Choosing Data Sources and Shared Semantic Models

This episode focuses on a practical PL-300 objective: Choosing Data Sources and Shared Semantic Models. A governed shared semantic model can preserve approved definitions, while separate sources may still be needed when the business requirement introduces data the model does not…

A sales director asks for a Power BI report that combines revenue, targets, and regional ownership. The organization already has a governed sales semantic model, but the targets also arrive in a separate business-managed source. Should you connect directly to every source and build a new model, query data without taking a separate copy, or reuse the existing semantic model? More importantly, what evidence would prove that your choice is correct for the requirement rather than merely a configuration that happens to work?

That decision sits inside a current skill in the PL-300 blueprint: identify and connect to data sources or a shared semantic model. The wording is short, but the analyst judgment behind it is substantial. Before choosing a connector or opening the transformation tools, you need to understand what outcome the business needs and which part of Power BI should own it.

Start by restating the requirement in operational terms. The sales director needs trusted revenue, approved regional ownership, and target data that may have a different owner. The report must also remain useful when the underlying information changes.

That gives you several questions. Who owns each set of data? How current must the report be?

Is there already a governed model that defines the business meaning? Does the analyst need to create new model behavior, or is the task mainly to present existing governed information? These questions influence whether you should ingest data, query it, or reuse an existing semantic model.

Ingesting, querying, and reusing aren’t interchangeable labels for the same design. They represent different ways for a Power BI solution to obtain or use information. Through connectors, Power BI can acquire data from sources for preparation and analysis.

Power Query is the data acquisition and transformation environment where source data can be selected, examined, cleaned, and shaped. In another design, the report can reuse a shared semantic model whose definitions and organization are already governed. The right choice depends on the requirement, not on which option feels most familiar.

Suppose the existing sales semantic model already provides the trusted revenue and regional definitions required by the director. Rebuilding those definitions from the underlying sources could duplicate work and separate the new report from the governed model. Reusing the shared semantic model may better preserve that ownership and governance.

But the separate target data raises another question. If the report requires modeling that the shared model doesn’t provide, you need to recognize that need before treating reuse as a complete answer. The existence of a reusable model is relevant evidence, not an automatic decision.

The opposite mistake is to begin importing every available table simply because Power Query makes a connection possible. A successful connection tells you that Power BI reached something and obtained a result. It doesn’t tell you that you chose the proper owner, freshness approach, governance boundary, or modeling strategy. Connection success is therefore only one observation in a larger workflow. The defensible sequence is to define the outcome, identify the responsible Power BI layer, choose the connection approach, and then verify the resulting behavior.

Those Power BI layers matter because each one controls a different kind of result. At source or query time, the concern is data acquisition and transformation. Power Query is where source information is prepared before it becomes model content. If a field arrives in an unsuitable form or unwanted records need to be excluded during preparation, the relevant evidence is in the query output. Changing a report visual wouldn’t correct the data delivered by that query.

The semantic model is concerned with the organized analytical structure used by reports. Model behavior and Data Analysis Expressions, usually called DAX, belong to a different layer from source preparation. If the data was acquired correctly but a calculation produces an unexpected result under a filter, reconnecting to the source may leave the real problem untouched. The analyst should examine the model or the calculation context that controls that result.

Report behavior is another layer. A visual interaction or filter state can make a correct model appear to answer a different question. The Power BI service adds concerns such as refresh and access state. A report may behave as expected for its author yet present a different experience to a consumer because the relevant refresh, role, permission, or consumer condition differs. This is why troubleshooting begins by identifying the observable symptom and locating the layer that directly owns it.

Return to the sales solution. Imagine that one analyst connects the report to the governed semantic model and initially sees plausible regional revenue. During review, however, the west region shows revenue but no corresponding target. The analyst could react by changing a visual, replacing the semantic model, or adding an unrelated calculation. None of those actions is justified until the failure is reproduced and described precisely.

