 |
| Business Intelligence Series |
Independently whether standalone or encapsulated in database objects, the queries written can become complex in time, respectively difficult to comprehend and maintain. One can reduce the cognitive load by identifying the aspects that enable one’s intuition - order, similarity and origin being probably the most important aspects that help coping with the inherent complexity.
One should start with the table having the lowest level of detail, usually a transaction table that forms the backbone of a certain feature. For example, for Purchase Orders this could be upon case the distribution or line level. If Invoices are added to the logic, and there could be multiple invoice line for a record from the former logic, then this becomes the new level of detail. Further on, if General Ledger lines are added, more likely this becomes the lowest level of detail, and so on.
This approach allows to keep a consistent way of building the queries while enabling to validate the record count, making sure that no duplicates are added to the logic. Conversely, one can start also from the table with the lowest level of details, and add tables successively based on their level of detail, though the database engine may generate upon case a suboptimal plan. In addition, checking the cardinality may involve writing a second query.
One should try to keep the tables belonging to the same entity together, when possible (e.g. Purchase Order vs. Vendor information). This approach allows to reduce the volume of work required to manage, review, test and understand the query later. If the blocks are too big, then occasionally it makes sense to bring pieces of logic into CTEs (Common Table Expressions), or much better into views that allow to better encapsulate and partition the logic.
CTEs allow to split the logic into logical steps, allowing occasionally to troubleshoot the logic on pieces though one should keep a balance between maintainability and detail. In extremis, people may create unnecessarily an additional CTE for each join. The longer and the more fragmented a query, the more difficult it becomes to troubleshoot and even understand. Readability can be better achieved though indentation, by keeping things together that belong together, respectively partitioning the logic in logical blocks that derive from the model.
Keeping things together should be applied also to the other elements. For example, the join constraints should be limited only to the fields participating in the join (and, if possible, all the other constraints should be brought in the WHERE clause). Bringing the join constraints in the WHERE clause, an approach used in the past, decreases query readability no matter how well the WHERE clause is structured, and occasionally can lead to constraints’ duplication or worse, to missing pieces of logic.
The order of the attributes should follow a logic that makes sense. One approach is to start from the tables with lowest cardinality that identify entities uniquely and move to the attributes that have a higher level of detail. However, not all attributes are equally important, and thus one might have to compromise and create multiple groups of data coming from different levels.
One should keep in mind that the more random the order of the attributes is, the more difficult it becomes to validate the data as one needs to jump multiple times either in the query or in the mask. Ideally one should find a balance between the two perspectives. Having an intuitive logic of how the attributes are listed increases queries’ readability, maintainability and troubleshooting. The more random attributes’ listing, the higher the effort for the same.
