Power BI interviews should reveal whether a candidate can turn source data into an understandable, secure, and trustworthy analytical model. Recruiters need to assess more than visual design: the candidate should explain grain, relationships, calculation context, refresh, and how the same report behaves for different users. A polished dashboard is useful only when its numbers support the intended decisions.
This guide to 100+ Power BI Interview Questions and Answers contains 120 questions across beginner, intermediate, and experienced levels. It moves from semantic models and report interactions to DAX, Power Query, storage modes, security, performance, deployment, and operational diagnosis. Use the scenarios to distinguish familiarity with the interface from the ability to defend a business result and maintain it through change.
How can recruiters assess genuine Power BI reasoning when interview assistance can hide gaps in DAX context, data modeling, and security?
Give the candidate a small model with a misleading total or a slicer that follows the wrong relationship. Ask them to explain the expected result before changing the formula, then introduce a different user role or refresh boundary. Their investigation and verification plan reveal more than a memorized function definition.
Sherlock AI helps hiring teams assess signals of potential AI cheating, deepfakes, impersonation, and unauthorized assistance during interviews.
Evaluate those signals alongside the candidateโs practical work and explanation, not as conclusive proof of misconduct. A useful Power BI exercise rewards careful assumptions, correct results, and a clear account of what has actually been tested.
Power BI Interview Questions
The complete question bank appears first so interviewers can select prompts without exposing the answers. Match the questions to the roleโs responsibilities and state the available storage mode, service features, and data assumptions.
Power BI Interview Questions for Beginners
Establish the core vocabulary of models, data preparation, calculations, reports, and access. Ask candidates to explain what each feature means for a concrete business question.
What is Power BI, and what problems does it solve?
How do Power BI Desktop and the Power BI service differ?
How does Power BI relate to Microsoft Fabric?
What is a Power BI semantic model?
How do reports and dashboards differ?
What is a workspace?
What is a Power BI app?
What is a star schema?
What does a fact table's grain mean?
What is a dimension table used for?
What does relationship cardinality describe?
What is cross-filter direction?
How do active and inactive relationships differ?
Why is a dedicated date table useful?
What is Power Query?
How does M differ from DAX?
What is query folding?
Why should data types be set deliberately?
How do merge and append operations differ in Power Query?
What is a measure?
How does a calculated column differ from a measure?
What is row context in DAX?
What is filter context in DAX?
What does CALCULATE do?
How do SUM and SUMX differ?
Why is DIVIDE often useful for ratios?
How do BLANK and zero differ in a report?
What is a slicer?
How do visual, page, and report filters differ?
What is the difference between drill-down and drillthrough?
How do tooltips help a report?
What are bookmarks used for?
How would you choose an appropriate visual?
What makes a Power BI report accessible?
What is Import mode?
What is DirectQuery?
What is Direct Lake?
What is an on-premises data gateway used for?
What is row-level security?
How would you approach a first Power BI project?
Power BI Interview Questions for Intermediate
Explore filter behavior, transformations, refresh, security, performance, and deployment. Strong answers identify the assumptions and checks that make a proposed solution reliable.
What is context transition in DAX?
How does KEEPFILTERS change a CALCULATE filter?
How would you calculate a percentage of a defined total?
When is SELECTEDVALUE useful?
Why use variables in DAX?
Why can a grand total differ from the sum of visible rows?
When would you use an iterator over a summarized table?
What should you verify before writing classic time-intelligence measures?
How would you calculate sales by ship date when order date is the active relationship?
How would you troubleshoot a slicer that affects the wrong facts?
How would you model a many-to-many business relationship?
What is a disconnected table used for?
What are calculation groups useful for?
What are field parameters useful for?
How do visual calculations differ from model measures?
How would you handle locale-sensitive conversions in Power Query?
How would you handle errors and nulls in M?
How would you diagnose whether a query folds?
Why is Table.Buffer not a universal performance fix?
What are Power Query privacy levels for?
How do referencing and duplicating a Power Query query differ?
How would you use parameters in Power Query safely?
What does incremental refresh do?
How should RangeStart and RangeEnd filters be written?
How would you troubleshoot a refresh that works in Desktop but fails in the service?
How would you handle a source schema change?
How would you design dynamic row-level security?
What happens when a user belongs to multiple RLS roles?
How does object-level security differ from row-level security?
What does Build permission on a semantic model enable?
How would you assess data export permissions?
Why is Publish to web unsuitable for confidential data?
How would you use Performance Analyzer?
Why does high column cardinality matter in Import models?
How would you improve a slow DirectQuery report?
What is a composite model?
What is Dual storage mode useful for?
Does every Direct Lake model fall back to DirectQuery?
How would you use deployment pipelines?
How would you test a semantic model before release?
Power BI Interview Questions for Experienced
Use these scenarios to evaluate architectural trade-offs, governance, failure recovery, and cross-system reconciliation. Expect explicit ownership, evidence, and limits on guarantees.
How would you define a trustworthy enterprise semantic model?
How would you diagnose a report that shows plausible but incorrect totals?
How would you model a slowly changing customer dimension?
How would you handle multiple currencies in a model?
How would you model balances that should not be summed across time?
How would you handle multiple date roles across several fact tables?
When can bidirectional filtering create serious modeling problems?
How would you design aggregation tables without changing results?
How would you choose among Import, DirectQuery, and Direct Lake?
How do Direct Lake on OneLake and Direct Lake on SQL differ operationally?
Why must Direct Lake security be designed for its actual access path?
What does refreshing a Direct Lake semantic model mean?
How would you diagnose a capacity-related performance incident?
How would you investigate a measure that is slow only at the grand total?
How would you optimize DAX without relying on folklore?
How would you design incremental refresh for late-arriving corrections?
How would you coordinate upstream ingestion with semantic-model refresh?
How would you recover after a model refresh fails halfway through?
How would you design gateway resilience?
How would you migrate a model to a different data source?
How would you design a reliable Power Query transformation pipeline?
How would you manage credentials and ownership when a report author leaves?
How would you test RLS across a complex organizational hierarchy?
Why is hiding a field insufficient to secure sensitive data?
How would you review a request to share a report externally?
How would you manage semantic-model changes that affect many reports?
What should a Power BI deployment validation plan include?
How would you use version control with Power BI projects?
How would you govern self-service BI without preventing useful exploration?
How would you evaluate whether a report is actually useful?
How would you diagnose inconsistent results between Desktop and the service?
How would you handle timezone-sensitive reporting?
How would you assess a new or preview Power BI feature?
How would you design an accessible report with dense analytical content?
How would you investigate an unexpected increase in refresh duration?
How would you reconcile Power BI results with a financial source system?
How would you respond to a suspected Power BI data exposure?
How would you design a practical Power BI developer interview exercise?
What evidence would you collect before declaring a Power BI migration successful?
What does a strong end-to-end Power BI operating model look like?
Power BI Interview Questions and Answers
The answers follow the exact wording, numbering, and order of the bank. Power BI and Fabric features depend on the model type, service configuration, capacity, licensing, and current release; confirm the applicable documentation before implementation. The DAX and Power Query M examples state their model or input assumptions. They received static review only: no DAX engine, Power Query runtime, Desktop session, service refresh, or tenant security test was executed. The examples are interview illustrations, not evidence of a deployed modelโs correctness.
Power BI Interview Questions and Answers for Beginners
Establish the core vocabulary of models, data preparation, calculations, reports, and access. Ask candidates to explain what each feature means for a concrete business question.
1. What is Power BI, and what problems does it solve?
Power BI is Microsoft's analytics platform for connecting data, building semantic models, creating reports, and sharing insights. It helps people explore consistent business measures rather than repeatedly calculate them in disconnected files. I would begin with the decisions users need to make and the reliability of the underlying data. A visually polished report can still mislead if its grain, relationships, or metric definitions are wrong. Power BI supports analysis; it does not automatically establish data quality or business meaning.
