Databricks Certified Data Analyst Associate Study Guide¶

Independent AI-assisted resource — SOURCES + OBJECTIVES CHECKED; HUMAN REVIEW PENDING. Objective coverage, citations, volatility labels, links, and exam-integrity compliance were reviewed on October 28, 2026 against all 39 detailed PDF objectives and the nine live weighted domains. This is not a guarantee that the guide is error-free or current after that date. See the sources-and-objectives record. The official certification page and its linked exam guide are authoritative.

Study guide at a glance

Start here

Report an issue with DATABRICKS-DATA-ANALYST-ASSOCIATE

Blueprint checked2026-10-28

Review stateSources + objectives checked — human review pending

Guide-only active reading45 min–1 hr 15 min

Prerequisites: No formal prerequisite; the official exam guide recommends related training and six months of hands-on Databricks experience. Practice ANSI SQL, relational and dimensional modeling, data quality, SQL warehouses, query diagnosis, AI/BI dashboards, Genie and Unity Catalog governance.

Exam essentialsRead the objective map, key distinctions, and readiness checklist.

Deep understandingWork through every domain, including decisions, failure modes, and related items.

Hands-on labsBegin with Hands-on lab sequence, then use the remaining labs to produce evidence you can inspect and explain.

Official objective domains

See the objective map in this guide

The active-reading range uses 90–130 words per minute for technical material. Labs, troubleshooting, note-taking, spaced review, and prerequisite gaps add time; it is not a total preparation promise.

Library identifier: DATABRICKS-DATA-ANALYST-ASSOCIATE; Databricks does not publish a short exam code on the official page checked.
Current baseline: Detailed official exam guide current as of October 30, 2025, with 39 objectives; live nine-domain weighted coverage page and its linked PDF checked September 28, 2026.
Upcoming blueprint change: None announced as of September 28, 2026. The detailed PDF says its self-paced Data Analysis with Databricks course was being replaced by AI/BI for Data Analysts and SQL Analytics on Databricks; that is a learning-catalog transition, not evidence of a scheduled exam change.
Lifecycle status: Active; valid for two years, with the currently live exam required for recertification.
Assessment: 45 scored multiple-choice questions, 90 minutes, USD 200, no test aids, English, online or test-center delivery.
Prerequisite: None required. The official guide recommends related training and six months of hands-on Databricks experience.
Code convention: Exam SQL is ANSI SQL. Practice writing and reading SQL without depending on a vendor-specific shortcut when a standard construct is available.

Product terminology — VERIFY CURRENT: The exam still names AI/BI Genie spaces. Current setup documentation calls them Genie Agents and explicitly identifies the former name. Keep the published exam wording when mapping objectives, and use the current name when finding product instructions. This product change does not establish a new exam blueprint. The deep-review record preserves the distinction.

Living-guide watch — October 28, 2026¶

Current platform context: The published exam guide remains the October 30, 2025 version, with AI/BI Genie in scope. October 2026 Databricks release notes extend AI Search sources to streaming tables, materialized views and managed Iceberg V3 tables, while platform releases are staged and may reach accounts later. Use the live guide for exam scope and current documentation for feature availability. See the current exam page and October release notes.

How to use this guide¶

Treat the percentages as a study-budget signal, not a prediction of a particular form. Query execution, query analysis, dashboards, and Genie total 63%; nevertheless, those capabilities depend on correct data import, modeling, catalog discovery, and security.

business question -> governed source -> grain and model -> ANSI SQL result
-> profile and validate -> visualization or Genie answer -> permission boundary
-> refresh or alert -> usage and feedback evidence -> measured improvement

For each exercise, record the question, intended grain, source objects, warehouse, SQL text, query ID, result checks, visualization choice, refresh state, principal and grants, and one failure/recovery observation. Product names, licensing, sharing modes, serverless behavior, AI/BI features, and UI locations change quickly; recheck the linked documentation in the workspace where you practice.

About related items: A Related item: callout adds prerequisite, architectural, migration, security, operational, or adjacent context that makes an objective easier to understand. It is useful supporting knowledge, not a claim that the item appears verbatim in Databricks' published exam objectives.

