Task 4.1: Perform exploratory analytics

• AThe semantic link Spark native connector's _Metrics table with Spark SQL

  Incorrect

  The connector does expose measures through _Metrics, but it's for Spark users; add_measure in SemPy is the pandas route.

• BA Spark job that recreates the measures as DataFrame aggregations

  Incorrect

  Rewriting the logic duplicates the business rules, which then drift whenever finance changes the measures.

• CAn export of a report visual with both measures to a CSV file in the lakehouse

  Incorrect

  A one-time export goes stale and depends on what the visual shows, not on the model's current definitions.

• DSemPy's add_measure on a FabricDataFrame, naming the model's two measures

  Correct

  add_measure augments the data with measures computed by the semantic model, so the model's definitions are reused as they change.

Semantic link connects Fabric notebooks to Power BI semantic models so that data scientists don't have to reimplement business logic embedded in measures. In the SemPy Python library, the FabricDataFrame exposes add_measure, which augments data with measures from a semantic model; this operation isn't supported in the Spark native connector, where measures are reached through the virtual _Metrics table instead.

• AA Python visual in the report that loads the model and scores the visible customers

  Incorrect

  That puts code in the report, scores only what the visual receives, and runs at view time instead of nightly.

• BA T-SQL query with PREDICT in the warehouse, run by a Script activity every night

  Incorrect

  PREDICT isn't supported in Fabric Data Warehouse T-SQL; scoring runs in Spark.

• CA scheduled notebook that scores with PREDICT and writes a Delta table

  Correct

  PREDICT does batch scoring of MLflow models in Fabric, and a Delta table of scores is what a Direct Lake model reads.

• DA calculated column in the Direct Lake model that applies the trained model's coefficients

  Incorrect

  Calculated columns that reference Direct Lake tables aren't supported on Direct Lake on SQL, and copying coefficients duplicates the model.

Fabric operationalizes machine learning models with the scalable PREDICT function, which performs batch scoring from a notebook or from the ML model's item page. Writing the predictions to a Delta table in OneLake lets Power BI read them through Direct Lake alongside the rest of the model. PREDICT isn't part of the T-SQL surface area of Fabric Data Warehouse.

• AA numeric range parameter

  Correct

  A what-if parameter stores the variable, adds a slicer and a value measure, and other measures can use that value.

• BA field parameter that lists the discount rates

  Incorrect

  Field parameters let readers switch which fields or measures a visual uses, not supply a numeric variable.

• CA calculated column on Sales that applies each discount rate

  Incorrect

  Calculated columns are computed at refresh, so readers can't change the rate from the report.

• DA calculation group with one item per discount rate

  Incorrect

  Calculation items hard-code each rate in the model, so any change to the range means editing the model, and the group applies to every measure in the visual instead of supplying one rate to a projection measure.

In Power BI Desktop, a numeric range parameter stores a variable that readers adjust with a slicer. Creating the parameter also creates a measure that returns its current value, and you write other measures, such as projected revenue after discount, that use it. Parameters are designed for measures in visuals and can have at most 1,000 unique values.

• AShe saw the default 5,000-row sample; she should open a custom sample of all rows

  Correct

  The default sample is the first 5,000 rows, and Choose custom sample sets the size and sampling method.

• BData Wrangler hides nulls in Spark DataFrames; she should convert the DataFrame to pandas first

  Incorrect

  Data Wrangler reports missing values for both pandas and Spark DataFrames; the issue is the sample.

• CSummary statistics ignore columns of type string; she should cast Email to another type

  Incorrect

  Statistics depend on the data type, but missing values are reported for string columns too.

• DThe notebook kernel was busy; she should rerun the cell and reopen Data Wrangler

  Incorrect

  Data Wrangler can't open while the kernel is busy at all; it opened, so the summary came from the sample.

When Data Wrangler opens a DataFrame, it uses a sample, by default the first 5,000 rows, and its Summary panel shows dimensions, missing values and per-column statistics for that sample. To profile a different part or more of the data, select Choose custom sample and set the number of rows and the method (first records, last records or random). Data Wrangler works with both pandas and Spark DataFrames.

• Afind_dependencies on the DataFrame, then a filter of its output to the ZIP and CITY columns

  Incorrect

  It discovers which columns depend on which; it doesn't list the violating values.

• Bplot_dependency_metadata on the output of find_dependencies

  Incorrect

  It visualizes groups of dependent columns, not the rows that break a dependency.

• Clist_dependency_violations(determinant_col="ZIP", dependent_col="CITY")

  Correct

  It shows the values that violate a functional dependency between two columns, with their occurrence counts.