2. How do Power BI Desktop and the Power BI service differ?
Desktop is a Windows authoring application used for data preparation, modeling, and report creation, while the service provides cloud-based collaboration, distribution, administration, and other authoring capabilities. Their features overlap but are not identical. I would choose a workflow based on supported modeling features, governance, and deployment needs rather than assume every Desktop action exists in the service. Publishing also introduces permissions, credentials, refresh, and capacity considerations that do not arise in the same way in a local file.
3. How does Power BI relate to Microsoft Fabric?
Power BI is a workload within Microsoft's broader Fabric analytics platform, which includes additional data engineering, storage, and other capabilities. A Power BI report and a Fabric lakehouse are different artifacts with different responsibilities. I would identify which services the solution actually uses and verify their capacity, licensing, and security dependencies. Fabric integration can support a shared analytics architecture, but using Power BI does not mean every Fabric feature is available or necessary for the report's requirements.
4. What is a Power BI semantic model?
A semantic model organizes data, relationships, measures, and other metadata into a reusable analytical layer. Older material often calls it a dataset. It should expose meaningful business concepts so multiple reports can use consistent definitions. I would distinguish the model from the visuals built on top of it and document its grain, refresh behavior, and access rules. A model can be technically valid while still containing ambiguous relationships or misleading measures, so usability and correctness need deliberate review.
5. How do reports and dashboards differ?
A report provides interactive pages built around a semantic model, while a dashboard is a service experience that combines tiles for monitoring and navigation. Their filtering, authoring, and interaction capabilities differ. I would choose based on the user's task rather than use the terms interchangeably. A dashboard tile may not behave like the report visual it originated from, and its freshness can have separate considerations. Always verify what the audience actually sees and how they reach the underlying detail.
6. What is a workspace?
A workspace organizes Power BI and, where applicable, other Fabric content for collaboration and administration. Its roles affect what users can create, change, and consume. I would separate development collaboration from broad report distribution and grant the minimum role appropriate to each audience. A workspace is not just a folder: membership can materially affect security behavior. For example, testing row-level security as a workspace editor does not represent the experience of an ordinary Viewer.
7. What is a Power BI app?
An app packages selected workspace content for an intended audience, providing a managed distribution experience. It is distinct from the workspace used by authors to collaborate. I would define audiences, navigation, permissions, and the update process so users receive the right content without unnecessary edit access. An app is not a substitute for semantic-model security or source permissions. Verify the consumer's effective access path and any licensing or capacity requirements rather than assume publication makes everything accessible to everyone.
8. What is a star schema?
A star schema separates fact tables containing events or measurements from dimensions that describe how those facts are filtered and grouped. It often makes Power BI models easier to understand and optimize. I would define the fact grain and dimension keys before creating relationships. For example, sales lines can relate to Product, Customer, and Date dimensions. The pattern is a useful starting point, not a reason to ignore valid exceptions or force unrelated business concepts into one table.
9. What does a fact table's grain mean?
Grain states what one row represents, such as one sales line, one account balance per day, or one support event. It determines valid keys and aggregations. If an order-level shipping amount is repeated on every line, summing it can overstate shipping costs. I would document the grain and test representative cases before building measures. Column names and row counts alone do not establish the business meaning of a table or whether two fact tables can be combined safely.
10. What is a dimension table used for?
A dimension provides descriptive attributes for filtering and grouping facts, such as product category or customer region. Its key should identify the appropriate row for the modeled relationship. I would check uniqueness, missing keys, and any historical versioning rules. A dimension is not simply any table with text columns. Its purpose is to provide consistent analytical context, and incorrect or duplicated attributes can distort results even when the fact data itself is accurate.
11. What does relationship cardinality describe?
Cardinality describes whether each side of a relationship contains unique or repeated key values, such as one-to-many or many-to-many. In a typical star schema, a unique dimension key relates to repeated foreign keys in a fact table. I would validate the data rather than select a cardinality merely to make the relationship dialog accept it. Unexpected duplicates can signal a grain or data-quality problem. Cardinality also interacts with filter direction and storage-mode behavior, so the full model matters.
12. What is cross-filter direction?
It controls how filters propagate through a relationship. A common design lets dimension filters flow to fact tables in one direction, supporting predictable analysis. Bidirectional filtering can be appropriate for specific scenarios but may introduce ambiguous paths, performance costs, or unintended behavior. I would start with the simplest valid design and test the required interactions. Choosing Both everywhere is not a reliable way to fix a visual, and security propagation needs its own explicit review.
13. How do active and inactive relationships differ?
An active relationship normally participates in filter propagation, while an inactive relationship does not do so unless a calculation uses an appropriate mechanism to activate it. This is useful when a fact has multiple date roles, such as order date and ship date. I would decide which relationship should be the default and document measures that use alternatives. An inactive line is not evidence that the relationship is unnecessary; it may support a deliberate analytical role with important restrictions.
14. Why is a dedicated date table useful?
It provides a consistent calendar for grouping and time calculations, including business periods and dates without transactions. I would define its range, uniqueness, completeness, and relevant calendar attributes, then connect it appropriately to facts. A date table helps avoid inconsistent automatic date structures across many fields. Requirements differ between classic and newer calendar-based time-intelligence features, so I would identify the approach explicitly. A date-looking column alone is not enough to guarantee correct year-to-date or prior-period results.
15. What is Power Query?
Power Query connects to data and applies transformations that prepare it for use in the model or another destination. Its M language describes those transformations, including steps generated through the interface. I would use it for tasks such as parsing, selecting columns, joining preparation tables, and assigning types, while considering which work belongs upstream. Power Query and DAX solve different problems. A transformation that succeeds on a small preview can still fail or perform poorly during a full service refresh.
16. How does M differ from DAX?
M is the language used by Power Query to obtain and shape data, while DAX defines analytical calculations over a model and, in supported contexts, visual calculations. M transformations commonly run during data preparation or refresh; measures are evaluated for query context. I would choose based on the required timing and semantics, not syntax preference. Moving a calculation between them can change storage, refresh cost, filter behavior, and results, so apparently similar formulas are not automatically interchangeable.
17. What is query folding?
Query folding is Power Query's attempt to translate supported transformations into operations performed by the source. It can reduce data movement and local processing, but depends on the connector, source, and transformation sequence. Not every step folds, and lack of a visible native-query option is not universal proof that nothing folds. I would inspect the available diagnostics and actual source behavior. Folding improves execution placement; it does not prove that the transformation's business logic is correct.
18. Why should data types be set deliberately?
Types affect parsing, storage, sorting, relationships, and calculation behavior. A customer code with leading zeros should not automatically become a number, and a date interpreted with the wrong locale can remain plausible while being incorrect. I would define types according to the data contract and test boundary values. Type conversion errors should be investigated rather than silently replaced with blanks. A report can look normal while totals or relationships are wrong because a key or amount was interpreted incorrectly.
19. How do merge and append operations differ in Power Query?
Merge combines tables using matching keys and a selected join kind, while append stacks rows from tables into one result. Appending does not match records by business identity, and merging can multiply rows when lookup keys are not unique. I would check column alignment, data types, and intended grain before either operation. Afterward, reconcile row counts and important values. The operation name alone does not establish that two datasets are semantically compatible.
20. What is a measure?
A measure is a calculation evaluated in the context of a query, such as the filters created by a visual or slicer. It is useful for reusable business metrics like revenue or margin percentage. Unlike a stored row-by-row column value, a measure can return different results as the filter context changes. I would define explicit measures with clear names and formats. A correct formula still needs a clear business definition, including which transactions, dates, and units belong in the calculation.
DAX measure example: Sales has a numeric Amount column.