Objective map¶

Published domain	Weight	What you must be able to decide or do
Understanding of Databricks Data + AI Platform	11%	Relate platform components, Catalog Explorer, Marketplace, lineage and certification to an analytics task.
Managing Data	8%	Discover, query, tag, trace and clean governed tables.
Importing Data	5%	Select a supported ingestion or sharing route and use the upload UI for small files.
Executing queries using Databricks SQL and Databricks SQL Warehouses	20%	Choose a warehouse and express joins, sets, aggregates, table creation, history and cross-system analysis in ANSI SQL.
Analyzing Queries	15%	Diagnose correctness and performance using query evidence, Photon, caching, history and clustering.
Creating Dashboards and Visualizations in Databricks	16%	Build, parameterize, share, schedule, embed and alert on an AI/BI dashboard.
Developing, Sharing, and Maintaining AI/BI Genie spaces	12%	Ground a Genie space, govern access, validate answers and improve it from feedback.
Data Modeling with Databricks SQL	5%	Match star, snowflake and Data Vault patterns to grain, change and medallion-layer needs.
Securing Data	8%	Apply the three-level namespace, ownership and least-privilege Unity Catalog access.

1. Understanding the Databricks Data + AI Platform (11%)¶

Map components to responsibilities¶

Delta Lake supplies table transactions, schema controls and history over object storage. It is a storage/table contract, not the governance catalog or query engine.
Unity Catalog supplies the catalog.schema.object namespace, discovery, permissions, lineage, auditing and governance across supported data and AI objects.
Databricks SQL supplies SQL warehouses, the SQL editor, query history, dashboards, alerts and SQL-facing analytics experiences.
Mosaic AI is the broader AI tooling family. The exam's analyst-specific generative experience is AI/BI Genie, not a request to master the full model-development stack.
Lakeflow Jobs orchestrates tasks; Lakeflow Spark Declarative Pipelines manages declarative batch/stream pipelines. The 2025 PDF uses the older wording “Delta Live Tables.” Translate old course material to the current product name.
The Data Intelligence Engine uses platform metadata and context to improve experiences such as natural-language analytics. Do not confuse an AI-generated answer with validated business semantics.

Use the platform introduction to trace an analyst request through identity, workspace UI, compute, governance, table storage and an output. A sound answer says which component owns each decision.

Operate Catalog Explorer as a discovery and control surface¶

Catalog Explorer lets you browse catalogs and schemas; inspect tables, views, volumes, models and functions; review owners, tags, permissions, history and lineage where available; and create or manage supported objects. Know the distinction:

Concept	Meaning	Decision consequence
Managed table	Unity Catalog manages governance and the underlying data lifecycle	Dropping the table can delete managed data after the retention behavior applies.
External table	Unity Catalog governs access while files remain at an external location	Dropping the table removes metadata, not the underlying files.
View	Saved query that resolves its source on use	Useful for abstraction/security, but not precomputed merely because it is saved.
Certified asset	A trusted-data marker applied through Unity Catalog	Helps discovery; it does not grant access or prove every query is correct.
Tag	Searchable metadata on a securable object	Useful for classification and discovery; privileges still govern access.
Lineage	Recorded upstream/downstream relationships for supported workloads	Supports impact analysis and troubleshooting; absence is not proof of no dependency.

The Unity Catalog overview and data-lineage documentation are the current baselines.

Understand Marketplace without treating it as anonymous download¶

Databricks Marketplace is a governed discovery and exchange surface for data products and other assets. A consumer searches listings, evaluates provider and terms, obtains instant-access data or requests access, then queries the shared asset in the recipient environment. Marketplace commonly uses Delta Sharing, so the provider can share without copying a static export into the consumer's workspace.

Related item: Marketplace is a catalog and commercial/distribution experience; Delta Sharing is the open sharing protocol and governed delivery mechanism. A listing can make a product discoverable, while the share, recipient and permissions determine access.

2. Managing Data (8%)¶

Discover and query trusted datasets¶

Begin with the business definition and required grain, then inspect certification, owner, description, tags, schema, freshness and lineage before querying. A certified table is a useful trust signal, not permission to ignore the source definition. Fully qualify objects when ambiguity matters:

SELECT order_date, region, SUM(net_amount) AS revenue
FROM analytics.gold.fact_order
WHERE order_date >= DATE '2026-01-01'
GROUP BY order_date, region;

Record whether the query reads a managed table, external table, view, materialized view or federated source. That affects freshness, performance, lifecycle and the evidence needed for a result.

