Implementing data snapshots

Build point-in-time snapshots from change-history data so you can report the latest status of an entity as of any date.

Use case

For a dataset that contains a sequence of changes to a property over time, we want to be able to get the most recent state of said property at any given date. In this recipe, we’ll learn how to calculate snapshots of statuses at any given date for a cube with Product Id, Status, and Changed At dimensions.

We can consider the status property to be a slowly changing dimension (SCD) of type 2. Modeling data with slowly changing dimensions is an essential part of the data engineering skillset.

Data modeling

Let’s explore the statuses cube that contains data like this:

order_id	status	changed_at
1	shipped	2019-01-19 00:00:00
1	processing	2019-03-14 00:00:00
1	completed	2019-01-25 00:00:00
2	processing	2019-08-21 00:00:00
2	completed	2019-04-13 00:00:00
2	shipped	2019-03-18 00:00:00

We can see that statuses change occasionally. How do we count orders that remained in the shipped status at a particular date?

First, we need to generate a range with all dates of interest, from the earliest to the latest. Second, we need to join the dates with the statuses and leave only the most recent statuses to date.

YAML

cubes:
  - name: status_snapshots
    extends: statuses
    sql: |
      -- Create a range from the earlist date to the latest date
      WITH range AS (
        SELECT date
        FROM GENERATE_SERIES(
          (SELECT MIN(changed_at) FROM {statuses.sql()} AS statuses),
          (SELECT MAX(changed_at) FROM {statuses.sql()} AS statuses),
          INTERVAL '1 DAY'
        ) AS date
      )

      -- Calculate snapshots for every date in the range
      SELECT range.date, statuses.*
      FROM range
      LEFT JOIN {statuses.sql()} AS statuses
        ON range.date >= statuses.changed_at
        AND statuses.changed_at = (
          SELECT MAX(changed_at)
          FROM {statuses.sql()} AS sub_statuses
          WHERE sub_statuses.order_id = statuses.order_id
        )
    dimensions:
      - name: date
        sql: date
        type: time

cube(`status_snapshots`, {
  extends: statuses,

  sql: `
    -- Create a range from the earlist date to the latest date
    WITH range AS (
      SELECT date
      FROM GENERATE_SERIES(
        (SELECT MIN(changed_at) FROM ${statuses.sql()} AS statuses),
        (SELECT MAX(changed_at) FROM ${statuses.sql()} AS statuses),
        INTERVAL '1 DAY'
      ) AS date
    )

    -- Calculate snapshots for every date in the range
    SELECT range.date, statuses.*
    FROM range
    LEFT JOIN ${statuses.sql()} AS statuses
      ON range.date >= statuses.changed_at
      AND statuses.changed_at = (
        SELECT MAX(changed_at)
        FROM ${statuses.sql()} AS sub_statuses
        WHERE sub_statuses.order_id = statuses.order_id
      )
  `,

  dimensions: {
    date: {
      sql: `date`,
      type: `time`
    }
  }
})

To generate a range of dates, here we use the GENERATE_SERIES function which is Postgres-specific. Other databases have similar functions, e.g., GENERATE_DATE_ARRAY in BigQuery.

Please note that it makes sense to make the status_snapshots cube extend the original statuses cube in order to reuse the dimension definitions. We only need to add a new dimension that indicates the date of a snapshot. We’re also referencing the definition of the statuses cube with the sql() property.

Result

To count orders that remained in the shipped status at a particular date, filter on both status_snapshots.date and status_snapshots.status. Running the same query for two different dates shows how the snapshot changes over time:

date	status	count
2019-04-01	shipped	16
2019-05-01	shipped	25