Question 420 A gold query has fast compute but spends excessive time opening many Delta data files. Which issue should be evaluated? Change access grants for the query user while ignoring physical file fragmentation Rewrite the notebook comments while leaving the many small Delta files unchanged Small-file fragmentation and compaction state Increase SQL warehouse size without inspecting file fragmentation Explanation Correct Answer C:Opening a very large number of tiny files can dominate I/O and metadata overhead. Incorrect Answers A:Authorization changes do not correct file-open overhead once queries already execute successfully. B:Comment updates do not change the number of files the engine must open and schedule during the scan. D:Compute scaling alone does not remove per-file open and metadata overhead from many tiny files.
Question 421 A Lakeflow Jobs task is skipped after its upstream task fails. The task's own notebook has no error log. What is the likely reason? Unity Catalog automatically removed the table The downstream notebook necessarily ran and crashed silently The downstream task did not execute because its dependency condition was unsatisfied Git automatically deleted the notebook output Explanation Correct Answer C:A skipped task can have no application exception because the orchestrator never started its notebook. Incorrect Answers A:A skip state does not equate to data-object deletion. B:No execution may have happened at all after upstream failure. D:Job dependency behavior does not imply Git file deletion.
Question 422 A job's median runtime remains stable, but its p95 runtime has doubled with occasional severe slow executions. What should the engineer investigate? Only the average row count over the entire month Compare only the average successful runtime instead of tail distributions Inspect a representative fast execution only and disregard slow outlier run IDs Outlier run histories, skewed stages and intermittent resource contention Explanation Correct Answer D:Tail-latency regression can hide behind an unchanged median; outlier-level evidence is needed. Incorrect Answers A:Averages can conceal occasional data spikes and slow tasks. B:A stable median can conceal the few extreme slow runs responsible for the doubled p95. C:Typical fast runs cannot isolate the cause of occasional severe slow tasks; outlier evidence is necessary.
Question 423 A Spark executor crashes repeatedly due to memory pressure during an unbounded collect_list aggregation on a hot key. What is the root issue? Accumulating an extremely large per-key collection in one task Raise executor memory without examining the hot key's unbounded aggregation Increase SQL warehouse concurrency for the failing Spark aggregation Disable all retries so the oversized per-key aggregate fails only once Explanation Correct Answer A:An unbounded aggregation for a skewed key may exceed executor memory, requiring a different aggregation design. Incorrect Answers B:More memory may postpone a crash but does not address the pathological per-key collection that grows without bound. C:The Spark executor crashes while materializing a hot key; unrelated SQL serving capacity will not correct that operation. D:Fewer retries reduce repeated failures but do not bound the collection or fix the memory pressure.
Question 424 A team wants to compare current job failure rates with those last month. Which denominator is necessary for a meaningful rate? Only the latest single run status Number of failed runs divided by total relevant runs for each period Compare failure counts across periods without normalizing by the number of attempted runs Measure failed tasks but never count the job runs that were eligible to fail Explanation Correct Answer B:Rates need comparable exposure counts rather than raw failed-run numbers alone. Incorrect Answers A:One run cannot establish a monthly failure rate. C:Raw failure counts can mislead if execution volume increased or decreased; use per-period failures divided by relevant runs. D:A rate requires an appropriate run-level denominator, not just the number of failures.
Question 425 A job run-history graph shows failures spike only after a change to the selected compute policy. What is the targeted investigation? Only change dashboard layout Remove Unity Catalog read access for every user Compare current and previous compute-policy constraints and startup error messages Immediately rewrite all SQL joins Explanation Correct Answer C:A policy change correlated with startup failures is a grounded diagnostic lead. Incorrect Answers A:Layout changes do not affect cluster policy enforcement. B:Revoking data rights does not address compute startup policy. D:Join semantics are not the initial suspect for provisioning policy failures.
Question 426 A Databricks SQL warehouse is consistently saturated during dashboard hours, while Spark ETL jobs execute on separate compute. What should be tuned? Only catalog table descriptions Only the ETL Spark driver memory The SQL warehouse sizing/concurrency behavior for that dashboard demand Only ingestion source file naming Explanation Correct Answer C:The bottleneck is in SQL serving capacity, not the isolated ETL clusters. Incorrect Answers A:Descriptions do not allocate compute. B:ETL resources are separate and do not directly add SQL warehouse serving capacity. D:Object names do not increase warehouse concurrency.
Question 427 A team applies Liquid Clustering to a large table but continues manually enforcing old directory partition assumptions. What should it reconsider? Assume Liquid Clustering requires a fixed partition per every distinct key Keep both rigid static partitioning conventions and Liquid Clustering as independent layouts Only enable notebook line numbers The intended clustering layout and supported maintenance model rather than rigid manual partitions Explanation Correct Answer D:Liquid clustering changes the layout strategy; legacy partitioning assumptions may conflict with the intended optimization. Incorrect Answers A:Rigid overpartitioning is not its core approach. B:The team should reassess manual partition assumptions under a supported clustering model rather than maintain incompatible physical layout constraints. C:Editor settings are not physical file management.
Question 428 An executor OOM happens only when a particular region processes unusually many events. Which finding most strongly supports data skew? The slow partition contains most of that region's records while others remain small The notebook output header is long The cluster has enough total aggregate memory but one task is large All regions contribute identical row counts Explanation Correct Answer A:Concentrated key volume produces imbalanced tasks and memory pressure. Incorrect Answers B:Presentation has no causal link to partition distribution. C:Total memory does not eliminate per-task hotspots, so this is compatible with skew rather than proof against it. D:Uniform distribution would weaken a region-specific skew hypothesis.
Question 429 A job stage spends substantial time garbage-collecting objects after processing many very large per-task collections. Which class of issue should be checked? Increase parallel job concurrency while retaining the same oversized per-task object collections Tune only SQL warehouse capacity when the observed GC occurs inside Spark job executors Executor heap pressure and intermediate object retention Assume slow task CPU utilization proves memory is abundant Explanation Correct Answer C:High garbage-collection time can indicate excessive in-memory working sets or memory allocation pressure. Incorrect Answers A:More parallel tasks can increase aggregate memory pressure; it does not address excessive heap allocations and GC per task. B:Warehouse capacity changes BI compute, not heap allocation and object retention on the Spark executors running this stage. D:Frequent garbage collection can dominate CPU even when the heap is under severe allocation pressure.
Question 430 A file ingestion pipeline grows from 1 GB to 100 GB per day and its run duration rises proportionally. Which evidence should guide tuning? Only absolute run duration without input size Input-size-normalized throughput, task utilization and I/O metrics Only notebook source line count Only the count of catalog folders Explanation Correct Answer B:Comparing bytes processed per unit time distinguishes normal scaling from deteriorating efficiency. Incorrect Answers A:More input naturally takes longer, so duration alone may falsely imply a regression. C:Code line count is not the measured data processing rate. D:Metadata inventory does not characterize the throughput change.