• Dadd_measure with a distinct count of CITY

  Incorrect

  It doesn't test dependencies between the DataFrame's columns or list the values that break one.

Semantic link's find_dependencies detects approximate functional dependencies by using a threshold on conditional entropy, and plot_dependency_metadata visualizes them. list_dependency_violations takes a determinant column and a dependent column and shows the values that violate the dependency, with their counts, which helps identify data quality problems such as ZIP codes linked to several cities.

• APython visuals in the service sample 10% of the rows at random

  Incorrect

  No random sampling is applied; the limit is the top 150,000 rows.

• BA Python visual uses at most 150,000 rows, the top rows of the selection

  Correct

  When more than 150,000 rows are selected, only the top 150,000 are used and a message appears on the image.

• CPython visuals ignore rows with blank values in any field

  Incorrect

  Blank handling isn't what thins the plot here; the row limit and grouping are.

• DFields are grouped like a table visual, so duplicate rows appear only once unless an index field makes them unique

  Correct

  Adding a unique index field prevents the grouping so that every row reaches the script.

• EThe service converts the data to an aggregate table before running the script

  Incorrect

  The script receives grouped rows, not an aggregate table built by the service.

The data a Python visual uses for plotting is limited to 150,000 rows; if more are selected, only the top 150,000 are used and a message appears on the image, and the input data also has a 250 MB limit. As in table visuals, fields are grouped and duplicate rows appear once, so add an index field to make all rows unique when you need every row.

• AThe semantic-link package with %pip install in every session

  Incorrect

  An install is only needed to update to the latest version, not to use it.

• BNothing; it's in the default runtime from Runtime 1.2 on

  Correct

  For Fabric Runtime 1.2 and above, semantic link is in the default runtime, with no install needed.

• CPower BI Desktop on the Spark cluster

  Incorrect

  Notebooks reach semantic models through semantic link, not through Desktop.

• DAn on-premises data gateway for the semantic model

  Incorrect

  Semantic link connects to semantic models from Fabric notebooks without a gateway.

Semantic link connects Power BI semantic models with Data Science in Microsoft Fabric. For Fabric Runtime 1.2 (Spark 3.4) and above, semantic link is available in the default runtime and doesn't need to be installed; %pip install -U semantic-link updates it to the latest version.

• AData Wrangler can't open Spark DataFrames

  Incorrect

  Data Wrangler opens both pandas and Spark DataFrames.

• BCustom code works only on a custom sample, not on the default view

  Incorrect

  How the DataFrame is sampled doesn't decide which operations are offered; its type does.

• CCustom code needs the notebook to be attached to an environment

  Incorrect

  Environments aren't the reason; the operation type depends on the DataFrame kind.

• DCustom code operations currently support only pandas DataFrames

  Correct

  Data Wrangler's custom code operations are limited to pandas DataFrames.

Data Wrangler opens pandas or Spark DataFrames from a Fabric notebook, shows summary statistics and offers data-cleaning operations that generate code automatically. One of its current limitations is that custom code operations support only pandas DataFrames.

• A5,000 by default, and at most 100,000

  Incorrect

  The documented default for display is 1,000 rows, and the widget's maximum is 10,000.

• B10,000 by default, and at most 1 million

  Incorrect

  10,000 is the maximum, not the default.

• C1,000 by default, and at most 10,000

  Correct

  display renders 1,000 rows by default, and the output widget can view and profile up to 10,000 rows.

• DAll 3 million rows

  Incorrect

  The display widget doesn't render whole large DataFrames.

The display() function renders Spark DataFrames, pandas DataFrames and SQL results as interactive tables or charts. You can specify the row count to render; the default is 1,000, and the output widget can view and profile at most 10,000 rows. The Inspect pane shows each column's distribution and statistics for the rendered data.

• ALower the conditional entropy threshold

  Correct

  A lower threshold makes find_dependencies more selective, so only stronger dependencies are detected.

• BRaise the conditional entropy threshold

  Incorrect

  A higher threshold lets weaker dependencies through.

• CSet dropna=True

  Incorrect

  dropna removes rows with NaN values from evaluation; it doesn't make the function stricter.

• DCall plot_dependency_metadata first

  Incorrect

  That function visualizes dependencies already found; it doesn't change detection.

Semantic link's find_dependencies uses a threshold on conditional entropy to discover approximate functional dependencies, where low conditional entropy indicates strong dependence. To make it more selective, set a lower threshold so that only stronger dependencies are detected. With dropna=True, rows with NaN values in either column are left out of the evaluation.