Pass Databricks Certified Data Engineer Professional Exam in First Attempt Guaranteed!

Monitor cache hit rate and eviction pressure. This monitor action addresses a different case (cache reuse), where repeated recomputation dominates each branch's execution. Persistence can avoid repeated work if reused data fits the memory budget. In the stated problem, an unsupported execution mode limits vectorized engine use; this choice would not establish the required photon query execution decision.

Monitor serverless task failures and cost per run. This monitor action addresses a different case (serverless job fit), where provisioning overhead exceeds useful task execution time. Supported serverless jobs can reduce infrastructure management and startup overhead. In the stated problem, an unsupported execution mode limits vectorized engine use; this choice would not establish the required photon query execution decision.

Monitor query scans and available column stats. This monitor action addresses a different case (data skipping statistics), where file-level statistics are missing for the filter column. Delta data skipping uses collected column statistics to eliminate files. In the stated problem, an unsupported execution mode limits vectorized engine use; this choice would not establish the required photon query execution decision.

Inspect raw source and the rescued-data column distribution. This diagnose action addresses a different case (auto loader rescued data), where unknown or mismatched fields disappear during the bronze load. The rescued data column can preserve records with unexpected schema content for later remediation. In the stated problem, a merge step fails after a producer changes an integer field to string; this choice would not establish the required parquet schema mismatch decision.

Inspect duplicate file IDs and the ingestion checkpoint state. This diagnose action addresses a different case (retry-safe file loads), where a partially processed input file is loaded twice into Delta. Checkpointed incremental ingestion or stable event-level deduplication prevents repeated effects. In the stated problem, a merge step fails after a producer changes an integer field to string; this choice would not establish the required parquet schema mismatch decision.

Inspect writer and reader schemas and decoding failure examples. This diagnose action addresses a different case (avro schema evolution), where an older consumer cannot deserialize records from a newer producer schema. Avro writer-reader schema compatibility must be maintained for evolution. In the stated problem, a merge step fails after a producer changes an integer field to string; this choice would not establish the required parquet schema mismatch decision.

Publish a minimal secure view with only approved attributes. This design action addresses a different case (sensitive data suppression), where a broad SELECT still includes unstructured notes. Column suppression and data minimization prevent unnecessary sensitive-field release. In the stated problem, expired records remain available in historical serving tables; this choice would not establish the required retention purge decision.

Plan secret rotation with a new value and safe rollback. This design action addresses a different case (credential rotation procedure), where the hard-coded service password has not rotated for a year. Secret management and controlled rotation reduce long-lived credential exposure. In the stated problem, expired records remain available in historical serving tables; this choice would not establish the required retention purge decision.

Assign distinct deployment operators and data stewards with scoped privileges. This design action addresses a different case (separation of duties), where all developers currently own production catalogs and can grant access. Separate deploy and governance roles reduce accidental privilege escalation. In the stated problem, expired records remain available in historical serving tables; this choice would not establish the required retention purge decision.

Apply the data-skipping-friendly predicate or clustering update. This implement action addresses a different case (query profile scans), where operators assume the query is always CPU-bound. Query Profile identifies scan, pruning and shuffle behavior for SQL workloads. In the stated problem, operators only see the most recent run in the Jobs UI; this choice would not establish the required job operational history decision.

Query the pipeline event log for the failed dataset and update ID. This implement action addresses a different case (pipeline event log), where the Jobs status only reports a high-level failure. Lakeflow pipeline event logs include update-level errors and progress events. In the stated problem, operators only see the most recent run in the Jobs UI; this choice would not establish the required job operational history decision.

Record stage timestamps and expose a freshness query. This implement action addresses a different case (end-to-end sla), where operators monitor compute uptime rather than data freshness. SLA monitoring should measure source event time to published table availability. In the stated problem, operators only see the most recent run in the Jobs UI; this choice would not establish the required job operational history decision.

Audit share membership changes against approvals. This monitor action addresses a different case (share catalog scope), where the recipient can discover unintended datasets. A Delta Sharing share should contain only explicitly approved assets. In the stated problem, catalog discovery was mistaken for full table access; this choice would not establish the required foreign catalog permission decision.

Monitor recipient access and credential expiry. This monitor action addresses a different case (open protocol recipient), where the partner cannot use an internal Databricks-to-Databricks credential flow. Open Delta Sharing can support non-Databricks recipients using the sharing protocol. In the stated problem, catalog discovery was mistaken for full table access; this choice would not establish the required foreign catalog permission decision.

Audit share membership and newly exposed tables. This monitor action addresses a different case (d2d managed recipient), where a recipient team copies files manually and loses timely updates. Databricks-to-Databricks Delta Sharing supports governed live access without manual copying. In the stated problem, catalog discovery was mistaken for full table access; this choice would not establish the required foreign catalog permission decision.

Run databricks bundle validate for the intended target in CI. This implement action addresses a different case (bundle validation), where an invalid resource key would be rejected only after production change. Bundle validation detects many configuration problems before deployment. In the stated problem, the notebook cell shows only a generic stage failure; this choice would not establish the required cluster log diagnostics decision.

Add relevant test jobs and environment approval gates to the pipeline. This implement action addresses a different case (ci quality gates), where the deployment pipeline only checks YAML syntax. CI must run transformation tests and controlled integration checks before deployment. In the stated problem, the notebook cell shows only a generic stage failure; this choice would not establish the required cluster log diagnostics decision.

