reference semantics” and “Keys and fast binary search based subset” vignettes first.


We will use the same flights data as in the “Introduction to data.table” vignette.

flights <- fread("flights14.csv")
#     year month   day dep_delay arr_delay carrier origin   dest air_time distance  hour
#  [ reached getOption("max.print") -- omitted 7 rows ]
# [1] 253316     11


In this vignette, we will

1. Secondary indices

a) What are secondary indices?

Secondary indices are similar to keys in data.table, except for two major differences:

b) Set and get secondary indices

– How can we set the column origin as a secondary index in the data.table flights?

  • setindex and setindexv() allows adding a secondary index to the data.table.

  • Note that flights is not physically reordered in increasing order of origin, as would have been the case with setkey().

  • Also note that the attribute index has been added to flights.

  • setindex(flights, NULL) would remove all secondary indices.

– How can we get all the secondary indices set so far in flights?

  • The function indices() returns all current secondary indices in the data.table. If none exists, NULL is returned.

  • Note that by creating another index on the columns origin, dest, we do not lose the first index created on the column origin, i.e., we can have multiple secondary indices.

c) Why do we need secondary indices?

– Reordering a data.table can be expensive and not always ideal

Consider the case where you would like to perform a fast key based subset on origin column for the value “JFK”. We’d do this as:

setkey() requires:

  1. computing the order vector for the column(s) provided, here, origin, and

  2. reordering the entire data.table, by reference, based on the order vector computed.

Computing the order isn’t the time consuming part, since data.table uses true radix sorting on integer, character and numeric vectors. However reordering the data.table could be time consuming (depending on the number of rows and columns).

Unless our task involves repeated subsetting on the same column, fast key based subsetting could effectively be nullified by the time to reorder, depending on our data.table dimensions.

– There can be only one key at the most

Now if we would like to repeat the same operation but on dest column instead, for the value “LAX”, then we have to setkey(), again.

And this reorders flights by dest, again. What we would really like is to be able to perform the fast subsetting by eliminating the reordering step.

And this is precisely what secondary indices allow for!

– Secondary indices can be reused

Since there can be multiple secondary indices, and creating an index is as simple as storing the order vector as an attribute, this allows us to even eliminate the time to recompute the order vector if an index already exists.

– The new on argument allows for cleaner syntax and automatic creation and reuse of secondary indices

As we will see in the next section, the on argument provides several advantages:

on argument

2. Fast subsetting using on argument and secondary indices

a) Fast subsets in i

– How can I subset based on origin and dest columns?

For example, if we want to subset "JFK", "LAX" combination, then:

  • on argument accepts a character vector of column names corresponding to the order provided to i-argument.

  • Since the time to compute the secondary index is quite small, we don’t have to use setindex(), unless, once again, the task involves repeated subsetting on the same column.

b) Select in j

All the operations we will discuss below are no different to the ones we already saw in the Keys and fast binary search based subset vignette. Except we’ll be using the on argument instead of setting keys.

– Return arr_delay column alone as a data.table corresponding to origin = "LGA" and dest = "TPA"

d) Compute or do in j

– Find the maximum arrival delay corresponding to origin = "LGA" and dest = "TPA".

e) sub-assign by reference using := in j

We have seen this example already in the Reference semantics and Keys and fast binary search based subset vignette. Let’s take a look at all the hours available in the flights data.table:

We see that there are totally 25 unique values in the data. Both 0 and 24 hours seem to be present. Let’s go ahead and replace 24 with 0, but this time using on instead of setting keys.

Now, let’s check if 24 is replaced with 0 in the hour column.

  • This is particularly a huge advantage of secondary indices. Previously, just to update a few rows of hour, we had to setkey() on it, which inevitably reorders the entire data.table. With on, the order is preserved, and the operation is much faster! Looking at the code, the task we wanted to perform is also quite clear.

f) Aggregation using by

– Get the maximum departure delay for each month corresponding to origin = "JFK". Order the result by month

  • We would have had to set the key back to origin, dest again, if we did not use on which internally builds secondary indices on the fly.

g) The mult argument

The other arguments including mult work exactly the same way as we saw in the Keys and fast binary search based subset vignette. The default value for mult is “all”. We can choose, instead only the “first” or “last” matching rows should be returned.

h) The nomatch argument

We can choose if queries that do not match should return NA or be skipped altogether using the nomatch argument.

– From the previous example, subset all rows only if there’s a match

  • There are no flights connecting “JFK” and “XNA”. Therefore, that row is skipped in the result.

3. Auto indexing

First we looked at how to fast subset using binary search using keys. Then we figured out that we could improve performance even further and have more cleaner syntax by using secondary indices.

That is what auto indexing does. At the moment, it is only implemented for binary operators == and %in%. An index is automatically created and saved as an attribute. That is, unlike the on argument which computes the index on the fly each time (unless one already exists), a secondary index is created here.

Let’s start by creating a data.table big enough to highlight the advantage.

When we use == or %in% on a single column for the first time, a secondary index is created automatically, and it is used to perform the subset.

The time to subset the first time is the time to create the index + the time to subset. Since creating a secondary index involves only creating the order vector, this combined operation is faster than vector scans in many cases. But the real advantage comes in successive subsets. They are extremely fast.

In recent version we extended auto indexing to expressions involving more than one column (combined with & operator). In the future, we plan to extend binary search to work with more binary operators like <, <=, > and >=.

We will discuss fast subsets using keys and secondary indices to joins in the next vignette, “Joins and rolling joins”.