--- title: "Joins in data.table" date: "`r Sys.Date()`" output: markdown::html_format vignette: > %\VignetteIndexEntry{Joins in data.table} %\VignetteEngine{knitr::knitr} \usepackage[utf8]{inputenc} editor_options: chunk_output_type: console --- ```{r, echo = FALSE, message = FALSE} require(data.table) knitr::opts_chunk$set( comment = "#", error = FALSE, tidy = FALSE, cache = FALSE, collapse = TRUE ) ``` In this vignette you will learn how to perform any join operation using resources available in the `data.table` syntax. It assumes familiarity with the `data.table` syntax. If that is not the case, please read the following vignettes: - `vignette("datatable-intro", package="data.table")` - `vignette("datatable-reference-semantics", package="data.table")` - `vignette("datatable-keys-fast-subset", package="data.table")` *** ## 1. Defining example data To illustrate how to use the method available with real life examples, let's simulate a **normalized database** from a little supermarket by performing the following steps: 1. Defining a `data.table` where each product is represented by a row with some qualities, but leaving one product without `id` to show how the framework deals with ***missing values***. ```{r} Products = data.table( id = c(1:4, NA_integer_), name = c("banana", "carrots", "popcorn", "soda", "toothpaste"), price = c(0.63, 0.89, 2.99, 1.49, 2.99), unit = c("unit", "lb", "unit", "ounce", "unit"), type = c(rep("natural", 2L), rep("processed", 3L)) ) Products ``` 2. Defining a `data.table` showing the proportion of taxes to be applied for processed products based on their units. ```{r} NewTax = data.table( unit = c("unit","ounce"), type = "processed", tax_prop = c(0.65, 0.20) ) NewTax ``` 3. Defining a `data.table` simulating the products received every Monday with a `product_id` that is not present in the `Products` table. ```{r} set.seed(2156) ProductReceived = data.table( id = 1:10, date = seq(from = as.IDate("2024-01-08"), length.out = 10L, by = "week"), product_id = sample(c(NA_integer_, 1:3, 6L), size = 10L, replace = TRUE), count = sample(c(50L, 100L, 150L), size = 10L, replace = TRUE) ) ProductReceived ``` 4. Defining a `data.table` to show some sales that can take place on weekdays with another `product_id` that is not present in the `Products` table. ```{r} sample_date = function(from, to, size, ...){ all_days = seq(from = from, to = to, by = "day") weekdays = all_days[wday(all_days) %in% 2:6] days_sample = sample(weekdays, size, ...) days_sample_desc = sort(days_sample) days_sample_desc } set.seed(5415) ProductSales = data.table( id = 1:10, date = ProductReceived[, sample_date(min(date), max(date), 10L)], product_id = sample(c(1:3, 7L), size = 10L, replace = TRUE), count = sample(c(50L, 100L, 150L), size = 10L, replace = TRUE) ) ProductSales ``` ## 2. `data.table` joining syntax Before taking advantage of the `data.table` syntax to perform join operations we need to know which arguments can help us to perform successful joins. The next diagram shows a description for each basic argument. In the following sections we will show how to use each of them and add more complexity little by little. ``` x[i, on, nomatch] | | | | | | | \__ If NULL only returns rows linked in x and i tables | | \____ a character vector o list defining match logict | \_____ primary data.table, list or data.frame \____ secondary data.table ``` > Please keep in mind that the standard argument order in data.table is `dt[i, j, by]`. For join operations, it is recommended to pass the `on` and `nomatch` arguments by name to avoid using `j` and `by` when they are not needed. ## 3. Equi joins This the most common and simple case as we can find common elements between tables to combine. The relationship between tables can be: - **One to one**: When each matching value is unique on each table. - **One to many**: When some matching values are repeated in one of the table both unique in the other one. - **Many to many**: When the matching values are repeated several times on each table. In most of the following examples we will perform *one to many* matches, but we are also going to take the time to explain the resources available to perform *many to many* matches. ### 3.1. Right join Use this method if you need to combine columns from 2 tables based on one or more references but ***keeping all rows present in the table located on the right (in the the square brackets)***. In our supermarket context, we can perform a right join to see more details about the products received as this is relation *one to many* by passing a vector to the `on` argument. ```{r} Products[ProductReceived, on = c(id = "product_id")] ``` As many things have changed, let's explain the new characteristics in the following groups: - **Column level** - The *first group* of columns in the new data.table comes from the `x` table. - The *second group* of columns in the new data.table comes from the `i` table. - If the join operation presents a present any **name conflict** (both table have same column name) the ***prefix*** `i.