Slowly Changing Dimension methodology
Source:vignettes/slowly-changing-dimension.Rmd
slowly-changing-dimension.Rmd
library(tidyverse)
#> ── Attaching core tidyverse packages ──────────────────────── tidyverse 2.0.0 ──
#> ✔ dplyr 1.1.4 ✔ readr 2.1.5
#> ✔ forcats 1.0.0 ✔ stringr 1.5.1
#> ✔ ggplot2 3.5.1 ✔ tibble 3.2.1
#> ✔ lubridate 1.9.3 ✔ tidyr 1.3.1
#> ✔ purrr 1.0.2
#> ── Conflicts ────────────────────────────────────────── tidyverse_conflicts() ──
#> ✖ dplyr::filter() masks stats::filter()
#> ✖ dplyr::lag() masks stats::lag()
#> ℹ Use the conflicted package (<http://conflicted.r-lib.org/>) to force all conflicts to become errorsA slowly changing dimension is a concept in data warehousing which refers to data which may change over time, but at an irregular schedule.
Type 1 and Type 2 history
For example, consider the following table of forecasts for a number of cities:
# Current date: 2023-09-28
forecasts
#> # A tibble: 4 × 2
#> City Forecast
#> <chr> <dbl>
#> 1 New York 20
#> 2 Los Angeles 23
#> 3 Seattle 16
#> 4 Houston 34The following day, the forecasts will have changed, and — barring the occasional data hoarder — the existing data is no longer relevant.
In this example, most (if not all) of the values of the
Forecast column will change with each regular update.
Putting it into other words, the table is a snapshot1 of
forecasts at the last time of update.
The following day, the forecasts naturally change:
# Current date: 2023-09-29
forecasts2
#> # A tibble: 4 × 2
#> City Forecast
#> <chr> <dbl>
#> 1 New York 18
#> 2 Los Angeles 25
#> 3 Seattle 17
#> 4 Houston 34We could choose to update the forecasts table so that it would always contain the current data. This is what is referred to as Type 1 methodology (Kimball and Ross 2013).
Databases are thankfully a rather efficient way of storing and
accessing data, so instead of discarding the values from the previous
day, we append the new data to those of the previous day. Also, in order
to keep our data organized, we add a column with the date of the
forecast, aptly named ForecastDate.
The full table of forecasts for the two days now looks like below, and we are slowly building a full history of forecasts:
forecasts_full
#> # A tibble: 8 × 3
#> City Forecast ForecastDate
#> <chr> <dbl> <date>
#> 1 New York 20 2023-09-28
#> 2 Los Angeles 23 2023-09-28
#> 3 Seattle 16 2023-09-28
#> 4 Houston 34 2023-09-28
#> 5 New York 18 2023-09-29
#> 6 Los Angeles 25 2023-09-29
#> 7 Seattle 17 2023-09-29
#> 8 Houston 34 2023-09-29Managing historical data by inserting new data in this manner is often referred to as Type 2 methodology or Type 2 history.
Our table now provides much more information for the user through filtering:
# Current forecasts
forecasts_full %>%
slice_max(ForecastDate, n = 1) %>%
select(!"ForecastDate")
#> # A tibble: 4 × 2
#> City Forecast
#> <chr> <dbl>
#> 1 New York 18
#> 2 Los Angeles 25
#> 3 Seattle 17
#> 4 Houston 34
# Forecasts for a given date
forecasts_full %>%
filter(ForecastDate == "2023-09-28")
#> # A tibble: 4 × 3
#> City Forecast ForecastDate
#> <chr> <dbl> <date>
#> 1 New York 20 2023-09-28
#> 2 Los Angeles 23 2023-09-28
#> 3 Seattle 16 2023-09-28
#> 4 Houston 34 2023-09-28
# Full history for a given city
forecasts_full %>%
filter(City == "New York")
#> # A tibble: 2 × 3
#> City Forecast ForecastDate
#> <chr> <dbl> <date>
#> 1 New York 20 2023-09-28
#> 2 New York 18 2023-09-29Now, we note that the forecast for Houston has not changed between the two days.
In order to keep our data as minimized as possible, we modify the
table again, now expanding ForecastDate into
ForecastFrom and ForecastUntil.
Our table of forecasts now looks like this:
forecasts_scd
#> # A tibble: 7 × 4
#> City Forecast ForecastFrom ForecastUntil
#> <chr> <dbl> <date> <date>
#> 1 New York 20 2023-09-28 2023-09-29
#> 2 Los Angeles 23 2023-09-28 2023-09-29
#> 3 Seattle 16 2023-09-28 2023-09-29
#> 4 Houston 34 2023-09-28 NA
#> 5 New York 18 2023-09-29 NA
#> 6 Los Angeles 25 2023-09-29 NA
#> 7 Seattle 17 2023-09-29 NAFor now, the ForecastUntil value is set to
NA, as it is not known when these rows will “expire” (if
ever). This also makes it easy to identify currently valid data.