Create this as a measure, not a calculated column. It aggregates Amount under the current filter context. In a visual filtered to one product, it returns that productโs visible sales rather than an unconditional whole-table total.
21. How does a calculated column differ from a measure?
A calculated column produces a value for each row and, in Import models, is stored and updated during model processing. A measure is evaluated for the current query context. Columns can support grouping, relationships, or sorting, while measures typically summarize or derive analytical results. I would consider storage mode and feature restrictions before choosing. Creating a column merely to avoid understanding filter context can increase model size and produce a value that does not respond to report selections as intended.
22. What is row context in DAX?
Row context identifies the current row during an operation such as a calculated column or an iterator. It allows an expression to access values from that row. It does not automatically filter the entire model in the same way as filter context. I would distinguish the current row from the set of rows visible to an aggregation. Confusing these contexts often explains why a formula repeats a grand total or behaves differently when moved from a calculated column to a measure.
23. What is filter context in DAX?
Filter context is the set of filters affecting a calculation, including those from visuals, slicers, relationships, and DAX expressions. The same measure can return a total for one region or the entire visible population depending on that context. I would trace where filters originate and how they propagate before changing the formula. Filter context is not simply the rows shown on the screen; hidden filters, model relationships, and security rules can also influence the result.
24. What does CALCULATE do?
CALCULATE evaluates an expression in a modified filter context. It can add, replace, or otherwise change filters according to its arguments, and it performs context transition when row context is present. I would explain the intended filter change before writing the formula. A filter on an already filtered column normally replaces that column's existing filter unless a modifier such as KEEPFILTERS changes the behavior. CALCULATE is powerful, but it is not a generic fix for every incorrect total.
25. How do SUM and SUMX differ?
SUM aggregates a single numeric column, while SUMX iterates a table expression and evaluates an expression for each row before aggregating. For example, SUMX can calculate quantity times unit price when that line amount is not already stored. I would choose the simplest expression that matches the model and required grain. An iterator is not automatically slow, but unnecessary iteration or a large intermediate table can be costly. Verify the result and performance in the actual context.
DAX measure example: each Sales row is one line with Quantity and UnitPrice.
The expression multiplies values within each row before summing. Two lines with quantities 2 and 3 and prices 10 and 20 contribute 20 and 60, totaling 80. Multiplying the separate column sums would give 150 and answer a different question. Taxes, discounts, and currency conversion are deliberately outside this example.
26. Why is DIVIDE often useful for ratios?
DIVIDE handles division with a zero or blank denominator using defined DAX behavior and an optional constant alternate result. It can make a measure clearer than repeating manual checks. I would decide whether an undefined ratio should be blank or an explicit value based on its meaning, not use zero merely to make a visual look complete. The numerator and denominator must also represent compatible populations. Safe division cannot repair an incorrect relationship or inconsistent filter scope.
DAX measure example: Sales Amount and Total Cost are existing compatible measures.
With sales of 100 and cost of 60, the result is 0.4; format it as a percentage. The omitted alternate result preserves BLANK when the denominator is zero or blank. Returning zero instead would be a deliberate business choice, not merely cosmetic error handling.
27. How do BLANK and zero differ in a report?
BLANK can represent no value or an undefined result, while zero is an actual numeric value. Treating them interchangeably can change visual density and business interpretation. I would ask whether the absence of transactions means zero activity, missing data, or an unavailable calculation. DAX operators and functions can handle blanks in ways that require care, so test the exact expression. Replacing every blank with zero may conceal refresh or relationship problems that users need to see.
28. What is a slicer?
A slicer is a report visual that lets users select values to filter other content according to configured interactions and model relationships. It is a convenience for analysis, not an access-control mechanism. I would make its scope clear and test single, multiple, and empty selections. A slicer can appear to do nothing when relationships or interactions are not configured as expected. Hidden or synchronized slicers can also affect a page, so diagnosis should include the full filter state.
29. How do visual, page, and report filters differ?
They apply at different scopes: one visual, a report page, or the report's pages more broadly. Their effects combine with slicers, interactions, model relationships, and security. I would inspect all relevant scopes when a number seems inconsistent between visuals. Clearing one visible slicer does not necessarily remove every filter. Filters help tailor analysis, but they should not be relied on to protect data from a user who has broader model access.
30. What is the difference between drill-down and drillthrough?
Drill-down navigates levels within a hierarchy in a visual, while drillthrough takes a user to a detail page using configured filter context. They support different analysis paths. I would design them around a clear question and verify which filters are carried into the destination. A detail page can be confusing if its title does not reveal the active context. Neither feature changes the underlying authorization rules or guarantees that the destination contains the intended population without testing.
31. How do tooltips help a report?
Tooltips provide focused context without crowding the main visual, such as a definition, breakdown, or supporting measure. I would keep them concise and ensure important information is not available only through pointer hover. Report-page tooltips have their own filter and design considerations. Test them with representative selections and accessibility paths. A tooltip should clarify the metric or comparison, not hide essential caveats that users need in order to interpret the main chart correctly.
32. What are bookmarks used for?
Bookmarks capture aspects of a report's state, such as page, filters, and display settings, according to their configuration. They can support navigation or guided views, but can also unexpectedly restore old selections if their scope is not understood. I would decide whether data state should be included and test the experience after report changes. A bookmark is not a security boundary. It should make navigation clearer rather than conceal complex state that users cannot explain or reset.
33. How would you choose an appropriate visual?
I would start with the analytical task: comparison, trend, distribution, relationship, or detail lookup. Then choose a visual that represents the data honestly at the right grain. For example, a line chart can show change over ordered time, while a table may be better for exact values. Avoid unnecessary decoration and overloaded categories. I would verify axes, units, sorting, and aggregation because a suitable chart type can still mislead when those choices are wrong.
34. What makes a Power BI report accessible?
It should support understandable labels, sufficient contrast, meaningful alternative text, keyboard navigation, and a logical reading and tab order where the platform allows. Do not rely only on color to convey meaning. I would test representative users' tasks and assistive-technology behavior rather than assume an attractive layout is accessible. Custom visuals may have different capabilities. Important definitions and warnings should remain available through more than a hover interaction, and accessibility should be reviewed after design changes.
35. What is Import mode?
Import mode loads data into the semantic model's storage for analytical queries, usually providing fast interactive performance. The imported data reflects the latest successful processing of the relevant tables or partitions, not every live source change. I would plan refresh, model size, and capacity requirements alongside the report. Import is often a strong default when its freshness and governance constraints fit, but it is not universally appropriate for every volume, source, or security requirement.
36. What is DirectQuery?
DirectQuery keeps the relevant table data at the source and sends queries as users interact, subject to caching and other platform behavior. It can support scenarios where importing is unsuitable, but performance and availability depend heavily on the source, network, model, and report design. I would not describe it as automatically real-time or free of data movement. Feature restrictions and connection behavior vary by source and model, so test the actual workload and identity configuration.
37. What is Direct Lake?
Direct Lake is a Fabric-based semantic-model storage mode that can load data into memory from supported Delta tables in OneLake for analysis. It differs from both a full Import refresh and sending every analytical query to a source through DirectQuery. I would identify whether the model uses Direct Lake on OneLake or on SQL, because their capabilities, security paths, and fallback behavior differ. Capacity and supported-source requirements matter; the name alone does not establish a universal performance or freshness guarantee.
38. What is an on-premises data gateway used for?
A gateway provides a supported connection path between cloud services and data sources that require it, such as certain on-premises sources. It needs appropriate credentials, configuration, availability, and maintenance. I would distinguish connectivity from permission to read the source and from the semantic model's refresh schedule. Installing a gateway does not automatically make every local file or database available. Verify the data-source mapping and the identity used by the actual service operation.