` is added to column names from the **right-hand table** (table on `i` position). - **Row level** - The missing `product_id` present on the `ProductReceived` table in row 1 was successfully matched with missing `id` of the `Products` table, so `NA` ***values are treated as any other value***. - All rows from in the `i` table were kept including: - Not matching rows like the one with `product_id = 6`. - Rows that repeat the same `product_id` several times. #### 3.1.1. Joining by a list argument If you are following the vignette, you might have found out that we used a vector to define the relations between tables in the `on` argument, that is really useful if you are **creating your own functions**, but another alternative is to use a **list** to define the columns to match. To use this capacity, we have 2 equivalent alternatives: - Wrapping the related columns in the base R `list` function. ```{r, eval=FALSE} Products[ProductReceived, on = list(id = product_id)] ``` - Wrapping the related columns in the data.table `list` alias `.`. ```{r, eval=FALSE} Products[ProductReceived, on = .(id = product_id)] ``` #### 3.1.2. Alternatives to define the `on` argument In all the prior example we have pass the column names we want to match to the `on` argument but `data.table` also have alternatives to that syntax. - **Natural join**: Selects the columns to perform the match based on common column names. To illustrate this method, let's change the column of `Products` table from `id` to `product_id` and use the keyword `.NATURAL`. ```{r} ProductsChangedName = setnames(copy(Products), "id", "product_id") ProductsChangedName ProductsChangedName[ProductReceived, on = .NATURAL] ``` - **Keyed join**: Selects the columns to perform the match based on keyed columns regardless of their names.To illustrate this method, we need to define keys in the same order for both tables. ```{r} ProductsKeyed = setkey(copy(Products), id) key(ProductsKeyed) ProductReceivedKeyed = setkey(copy(ProductReceived), product_id) key(ProductReceivedKeyed) ProductsKeyed[ProductReceivedKeyed] ``` #### 3.1.3. Operations after joining Most of the time after a join is complete we need to make some additional transformations. To make so we have the following alternatives: - Chaining a new instruction by adding a pair of brakes `[]`. - Passing a list with the columns that we want to keep or create to the `j` argument. Our recommendation is to use the second alternative if possible, as it is **faster** and uses **less memory** than the first one. ##### Managing shared column Names with the j argument The `j` argument has great alternatives to manage joins with tables **sharing the same names for several columns**. By default all columns are taking their source from the the `x` table, but we can also use the `x.` prefix to make clear the source and use the prefix `i.` to use any column form the table declared in the `i` argument of the `x` table. Going back to the little supermarket, after updating the `ProductReceived` table with the `Products` table, it seems convenient apply the following changes: - Changing the columns names from `id` to `product_id` and from `i.id` to `received_id`. - Adding the `total_value`. ```{r} Products[ ProductReceived, on = c("id" = "product_id"), j = .(product_id = x.id, name = x.name, price, received_id = i.id, date = i.date, count, total_value = price * count) ] ``` ##### Summarizing with on in data.table We can also use this alternative to return aggregated results based columns present in the `x` table. For example, we might interested in how much money we expend buying products each date regardless the products. ```{r} dt1 = ProductReceived[ Products, on = c("product_id" = "id"), by = .EACHI, j = .(total_value_received = sum(price * count)) ] dt2 = ProductReceived[ Products, on = c("product_id" = "id"), ][, .