Use tags and lineage as investigation tools¶

Search tags to locate domain, sensitivity, quality or ownership metadata, then confirm the tag definition and who may apply it.
Use upstream lineage to identify source tables and transformations behind a questionable metric.
Use downstream lineage before changing a column, table or view.
Compare lineage with query history and job/pipeline evidence. Each answers a different question: dependency, execution and production state.

Clean data without hiding quality defects¶

Profile missing, invalid, duplicate and out-of-range values before changing them. Define a business rule and retain rejected counts. Common SQL tools include CASE, COALESCE, NULLIF, TRY_CAST, TRIM, regular-expression functions, window functions and deduplication with ROW_NUMBER().

WITH ranked AS (
  SELECT *, ROW_NUMBER() OVER (
    PARTITION BY customer_id
    ORDER BY updated_at DESC, ingest_id DESC
  ) AS rn
  FROM analytics.silver.customer_candidate
)
SELECT customer_id,
       NULLIF(TRIM(email), '') AS email,
       NULLIF(TRIM(postal_code), '') AS postal_code
FROM ranked
WHERE rn = 1;

Postal codes are identifiers: preserve leading zeros and alphabetic characters rather than casting them to integers. This query also assumes updated_at and the tie-breaker ingest_id define a unique ordering within each customer. Quarantine missing business keys before deduplication; otherwise unrelated null keys can collapse into one partition.

For a genuinely numeric field, TRY_CAST can turn malformed or overflowing supported conversions into null. It still rejects unsupported source/target combinations and has documented nested-type restrictions. Count failed conversions separately from source nulls, retain original values for diagnosis, and apply a business rule before substituting zero.

Related item: Medallion names are quality contracts, not mandatory database names. Bronze normally preserves reproducible source evidence, silver applies validated conformance, and gold serves a consumer-shaped model. A cleanup query belongs where its ownership, replay and quality evidence are clear.

3. Importing Data (5%)¶

Choose an intake route¶

Need	Suitable starting mechanism	Key check
Tiny local CSV/JSON/Parquet for exploration	Workspace upload/create-table UI	size limits, inferred types, reproducibility and sensitive-data policy
Incremental cloud files	Auto Loader or `COPY INTO`, depending scale and flow	discovery/checkpoint state, schema behavior and replay
Supported database or SaaS source	Lakeflow Connect or governed connector	release state, CDC semantics, identity and destination
Query without copying	Lakehouse Federation	latency, pushdown, source load, credentials and feature support
Provider-to-recipient governed exchange	Delta Sharing or Marketplace	provider, recipient, share, refresh and permissions
Custom REST/API intake	controlled code/job that lands raw evidence	pagination, rate limits, secrets, watermark and idempotency

The data-sharing documentation explains provider and recipient concepts. Upload is valuable for learning and one-off exploration, but it does not create a production ingestion design: capture the original file, type decisions, table ownership and how a repeatable pipeline would replace the manual action.

Related item: Importing and federation solve different problems. Importing establishes a Databricks-side copy and freshness process; federation leaves data in the remote system and introduces remote availability, pushdown and source-capacity dependencies.

4. Executing queries with Databricks SQL and SQL warehouses (20%)¶

Select and reason about a SQL warehouse¶

A SQL warehouse is the compute resource for Databricks SQL workloads. Choose size, scaling, channel, serverless/classic option, auto-stop and permissions from concurrency, latency, feature, isolation, network and cost requirements. Do not assume a bigger warehouse corrects a poor join or scans less data. Review current SQL warehouse documentation.

Separate these permissions:

Ability to use the warehouse.
Ability to use the catalog and schema.
Ability to select source objects.
Ability to create or modify target objects.
Ability to view/share the query, dashboard or Genie space.

Join at a declared grain¶

Before a join, state the grain and cardinality of both sides. An inner join removes unmatched rows; a left join retains every left row; right and full joins preserve their respective unmatched sides. Multiple conditions normally belong in ON, while a post-join WHERE filter can unintentionally remove null-extended rows.

SELECT d.calendar_month,
       p.category,
       SUM(f.quantity * f.net_unit_price) AS revenue
FROM analytics.gold.fact_order_line AS f
JOIN analytics.gold.dim_date AS d
  ON f.order_date_key = d.date_key
LEFT JOIN analytics.gold.dim_product AS p
  ON f.product_key = p.product_key
GROUP BY d.calendar_month, p.category
ORDER BY d.calendar_month, revenue DESC;

Validate row count, distinct business keys, unmatched keys and aggregate totals before and after. UNION removes duplicate result rows; UNION ALL preserves them and usually avoids the deduplication work. Pick from semantics, not speed alone.

