I have seen many a data warehouses which are completely emptied very night and reloaded again. This is a perfectly fine ETL strategy. Many would argue that those are not data warehouses in strict sense but I will leave it out from this blog.
The other strategy is, obviously, to do incremental load or merge load where data is incrementally modified in the data warehouse (I will stick to calling it Merge strategy). So which one is better? Well, like answers to most questions in IT, “it depends”. Here are few counter questions I ask before deciding on the strategy. They are in no particular order and certainly not ordered by importance. Before making any decision, each one needs to be considered carefully.
How much data are you expecting from the source?
With smaller data sizes, Truncate/Reload generally performs better, plus ETL tends to be simpler. For larger data sizes: it’s difficult to say which one will perform better but I have found that in the longer run Merge starategy pays off in terms of performance.
How many data sources are there?
With multiple data sources, I incline towards Merge. I can import raw data from all the sources in one central place (variedly called imports database, staging database, holding database, raw_datastore etc.), baseline it and then continue with data loading to data warehouse. If for some reason one of the data sources is not available, I can handle that gracefully in ETL and when the data does becomes available I can retrospectively amend it. With Truncate/Reload, it kind of become manadatory that all the data sources are available during ETL and data across them is synchronised.
What are the types of the data sources?
If the data source is in the form of flat file, doing Truncate/Reload may not be such a good idea. What if the file becomes unavailable during ETL? Also reading from set of files everyday, doesn’t sound right.
Is ETL data source a complete data set or delta load?
Obviously, if the data source is delta load, then Merge is only option.
Do you need to maintain history?
If there are Type II and Type III attributes in data warehouse and data sources doesn’t store audit information, then Merge is the only option. A different take on the same question is, ‘Would there be a future need to show historical data?’ ,in which case, I use Merge even though upfront ETL efforts are more.
How complex are the ETL rules?
If ETL only consists of moving data from source into star schema structure to ease reporting, Truncate/Reload can be a suitable options. However, for complex ETL, I prefer Merge .
Does the source have data modification timestamps?
Having these can be very helful in Merge strategy.
Would you need to extend the solution?
If there are chances of adding more data sources, the Merge is better option. It’s usually difficult to add new data source if ETL is designed around Truncate/Reload strategy.
How much is your ETL operation window?
In most cases, this isn’t an issue but still worthwhile considering upfront.
Hope that was helpful. Feel free to let me know any other issues that you might consider before selecting your ETL strategy.