Databricks
Each of these is a real decision Databricks is working through, written so you can make the call yourself — incomplete information, someone who pushes back, no answer key.
The Databricks Data Intelligence Platform is a unified, cloud-based lakehouse on AWS, Azure, and GCP, built on open-source projects including the Databricks-created Apache Spark, Delta Lake, and MLflow, plus Unity Catalog, the metadata and governance foundation, combining data engineering, ML, AI, data warehousing, BI, and real-time streaming analytics for enterprise data engineering, data science, ML, and BI teams. It processes exabytes of data per day across its fleet of virtual machines. It includes Lakebase, a fully-managed Postgres for intelligent applications and operational workloads; Customer Lake, an agentic Customer Data Platform; Omnigent, a platform to build, compose, govern, and share AI coding agents; Lakewatch, an agentic Security Lakehouse unifying security, IT, and business data in one governed environment for AI-powered detection, investigation, and response; Model Serving, a unified, scalable, governed platform to deploy and manage AI/ML models, from traditional ML to fine-tuned and proprietary LLMs; AI Runtime (AIR), a managed large-scale GPU training and fine-tuning platform; and Unity AI Gateway, moving trillions of tokens per week for thousands of customers.
Your first one is free. No card.
20 live right now.
The rest
Prove incremental views stay correct before launch
Production-like runs forecast 46% lower compute but show 17 rare result mismatches across 60 million update sequences. You have to weigh launch savings against correctness evidence that is expensive to obtain.
Prove failback keeps pipeline state correct
Rare failback simulations show ordering conflicts across stateful pipeline metadata before a committed production cutover. You have to weigh a six-hour recovery objective against proof that recovery never silently loses or duplicates data.
Set the review gate for production pipeline promotion
Seven percent of proposed promotions show conflicting review evidence across Git providers before a committed launch. You have to weigh broad provider coverage against a release boundary that never promotes unverified changes.
Enforce budgets on every governed model request
2.7% of sampled governed requests lack a conclusive budget decision before a committed migration. You have to weigh zero spend leakage against false denials that interrupt production applications.
Rebuild the request flow for governed data
29% of recent governed-dataset requests ended in workspace-wide or owner-level grants, with purpose and scope unclear at decision time. You have to keep access fast without locking in permissions that are expensive to unwind.
Choose one default for cross-region sharing
Two architectures both pass the launch review, each on a different subset of latency, governance, and regional tests, and some findings are machine-generated. You have to commit to one and show the evidence standard behind it.
Keep platform knowledge through the hub move
The consolidation is fixed at 90 days, and 38 people are flagged as sole owners of platform knowledge, though some of those matches are wrong. You have to apply the policy the same way for everyone before notifications go out.
Decide what the export block must catch
Two weeks before the committed control date, 14% of export events in scope cannot be classified either way. You have to weigh broken analytics against the promise you made to a regulator and a customer.
Set the service-credit reserve before launch
A signed contract exposes up to $4.8M in annual service credits, and the reserve is due before the customer goes live. You weigh machine-generated evidence against a wider scenario range, and the terms hold for the year.
Decide if the identity merges are safe to run
Automated matching found likely duplicates across enterprise workspaces, but manual sampling overturned 11% of proposed merges. You have one window before the billing cycle, and a wrong merge breaks access and reporting.
Cut retention to 30 days without breaking audits
The contract now caps customer-level records at 30 days, but copies sit in bronze tables, checkpoints, aggregates, and 17 downstream paths. You have 45 days to make deletion reach all of them and prove the audits still add up.
Decide who gets routed into the new path
Next month, regulated and global accounts can only start through the new assessment, and its routing is not yet proven for them. You weigh speed against tighter access, because retiring the old paths is hard to undo.
Decide what counts as sensitive before enforcement
Final validation shows recall below target on compressed files and multilingual columns, and a dry run of enforcement blocked jobs that should have run. You have to weigh what the model misses against the work it wrongly stops.
Switch pipelines off legacy authentication in 21 days
One authentication mode has a confirmed flaw and has to be off in every managed workspace within 21 days. You have to sequence migration and enforcement so exposure closes without breaking data delivery.
Decide whether to sign a three-year deal with the integrator
The customer won't release procurement without a joint commitment, and its approval meeting just moved up three weeks. You have to weigh consumption commitments against the fixed services terms the integrator wants.
Decide whether the claims cutover can go live
The rehearsal hit target throughput, but lineage checks flag 14 traceability gaps in regulated fields and manual sampling confirms only nine. With 10 weeks to the contract end, you have to decide how much evidence is enough.
Move query results into a regional cache
Dashboard queries that repeat take 29% of SQL compute, and local caches are uneven from one workspace to the next. You have to show the speed gain is real before customers move, and that no one can read another team's results.
Set a retention default before teams expand
In 34% of workspaces, intermediate artifacts sit past the approved window, and regulated accounts are waiting to expand. You have to set a default that cuts stored data without deleting what audits and recovery need.
Decide whether to cut over on the fixed date
The final rehearsal found three cases where every component check passed but dashboards came back stale or duplicated after recovery. You weigh a delay against a cutover that could break data.