DP-750 — frequently asked questions

Shared compute is interactive and mixes workloads across users, so it cannot guarantee production isolation or per-run autoscaling for scheduled, non-interactive telemetry pipelines.

No. Photon is a performance accelerator; autoscaling comes from running ingestion on an autoscaling job compute cluster instead of a fixed-node all-purpose cluster.

The assets table must be created in finance.procurement, so the session must first switch the current catalog to finance, then set the current schema to procurement.

It is valid fully qualified syntax, but the exhibit provides separate USE CATALOG and USE SCHEMA targets, so the expected drag values are finance, procurement, and assets.

Disabling Photon can significantly slow down query execution, leading to longer runtimes and potentially higher total compute costs, offsetting any DBU savings.

No, fixing the worker count prevents the cluster from scaling down during low activity, leading to wasted resources and higher costs compared to using autoscaling.

Liquid clustering handles high-cardinality columns and evolving filter patterns without re-partitioning, while partitioning on CustomerId would create too many small files and degrade performance.

Deletion vectors mark rows as deleted without rewriting full Parquet files, so compliance updates and deletes complete faster and with less write amplification.

Spot Instances can save up to 90% but introduce interruption risks. For critical batch workloads where reliability is key, right-sizing reserved/on-demand nodes via fewer workers is safer.

No, because the workload is predictable and does not spike. Autoscaling is designed for variable loads; removing it for stable loads simply stops paying for unused reserve capacity.

Job clusters require manual management and setup. Serverless compute automates provisioning and scaling, significantly reducing administrative effort.

Yes, serverless compute supports continuous processing modes required by streaming pipelines in Lakeflow SDP while optimizing costs via pay-per-use.

Terraform manages cloud infrastructure resources. While it can deploy some Databricks objects, Asset Bundles are the native, optimized tool for managing notebooks, jobs, and workflows via Git.

It allows the same bundle definition to be deployed to different environments (dev, prod) with specific variables (like cluster sizes or endpoints) without changing the core code.

No. Databricks Runtime ML includes GPU-capable libraries, but GPU acceleration requires a GPU-enabled VM node type; a general purpose node has no GPU to accelerate.

A single-node cluster runs only a driver with no worker nodes, so there are no executors to spread tasks across; distributed training needs a multi-node cluster with workers.

Genie instructions are stored on the Genie space, not on tables or schemas, so guidance placed on Sales1 or Schema1 would not scope queries to Space1 and could leak across spaces.

Instructions are private to each Genie space, so editing Space1's instructions changes only that space's query routing while all other spaces keep their own settings.

No. dbutils.secrets.get resolves storageKey from kv-secure on each call, so Notebook1 always sees the value the cloud security team last published.

Only read access to the secret (Get). List, write, delete or Secrets Officer roles are not needed for retrieval and violate least privilege.

The hotspot allows only one property; delta.deletedFileRetentionDuration is the one that lets VACUUM reclaim deleted data files. Log retention can be managed separately.

No. VACUUM runs online and readers can query the table; unreferenced files are removed after the retention threshold.

The transaction log (_delta_log) is a separate retention track with its own 30-day clock. Without delta.logRetentionDuration, the history of the deletions can age out independently, breaking the audit trail for the deleted customer data.

Deletion vectors mark rows as deleted but keep the underlying files indefinitely, so nothing is removed permanently without a manual REORG TABLE step. That fails the automatic 30-day permanent-removal requirement and increases administrative effort.

ALL PRIVILEGES includes MANAGE, which allows group1 to grant permissions, violating the 'cannot grant permissions' requirement.

Yes, USE CATALOG on catalog1 is required to access the catalog before using schema1 and creating tables.

No — the connector is the Azure resource that carries the identity. The authentication is the managed identity it uses, which is why that is the credential-free dropdown choice.

READ VOLUME on Volume1 (plus USE CATALOG and USE SCHEMA on its parents). WRITE VOLUME would allow modification and breaks least privilege.

Use Unity Catalog Foreign Catalogs or create external tables using the SQL Server JDBC connection.

It is used for ingesting and transforming data by moving it into Databricks-managed Delta tables.

Both sources are file-based in ADLS Gen2 and require schema drift handling; Auto Loader provides incremental ingestion with schema evolution and exactly-once semantics.

The weather data comes from a custom REST API; a Databricks notebook allows custom HTTP calls, while Lakeflow Connect is typically for supported databases and SaaS apps.

The two options conflict. ignoreChanges is the deprecated Delta streaming setting and Databricks raises an error when it is combined with a CDF read; use skipChangeCommits instead.

insert, update_preimage, update_postimage and delete. The preimage/postimage pair lets the downstream pipeline replay each update deterministically.

A snapshot only provides a static view of data at one moment. It cannot inherently track or display changes over time as new records are added.

While ideal, the exam scenario focuses on the functional need to track progression. In practice, time series profiles leverage versioning or timestamps to show metric evolution.

No. In Lakeflow SDP, failed expectations prevent the specific invalid record from being written to the target table and send it to quarantine, allowing the rest of the pipeline to complete successfully.

Option D requires writing custom code to identify, filter, and store bad records. Option C uses built-in declarative syntax, which is significantly faster to implement and maintain.