The expected result is that the west region shows both revenue and its applicable target. The actual result is that revenue appears while the target does not. Now hold the report state steady and inspect where the target enters the workflow. If the target is missing from the output of the acquisition and preparation process, the problem is at source or query time.

If the target reaches the model but fails only under a particular regional filter, inspect the model relationship or DAX context governing that result. If the model returns the correct target but the report suppresses or filters it, inspect the report interaction. If the author sees current targets while consumers see older ones, refresh or access state becomes relevant.

Notice the discipline here. You don’t jump across all layers making speculative changes. You capture the expected and actual result, reproduce the failing condition, and inspect the configuration that controls that condition.

Once you make a correction, test the original failing region and a normal region that already worked. Fixing the west region isn’t enough if the same change causes the north region to display an incorrect target. Observable behavior, not the fact that a settings dialog accepted your choice, establishes whether the correction worked.

This also explains why there’s no universally correct connection option. Two reports may both concern sales and still need different designs. One may be intended to present governed measures exactly as the organization defines them. Reusing the shared semantic model aligns naturally with that ownership.

Another may require preparation or modeling that isn’t supplied by the shared model. That requirement may justify acquiring data through connectors instead. The decision changes with ownership, freshness, governance, and modeling needs.

Make that decision before transforming data. Otherwise, you can spend time cleaning and reshaping information that shouldn’t have been independently acquired in the first place. You may also discover late in the process that a governed semantic model already contained the approved definitions your report was expected to use. Transformation skill is important, but transformation can’t rescue a connection strategy that conflicts with the requirement.

For the sales solution, separate the requested outcome into parts without losing sight of the complete report. Revenue and regional ownership come from the governed semantic model in this hypothetical design. Targets come from a separately managed source.

For each part, record the owner, the required freshness, the governance expectation, and any needed modeling behavior. You aren’t choosing based on source names alone. You’re deciding what the report may reuse, what it must acquire, and where any preparation belongs.

If source data needs cleaning or shaping, Power Query is the appropriate environment to inspect that process. Its formula language is called M, although knowing that name doesn’t replace understanding the output. The important evidence is what each query delivers. Does the target query contain the expected regions?

Are the necessary values present after preparation? Does its output match the business requirement? A query that completes without an error can still return incomplete or unsuitable data. Successful execution and correct analytical content are different tests.

Once data reaches the semantic model, examine the behavior produced there. Does regional filtering connect the intended data? Do calculations return the expected values under relevant conditions? If a result changes incorrectly when a filter is applied, that’s evidence about model or DAX behavior rather than proof that the source connection failed. The investigation follows the result back to its relevant Power BI layer.

Then examine the report as a consumer would use it. Apply the relevant filter state and observe whether the presentation reflects the model output. A correct number hidden by a report interaction is still a failed consumer outcome. Conversely, a visually plausible chart doesn’t prove that its underlying query or calculation is correct. Plausibility is especially weak evidence when the analyst only tests the most common region or period.

Freshness needs the same deliberate treatment. If current information is a business requirement, verify the refresh or access state that affects what consumers receive. Don’t infer freshness from the report opening successfully.

Opening proves that a report was available. It doesn’t, by itself, establish that every underlying result reflects the required state. Compare the expected condition with the observable consumer experience.

A useful validation chain now emerges. First, inspect query output when the question concerns acquisition or preparation. Next, inspect semantic-model or DAX results when the question concerns analytical behavior. Then test filters and interactions when the question concerns the report.

Finally, check refresh, access, role, or consumer state when the symptom occurs in the service experience. You won’t always need every check. You select the evidence that corresponds to the suspected layer.

Suppose the initial sales report works for the default view. Total revenue and total targets both appear, so the analyst calls the work complete. A regional filter later reveals that one region has a blank target. The total was only a normal case.

It didn’t exercise the condition that exposed the defect. A better test includes both the unfiltered result and at least one meaningful filtered result. If roles or consumer access affect the requirement, the relevant experience must also be tested under that condition rather than inferred from the author’s view.