(total_value_received = sum(price * count)), by = "product_id" ] identical(dt1, dt2) ``` #### 3.1.4. Joining based on several columns So far we have just joined `data.table` base on 1 column, but it's important to know that the package can join tables matching several columns. To illustrate this, let's assume that we want to add the `tax_prop` from `NewTax` to **update** the `Products` table. ```{r} NewTax[Products, on = c("unit", "type")] ``` ### 3.2. Inner join Use this method if you need to combine columns from 2 tables based on one or more references but ***keeping only rows matched in both tables***. To perform this operation we just need to add `nomatch = NULL` or `nomatch = 0` to any of the prior join operations to return the same results. ```{r} # First Table Products[ProductReceived, on = c("id" = "product_id"), nomatch = NULL] # Second Table ProductReceived[Products, on = .(product_id = id), nomatch = NULL] ``` Despite both tables have the same information, they present some relevant differences: - They present different order for their columns - They have some name differences on their columns names: - The `id` column of first table has the same information as the `product_id` in the second table. - The `i.id` column of first table has the same information as the `id` in the second table. ### 3.3. Not join This method **keeps only the rows that don't match with any row of a second table**. To apply this technique we just need to negate (`!`) the table located on the `i` argument. ```{r} Products[!ProductReceived, on = c("id" = "product_id")] ``` As you can see, the result only has 'banana', as it was the only product that is not present in the `ProductReceived` table. ```{r} ProductReceived[!Products, on = c("product_id" = "id")] ``` In this case, the operation returns the row with `product_id = 6,` as it is not present on the `Products` table. ### 3.4. Semi join This method extract **keeps only the rows that match with any row in a second table** without combining the column of the tables. It's very similar to subset as join, but as in this time we are passing a complete table to the `i` we need to ensure that: - Any row in the `x` table is duplicated due row duplication in the table passed to the `i` argument. - All the renaming rows from `x` should keep the original row order. To make this, you can apply the following steps: 1. Perform a **inner join** with `which = TRUE` to save the row numbers related to each matching row of the `x` table. ```{r} SubSetRows = Products[ ProductReceived, on = .(id = product_id), nomatch = NULL, which = TRUE ] SubSetRows ``` 2. Select and sort the unique rows ids. ```{r} SubSetRowsSorted = sort(unique(SubSetRows)) SubSetRowsSorted ``` 3. Selecting the `x` rows to keep. ```{r} Products[SubSetRowsSorted] ``` ### 3.5. Left join Use this method if you need to combine columns from 2 tables based on one or more references but ***keeping all rows present in the table located on the left***. To perform this operation, we just need to **exchange the order between both tables** and the columns names in the `on` argument. ```{r} ProductReceived[Products, on = list(product_id = id)] ``` Here some important considerations: - **Column level** - The *first group* of columns now comes from the `ProductReceived` table as it is the `x` table. - The *second group* of columns now comes from the `Products` table as it is the `i` table. - It didn't add the prefix `i.` to any column. - **Row level** - All rows from in the `i` table were kept as we never received any banana but row is still part of the results. - The row related to `product_id = 6` is no part of the results any more as it is not present in the `Products` table. #### 3.5.1. Joining after chain operations One of the key features of `data.table` is that we can apply several operations before saving our final results by chaining brackets. ```r DT[ ... ][ ... ][ ... ] ``` So far, if after applying all that operations **we want to join new columns without removing any row**, we would need to stop the chaining process, save a temporary table and later apply the join operation. To avoid that situation, we can use special symbols `.SD`, to apply a **right join based on the changed table**. ```{r} NewTax[Products, on = c("unit", "type") ][, ProductReceived[.SD, on = list(product_id = id)], .SDcols = !c("unit", "type")] ``` ### 3.6. Many to many join Sometimes we want to join tables based on columns with **duplicated `id` values** to later perform some transformations later. To illustrate this situation let's take as an example the `product_id == 1L`, which have 4 rows in our `ProductReceived` table. ```{r} ProductReceived[product_id == 1L] ``` And 4 rows in our `ProductSales` table. ```{r} ProductSales[product_id == 1L] ``` To perform this join we just need to filter `product_id == 1L` in the `i` table to limit the join just to that product and set the argument `allow.cartesian = TRUE` to allow combining each row from one table with every row from the other table. ```{r} ProductReceived[ProductSales[list(1L), on = "product_id", nomatch = NULL], on = "product_id", allow.cartesian = TRUE] ``` Once we understand the result, we can apply the same process for **all products**. ```{r} ProductReceived[ProductSales, on = "product_id", allow.cartesian = TRUE] ``` > `allow.cartesian` is defaulted to FALSE as this is seldom what the user wants, and such a cross join can lead to a very large number of rows in the result. For example, if Table A has 100 rows and Table B has 50 rows, their Cartesian product would result in 5000 rows (100 * 50). This can quickly become memory-intensive for large datasets. #### 3.6.1. Selecting one match After joining the table we might find out that we just need to return a single join to extract the information we need. In this case we have 2 alternatives: - We can select the **first match**, represented in the next example by `id = 2`. ```{r} ProductReceived[ProductSales[product_id == 1L], on = .(product_id), allow.cartesian = TRUE, mult = "first"] ``` - We can select the **last match**, represented in the next example by `id = 9`. ```{r} ProductReceived[ProductSales[product_id == 1L], on = .(product_id), allow.cartesian = TRUE, mult = "last"] ``` #### 3.6.2. Cross join If you want to get **all possible row combinations** regardless of any particular id column we can follow the next process: 1. Create a new column in both tables with a constant. ```{r} ProductsTempId = copy(Products)[, temp_id := 1L] ``` 2. Join both table based on the new column and remove it after ending the process, as it doesn't have reason to stay after joining. ```{r} AllProductsMix = ProductsTempId[ProductsTempId, on = "temp_id", allow.cartesian = TRUE] AllProductsMix[, temp_id := NULL] # Removing type to make easier to see the result when printing the table AllProductsMix[, !c("type", "i.type")] ``` ### 3.7. Full join Use this method if you need to combine columns from 2 tables based on one or more references ***without removing any row***. As we saw in the previous section, any of the prior operations can keep the missing `product_id = 6` and the **soda** (`product_id = 4`) as part of the results. To save this problem, we can use the `merge` function even thought it is lower than using the native `data.table`'s joining syntax. ```{r} merge(x = Products, y = ProductReceived, by.x = "id", by.y = "product_id", all = TRUE, sort = FALSE) ``` ## 4. Non-equi join A non-equi join is a type of join where the condition for matching rows is not based on equality, but on other comparison operators like <, >, <=, or >=. This allows for **more flexible joining criteria**. In `data.table`, non-equi joins are particularly useful for operations like: - Finding the nearest match - Comparing ranges of values between tables It's a great alternative if after applying a right of inner join: - You want to decrease the number of returned rows based on comparing numeric columns of different table. - You don't need to keep the columns from table `x`*(secondary data.table)* in the final table. To illustrate how this work, let's center over attention on how are the sales and receives for product 2. ```{r} ProductSalesProd2 = ProductSales[product_id == 2L] ProductReceivedProd2 = ProductReceived[product_id == 2L] ``` If want to know, for example, if can find any receive that took place before a sales date, we can apply the next code. ```{r} ProductReceivedProd2[ProductSalesProd2, on = "product_id", allow.cartesian = TRUE ][date < i.date] ``` What does happen if we just apply the same logic on the list passed to `on`? - As this opperation it's still a right join, it returns all rows from the `i` table, but only shows the values for `id` and `count` when the rules are met. - The date related `ProductReceivedProd2` was omited from this new table. ```{r} ProductReceivedProd2[ProductSalesProd2, on = list(product_id, date < date)] ``` Now, after applying the join, we can limit the results only show the cases that meet all joining criteria. ```{r} ProductReceivedProd2[ProductSalesProd2, on = list(product_id, date < date), nomatch = NULL] ``` ## 5. Rolling join Rolling joins are particularly useful in time-series data analysis. They allow you to **match rows based on the nearest value** in a sorted column, typically a date