Question 440 A job runs on a small node and repeatedly crashes during an expensive aggregation. Which measured changes should be considered before scaling blindly? Only changing the Spark UI date format Per-task memory, shuffle size, input growth and partition skew Only the table namespace length Select a larger node immediately without comparing spills or task input sizes Explanation Correct Answer B:The cause can be insufficient memory or poor distribution; measurements inform whether tuning or resources help. Incorrect Answers A:Display settings cannot resolve the physical resource bottleneck. C:Table identifier size does not reveal aggregation pressure. D:Blind scaling may mask or worsen cost without identifying whether skew, shuffle, or per-task memory is the actual bottleneck.
Question 441 A groupBy task spills to disk after a recent change that increased the width of every row by adding a large text field. What is a likely remediation? Project only needed columns before the aggregation when semantics permit Increase shuffle partitions blindly while retaining all large unused text columns Keep all unused text columns through every shuffle Increase driver memory without projecting the unused large text field before grouping Explanation Correct Answer A:Removing unused wide fields can reduce intermediate row size and memory pressure. Incorrect Answers B:More partitions may reduce per-task payload, but removing unnecessary wide columns before shuffle directly addresses the newly inflated row width. C:Unnecessary large columns can worsen data movement and spilling. D:The wide rows move through executor shuffle tasks, so driver heap is not the targeted fix for avoidable per-row data volume.
Question 442 An ETL job's run history indicates intermittent failures following cloud network errors, but stable runs show correct results. What mitigation is most appropriate? Bounded retry with checkpointed or idempotent task semantics and monitoring Disable every retry and permanently accept missing batches Increase job parallelism without storing source progress for safe retries Ignore all failures if a later run succeeds Explanation Correct Answer A:Transient network faults may recover, but retry safety and persistent failures still need visibility. Incorrect Answers B:Avoiding retries can leave data gaps after temporary connectivity issues. C:More concurrent tasks cannot protect completeness after network interruptions; bounded idempotent retries and checkpointing are needed. D:Unobserved missed intervals may still contain missing source records.
Question 443 A SQL query becomes slower after a partitioning change but the data volume remains similar. Which comparison is most useful? Compare only overall job duration without inspecting file pruning or the new scan plan Compare only the configured number of table partitions before and after Execution plans, files scanned, pruning effectiveness and shuffle metrics before/after Only the query author's display name Explanation Correct Answer C:A physical-plan comparison can reveal whether the layout change increased scanned data or exchanges. Incorrect Answers A:Elapsed time identifies a regression but not whether the partitioning change altered scanned files, pruning or shuffles. B:Partition count alone does not show whether the query plan prunes files or incurs larger shuffles after the change. D:Identity labels are not an execution plan.
Question 444 A running Spark task has unusually high remote fetch wait time compared with its compute time. Which area is likely constraining it? Network/shuffle fetch behavior rather than pure arithmetic CPU Increase executor CPU cores while ignoring large remote shuffle fetch wait Increase upstream file retention while leaving remote shuffle-fetch delays uninvestigated Optimize CPU-only expression execution while ignoring shuffle transfer time Explanation Correct Answer A:Excess wait on remote shuffle blocks indicates data transfer latency or exchange pressure. Incorrect Answers B:More cores may help CPU-heavy tasks but cannot directly remove network or shuffle-read waits when transfers dominate execution. C:Retention policy affects source availability; it does not correct slow network transfers during Spark shuffle fetch. D:Arithmetic tuning cannot directly address a stage primarily stalled while retrieving remote shuffle blocks.
Question 445 An engineer uses the Spark UI after a join and sees task input rows quadruple unexpectedly. What should be examined? Join key cardinality and duplicate matches on each side Only cluster hostname labels Increase executor memory before checking the join keys and multiplicity Inspect only output Delta file sizes and assume they identify the join key causing row multiplication Explanation Correct Answer A:Many-to-many joins can multiply row counts and increase downstream work. Incorrect Answers B:Hostname naming does not determine relational matching. C:Memory changes can mask resource pressure but not correct unexpected one-to-many or many-to-many join multiplication. D:Output file sizes can show growth but cannot reveal which join key or nonunique match condition multiplied the joined rows.
Question 446 A team notices two job runs with identical inputs have very different durations. Which variables should be controlled when benchmarking? Compute size, cache state, concurrency, data volume and runtime versions Only the reviewer's Git username Compare task execution time but leave cluster size and background concurrency uncontrolled Only the output table display name Explanation Correct Answer A:Comparable conditions are needed to attribute timing changes to a specific optimization. Incorrect Answers B:Review authorship is not the runtime configuration. C:A benchmark that ignores resource contention and runtime conditions cannot attribute the measured difference to the intended change. D:A name does not capture execution environment differences.
Question 447 A large Delta query spends most of its time on metadata planning because it touches hundreds of thousands of tiny files. What should be addressed? Physical file layout and compaction/predictive optimization eligibility Increase broadcast join threshold while leaving hundreds of thousands of tiny files Only the notebook branch name Only table access privilege count Explanation Correct Answer A:Reducing fragmentation can lower metadata overhead and improve scan planning. Incorrect Answers B:Broadcast decisions do not directly repair the metadata and file-open overhead from fragmented Delta data. C:Source branch naming does not compact Delta files. D:Grants do not by themselves reduce file count.
Question 448 A task is failing with out-of-memory while repartitioning a large input, but the driver logs show no collect call. Where should attention focus? Increase driver memory but do not inspect executor memory or shuffle payload sizes Increase task retries while retaining the same oversized shuffle partitions Executor task payloads, shuffle partition sizes and memory settings Increase the number of table partitions without examining the actual shuffle data distribution Explanation Correct Answer C:Without driver collection, repartition OOM is more likely in distributed task memory and shuffle behavior. Incorrect Answers A:The stated failure arises during distributed repartitioning without driver collection; changing only driver memory is unlikely to correct executor task pressure. B:Retries repeat executor work on similarly sized partitions and do not fix the underlying memory footprint or partition skew. D:Table partition count alone does not show executor task payloads or whether the repartition stage creates oversized shuffle partitions.
Question 449 A team uses a SQL warehouse for an interactive report and notices heavy queueing at peak times. What is relevant for tuning? Add more source event-time columns without measuring SQL warehouse queue duration Raise the task retry limit while keeping warehouse size and concurrency unchanged Refresh the dashboard more frequently while leaving peak SQL warehouse queues unchanged Warehouse scaling and query concurrency settings based on actual peak load Explanation Correct Answer D:Queueing may reflect serving concurrency constraints that require resource sizing and workload management. Incorrect Answers A:Event-time metadata has no direct impact on the warehouse concurrency bottleneck reported during interactive peaks. B:Retries are designed for failed executions and do not address query queues caused by insufficient serving capacity at peak load. C:More frequent refreshes can increase load; warehouse capacity and concurrent query behavior are the relevant performance controls.
Question 450 A job succeeds when executed alone but fails on memory when many identical jobs overlap. What should be examined? Only the column names in the input schema Concurrency, shared compute capacity and per-job memory requirements Only notebook comment style Only dashboard formatting Explanation Correct Answer B:Overlapping executions can create resource contention that is absent in isolation. Incorrect Answers A:Column labels do not explain concurrency-dependent memory loss by themselves. C:Comment style does not prevent overlapping tasks. D:Presentation cannot allocate additional memory.