After a correction, preserve the distinction between cause and symptom. If a query omitted a region, changing a visual to hide blank values would conceal the symptom while leaving the query defect intact. If the query output contains the region and the model fails to associate it as expected, repeatedly editing the query may add complexity without addressing the model behavior. If model output is correct but a report interaction excludes the value, rebuilding the source connection would be an even larger detour. The nearest visible layer isn’t necessarily the layer causing the problem.

This requirement-to-layer mapping is useful beyond troubleshooting. It helps you evaluate design alternatives before building. Ask what the business is trying to preserve. If the priority is reuse of governed definitions, a shared semantic model deserves early consideration.

If the solution requires separate acquisition and preparation, examine the available source connection and its query output. If the requirement is a new analytical result, determine whether that work belongs in the model rather than disguising it as source cleaning. If the issue is how a consumer explores correct results, focus on the report experience.

Be cautious with broad statements such as importing is always better because it gives the analyst control, or reuse is always better because governance matters. Control without appropriate ownership can create an unnecessary competing solution. Reuse without checking modeling needs can leave the report unable to meet its requirement. Governance, freshness, ownership, and modeling needs must be considered together. The best choice is the one you can connect to the stated business outcome and verify at the layer where its behavior occurs.

The same reasoning protects the surrounding architecture. A report isn’t correct merely because its main number looks right once. Its acquisition approach should align with the responsible owner. Its prepared data should support the intended analytical structure.

Its calculations and filters should behave under relevant conditions. Its consumers should receive the expected experience. A design that satisfies one visible result while undermining those connected concerns hasn’t fully satisfied the requirement.

When an exam-style scenario presents several Power BI features, resist choosing the feature whose name most closely resembles the request. Translate the request into behavior. Is the task to acquire or prepare data, reuse governed analytical content, change model behavior, affect a report interaction, or manage what happens during refresh and access?

Then select the option that operates at that layer. The interface may expose many valid settings because Power BI solutions have different requirements. Availability doesn’t establish suitability.

For a working analyst, document the reason in plain terms. The report reuses a shared semantic model because the required sales definitions are governed there. A separate connection is considered only where the requirement calls for data or modeling not supplied by that model. Query output is checked for the expected records.

Model and DAX behavior are tested under relevant filters. Report and service conditions are checked from the consumer’s perspective. This explanation is much stronger than saying the chosen configuration worked.

If the report later fails, that record also gives troubleshooting a sensible starting point. You know which parts were reused, which were acquired, what each layer was expected to produce, and which observations originally supported the decision. Reproduce the failure, compare expected with actual behavior, and change the controlling layer. Then retest the failing condition and a previously successful condition. That cycle turns connection work from a one-time setup activity into an analyst workflow grounded in evidence.

The original sales request can now be answered without pretending there’s one universal design. The existing governed semantic model should be evaluated for the revenue and regional definitions it already owns. The target requirement should be evaluated against what that model supplies and what separate acquisition or modeling may be needed. Whichever approach is selected, the analyst proves it through the relevant outputs and consumer behavior. A green connection indicator or an error-free refresh may be useful evidence, but neither one alone proves that the business question has been answered correctly.

A sound Power BI connection decision begins with the business requirement, not with a familiar connector or setting. Decide whether the solution should ingest data, query it, or reuse a governed semantic model by considering ownership, freshness, governance, and modeling needs. Then prove the decision in the layer that controls the result.

Check query output for preparation issues, model or DAX behavior for analytical issues, and report or service conditions for consumer-facing issues. When you can explain both why the approach fits and what testable evidence confirms it, you’ve moved beyond making a connection. You’re managing the connection as part of a complete analyst workflow.

Choose a connection strategy for a source that must balance freshness, performance, and governance. Which observed result would tell you that your selected strategy is the wrong fit? Use the episode reference to review precise terminology and the checks that are easier to inspect than to hear. This lesson was produced by Bare Metal Cyber dot com.

Choosing Data Sources and Shared Semantic Models

headphones Listen Anywhere

More Options »
Broadcast by