39. What is row-level security?
Row-level security restricts which rows a user can access through a semantic model according to defined roles and filters. It should be designed and tested with the intended consumer identity. In the service, workspace Admin, Member, and Contributor roles are not restricted by model RLS in the same way as Viewers. I would not test only with an author account or use report filters as a substitute. Security also depends on sharing, role membership, and other access paths.
40. How would you approach a first Power BI project?
I would clarify the decisions, audience, data sources, metric definitions, freshness, and access requirements before designing visuals. Build a small reliable model with an explicit grain, validate measures against known examples, and then create focused report pages. Test the service experience, refresh, and permissions with representative users. A useful first release answers a few important questions correctly and has an owner, rather than contain many charts whose definitions or update process are unclear.
Power BI Interview Questions and Answers for Intermediate
Explore filter behavior, transformations, refresh, security, performance, and deployment. Strong answers identify the assumptions and checks that make a proposed solution reliable.
41. What is context transition in DAX?
Context transition converts an existing row context into filter context, most notably when CALCULATE is evaluated in row context. This can make an aggregation respond to the current row's values. A model measure invoked in row context receives an implicit context transition, which is easy to overlook inside iterators. I would trace both contexts explicitly when debugging. Row context alone does not filter an aggregation, and adding CALCULATE without understanding the resulting filters can produce a different error rather than a correct answer.
DAX calculated-column example: Customer has a unique key and an active one-to-many relationship to Sales.
Create this column on Customer in an appropriate Import model. CALCULATE transforms the current customer row context into filter context, which reaches related sales. The stored result is calculated during processing and does not recompute in response to report slicers. For interactive customer totals, the Sales Amount measure is usually the appropriate design.
42. How does KEEPFILTERS change a CALCULATE filter?
KEEPFILTERS makes the added filter intersect with existing filters on the affected columns rather than simply replace them under normal CALCULATE behavior. For example, a measure restricted to one product color can return blank when the current selection excludes that color. I would choose replacement or intersection based on the metric's meaning. The function is not a general instruction to preserve every aspect of report context; examine the particular columns and filters involved and test conflicting selections.
DAX measure example: Product filters Sales through an active relationship.
A Blue-only filter on Product[Color] intersects with Red and leaves no matching product rows. Without KEEPFILTERS, this Boolean filter would replace an existing filter on that column. Other product filters remain relevant; this example does not clear the entire Product table.
43. How would you calculate a percentage of a defined total?
I would define exactly which filters the denominator should retain and which it should remove, then use an appropriate filter modifier such as REMOVEFILTERS. A percentage of all products within the selected date and region is different from a percentage of the entire model. Removing filters from too broad a table can change the denominator unexpectedly. I would test row, subtotal, and grand-total contexts and label the measure so users understand the comparison population.
DAX measure example: Product contains Category, and Sales Amount is the base measure.
The denominator removes Category filters only, including a filter from a Category slicer. Date, region, and other product attributes remain in effect. This is not automatically a percentage of the visually selected categories; define that different requirement explicitly before choosing a different filter pattern.
44. When is SELECTEDVALUE useful?
SELECTEDVALUE returns the single distinct value in a column's current context, or an alternate result when there is not exactly one. It is useful for titles, parameter-like selections, or logic that requires one value. It does not mean the user clicked exactly one item, because other filters can leave one distinct value too. I would handle both no-value and multiple-value contexts deliberately. Assuming it always returns a value can make measures fail or display misleading labels at totals.
DAX measure example: Category is a text column in Product.
The fallback covers both zero and multiple distinct visible categories. A single visible blank category can still return BLANK rather than the fallback, because blank can be the one distinct value. Do not assume this function proves that a user explicitly selected one slicer item.
45. Why use variables in DAX?
Variables can improve readability, avoid repeating an expression, and make intermediate results easier to inspect. They are evaluated in the context where they are defined, so they are not textual substitutions that automatically recalculate under a later CALCULATE filter. I would choose their placement carefully. For example, storing a scalar total before changing filters can preserve the original total when the author expected a new one. Use variables to clarify context, not to hide it behind short names.
46. Why can a grand total differ from the sum of visible rows?
A measure is evaluated again in the total's filter context; the total is not always a sum of the displayed measure values. Ratios, distinct counts, and other nonadditive measures commonly behave this way. I would determine the intended business meaning before forcing a sum with an iterator. For example, a total margin percentage usually divides total margin by total revenue rather than average the row percentages. A different total can be correct, and a visually convenient total can be wrong.
47. When would you use an iterator over a summarized table?
I would use it when the business calculation must be evaluated at a specific grain and then aggregated, such as applying a rule per customer before summing results. The summarized table defines that grain, so its columns and filters need careful review. An iterator over the wrong population can double count or omit entities. I would test small explicit examples and performance at realistic cardinality. Do not add an iterator merely to make a total equal the visible rows without confirming that this is the intended metric.
48. What should you verify before writing classic time-intelligence measures?
I would confirm the calendar's date uniqueness, continuity, range, relationship, and the intended business calendar, including any required date-table marking. Then define whether the comparison concerns calendar periods, fiscal periods, or a custom operational window. Missing dates and inactive relationships can change results. Classic date-column patterns and newer calendar-based features have different requirements and capabilities, so I would state which is used. A familiar year-to-date formula does not establish that the source data or calendar semantics are appropriate.
49. How would you calculate sales by ship date when order date is the active relationship?
I would use a measure that activates the appropriate inactive ship-date relationship with USERELATIONSHIP, assuming the model and security configuration support that pattern. The base measure can remain shared while the date role changes for that calculation. I would test dates where order and shipment occur in different periods. Role-playing date dimensions may be clearer when users need simultaneous independent filtering by both dates. Also review RLS restrictions rather than assuming relationship activation works identically under every security setup.
DAX measure example: Date has active OrderDate and inactive ShipDate relationships to Sales.
The referenced ShipDate relationship must already exist. For this calculation it supplies the alternate date path in the described model. This does not create a relationship or establish a general security bypass; models with RLS have additional restrictions that must be tested. Normalize date keys if source timestamps contain time components.
50. How would you troubleshoot a slicer that affects the wrong facts?
I would inspect the relationship path, cardinality, active status, filter direction, and visual interaction settings, then test with a small known population. Hidden filters or a disconnected table may also explain the behavior. I would not immediately enable bidirectional filtering across the model. The fix should express the intended relationship without creating ambiguous paths elsewhere. Validate related measures and security after changes, because a visual improvement can conceal a broader modeling problem.
51. How would you model a many-to-many business relationship?
I would first distinguish a genuine many-to-many relationship from duplicate keys caused by poor data quality or mixed grain. A bridge table with appropriate unique dimensions can often make the relationship explicit, though native many-to-many cardinality supports some scenarios with its own semantics. I would test filter propagation and totals, especially when facts connect at different grains. The model should explain the business association; selecting many-to-many simply to bypass a uniqueness error is not a sound design.
52. What is a disconnected table used for?
A disconnected table has no relationship path to the relevant facts and can support user selections or scenarios interpreted by measures. For example, a measure can read a selected threshold without automatically filtering the sales table. I would make that behavior explicit and handle multiple or absent selections. Disconnection is not a security mechanism. When using functions such as TREATAS to apply its values as filters, verify the target columns and intended semantics rather than assume the table behaves like a normal dimension.
53. What are calculation groups useful for?
Calculation groups can centralize reusable calculation patterns, such as time comparisons, reducing duplicated measures when the model supports them. Their calculation items modify how selected measures are evaluated, and precedence or formatting interactions can become complex. I would test representative measures, including ratios and measures that should not receive a transformation. They are not a replacement for a clear base model. Feature support and authoring workflows depend on the environment, so confirm applicable capabilities before designing around them.