Aggregate and summarize correctly¶

COUNT(*) counts rows; COUNT(column) excludes nulls.
COUNT(DISTINCT x) is exact; approx_count_distinct(x) trades exactness for scale. State that tradeoff.
AVG ignores null values. Decide whether null means missing, zero or not applicable before using it.
Use GROUP BY at the requested grain. Window functions calculate over a partition without collapsing all detail rows.
Apply stable sorting when output order matters. A LIMIT without deterministic order is not a top-N definition.

NULL and anti-join failures¶

SQL uses three-valued logic. A comparison with null is unknown, and a WHERE clause retains only true conditions. Use IS NULL for missing values. The NULL semantics reference explains why NOT IN is dangerous when its comparison set contains null: a nonmatching value can still evaluate to unknown and disappear. Use a correlated NOT EXISTS when the requirement is absence of a matching row, and decide how null business keys should be handled.

With amounts 40, 60 and null, COUNT(*) is 3, COUNT(amount) is 2, SUM(amount) is 100 and AVG(amount) is 50. Replacing null with zero changes the average to about 33.33. Neither interpretation is automatically correct; the business definition must choose. Likewise, SUM(DISTINCT amount) is not a repair for join duplication: two legitimate orders worth 40 each must still total 80.

Create and query durable objects¶

Know how to create a managed or external table, a view, and a table from a query. Choose CSV, Parquet or Delta from the contract: Delta supplies transactional table behavior; Parquet is a file format; CSV requires explicit delimiter/header/type discipline. Use three-part names and explicit target locations only where the external-lifecycle requirement justifies them.

Lakehouse Federation uses connections and foreign catalogs to query external systems. Cross-system analysis may join a Delta table to a federated table, but the plan can move or push work across a network boundary. Inspect filtering, join location, supported types/functions and source load.

Distinguish views, materialized views and streaming tables¶

Object	Core behavior	Strong fit
Standard view	Stores a query definition; computes on use	semantic abstraction, reusable logic and some security patterns
Materialized view	Stores precomputed results and refreshes them	repeated analytical computation where controlled freshness is acceptable
Streaming table	Incrementally processes an append/streaming source	continuously arriving data and incremental pipeline semantics

Read the current materialized-view and streaming-table guidance. “Dynamic view” in older learning material often refers to a standard view whose results respond to user/context logic; do not confuse that pattern with persisted materialization.

Use history as evidence¶

Delta time travel lets you query a version or timestamp within retained history, while DESCRIBE HISTORY shows table operations. The table-history documentation explains retention interactions. Use history to reproduce a prior result or investigate a change; do not promise indefinite rollback after retention cleanup.

Related item: The SQL editor and Assistant can draft, explain and debug SQL, but the analyst owns the grain, authorization and validation. Preserve the generated query and test its row-level and aggregate behavior before publishing it.

5. Analyzing Queries (15%)¶

Diagnose before tuning¶

Use a repeatable loop:

Confirm the result is semantically correct and identify the query ID, warehouse and data version.
Inspect duration, queue time, compilation, execution and bytes/rows read in query history.
Use the query profile to find expensive operators, scans, shuffles, joins and skew.
Form one hypothesis: filter earlier, project fewer columns, correct a join, improve data layout, precompute, or right-size compute.
Change one factor and compare the same workload and result checks.

Query Insights may surface likely causes and recommendations; the profile supplies evidence. Query history supplies workload context. Neither replaces the business-correctness check.

Place Photon and caching correctly¶

Photon is Databricks' vectorized engine for supported SQL and DataFrame workloads. It can accelerate supported operators, joins, aggregations and Delta/Parquet processing. Verify compute and operation support; unsupported parts can fall back. Photon does not repair Cartesian joins, incorrect filters, bad grain or an overloaded remote federated source.

The query-caching reference distinguishes returned-result caches from local copies of source files. A local result cache ends with its cluster, while serverless remote result caching can survive a warehouse restart. Therefore, restarting a warehouse alone does not establish a cold benchmark. SET use_cached_result = false disables result reuse for a benchmark session; it does not establish that every disk or application cache is cold. Record cache settings, data version and repeated results, then restore normal settings. AI/BI dashboards have their own caching behavior; the legacy SQL UI cache is not their contract.

