Choosing a Cloud Data Platform: 10 Questions That Matter More Than the Demo
Judge a platform by what it can enforce for your business, not by how quickly it answers a demo query.
In this article
1The demo went perfectly. The meeting did not.
Picture a retail group. Its three-week proof of concept ended on a high. Dashboards loaded faster than anyone remembered, and the vendor’s engineers even showed a personalised-offer prototype.
Then the CFO asked two questions. Which of our two revenue figures will the platform treat as correct? And what will the monthly bill look like once marketing connects its offer engine? Nobody could answer either. The pilot had tested speed. It had not tested definitions, accountability or running cost, and those decide whether a platform pays back.
That gap is the subject of this guide. Most buying checklists compare features. This one asks what you must decide before features matter, and what evidence should convince you. Logesys is a Databricks and Microsoft Fabric partner, so weigh our perspective accordingly. The questions apply equally to platforms we do not partner with, and the last section explains when waiting is the right answer.
A successful platform decision is not the one that wins the demo; it is the one your organisation can govern, afford and operate after implementation.
2Start with your seat
Leaders carry different risks into the same decision. Begin with the questions that match your role, then read the rest.
| If you are | Start with | Because your exposure is |
|---|---|---|
| CIO or CTO | 389 | Integration effort, skills and lock-in |
| CFO or finance lead | 610 | Cost that grows after signing |
| CDO or head of analytics | 257 | Definitions, access, and AI use of data |
| Enterprise architect | 349 | Estate fit, freshness and reversibility |
| BI leader | 24 | Consistent numbers and responsive dashboards |
3Why this is a business decision
Capabilities across the major platforms overlap more each year. What differs is the operating model: who pays for compute, where governance lives, how open the storage is and how many specialists you need.
Think of a chain:
- Platform
- Pipelines
- Trusted data
- Business intelligence
- AI and automated decisions
A weakness upstream arrives downstream as two dashboards disagreeing, or an assistant quoting the wrong customer record.
The platform can enable each link. Only your organisation decides what each link means.
4The ten questions
Each question gives the decision, the evidence to request and the red flag.
1What decision will this platform change?
Name the business decision it must improve, such as replenishment, pricing or risk limits, and the metric that proves it. Ask for a written list of two or three use cases with owners, the data each needs and a baseline measured today. Include one structured use case and one that mixes data types.
If you are unsure what should move at all, settle that first with data estate triage.
The red flag is success defined as “migrated” or “dashboards are faster”.
2Which definitions must survive across departments?
Decide who owns “revenue” and “active customer”. A platform can store both versions faithfully, so ask the vendor to show one metric defined once and reused by the BI tool, SQL users and an AI assistant, with lineage (the record of where a figure came from) back to source. The trade-off is that central definitions slow local teams, and deliberate regional variants need recording.
Take two KPIs your leaders already dispute and reconcile them in the pilot.
The red flag is “each team defines measures in its own tool”.
3Where does your data live, and how does it arrive?
Check the connector list against your real sources, including change data capture (replicating only what changed) from your ERP, and estimate the cost of anything missing. Look at openness too. Microsoft documents that OneLake stores data as Delta Parquet by default, and that shortcuts can reference data in Amazon S3, Azure Data Lake Storage and other stores without copying it. Azure Databricks catalogs can be mirrored in the same way, but those shortcut tables are read-only and engine feature support varies. Tight ecosystem fit speeds the start and narrows later options.
Ingest your messiest source, not your cleanest.
The red flag is “custom connectors can be built” with no estimate.
4How fresh must data be, for whom, and at what concurrency?
Classify use cases by required freshness instead of defaulting everything to streaming. Test with your own concurrent users, refresh patterns and data volumes, and ask for latency under load as measured, not quoted. Fresher data costs more and needs monitoring.
The red flag is a benchmark run on a dataset unlike yours.
5Who can see what, and can you prove it?
Decide where access policy lives: in one place or tool by tool. Ask each vendor for the same demonstration on a sensitive table: row and column controls, masking, lineage and an audit trail. Databricks, for example, documents Unity Catalog as providing access control, lineage and auditing for data and AI assets, yet notes that its privileges are not enforced when external systems read the data files directly. Strict central policy slows onboarding.
Have your security team try to read a masked column through every route: SQL, notebook, BI tool, API and storage.
The red flag is separate permission models for tables, files and ML assets.
6What will it cost in year three, and who sees the bill?
Pricing mechanics differ, so model your workload under each. Snowflake bills credits while a virtual warehouse runs, even when idle, and multi-cluster rates (an Enterprise Edition feature) multiply by the clusters running. BigQuery offers on-demand pricing per TiB processed or capacity pricing per slot-hour, and its controls differ: daily quotas and per-query limits on demand, slot limits under capacity. Usage pricing follows demand but is harder to forecast. Commitments forecast well but risk idle spend. Add people, migration, parallel running and data movement.
For the AI layer, see our piece on cost governance for AI.
The red flag is a total cost of ownership that lists only storage and compute.
7What changes when AI and agents arrive?