54. What are field parameters useful for?
Field parameters let report users change the fields or measures used in a visual through a supported selection experience. They can reduce the need for many nearly identical pages. I would test combinations, labels, sorting, and the effect on accessibility and analytical clarity. A parameter that offers incompatible metrics can make a chart difficult to interpret. Field parameters change report behavior; they do not grant access or override model security, and their feature limitations should be checked for the intended scenario.
55. How do visual calculations differ from model measures?
Visual calculations operate over the data structure of a visual in supported scenarios, while model measures belong to the semantic model and are reusable across reports and queries. They can simplify certain visual-specific calculations, but their evaluation context and portability differ. I would choose them only when the intended calculation belongs to that visual and the feature is supported in the target environment. Do not move a governed business metric into a visual merely because its formula looks shorter there.
56. How would you handle locale-sensitive conversions in Power Query?
I would specify the expected culture and type rather than rely on the machine's default interpretation. A text value such as a day/month date can be parsed incorrectly under another locale without producing an obvious error. Test representative and ambiguous values and retain diagnostics for failures. The source's format contract should determine conversion. Formatting a column after loading does not repair an incorrectly parsed value, and replacing conversion errors with null can conceal the underlying mismatch.
Power Query M example: a self-contained conversion using the en-GB culture.
The intended dates are 31 December 2025 and 1 February 2026. Explicit culture avoids interpreting the second value as January 2 in an incompatible locale. This local table does not demonstrate query folding or a service refresh; malformed input still needs an intentional error policy.
57. How would you handle errors and nulls in M?
I would distinguish a missing value from an evaluation error and decide whether to reject, preserve, or explicitly substitute each case. M's try expression can expose whether an operation failed and retain error information for diagnosis. A blanket otherwise null can make bad input indistinguishable from genuinely absent data. I would test malformed values and monitor error counts before choosing a fallback. The correct policy depends on the field's role and whether partial data remains meaningful to consumers.
Power Query M example: preserve the raw value and a structured conversion outcome.
The Conversion record distinguishes a successful value from an error via HasError and the corresponding Value or Error field. The invalid text can be inspected without silently substituting zero. Null is a missing value rather than automatically a parsing error. A production query should route or report failed records according to an explicit data-quality policy.
58. How would you diagnose whether a query folds?
I would use the connector's supported indicators, native-query inspection where available, query plans or diagnostics, and source-side evidence as appropriate. Evaluate the relevant steps and full refresh path rather than infer folding solely from a disabled menu option. Some transformations fold partially or are reordered during evaluation. I would also verify that the generated source operation preserves the intended semantics. Performance measurements and source load matter more than assuming every folded query is automatically efficient.
59. Why is Table.Buffer not a universal performance fix?
Table.Buffer materializes a table in memory with documented shallow buffering behavior and can prevent downstream folding. It may help a specific repeated-evaluation scenario but can also increase memory use and slow a query. I would measure the actual problem and understand which values and operations are being evaluated before adding it. Buffering is not a guarantee of a single source read across every query or refresh. Use a targeted explanation and before-and-after evidence rather than copying it into every slow query.
60. What are Power Query privacy levels for?
Privacy levels help control how data from different sources is combined and potentially passed between them during evaluation. They are not a substitute for source authentication or access control. I would classify sources appropriately and investigate privacy-related behavior rather than disabling protections just to make a refresh succeed. Combining a sensitive source with another system can create a disclosure path. The available settings and execution behavior vary by host, so verify the actual Desktop or service environment.
61. How do referencing and duplicating a Power Query query differ?
A duplicate begins as a copy of the query's steps, while a reference depends on the result of another query. References can improve maintainability, but they do not guarantee that the source is physically read once for all consumers. Evaluation and caching depend on the host and query plan. I would use clear staging and output responsibilities and test refresh behavior. A change to a shared reference can affect multiple outputs, so dependency visibility matters.
62. How would you use parameters in Power Query safely?
I would use parameters for explicit configurable values such as environment-specific endpoints or bounded filters, with types and acceptable values appropriate to the use. They are not a secure place to hide credentials. Dynamic source construction can also affect service refresh support and folding. I would test the published behavior and verify that deployment changes select the intended source. A parameterized development query should not accidentally read production data because a default value was left unchanged.
63. What does incremental refresh do?
Incremental refresh manages partitions so eligible historical data can be retained while a defined recent period is refreshed, reducing repeated processing when configured appropriately. It depends on supported sources, filters, and service capabilities. I would define retention and refresh windows based on late changes and business requirements. It is not automatic CDC and does not guarantee that arbitrary old corrections or deletes will be detected. Verify the first service refresh, partition behavior, and actual source queries rather than rely only on Desktop preview results.
64. How should RangeStart and RangeEnd filters be written?
For an incremental-refresh pattern, I would use the required parameter names and compatible types, then apply an interval that includes one boundary and excludes the other, such as greater than or equal to RangeStart and less than RangeEnd. This avoids double inclusion at adjacent partition boundaries. Verify folding and source timestamp semantics. The parameters' sample values in Desktop are not the complete service policy. I would test records exactly at both boundaries and assess how late corrections outside the refresh window are handled.
Power Query M example: Source is a table with datetime OrderDate; RangeStart and RangeEnd are datetime parameters.
The lower-inclusive, upper-exclusive interval avoids counting a boundary timestamp in two adjacent windows. Keep the parameter names and types required by the incremental-refresh setup. This expression alone does not configure a refresh policy or prove folding; validate the supported source query and policy separately.
65. How would you troubleshoot a refresh that works in Desktop but fails in the service?
I would compare credentials, gateway mapping, source reachability, privacy settings, dynamic-source behavior, runtime differences, and the exact error. Desktop may use the author's local identity or file path, while the service uses configured connections and infrastructure. I would reproduce the failing stage with appropriate authorized diagnostics. Republishing repeatedly without identifying the difference can obscure the cause. Verify the complete refresh result and data freshness after the fix, not merely that the error message disappeared.
66. How would you handle a source schema change?
I would assess whether it changes required columns, types, relationships, or business meaning and identify affected queries, measures, visuals, and refresh behavior. A column rename can break more than the extraction step. Test the change in a controlled environment and preserve a recovery path. Do not silently fill every missing field with null just to keep refresh green. If the contract intentionally changes, update the model and consumers together and verify historical comparability where relevant.
67. How would you design dynamic row-level security?
I would map authenticated users or groups to the data they may access through an explicit security model, using supported identity functions and relationships. Validate the identity format actually returned in the service and avoid ambiguous mappings. Test allowed, denied, unmapped, and multi-assignment cases with representative consumer accounts. A filter expression alone is insufficient if relationship propagation or workspace roles bypass the intended restriction. The design should default appropriately when no authorized mapping exists and have a controlled maintenance process.
68. What happens when a user belongs to multiple RLS roles?
Role filters can combine additively, so membership in multiple roles can broaden the rows visible rather than intersect restrictions as an author might expect. I would test the effective access for real combinations and avoid using an additional restrictive role as a presumed denial override. Group membership and workspace permissions also matter. Document the intended access model and review changes to memberships. A role-by-role test does not prove that the combined user experience is appropriately restricted.
69. How does object-level security differ from row-level security?
OLS restricts access to model objects such as tables or columns, including relevant metadata, while RLS filters rows. Hiding a column in the model interface is not equivalent to securing it. I would choose the appropriate control and test visuals, queries, and consumer access paths that depend on restricted objects. Authoring and deployment support should be verified for the actual tools and environment. OLS can cause dependent visuals or calculations to fail for restricted users, so the report experience needs deliberate design.
70. What does Build permission on a semantic model enable?
Build permission supports creating new content and using supported analytical connections based on the model, rather than merely viewing an existing report. I would grant it according to the intended audience and data-use requirements. It does not automatically mean a Viewer bypasses RLS, but it expands how authorized data can be explored and used. Review workspace roles, sharing, export settings, and other capabilities together. A report's limited visible fields do not necessarily define everything a model consumer can access.