Related item — VERIFY CURRENT: SQL release notes distinguish the Current and Preview warehouse channels from ongoing web UI updates. Rollouts are staged, so a published release is not proof of availability in your workspace. Record the warehouse channel/version when reproducing a result.

Use history, clustering and audit evidence deliberately¶

The query-history UI supports filtering and inspection of executed queries. Table history answers what changed in a Delta table; query history answers what SQL ran; lineage answers dependencies. Combine them to validate or compare historical results.

Liquid clustering reorganizes data by selected clustering keys and can improve skipping as access patterns evolve. Choose keys from repeated selective predicates and join patterns, then measure. Review liquid clustering because runtime requirements and automatic clustering behavior are volatile.

Related item: Query performance is a system property: model grain, file/layout state, statistics, warehouse capacity/concurrency, cache state, remote-source behavior and SQL shape interact. A single fast run is weak evidence; record a representative comparison.

6. Creating dashboards and visualizations (16%)¶

Build from a question, not a chart palette¶

An AI/BI dashboard can use multiple datasets and pages with visualization, text and image widgets. Start with audience and decision, then select the visual:

Analytical need	Useful starting visual	Common failure
Trend over ordered time	Line	treating irregular categories as continuous time
Compare categories	Bar	too many categories or truncated axes
Single status against target	KPI/counter	hiding denominator, time window or uncertainty
Relationship/distribution	Scatter, histogram or box plot where supported	implying causation from correlation
Composition	Stacked bar/area, sparingly	unreadable categories or comparing angles
Detailed lookup	Table	substituting raw detail for prioritization

Notebook and SQL-editor visualizations are useful during exploration; dashboards add governed publication, filters, schedules and sharing. Validate totals and edge cases in SQL before polishing the display.

Parameterize safely¶

Dashboard filters constrain datasets or fields. Parameters can substitute controlled values into a dataset query. Use the dashboard-parameter guidance to distinguish field filters from query parameters, define defaults and test all permitted values. Do not concatenate untrusted input into SQL.

Publishing creates a consumable snapshot of dashboard configuration; sharing determines who may view or edit. Workspace users/groups, account users and external sharing modes have different identity and entitlement requirements. The current dashboard-sharing documentation should control the design.

Grant the minimum dashboard and underlying-data access required by the chosen credential model.
Use a shareable link only when its workspace/account and external-sharing behavior matches policy.
Treat embedding in an external application as an authentication, authorization, content-security and lifecycle integration—not merely an iframe.
A scheduled refresh changes data freshness; it does not email every consumer by itself.
An alert evaluates a query/threshold on a schedule and sends configured notifications. Define threshold, evaluation frequency, destination, empty/error behavior and owner.

Prove the publication identity¶

The sharing reference distinguishes two published data-permission modes:

Mode	Data permission used	Validation case
Shared data permissions	Publisher's permissions	A viewer without direct table access may still see the published data; approve the exposed result deliberately
Individual data permissions	Each viewer's permissions	Test viewers with different row/data access and confirm the intended differences

Compute access uses the publisher's credentials in both modes. Draft dashboards use the viewer's data permissions, so successful draft testing does not prove the published mode is correct. Publishing snapshots configuration; changing the draft does not update that snapshot until republished. Data refresh is a separate operation.

Account sharing requires registered account identities; a link does not by itself create anonymous access. Account-only users also have restrictions for workspace-bound securables. Check folder inheritance as well as direct dashboard grants.

VERIFY CURRENT: AI/BI dashboard publishing, external sharing, embedding, subscriptions and alert capabilities can change independently. Recheck exact entitlement and credential behavior before production use.

Related item: Dashboard access and Unity Catalog data access are separate layers. Document which identity executes the dataset and whether consumers need direct source privileges; otherwise a dashboard may be overexposed or fail after publication.

Ground a space before asking it questions¶

An AI/BI Genie space provides natural-language exploration over selected Unity Catalog data with a SQL warehouse. A useful space has a narrow domain, curated tables/views, clear column descriptions, representative sample questions, domain instructions and validated SQL examples or Trusted Assets.

domain boundary + governed data + business definitions + instructions
+ trusted examples + usable warehouse + least-privilege sharing
-> generated SQL -> answer validation -> feedback -> maintained space

Pick sources at a consistent grain. Hide ambiguous or unsafe columns behind governed views. Explain synonyms, date semantics, fiscal rules, exclusions and approved metrics. A broad catalog dump makes the model guess relationships and business meaning.

