Generate history from Snapshot data

Frequently, the Data Warehouse / Business Intelligence system encounters scenario wherein historical changes need to be captured and/or derived when the source provides only the latest data (i.e snapshot) as on date.

Typically, generating history is viewed from Slowly Changing Dimension (SCD) perspective. Logically, it is correct because fact data is immutable by nature as it refers to an event in transaction system. ETL tools today have predefined transformations that can identify a change on the data and flag (e.g. Table Comparison & History Preserve in BODS) them accordingly. Using the flag (i.e Insert, Update or Delete) in previous step, the ETL pipeline processes the data further based on business logic.

When we look deeper, these ETL transformations depend on two factors to perform the check and update. Among the two factors, “key attribute(s)” forms the premise for capturing history and the “tracking attribute(s)” that will be referenced during the check and update process.

Good use case is CRM Opportunity analysis when a user would analyze the history of changes and associated activities. For simplicity, the example here shows a single key attribute and one tracking attribute, which along with date attributes show various changes to the data over time. In real world scenarios, the key can be a composite key and tracking attributes can be more than one.

In a generic sense, consider a scenario wherein multiple key attributes are to be used to track change. In CRM context, we can have Order Value, Account Owner, Status, Activity, Contact Person, Close Date etc can change multiple times, which impacts the number of fields defined as “key attribute”. Standard ETL transformations can handle such scenarios, but there are few tricks that can be employed to improve performance. ETL tools execute multiple select queries against the database with target table data in order to perform the compare against new snapshot data. This might lead to inefficiency for large volumes of data.

The focus here is scenarios where latest snapshot is provided by source system. Without using ETL transformations, if we are to derive the flags (i.e. I, U, D), it is complicated without long procedures to scan all table content. Classic work around is to create a single field that is concatenated value of all key attribute fields and use the same as key attribute. This “key attribute” can be a large character field that is easier to manage and create indexes upon for performance, which need not be exposed to reporting layer.

Adding upon to that idea, a cool and nifty feature is the MD5 hash that can be generated on the “key attribute”. It is a standard function available in all databases (e.g. HASHBYTES(‘MD5′,’text value here’) in MS SQL Server) which has roots in cryptography. An MD5 hash has constant length and always stays the same for a string of any length. Keeping this field as “key attribute” is more efficient for processing lookups and managing indices. Sample text and MD5 hashes are below

  • Hash: 3f4190b90136c5ed60a56a558ff6bbf1
    • String: CUST001~JOHN~USA~MBA~12YEXP~20PROJ~AINC~Senior Consultant~100K
  • Hash: caa08e04e6c31b7483d5a0278dfb245a
    • String: CUST001~JOHN~USA~MBA

In this approach we are effectively keeping the ETL configuration simple by creating the “key attribute” column in delegated manner, which will aid in performance and management.

NOTE: SAP HANA HASH_MD5 function processes the MD5 in binary mode than character mode and type conversions are required before converting the data.


Be the first to comment

Leave a Reply

Your email address will not be published.