Pin or rebuild incompatible wheels and update the bundle library. This implement action addresses a different case (python dependency triage), where a package import error appears only on the target compute. Runtime-specific wheel compatibility and pinning must be tested in CI. In the stated problem, the notebook cell shows only a generic stage failure; this choice would not establish the required cluster log diagnostics decision.

Read old and new records and verify null semantics for the new field. This validate action addresses a different case (controlled schema evolution), where the silver writer rejects every new record after the upstream schema change. Schema evolution must be reviewed rather than enabling unrestricted acceptance for every column. In the stated problem, the pipeline continues despite critical null records; this choice would not establish the required expectation fail action decision.

Append null IDs and confirm they are excluded while valid IDs remain. This validate action addresses a different case (expectation drop action), where invalid rows currently reach downstream aggregates. A Lakeflow expectation configured to drop invalid rows filters violations from target output. In the stated problem, the pipeline continues despite critical null records; this choice would not establish the required expectation fail action decision.

Count retained fact rows and confirm null-key records are represented. This validate action addresses a different case (join with nullable keys), where rows with null external IDs disappear from an inner join. SQL null equality and join type determine whether unmatched rows survive. In the stated problem, the pipeline continues despite critical null records; this choice would not establish the required expectation fail action decision.

Use a left join plus a documented fallback label for unmatched keys. This design action addresses a different case (conditional enrichment), where an unconditional inner join discards rows with missing region codes. A left join with explicit CASE logic preserves rows while enriching those that match. In the stated problem, the join triggers a costly shuffle of the large fact dataset; this choice would not establish the required broadcast lookup table decision.

Preserve unmatched facts with a left join when business rules require them. This design action addresses a different case (join with nullable keys), where rows with null external IDs disappear from an inner join. SQL null equality and join type determine whether unmatched rows survive. In the stated problem, the join triggers a costly shuffle of the large fact dataset; this choice would not establish the required broadcast lookup table decision.

Choose fail semantics for non-negotiable constraints. This design action addresses a different case (expectation fail action), where the pipeline continues despite critical null records. Fail-on-violation expectations stop updates rather than silently filtering rows. In the stated problem, the join triggers a costly shuffle of the large fact dataset; this choice would not establish the required broadcast lookup table decision.

Monitor repeated dependency failures and overall critical-path duration. This monitor action addresses a different case (job task dag), where dependent tasks sometimes run before the bronze load finishes. Lakeflow Job task dependencies form a directed acyclic execution graph. In the stated problem, a transitive dependency updated to an incompatible version; this choice would not establish the required library version pinning decision.

Capture Arrow conversion failures following package upgrades. This monitor action addresses a different case (udf return schema), where returned values cannot be converted to the declared Spark return type. Pandas UDF return values must conform to the declared Spark data type. In the stated problem, a transitive dependency updated to an incompatible version; this choice would not establish the required library version pinning decision.

Monitor entrypoint failures after runtime or package upgrades. This monitor action addresses a different case (python task entrypoints), where the task exits before setting up its argument parser. A standard Python entrypoint and explicit arguments avoid notebook runtime assumptions. In the stated problem, a transitive dependency updated to an incompatible version; this choice would not establish the required library version pinning decision.

Use lower and trim native expressions for string standardization. This design action addresses a different case (avoid python udf for sql), where the Python UDF creates an avoidable cross-language execution boundary. Built-in Spark SQL functions can execute in optimized expressions without Python UDF overhead. In the stated problem, row-by-row Python invocation dominates stage runtime; this choice would not establish the required vectorized udf decision.

Preserve a stable checkpoint location for compatible restarts. This design action addresses a different case (streaming checkpoint identity), where a newly selected checkpoint location makes the stream reread historical input. Streaming checkpoint metadata records progress and state for restart continuity. In the stated problem, row-by-row Python invocation dominates stage runtime; this choice would not establish the required vectorized udf decision.

Define validated region and catalog parameters for each invocation. This design action addresses a different case (job parameters), where code contains hard-coded region IDs and catalog names. Task parameters allow runtime-specific values without changing notebook source. In the stated problem, row-by-row Python invocation dominates stage runtime; this choice would not establish the required vectorized udf decision.

Inspect per-key frequencies and Spark task duration distribution. This diagnose action addresses a different case (skewed group aggregation), where a single campaign dominates rows and one reduce task runs for minutes. A hot aggregation key causes skew despite ample parallelism. In the stated problem, the current INSERT-only target contains duplicate customer keys; this choice would not establish the required cdc merge semantics decision.

Inspect input schema, target schema and invalid decimal examples. This diagnose action addresses a different case (silver schema enforcement), where implicit type coercion risks invalid numeric values being stored. Schema validation and explicit casts make type conversion failures observable. In the stated problem, the current INSERT-only target contains duplicate customer keys; this choice would not establish the required cdc merge semantics decision.

Inspect resend delay distribution and state-store size. This diagnose action addresses a different case (streaming dedup watermark), where a plain dropDuplicates retains state without a practical bound for never-ending keys. Watermarked streaming deduplication can bound retained duplicate state under an explicit lateness policy. In the stated problem, the current INSERT-only target contains duplicate customer keys; this choice would not establish the required cdc merge semantics decision.