or time column. This is valuable when you need to align data from different sources **that may not have exactly matching timestamps**, or when you want to carry forward the most recent value. For example, in financial data, you might use a rolling join to assign the most recent stock price to each transaction, even if the price updates and transactions don't occur at the exact same times. In our supermarket example, we can use a rolling join to match sales with the most recent product information. Let's assume that the price for Bananas and Carrots changes at the first date of each month. ```{r} ProductPriceHistory = data.table( product_id = rep(1:2, each = 3), date = rep(as.IDate(c("2024-01-01", "2024-02-01", "2024-03-01")), 2), price = c(0.59, 0.63, 0.65, # Banana prices 0.79, 0.89, 0.99) # Carrot prices ) ProductPriceHistory ``` Now, we can perform a right join giving a different prices for each product based on the sale date. ```{r} ProductPriceHistory[ProductSales, on = .(product_id, date), roll = TRUE, j = .(product_id, date, count, price)] ``` If we just want to see the matching cases we just need to add the argument `nomatch = NULL` to perform an inner rolling join. ```{r} ProductPriceHistory[ProductSales, on = .(product_id, date), roll = TRUE, nomatch = NULL, j = .(product_id, date, count, price)] ``` ## 7. Taking advange of joining speed ### 7.1. Subsets as joins As we just saw in the prior section the `x` table gets filtered by the values available in the `i` table. Actually, that process is faster than passing a Boolean expression to the `i` argument. To filter the `x` table at speed we don't to pass a complete `data.table`, we can pass a `list()` of vectors with the values that we want to keep or omit from the original table. For example, to filter dates where the market received 100 units of bananas (`product_id = 1`) or popcorn (`product_id = 3`) we can use the following: ```{r} ProductReceived[list(c(1L, 3L), 100L), on = c("product_id", "count")] ``` As at the end, we are filtering based on a join operation the code returned a **row that was not present in original table**. To avoid that behavior, it is recommended to always to add the argument `nomatch = NULL`. ```{r} ProductReceived[list(c(1L, 3L), 100L), on = c("product_id", "count"), nomatch = NULL] ``` We can also use this technique to filter out any combination of values by prefixing them with `!` to negate the expression in the `i` argument and keeping the `nomatch` with its default value. For example, we can filter out the 2 rows we filtered before. ```{r} ProductReceived[!list(c(1L, 3L), 100L), on = c("product_id", "count")] ``` If you just want to filter a value for a single **character column**, you can omit calling the `list()` function pass the value to been filtered in the `i` argument. ```{r} Products[c("banana","popcorn"), on = "name", nomatch = NULL] Products[!"popcorn", on = "name"] ``` ### 7.2. Updating by reference The `:=` operator in data.table is used for updating or adding columns by reference. This means it modifies the original data.table without creating a copy, which is very memory-efficient, especially for large datasets. When used inside a data.table, `:=` allows you to **add new columns** or **modify existing ones** as part of your query. Let's update our `Products` table with the latest price from `ProductPriceHistory`: ```{r} copy(Products)[ProductPriceHistory, on = .(id = product_id), j = `:=`(price = tail(i.price, 1), last_updated = tail(i.date, 1)), by = .EACHI][] ``` In this operation: - The function `copy` prevent that `:=` changes by reference the `Products` table.s - We join `Products` with `ProductPriceHistory` based on `id` and `product_id`. - We update the `price` column with the latest price from `ProductPriceHistory`. - We add a new `last_updated` column to track when the price was last changed. - The `by = .EACHI` ensures that the `tail` function is applied for each product in `ProductPriceHistory`. *** ## Reference - *Understanding data.table Rolling Joins*: https://www.r-bloggers.com/2016/06/understanding-data-table-rolling-joins/ - *Semi-join with data.table*: https://stackoverflow.com/questions/18969420/perform-a-semi-join-with-data-table - *Cross join with data.table*: https://stackoverflow.com/questions/10600060/how-to-do-cross-join-in-r - *How does one do a full join using data.table?*: https://stackoverflow.com/questions/15170741/how-does-one-do-a-full-join-using-data-table - *Enhanced data.frame*: https://rdatatable.gitlab.io/data.table/reference/data.table.html