71. How would you assess data export permissions?
I would identify the export path, user permissions, tenant and report settings, and the sensitivity of the data that can be obtained. Exporting summarized data differs from access to underlying detail or an external analytical connection. Test the actual consumer experience and security rules, including downstream copies that no longer behave like an interactive report. A hidden visual or page is not a reliable control. The decision should reflect authorized data use, not only whether an export button is visible.
72. Why is Publish to web unsuitable for confidential data?
Publish to web creates publicly accessible content under the feature's supported behavior, so it is not an authenticated internal sharing mechanism. I would not use it for confidential or restricted information. Internal distribution should use appropriate authenticated access and model security, with tenant controls and licensing considered. A hard-to-guess link does not make public content private. Before choosing any sharing method, verify who can access it and what data the published experience can expose.
73. How would you use Performance Analyzer?
I would capture representative interactions and inspect the time spent in visual queries, rendering, and other reported categories to identify where investigation should focus. It helps compare visuals and obtain their DAX queries, but it is not a complete explanation of every source or capacity bottleneck. Test both typical and expensive filter states. I would make one targeted change at a time and compare results while preserving correctness. A faster visual that silently excludes required data is not an improvement.
74. Why does high column cardinality matter in Import models?
Columns with many distinct values can require more storage and affect compression and query cost, depending on their type and distribution. I would remove unnecessary columns and choose appropriate representations while preserving business requirements. A unique transaction identifier may be essential for detail but unnecessary in a particular analytical model. Measure the impact rather than assume every high-cardinality column is wrong. Splitting or transforming values can have usability and correctness consequences, so optimization should be evidence-based.
75. How would you improve a slow DirectQuery report?
I would inspect generated source queries, indexes or source design, network and gateway behavior, model relationships, measure complexity, and the number of visual queries. Reduce unnecessary work and test representative concurrency at the source. Import aggregations or a different storage design may help when supported, but should be evaluated against freshness and security needs. Increasing report refresh frequency can make the problem worse. The source's response time and workload capacity are part of the report's operating model.
76. What is a composite model?
A composite model combines supported table storage modes or model sources within one semantic model. It can balance performance and freshness, but introduces relationship, query, and security considerations that differ from a simple Import model. I would identify each table's mode and how queries cross source groups. Not every feature combination is supported in every environment. A composite design should solve a concrete requirement, with tests for totals and access, rather than become the default merely because several modes are available.
77. What is Dual storage mode useful for?
A Dual table can participate as cached Import data or through DirectQuery according to the query context and supported model behavior. It is often useful for dimensions that interact with both imported aggregations and DirectQuery facts. I would test the actual query paths and refresh needs. Dual does not mean the table is always simultaneously queried both ways, nor does it eliminate source consistency concerns. Its value depends on the model's relationships and workload, not simply enabling the setting.
78. Does every Direct Lake model fall back to DirectQuery?
No. Current documentation distinguishes Direct Lake on OneLake from Direct Lake on SQL. The SQL-based option can fall back under documented conditions and configuration, while the OneLake option does not use that same SQL-endpoint fallback path. I would identify the exact mode before diagnosing performance or security behavior. Source views, access controls, and capacity guardrails can have different consequences. A general statement that Direct Lake always falls back, or never falls back, is therefore misleading.
79. How would you use deployment pipelines?
I would use supported deployment stages to promote reviewed content through environments with controlled connections, parameters, and validation. Availability and supported artifact behavior depend on the Fabric or Power BI environment and capacity. Deployment is not the same as copying all data, credentials, memberships, and runtime state automatically. I would verify dependencies and environment-specific settings after promotion. A successful deployment status does not prove refresh, security, and business results are correct in the target environment.
80. How would you test a semantic model before release?
I would use small known datasets and expected results to test measures, relationships, date behavior, blanks, totals, and security, then test representative production-scale interactions. Include negative access cases and service refresh rather than only Desktop visuals. Compare business totals with an authoritative source and explain intentional differences. I would preserve the tested model version and deployment evidence. Mechanical validation catches structural issues, while meaningful acceptance tests establish whether users can make the intended decisions from the results.
Power BI Interview Questions and Answers for Experienced
Use these scenarios to evaluate architectural trade-offs, governance, failure recovery, and cross-system reconciliation. Expect explicit ownership, evidence, and limits on guarantees.
81. How would you define a trustworthy enterprise semantic model?
I would begin with agreed business definitions, fact grain, dimensional keys, ownership, and access requirements before writing measures. A shared model should make common analysis consistent while documenting exclusions and legitimate variants, such as booked revenue versus recognized revenue. Validate representative transactions and totals against authoritative records and assign owners to changes. Reuse alone does not establish correctness: a widely adopted model can propagate a subtle error quickly. Release notes, lineage, tests, and a controlled deprecation process make shared definitions maintainable.
82. How would you diagnose a report that shows plausible but incorrect totals?
I would trace the measure from its business definition through filter context, relationships, grain, and source records. Compare a small known population before investigating performance. Duplicated dimension keys, fact-to-fact joins, missing exchange rates, or inappropriate aggregation can all produce believable numbers. Inspect intermediate results and reconcile identities as well as totals, because offsetting errors can cancel out. The fix should address the semantic cause and add a regression case; changing formatting or forcing totals to sum visible rows can conceal the defect.
83. How would you model a slowly changing customer dimension?
I would decide whether analysis needs current attributes or the attributes that applied when a transaction occurred. For historical reporting, a type-two dimension commonly uses a surrogate key for each version and validity boundaries in the upstream warehouse. Facts must resolve to the correct version before the model relationship is applied. Joining sales only to a customer's current natural key can rewrite historical segmentation. I would test boundary dates, overlapping versions, unknown members, and late-arriving changes rather than assuming a relationship reconstructs history automatically.
84. How would you handle multiple currencies in a model?
I would specify source currency, reporting currency, rate type, effective date, rounding, and the intended treatment of missing rates. Conversion generally belongs at the grain at which the correct rate can be determined; multiplying a grand total by one arbitrary rate can be wrong. Preserve original amounts for reconciliation and distinguish operational conversion from legally prescribed accounting rules. Test mixed currencies and changing rates with finance-approved examples. A currency format string changes presentation, not the economic meaning or conversion of a number.
85. How would you model balances that should not be summed across time?
I would identify the measure as semi-additive and define what a period total means, such as the closing balance at the last relevant date. Summing daily inventory snapshots over a month usually answers a different question from month-end inventory. The appropriate date may vary by account or product if observations are sparse, so a global last-date shortcut needs testing. I would document missing-date behavior and compare known balances at multiple levels before choosing a DAX implementation or an upstream snapshot strategy.
86. How would you handle multiple date roles across several fact tables?
I would document each fact's date semantics and decide which comparisons must work together. Separate role-playing date dimensions can support simultaneous independent filtering, while inactive relationships and explicit measures may serve narrower alternate-date calculations. One universal date slicer can become ambiguous when users expect order, shipment, and payment dates to mean different things. I would test cross-fact visuals and security behavior, not merely individual measures. The design should make the selected date role understandable and avoid silently changing meaning between pages.
87. When can bidirectional filtering create serious modeling problems?
Bidirectional propagation can introduce ambiguous paths, unexpected intersections, and expensive queries, especially across several fact tables. It may be justified for a specific relationship pattern, but turning it on everywhere to fix a slicer can create new errors elsewhere. I would diagram filter paths and test representative combinations before choosing a narrowly scoped model or measure solution. Security propagation has additional settings and constraints; ordinary bidirectional filtering should not be assumed to establish the required row-level security boundary automatically.