Validate and improve answers¶

For representative questions:

State the expected grain and calculation.
Inspect the generated SQL, tables, joins, filters and warehouse execution.
Compare with an independently validated query and known edge cases.
Capture whether the miss came from metadata, instructions, data, a relationship, ambiguous language or SQL generation.
Improve the smallest appropriate artifact: descriptions, instructions, sample question, benchmark, curated source or Trusted Asset.

User thumbs-up/down feedback is a signal, not ground truth; it does not automatically train or update the agent. The current monitoring reference distinguishes benchmark Chat mode, which compares query results against a supplied SQL answer, from Agent mode, which uses an LLM judge and optional evaluation notes. Validate any generated answer key independently.

Benchmark questions start new conversations, so a passing single-question test does not validate a multi-turn conversation. Test follow-up context separately. Keep business definitions, data snapshot, evaluation mode and expected results with the test. Benchmarks supply a repeatable question/expected-answer evaluation set. Refresh Unity Catalog metadata when schemas or descriptions change, and retire instructions or trusted examples that no longer match the source.

Grant space access to intended users/groups and verify the warehouse and data-execution model. Current Genie Agent setup documentation specifies pro or serverless SQL warehouses, with the configuring author's embedded credentials granting compute access. Data authorization uses each end user's own Unity Catalog identity. A consumer does not need direct warehouse permission merely because they use the agent, but still needs the required data access and agent permission. This differs from a dashboard published with shared data permissions. Test both allowed and denied consumers.

External-app embedding needs an explicit supported integration and identity design. Do not paste secrets, personal data or unrestricted source tables into instructions to work around governance.

VERIFY CURRENT: Genie Trusted Assets, benchmarks, monitoring, embedding and sharing evolve rapidly. Recheck the product documentation and workspace UI near study and deployment time.

Related item: Genie is a semantic interface over data and SQL, not a replacement for a governed semantic model. Better table grain, definitions, relationships and security improve both human-authored dashboards and AI-generated answers.

8. Data Modeling with Databricks SQL (5%)¶

Select a model from the workload¶

Pattern	Core idea	Use when	Cost to remember
Star schema	Fact at a declared grain linked to denormalized dimensions	BI measures, understandable joins and broad query performance	conformed keys, slowly changing dimensions and fact grain must be managed
Snowflake	Dimensions normalized into related tables	shared hierarchical structures or stricter normalization matters	more joins and a less direct analyst experience
Data Vault	Hubs, links and satellites separate business keys, relationships and descriptive history	auditable integration of changing multi-source data	requires a downstream information-mart/star layer for convenient BI

Medallion describes progressive data quality; star, snowflake and Data Vault describe logical modeling patterns. They can coexist. For example, bronze retains source evidence, silver integrates a Raw/Business Vault, and gold publishes star-schema marts.

Declare fact grain before keys and measures. Distinguish additive, semi-additive and non-additive measures. Use surrogate keys where dimension history requires versioned members. Ensure an unknown member or explicit unmatched-key process prevents silent fact loss.

Related item: A materialized view can accelerate a consumer-shaped result, but it is not itself a complete dimensional model. Modeling defines meaning and relationships; materialization defines computation and refresh behavior.

9. Securing Data (8%)¶

Traverse the three-level namespace¶

Unity Catalog uses catalog.schema.object. To query a table, a principal normally needs USE CATALOG, USE SCHEMA and the object privilege such as SELECT, subject to ownership, inherited privileges and other controls. Apply grants to groups or workload identities rather than individuals where practical. Review the current privileges reference.

Ownership permits management beyond ordinary data access, so separate owner/steward roles from consumers. Catalog and schema organization should reflect domain, lifecycle and isolation requirements; workspace folders are not the data-security hierarchy.

Secure lifecycle and sensitive data¶

Prefer managed tables when Databricks should manage the data lifecycle; use external tables when an external lifecycle is a requirement, not as an automatic default.
Store files in governed volumes where appropriate rather than unmanaged workspace paths.
Classify and tag sensitive columns, but enforce protection with privileges and supported fine-grained controls.
Grant analysts the curated views or tables needed for the question, not blanket catalog access.
Verify who can use the SQL warehouse and who can share the dashboard or Genie space.
Record owner, business purpose, retention, recipients and review date for shared assets.

Related item: Object privileges, row filters, column masks, ABAC policies, workspace bindings and dashboard/Genie permissions operate at different layers. Begin with least privilege and a simple governed view; add dynamic controls only when the requirement and supported operations justify them.