Adding a new column to save a single row of data naturally seems a bit overkill, but as the number of rows in the data set increases indefinitely, this solutions scales much better.
A “timeline of timelines”
Let’s now introduce additional information and see how managing slowly changing dimensions enables us to easily navigate large amounts of data over large periods of time.
Imagine a town of several thousand citizens, with a town hall maintaining a civil registry of names and addresses of every citizen, updated daily with any changes submitted by the citizens, each of whom having an individual identification number.2
The data is largely static, as a very small fraction of citizens move on any given day, but it is of interest to keep data relatively up-to-date. This is where managing a slowly changing dimension becomes very powerful, compared to full incremental backups.
One day, Alice Doe meets Robert “Bobby” Tables, and they move in together:
addresses
#> # A tibble: 4 × 8
#> ID GivenName Surname Address MovedIn MovedOut ValidFrom ValidUntil
#> <dbl> <chr> <chr> <chr> <date> <date> <date> <date>
#> 1 1 Alice Doe Donut Pla… 1989-06-26 NA 1989-06-26 2021-03-08
#> 2 2 Robert Tables Rainbow R… 1989-12-13 NA 1989-12-13 NA
#> 3 1 Alice Doe Donut Pla… 1989-06-26 2021-03-01 2021-03-08 NA
#> 4 1 Alice Doe Rainbow R… 2021-03-01 NA 2021-03-08 NAFirst thing to notice is that the registry is not updated in
real-time, as citizens may have been late in registering a change of
address. This can be seen when comparing the values of
MovedIn and ValidFrom for row 4.
When using Type 2 history, this feature is correctly replicated when reconstructing historical data:
slice_timestamp <- "2021-03-02"
addresses %>%
filter(ID == 1,
ValidFrom < !!slice_timestamp,
ValidUntil >= !!slice_timestamp | is.na(ValidUntil)) %>%
select(!c("ValidFrom", "ValidUntil"))
#> # A tibble: 1 × 6
#> ID GivenName Surname Address MovedIn MovedOut
#> <dbl> <chr> <chr> <chr> <date> <date>
#> 1 1 Alice Doe Donut Plains 1 1989-06-26 NAIn other words, even though Alice’s address was subsequently updated in the registry, we can still see that she was registered as living in Donut Plains at this time. This modeling of “timelines of timelines” is also called bitemporal modeling.
By now, things are going well between Alice and Robert; they get married, with Alice taking Robert’s surname. It is the same person that has lived with Robert, but as of the day of marriage, she has a different name:
filter(addresses2,
ID == 1,
Address == "Rainbow Road 8") %>%
select(ID, GivenName, Surname, MovedIn, MovedOut, ValidFrom, ValidUntil)
#> # A tibble: 2 × 7
#> ID GivenName Surname MovedIn MovedOut ValidFrom ValidUntil
#> <dbl> <chr> <chr> <date> <date> <date> <date>
#> 1 1 Alice Doe 2021-03-01 NA 2021-03-08 2023-08-28
#> 2 1 Alice Tables 2021-03-01 NA 2023-08-28 NAThis is now also reflected in the data; the MovedIn date
is persistent across the date of the name change, only the
Surname changes:
slice_timestamp <- "2022-03-04"
addresses2 %>%
filter(Address == "Rainbow Road 8",
is.na(MovedOut),
ValidFrom < !!slice_timestamp,
ValidUntil >= !!slice_timestamp | is.na(ValidUntil)) %>%
select(ID, GivenName, Surname, MovedIn, MovedOut)
#> # A tibble: 2 × 5
#> ID GivenName Surname MovedIn MovedOut
#> <dbl> <chr> <chr> <date> <date>
#> 1 2 Robert Tables 1989-12-13 NA
#> 2 1 Alice Doe 2021-03-01 NA
slice_timestamp <- "2023-09-29"
addresses2 %>%
filter(Address == "Rainbow Road 8",
is.na(MovedOut),
ValidFrom < !!slice_timestamp,
ValidUntil >= !!slice_timestamp | is.na(ValidUntil)) %>%
select(ID, GivenName, Surname, MovedIn, MovedOut)
#> # A tibble: 2 × 5
#> ID GivenName Surname MovedIn MovedOut
#> <dbl> <chr> <chr> <date> <date>
#> 1 2 Robert Tables 1989-12-13 NA
#> 2 1 Alice Tables 2021-03-01 NA