88. How would you design aggregation tables without changing results?
I would define the detailed fact grain and the coarser aggregation grain, then verify that supported measures can be answered correctly from the aggregated data. Additive sums are simpler than distinct counts or arbitrary ratios, which may require additional structure or detailed data. Configure mappings and storage behavior according to the supported model design and verify which queries actually use the aggregation. Compare detailed and aggregated results across filters, security roles, and totals. Faster queries are useful only if users receive the same intended meaning.
89. How would you choose among Import, DirectQuery, and Direct Lake?
I would compare freshness, source architecture, query latency, concurrency, security requirements, capacity, and operational ownership using representative workloads. Import often provides predictable in-memory analysis but needs a refresh strategy. DirectQuery moves query work to a supported source and inherits its latency and load constraints. Direct Lake is a Fabric-specific option with supported sources and distinct OneLake and SQL-endpoint behaviors. I would benchmark the actual deployment and check current feature restrictions instead of treating any storage mode as universally fastest, cheapest, or most secure.
90. How do Direct Lake on OneLake and Direct Lake on SQL differ operationally?
They use different access and fallback paths. Direct Lake on SQL uses the SQL analytics endpoint for discovery and permission-related behavior and can fall back to DirectQuery in supported situations when allowed by its configuration. Direct Lake on OneLake accesses supported OneLake tables directly and does not provide that SQL DirectQuery fallback. I would verify the exact model type before diagnosing a query or designing security. A limitation that produces fallback in one mode can produce a failure or require redesign in the other.
91. Why must Direct Lake security be designed for its actual access path?
SQL endpoint permissions and row-level policies should not be assumed to govern a separate direct OneLake file-access path. I would identify the model type, effective identity, supported OneLake controls, and semantic-model security, then test the precise route used by consumers. For Direct Lake on SQL, some SQL security conditions can affect fallback behavior; those details do not transfer automatically to Direct Lake on OneLake. Test representative authorized and unauthorized users rather than declaring protection based on one layer's configuration screen.
92. What does refreshing a Direct Lake semantic model mean?
Direct Lake refresh includes framing: updating the model's reference to the relevant table state rather than importing all source rows into a conventional Import copy. Queries still load needed data into memory, so Direct Lake is not a promise of zero memory consumption. I would distinguish source ingestion, table updates, framing, and query-time loading when investigating freshness. Test the actual change and read path, including supported automatic behavior and capacity constraints, instead of assuming that every committed upstream change is instantly visible in every report.
93. How would you diagnose a capacity-related performance incident?
I would correlate the incident window with workload, query durations, refresh activity, concurrency, and available capacity monitoring. Determine whether the issue is one expensive model or widespread contention before changing capacity. Compare representative queries and source timings, and check recent deployments or unusual exports. Capacity changes can mask an inefficient design and have commercial implications, so I would first establish the bottleneck and workload objective. Any remediation needs before-and-after measurements under comparable load, with honest limitations on the observation window.
94. How would you investigate a measure that is slow only at the grand total?
I would inspect how the total's broader filter context changes the work performed by the measure. An iterator or virtual table that is small for one product can become much larger at the total, and branching may select a different expression. Capture the relevant query and compare simplified measures while preserving semantics. I would avoid assuming that fewer characters or a particular DAX function is always faster. Evaluate correctness alongside performance because replacing a calculation with a simpler aggregation can alter the intended result.
95. How would you optimize DAX without relying on folklore?
I would establish a correct baseline and measure the actual visual query on representative model sizes and filter states. Isolate expensive expressions, repeated work, and unnecessary iteration, then test one justified change at a time. Variables can improve readability and sometimes avoid repeated evaluation, but they are not a universal speed guarantee. Model design and storage mode may dominate formula changes. Preserve regression cases and document measured results so an optimization is evidence for this workload rather than a claimed law of DAX.
96. How would you design incremental refresh for late-arriving corrections?
I would compare the correction window with the configured refresh window and the source's change behavior. A correction to an older period can remain invisible if that partition is no longer refreshed. Change detection has documented assumptions and does not magically discover every deletion or arbitrary historical edit. Define an approved historical reprocessing path and test boundary timestamps, updates, and deletes. Reconcile selected old periods periodically when warranted. The policy must reflect business correction behavior rather than being chosen only to shorten refresh duration.
97. How would you coordinate upstream ingestion with semantic-model refresh?
I would define when an upstream dataset is complete and ready, then trigger or schedule refresh against that stable boundary. A refresh that reads halfway through a multi-table load can combine inconsistent states even when it succeeds technically. Use supported orchestration and source consistency mechanisms, with run identifiers and failure handling appropriate to the platform. Monitor freshness and reconciliation after the operation. I would not equate a green refresh history with proof that every source table represents the same intended business cutoff.
98. How would you recover after a model refresh fails halfway through?
I would inspect the refresh mode, affected objects, error details, and documented transaction behavior before assuming the model contains a partial update. Different processing paths can have different commit semantics. Preserve the last known-good state where supported and resolve the root cause, such as schema drift, memory pressure, or credential failure, before retrying. Check timestamps and representative results afterward. A retry plan should account for concurrent jobs and upstream readiness instead of repeatedly launching expensive refreshes against an unchanged failure.
99. How would you design gateway resilience?
I would evaluate supported gateway clustering, network paths, identity ownership, version maintenance, and source dependencies against the required recovery objective. A cluster does not remove a single point of failure in the database, credential store, or network route. Test a controlled failover with representative refresh and DirectQuery workloads, and monitor the actual gateway used. Document ownership and recovery procedures. I would avoid treating a successful connectivity test as evidence that the service can sustain production load or survive every infrastructure failure.
100. How would you migrate a model to a different data source?
I would inventory schema, types, collation, timezone behavior, folding, credentials, privacy settings, relationships, and refresh policies before switching connections. Similar SQL tables can still produce different null, decimal, or text-comparison behavior. Test measures and source reconciliation on a representative snapshot and compare performance under the intended storage mode. Keep a rollback plan and explicitly validate security and service credentials after deployment. A connection string update is not sufficient proof that the model retains its meaning or operational reliability.
101. How would you design a reliable Power Query transformation pipeline?
I would define source contracts, explicit types, predictable error handling, and stable keys while keeping transformations understandable and observable. Prefer supported foldable steps when appropriate, but preserve correct semantics and privacy boundaries. Avoid deriving business rules from incidental row order or silently discarding conversion failures. Test empty input, unexpected columns, locale differences, and duplicate keys. Expensive shared preparation may belong upstream when many models reuse it; a sequence that works for a small Desktop sample is not automatically suitable for a production refresh.
102. How would you manage credentials and ownership when a report author leaves?
I would inventory semantic-model ownership, gateway connections, refresh credentials, workspace access, deployment automation, subscriptions, and upstream permissions. Transfer responsibility through supported procedures and test scheduled operations using the intended durable identities. A report remaining visible does not prove its next refresh will work. Review the departing person's access separately from operational continuity, and avoid exposing secrets in documentation. The handover should identify who responds to failures and who is authorized to change definitions, permissions, and production connections.
103. How would you test RLS across a complex organizational hierarchy?
I would define expected access for representative users, including managers, cross-functional roles, contractors, and users with no mapping. Test the effective service identity and relevant workspace role, not only Desktop's role simulation. Multiple roles are additive, so membership combinations require attention. Check relationships and unmatched identities for accidental broad access, and test export or connected analysis paths under the same permissions. Keep negative cases and expected record identifiers so a plausible total cannot hide an unauthorized subset.
104. Why is hiding a field insufficient to secure sensitive data?
Hiding is a modeling and presentation choice, not a security boundary. A user with appropriate model access may use another analysis path that exposes hidden objects. I would use supported object-level security or an architecture that excludes the sensitive data, together with correct workspace and model permissions. Test the intended consumer role through realistic access methods. Row-level security restricts rows rather than automatically concealing column metadata, and neither control should be assumed to constrain highly privileged workspace editors in the same way as viewers.