Integrated decision scenarios¶

Scenario A — governed executive revenue dashboard¶

A global retailer needs daily revenue, margin and return trends by region. Define the fact grain, validate currency and late-return rules, identify certified gold sources in Catalog Explorer, trace their lineage, and query with ANSI SQL. Check unmatched dimensions and reconcile totals. Publish an AI/BI dashboard with a region filter and controlled date parameter, schedule refresh, share to an executive group and configure an alert with an owner. Prove source, warehouse, dashboard and catalog permissions separately.

Scenario B — supplier data without uncontrolled copying¶

A supplier offers governed data through Marketplace/Delta Sharing, while an operational attribute remains in a supported federated database. Compare freshness, latency and access boundaries. Use fully qualified names, filter early, inspect the cross-system plan and materialize only when repeated computation and freshness justify it. Record provider/recipient ownership, query history, data version and recovery when the remote source is unavailable.

Scenario C — trusted sales Genie space¶

Business users want natural-language pipeline analysis. Curate a narrow star-shaped model, define opportunity stage, fiscal time and currency rules, select a warehouse, add sample questions and validated Trusted Assets, and restrict sharing. Build benchmark questions for total pipeline, conversion and aging; inspect generated SQL and edge cases. Use feedback to correct metadata or instructions and re-test rather than accepting popular answers as truth.

Worked analyst decisions and current reading¶

These are original learning cases, not vendor sample questions or recalled exam items.

Find a fanout error: Orders 1 and 2 contain amounts 40 and 60, and order 3 has a null amount. Their total is 100. If customer 10 has two dimension rows, an unconstrained left join duplicates order 1 and produces four rows totaling 140. Check dimension-key uniqueness or use a valid effective-date join; do not conceal the error with distinct measures.
Preserve unmatched rows: A left join followed by WHERE dimension.region = 'East' removes unmatched rows. Put that predicate in ON if the requirement is to keep every fact and attach only matching East attributes. Decide this from the requested population before tuning performance.
Choose the sharing contract: A viewer lacks SELECT on a sensitive table. A published dashboard using publisher data permissions can still expose its result. A current Genie Agent evaluates the viewer's own data identity. Verify each surface separately and choose the publication mode deliberately.
Evaluate an apparent Genie improvement: A new instruction changes a total from 100 to 140 and receives positive user feedback. Reconcile the join and independently verified answer key before marking the benchmark good. Repeat under the same data snapshot, then test paraphrases and follow-up questions separately.

Related item — VERIFY CURRENT: The June 16, 2026 Genie One, Agents and Ontology announcement, by Sydney Sundell, Ken Wong and Elise Georis, explains the broader product direction. Spend 20–30 minutes mapping its context sources, authorization and action boundaries. The current setup documentation corroborates the Spaces-to-Agents terminology. Broader integrations and autonomous actions are adjacent context, not additions to this exam's October 2025 objectives. Vendor benchmark results do not establish accuracy for your own dataset; build a representative evaluation set.

Diagnostic questions with answers¶

Why keep a postal code as text? It identifies a location and can contain leading zeros or letters; arithmetic is not its meaning.
Why can deduplication discard unrelated customers? Missing keys group together, or incomplete ordering chooses an arbitrary survivor. Reject missing keys and supply a stable tie-breaker.
Why can a larger warehouse leave an incorrect total unchanged? Compute capacity does not repair fact grain, fanout or business rules.
Why does NOT IN sometimes return no rows? A null in the comparison set can turn otherwise nonmatching results into unknown; define null behavior and consider NOT EXISTS.
Does restarting serverless prove a cold query? No. Remote result caching may survive; explicitly control result reuse and record other cache assumptions.
Does a dashboard link grant direct SELECT? No. Published shared-permission results may be visible without direct table grants; the permission mode determines data execution.
Do shared dashboard and Genie credentials mean the same thing? No. Genie embeds compute access while using the end user's data permissions; a dashboard can share publisher data permissions.
Does a thumbs-up teach Genie a new rule? Feedback signals a review opportunity; improve validated instructions/examples and retest.
Does a benchmark passing one question prove follow-up handling? No. Benchmark questions start fresh conversations; test contextual follow-ups separately.
Does the new Genie product name change the exam outline? The checked exam PDF and weighted page still use spaces; preserve their scope and cross-reference current terminology.

