This is a matter of taste. The main reason I see to break things up at all is for security, and secondarily for backups.
I think of ETL activity, including loading the final star schema tables, as an aspect of the
Staging layer. That's where all the "heavy lifting" happens in my warehouses.
Here's how I typically break up a data warehouse. These should be distinct physical databases, and may be separate VMs or physical machines if the volume of data requires it.
- A separate database for each important data source (CRM, OLTP database, telephony system, etc.). Here, I store a straightforward copy of data from the source system, with a minimum of modification. E.g., I will record
DateModified for each record, but perform no cleanup of data values or types. The idea is to grab data as quickly as possible, to minimize load on the source system.
- In principle, these databases can be purged at will and repopulated from the source systems with no change to the rest of the warehouse.
- Using separate databases for each source lets users (i.e., service accounts) have limited access. My
Service_SalesforceReader user has read and write privs in the
Salesforce database, and nowhere else.
- Writing to these databases typically happens via dedicated ETL applications, with a few SQL procedures supporting them.
Staging database for the intermediate steps; the "T" in "ETL". This includes lookup tables for data cleaning, tables for allocating synthetic keys, and log tables for ETL processes.
- Data here is mostly static which can be recreated from source control (e.g., country codes) or transformations of the source data, which can be recreated from those systems, albeit perhaps with different synthetic keys being created. Only people on the warehouse team should need access here.
- This database has 90% of the SQL procedures. They format and pivot the data into the star schema layout. There's one objective reason to keep the ETL logic in Staging: security. Only warehouse admins need access to those procedures, whereas many people may need access to the Warehouse layer.
Warehouse database, with tidy fact tables and dimension tables. This layer may have views to help with specific consumers. E.g., if you have a visualization tool that doesn't play nice with joins, there might be a view (possibly materialized) for each star, with all dimensions joined in.
- The only sprocs I have in this layer, in my current warehouse, are for logging, and I don't believe they're actually used.
In practice, you might need to grant some power users access to specific source databases, if there's data they need that doesn't rate being incorporated into the star schema. In such a case, you may also need to grant access to the synthetic key tables in
Staging that link them together, so they can traverse from (for example) a
CustomerSK integer to a Salesforce
AccountID string. Doing so makes it difficult for you to change anything without breaking people's processes, so this should be a last resort.