105. How would you review a request to share a report externally?
I would identify the recipient, authorized data scope, identity model, tenant settings, licensing or capacity requirements, and supported sharing mechanism. Evaluate the underlying model and access paths as well as the report page. Public publishing is a different disclosure decision from authenticated external sharing and is inappropriate for confidential data. Test a representative recipient's actual access and document ownership. I would not send a screenshot, export, or public link as a workaround for an unresolved permission requirement.
106. How would you manage semantic-model changes that affect many reports?
I would use lineage and dependency information to identify consumers, then classify changes as compatible, behavior-changing, or breaking. Renaming or removing a measure can disrupt thin reports and external analysis even if the main report still opens. Test representative consumers, communicate semantic changes, and provide a migration or deprecation period where appropriate. Version-controlled definitions and release evidence help investigation, but they do not replace data reconciliation. A deployment is complete only when affected audiences can use the intended updated behavior.
107. What should a Power BI deployment validation plan include?
I would validate model structure, measures, security roles, source bindings, refresh settings, gateway configuration, report interactions, and representative outputs in the target stage. Some deployment artifacts and settings have specific support rules, so verify the current pipeline behavior rather than assuming every operational setting moves automatically. Use test identities and known data scenarios before promoting broadly. Record the deployed version and recovery path. A successful pipeline operation confirms that an operation completed, not that production credentials, permissions, and business results are correct.
108. How would you use version control with Power BI projects?
I would use supported project formats and model definition files where appropriate, review changes in the context of the actual model, and keep secrets out of the repository. Text diffs improve collaboration but can be difficult to interpret without understanding dependencies and generated structure. Validate that the team's tools and deployment paths support the selected features and versions. Pair source review with refresh, security, and report tests; a clean merge does not prove that a measure is correct or that target-environment bindings are valid.
109. How would you govern self-service BI without preventing useful exploration?
I would separate exploratory work from trusted shared models and define clear ownership, endorsement criteria, data access, and support expectations. Give teams reusable definitions and training while retaining a route to propose improvements or new metrics. Excessive central bottlenecks can encourage unmanaged exports, while unrestricted duplication fragments business meaning. Measure adoption and correction needs, not merely the number of certified objects. Governance should help users understand which data is appropriate for a decision and whom to contact when results differ.
110. How would you evaluate whether a report is actually useful?
I would connect each page to a decision or task and assess whether users can find the necessary information accurately and efficiently. Usage metrics provide evidence of activity, but frequent visits do not prove business value and low visits may be appropriate for an exception report. Observe representative users, review accessibility, and examine resulting actions or recurring misunderstandings. Remove redundant visuals deliberately and validate changes. A visually impressive dashboard can still fail if its metric definitions, freshness, or next actions are unclear.
111. How would you diagnose inconsistent results between Desktop and the service?
I would compare model versions, refresh timestamps, source connections, credentials, locale or timezone assumptions, security context, and filters. The service may run against a different gateway or identity than the author used locally. Reproduce a small concrete discrepancy and trace it through the same measure and records. Avoid assuming that a publishing action refreshed all data or replaced every dependent artifact. The fix should address the identified difference, followed by verification under an ordinary consumer identity rather than only the owner's account.
112. How would you handle timezone-sensitive reporting?
I would define whether a timestamp represents UTC, source-local time, or a business reporting timezone before deriving dates and periods. Daylight-saving changes and timezone conversion can move transactions across date boundaries. Store sufficient source context and use a deliberate conversion strategy appropriate to the platform rather than relying on whichever machine refreshes the model. Test boundary events and daylight-saving transitions where relevant. Date-only business attributes should not automatically be treated like instants, and a formatted timestamp does not prove its timezone semantics.
113. How would you assess a new or preview Power BI feature?
I would check its current documentation, supported environments, limitations, security implications, and release status, then run a bounded proof of concept against a real requirement. A feature available in Desktop may have additional service, capacity, or deployment restrictions. Plan for changed behavior and a fallback when adopting preview functionality. I would avoid labeling the feature universally production-ready based on a demonstration. The decision should reflect the organization's tolerance for change and its ability to validate and support the specific workload.
114. How would you design an accessible report with dense analytical content?
I would prioritize the essential tasks, use meaningful titles and labels, provide suitable alternative text, and establish a logical keyboard navigation order. Do not communicate critical distinctions through color alone. Check contrast, focus behavior, screen-reader experience, and whether users can obtain the necessary data without interpreting a complex visual. Dense information may need several focused views rather than one overloaded page. Test with representative assistive technologies and users where possible; enabling an accessibility setting alone does not establish that the experience is usable.
115. How would you investigate an unexpected increase in refresh duration?
I would compare data volume, partition coverage, query folding, source execution plans or timings, gateway behavior, capacity contention, and recent transformation changes. An innocuous-looking Power Query step can move work out of the source or expand intermediate data. Determine whether the delay is extraction, transformation, processing, or waiting before optimizing. Test a justified change with the same scope and reconcile results. Increasing timeouts can be appropriate temporarily but does not explain why a previously stable refresh became slower.
116. How would you reconcile Power BI results with a financial source system?
I would align the population, cutoff, currency, recognition rules, adjustments, and aggregation grain before comparing totals. Reconcile stable transaction identifiers and selected detail as well as sums, because omissions and duplicates can offset one another. Document expected differences and obtain the responsible business owner's agreement on definitions. Preserve evidence for the compared versions and timestamps. A report matching one grand total is insufficient assurance that every segment, period, or security-filtered view is correct, especially when financial decisions depend on it.
117. How would you respond to a suspected Power BI data exposure?
I would follow the authorized incident process, restrict the affected sharing path where appropriate, preserve audit evidence, and identify the actual audience, objects, and time window. Check reports, semantic models, exports, public links, and connected analysis paths rather than focusing only on the visible page. Coordinate containment and notification decisions with the responsible teams. Revoking access does not retrieve copies already exported. After correction, test representative denied and allowed cases and document the underlying configuration or governance failure.
118. How would you design a practical Power BI developer interview exercise?
I would provide a small model with a known grain, a deliberately misleading total, and a realistic filtering or access requirement. Ask the candidate to explain the model, write or critique a measure, diagnose the discrepancy, and propose verification cases. Include a source-to-model transformation question so DAX is not treated as the solution to every problem. Assess reasoning and assumptions rather than memorized interface locations. The exercise should use non-sensitive data and clearly state which features or tools are available.
119. What evidence would you collect before declaring a Power BI migration successful?
I would reconcile data scope, critical measures, model relationships, report behavior, access controls, refresh, gateway or source bindings, and downstream consumers against agreed acceptance criteria. Compare representative normal and edge cases under actual user roles, and record known exclusions. Validate operational ownership, monitoring, and recovery as well as visual appearance. Counts of migrated artifacts and a successful deployment status are useful signals but do not prove semantic equivalence. Retain the recovery plan until responsible owners accept the business and operational results.
120. What does a strong end-to-end Power BI operating model look like?
It connects governed sources, clear semantic definitions, appropriate storage choices, secure access, usable reports, dependable refresh, and accountable change management. I would define who owns data quality, model logic, capacity, access, and incident response, with evidence-based tests at the boundaries. Users should understand freshness and limitations rather than infer certainty from a polished visual. Success means people can make the intended decisions from trustworthy information and maintain that trust through source changes, personnel changes, deployments, and failures.
Power BI Documentation and Further Reading
Use these Microsoft references to verify the semantics, supported configurations, and restrictions applicable to the actual model and environment. Storage behavior, security, and deployment should be checked together rather than inferred from a single feature name.
Similar Posts
