Configuring Cross-Filter Direction
A sales report appears to be working. Selecting a product category changes the revenue total, and selecting a region changes it again. The relationships are active, the visuals respond, and Power BI shows no obvious configuration problem.
Is that enough evidence that the model is correct? Not quite. The real test is whether filters travel through the model in the direction the business requirement needs, without creating competing routes or unintended effects elsewhere.
That distinction matters for the PL-300 skill of defining a relationship’s cardinality and cross-filter direction. Cardinality describes how rows on the two sides of a relationship can match, such as one row on one side matching many rows on the other. Cross-filter direction determines how filter context can propagate across that relationship. Together, those properties shape what a report calculates when someone selects a category, customer, region, or another value.
This behavior belongs primarily to the semantic model, which contains the tables, relationships, metadata, and calculations used by reports. It isn’t normally a Power Query transformation. Power Query can prepare and reshape data at refresh time, but changing a query step doesn’t replace a model relationship’s filter propagation. A report interaction can decide whether one visual affects another, but that interaction still depends on the model producing the right filtered result. Service permissions govern access to published content, but they don’t repair a relationship path that was modeled incorrectly.
Suppose a company has a Product table, a Customer table, and a Sales table. Each product can appear in many sales rows, and each customer can also appear in many sales rows. Product and Customer are descriptive tables, often called dimensions, while Sales records business events at a more detailed grain. The relationships allow filters from Product and Customer to affect the matching rows in Sales.
With the common single-direction design, a filter on Product can propagate to Sales. If someone selects the Bikes category, the sales total can be calculated from rows associated with bikes. A filter on Customer can also propagate to Sales. If someone selects customers in a particular region, the total can be calculated from those customers’ sales. This direction supports a straightforward analytical pattern in which descriptive tables filter a fact table.
The reverse journey isn’t automatically allowed. A Product filter that reaches Sales doesn’t necessarily continue backward from Sales into Customer. That means a Bikes selection may change the sales amount while leaving a customer slicer showing every customer, including customers who bought no bikes. The calculation can still be correct. The remaining customer choices simply haven’t been reduced through the Sales table.
Now the business asks for a different experience. When a user chooses Bikes, the customer slicer should show only customers with bike sales. That requirement needs filter context to travel from Product to Sales and then from Sales toward Customer. A bidirectional relationship, displayed as Both in the relationship setting, can permit filtering in both directions across the selected relationship. In this case, it might produce the requested slicer behavior.
The important word is might. Both isn’t a general upgrade from single direction. It changes the routes available to filter context, so it can affect calculations and interactions beyond the one slicer that motivated the change. It can also introduce performance or security complexity. The responsible decision starts with the required outcome, then asks whether enabling both directions creates a clear and unambiguous path through the whole model.
Consider an expanded version of the same model. Product and Customer are both related to Sales. They’re also related through a Returns table, because each return records a product and a customer. If relationships around both transaction tables are allowed to filter in both directions, a Product filter may be able to reach Customer through Sales and also through Returns. Those are two different relationship paths between the same parts of the model.
That’s the heart of an ambiguous filter path. Power BI relationships propagate filters along relationship paths. If the model offers more than one route that can carry a filter between tables, the intended meaning may no longer be clear. Should a product selection restrict customers based on sales, returns, or the combined consequences of both routes? The relationship settings alone don’t state the business interpretation.
This isn’t just a diagramming concern. The paths determine filter context, and filter context determines model results. A measure written in Data Analysis Expressions, usually called DAX, is evaluated within that context. If the available paths cause a customer filter to propagate more broadly than intended, a measure can receive a different set of sales or return rows. A visual may then display a plausible number that answers the wrong business question.
Start with single-direction relationships because they normally make the flow easier to reason about. In the sales model, Product filters Sales, Customer filters Sales, Product filters Returns, and Customer filters Returns. Each dimension affects the applicable transaction table, but a transaction table doesn’t automatically filter a dimension. That structure preserves clear routes for the usual revenue and return analysis.
Use Both only when the analysis actually requires reverse propagation and the resulting paths remain unambiguous. If the requirement is merely to calculate bike revenue by customer region, single-direction relationships already support it. Product and Customer filters both reach Sales, and the revenue measure is evaluated over the intersection of those filters. Bidirectional filtering isn’t needed just because two dimensions participate in the same calculation.
The customer-slicer requirement is different because it asks the selected product to alter the available customers. Before enabling Both, restate that outcome precisely. Is the slicer supposed to show customers with sales for the selected product, customers with returns for it, or either group? If the answer is sales, enabling reverse filtering through both Sales and Returns would be broader than the requirement. A setting that creates the visible effect in one normal case can still encode the wrong meaning.
This is why changing unrelated layers rarely solves the underlying problem. Removing customers in Power Query would alter the data loaded at refresh time, not respond dynamically to a report selection. Editing the visual interaction might stop or permit one visual’s influence, but it wouldn’t redefine how filters propagate through the semantic model. Changing a DAX measure might alter one result while leaving slicers and other measures subject to the original relationship behavior. Each layer controls something different.
Imagine that another analyst enables Both and reports that the solution works. You select Bikes, and the customer slicer becomes shorter. That observation is useful, but it’s only the beginning of validation.
Record what the expected customer set should be from the underlying business condition. Then compare the displayed customers with that expectation. A reduced list isn’t automatically the correct list.
Test a customer who bought a bike and returned it, a customer who bought one without returning it, and a customer who returned another product. These are hypothetical data conditions, not special Power BI categories. They help reveal which route is controlling the result. If the slicer includes or excludes a customer for the wrong reason, the model’s filter paths don’t represent the stated requirement even though the visual responds.