Hands-on lab sequence¶

Execution boundary, September 28: Eighteen portable SQL assertions ran in a local in-memory SQLite database, covering a subset of Lab 3 and the deduplication example. They checked nulls, aggregates, anti-joins, outer joins, fanout, set operations, ordering and preservation of 02108. This validates those relational examples only. No Databricks workspace, Delta operation, warehouse performance, Unity Catalog grant, dashboard or Genie feature was executed; the service labs below remain proposed.

Catalog evidence: In Catalog Explorer, find a certified dataset; record owner, type, tags, schema, history and upstream/downstream lineage. Explain which signals imply trust and which do not.
Controlled upload: Upload a small CSV, override at least one inferred type, create a governed table and document how production intake would replace the manual route. Test a malformed and duplicate row.
ANSI SQL correctness: Join a fact to two dimensions, calculate exact and approximate distinct counts, use a window calculation, and reconcile row counts, unmatched keys and totals.
Objects and history: Create a managed table, external table, standard view and materialized result where available. Change a Delta table, inspect history and reproduce an earlier result within retention.
Query investigation: Run a deliberately inefficient but safe query. Capture query history/profile evidence, change one factor and compare correctness, scan/shuffle and runtime under similar cache conditions.
AI/BI dashboard: Build a multi-page dashboard with two datasets, three justified visual types, a filter, a parameter and clear titles/units. Publish, share to a test group, schedule refresh and test an alert failure.
Genie space: Curate a narrow dataset, write domain instructions and sample questions, add a trusted query, create five benchmark questions, inspect generated SQL and correct one grounded failure.
Security proof: Create analyst and steward groups or a written simulation. Build a catalog/schema/object grant matrix, test allowed and denied SQL/dashboard/Genie actions, then remove temporary access and artifacts.

Readiness checks¶

Platform, management and import¶

I can distinguish Delta Lake, Unity Catalog, Databricks SQL, Lakeflow Jobs/pipelines, Mosaic AI and the Data Intelligence Engine.
I can explain managed versus external table lifecycle.
I can use certification, ownership, tags and lineage without treating any one signal as proof of correctness.
I can explain provider, share and recipient roles in Delta Sharing/Marketplace.
I can clean null, invalid and duplicate data with an explicit survivor and rejection rule.
I can select upload, file ingestion, connector, federation or sharing from ownership and freshness requirements.
I can document why a UI upload is not a production pipeline.
I can distinguish medallion quality layers from physical or dimensional model patterns.

SQL and query analysis¶

Dashboards, Genie, modeling and security¶

Places to learn¶

This is not a complete list, and it is not a recommendation to consume everything. Pick the explanation, lab environment, course, or assessment that closes a demonstrated gap; spend at least as much time producing and testing SQL, dashboards and Genie evidence as watching video. Durations are library planning estimates unless a provider runtime is identified. Public metadata was checked September 28, 2026; signed-in catalogs and paid lessons were not inspected.

Resource	Access	Estimated time
Official certification page and detailed exam guide	Free	1–2 hours to map objectives and inspect linked sample format; do not reproduce vendor questions
Databricks Academy — search for AI/BI for Data Analysts and SQL Analytics on Databricks	Free account or customer/partner entitlement varies	Approximately 8–16 hours selected learning, plus labs; verify current catalog durations after sign-in
Databricks Free Edition	Free account	12–24 hours across the eight labs and targeted experiments
Databricks SQL documentation	Free	4–8 hours selected reading and reproduction
Databricks YouTube	Free	2–5 hours of selected current SQL, dashboards, Genie and Unity Catalog sessions
Pluralsight: Databricks Certified Data Analyst Associate path	Paid/trial; includes practice exam	Provider lists five courses / 7 hours (listed lesson runtimes total 6h 47m), plus 1–2 hours library-estimated review; Jan–Mar 2025 coverage needs a Genie/nine-domain gap check
Whizlabs: Databricks Certified Data Analyst Associate	Paid; training/practice product	Current retrieval returned no usable course details; prior library budget 4–10 hours is unverified. Verify syllabus/runtime before purchase and reject recalled-question claims
Databricks Community certification forum	Free	30–90 minutes for current program announcements; community answers are secondary evidence

Use practice questions to diagnose a domain and explain every option from the official guide and documentation. Do not memorize recalled live-exam content. Recheck the live weighted page, linked PDF, course replacement note and volatile AI/BI behavior near the appointment.