Retest the ordinary calculations too. Select a region and product, then observe the sales amount, return amount, customer count, and any other result that depends on those tables. Compare expected and actual values rather than relying on animation or visual changes. A responsive report can still propagate more filter context than intended. The evidence you need is the resulting row set or calculation, not simply proof that a selection triggered something.
Performance also belongs in the evaluation because bidirectional filtering can add complexity to how filters move through a model. That doesn’t mean every use of Both will produce an unacceptable report. It means the modeler shouldn’t ignore performance merely because the result is logically correct. Test representative report interactions and data conditions, especially those that activate the relevant paths. The useful comparison is whether the design satisfies the requirement while maintaining acceptable behavior for the consumer.
Security deserves the same caution. Bidirectional filtering may create security complexity because filter propagation can reach tables through directions the modeler didn’t initially intend. A result that looks right during ordinary testing doesn’t establish that every governed experience is right. Where roles or restricted consumer experiences apply, validation needs to occur under those conditions. Don’t infer secure behavior from the unrestricted authoring view.
A disciplined troubleshooting process begins with one failing condition. Suppose the regional sales total is correct until a product is selected, after which the customer count becomes unexpectedly small. Preserve that combination of selections so you can reproduce it. State the expected result, the actual result, and which tables should be affected. Without that baseline, changing relationship settings becomes guesswork.
Next, inspect the relationships that directly govern the filter journey. Trace the intended route from Region or Customer into Sales and from Product into Sales. Then look for another active route through Returns or another related table.
Check direction as well as connectivity. Two tables being connected doesn’t tell you which way a filter is allowed to travel. A single-direction relationship may stop a route that a Both setting would open.
Resist the temptation to change every nearby relationship to Both. That can make the original symptom disappear by allowing more propagation, while creating new ambiguity somewhere else. A relationship option is appropriate only when its effect matches the business rule for the model. The fact that Power BI offers the option means it supports valid scenarios, not that the option is universally preferable.
After correcting the relationship design, repeat the original test without changing the expected result. If the product selection should affect revenue but not reduce the customer slicer, confirm both observations. If the approved requirement is that it should reduce the slicer to customers with matching sales, confirm the exact membership of that list. Then remove the product filter and make sure the report returns to the broader state you expect.
An edge condition is valuable because normal data can hide a poor design. If most customers have both sales and returns, the two routes may produce nearly identical customer lists. A customer who appears in only one transaction table exposes the distinction. Refreshing with a changed data condition can reveal the same issue, but refresh success by itself proves only that processing completed. It doesn’t prove that filter propagation still matches the business rule.
The same principle applies to DAX results. A measure returning a number without an error isn’t proof of correctness. Compare its output under controlled filter states.
Begin with no product selection, then one product category, then a customer or region that has a known relationship to the selected products. If the measure changes when its relevant rows change, that’s observable evidence. If an unrelated selection changes it, investigate the filter path that allowed the context to reach it.
You can also compare two visuals that should share the same context. If a sales total and a customer count are both intended to reflect the selected product, verify that they respond according to their definitions. If only the sales total should respond, a customer count that changes may reveal unintended reverse propagation. What counts as correct depends on the stated requirement, not on a general preference for more interactive behavior.
This requirement-first approach keeps the modeling decision separate from cosmetic convenience. A stakeholder might say that every slicer should always narrow every other slicer. That sounds attractive, but it’s not yet a complete model rule. Ask what population each slicer is supposed to represent and through which business events those populations are related. Sales and returns can connect the same products and customers while expressing different facts.
Cardinality remains part of the explanation because it describes the matching structure at each relationship. Cross-filter direction then controls how context is permitted to move across that structure. Neither property should be considered in isolation. A direction that seems useful on one relationship can create a second route when combined with the rest of the model. The model, not the individual settings viewed separately, produces the report behavior.
For an applied PL-300 decision, you should be able to identify the owning layer, explain the propagation mechanism, and name the evidence that would establish correctness. Here, the responsible layer is the semantic model relationship design. The mechanism is filter context moving along paths permitted by cardinality and cross-filter direction. The evidence comes from controlled selections, expected row populations, calculation results, representative performance, and any governed consumer state that matters to the requirement.
That also gives you a reliable way to reject tempting alternatives. A Power Query change is wrong when the requirement is dynamic report filtering because query transformations occur during data preparation and refresh. A visual-only change is incomplete when the model itself supplies the wrong filter context. A rewritten measure may solve one calculation while leaving other model behavior unchanged. Choosing the nearest available feature isn’t the same as choosing the layer that owns the behavior.
Return to the original sales report. Its changing revenue total showed that at least one filter path was active. It didn’t tell you whether the route was unique, whether reverse filtering was intentional, or whether other calculations and consumer states remained correct. Once those questions are tested, the report is no longer accepted merely because it appears responsive. It’s accepted because its relationship paths represent the business meaning and produce repeatable, observable results.
Cross-filter direction is a model decision about where filter context is allowed to travel, not a switch that should be made more permissive by default. Single direction usually provides the clearest starting point. Both can satisfy a genuine reverse-filtering requirement, but only when the resulting relationship paths remain unambiguous and the wider consequences are acceptable. The practical habit is to state the business outcome, trace the permitted paths through the semantic model, and verify the resulting customers, rows, calculations, interactions, and governed experiences. When those observations match the requirement under normal and edge conditions, you have evidence that the design is correct, rather than a configuration that merely happens to look convincing.
Predict how a filter should travel through the model before changing cross-filter direction. Which visual interaction would expose an ambiguous or overly broad filter path? The matching episode reference turns the spoken reasoning into a compact set of terms and hands-on tests. This lesson was produced by Bare Metal Cyber dot com.