AI-ready infrastructure is not AI-ready data. Decide which data AI may use and who approves it. Ask for permission-aware retrieval: a document chunk that carries its source’s access rights, and an audit trail of what an agent read. The NIST AI Risk Management Framework, a voluntary US framework, organises this around four functions: Govern, Map, Measure and Manage.
Our AI-ready data foundations guide covers the architecture layers.
The red flag is vector search sitting outside the governed catalog.
8Who will run it, and do those people exist?
Match the platform to your skills: SQL-first teams and Spark or Python teams are suited by different products. Ask for a support model, runbooks and an on-call plan. Managed and serverless options reduce operational load but give you less control and less cost visibility.
The red flag is a plan that assumes hiring nobody has approved.
9How hard is it to leave?
Choose your lock-in deliberately. Ask about table formats (Delta or Iceberg), catalog portability, export paths and contract terms. Openness costs convenience, because proprietary features are often the easiest route. Consolidation also concentrates operational risk in one place.
The red flag is critical logic that lives only in a proprietary feature with no export route.
10What would make us stop?
Write exit criteria before the pilot starts: the disputed KPI reconciles, the security test passes on every route, and projected cost sits inside an agreed band. Criteria set in advance feel bureaucratic, but they prevent sunk-cost momentum.
The red flag is judging pilot success after the results arrive.
5How the operating models differ
This table points to what to verify, not a ranking. Capabilities, pricing and preview status change often, so confirm each against current vendor documentation. If you are still choosing between warehouse, lake and lakehouse patterns, start with our architecture comparison.
| Platform | Operating model to examine | Verify in your pilot |
|---|---|---|
| Databricks | Lakehouse on open table formats, with Unity Catalog governing data and AI assets | Team skills, cost attribution |
| Microsoft Fabric | OneLake with Delta Parquet by default, and shortcuts to external storage | Fit with your Microsoft footprint, capacity costs |
| Snowflake | Credit-based virtual warehouses, plus separately billed serverless features | Warehouse sizing, idle time, multi-cluster use |
| Google BigQuery | Managed service with on-demand or capacity pricing | Scan costs versus reservations |
| Amazon Redshift | AWS-native warehouse, provisioned or serverless | AWS integration, tuning effort or serverless capacity cost |
Scroll the table sideways to see all columns
For platform detail, see our Databricks and Microsoft Fabric pages.
6Test your shortlist with one real workload
Hyper-personalisation, meaning offers tailored to one customer in near real time, is a demanding test because it exercises several questions at once. It forces an answer to:
- Which “customer” is authoritative question 2
- How fresh cart and browsing data must be 4
- Whether consent travels with the profile 5
- What each personalised decision costs at scale 6
- What a model may see 7
Run your shortlist against it. Our article on why Customer 360 projects fail shows where such efforts stall.
7Proceed, investigate or pause
Use this as a conversation guide, not a validated score.
| Signal | Proceed | Investigate further | Pause |
|---|---|---|---|
| Use cases | Named, with owners and baselines | Named, no baselines | Cannot name the decision |
| Definitions | Disputed KPIs reconciled in the pilot | Reconciled only on vendor data | Disputed, with no owner |
| Access | Security test passed on every route | Passed on some routes | No test run |
| Cost | Modelled on your workload, with an owner | Modelled on vendor assumptions | Storage and compute only |
| Operations | Operators identified and funded | Hiring planned, not approved | Nobody to run it |
| Exit | Format and export route documented | Partly documented | Unknown |
Scroll the table sideways to see all columns
If your current warehouse already meets these tests, the right answer may be to improve it, or to change nothing yet.
Before any vendor call
Write one page: your two disputed KPIs, two real workloads and exit criteria. That page is your evaluation brief.
A sensible next step
If you want a second opinion on your shortlist, book a discovery call with the Logesys data engineering team. Bring that page and your current cost baseline, and we will use the call to test your shortlist against a real workload and your two disputed KPIs, including whether a platform change, a better-run current platform or no change is the sounder decision.
8FAQ
How do I choose a cloud data platform?
Start from the decisions the platform must improve, not the feature list. Define who owns key metrics, what freshness each use case needs, how access will be enforced and what the workload costs under each vendor’s pricing model. Then test your shortlist on your own data, including the messiest source.
What is the difference between a cloud data warehouse, a data lake and a lakehouse?
A warehouse is optimised for structured, governed analytics. A lake stores data in open formats, including unstructured data. A lakehouse aims to support both on one foundation. Which fits depends on your workloads, skills and governance needs; see our architecture comparison.
How should I estimate the total cost of ownership?
Model your actual workload under each vendor’s billing mechanism, then add people, migration, parallel running, data movement and governance effort. A figure that includes only storage and compute understates the real cost.
Does one platform reduce complexity?
Sometimes, but consolidation can also relocate complexity and concentrate operational risk. Judge it by whether definitions, access and ownership become simpler to enforce, not by the number of tools removed.
How do I avoid being locked in?
You cannot remove lock-in, but you can choose it deliberately. Check open table formats, catalog portability, export paths and contract terms before signing, and document any critical logic that depends on a proprietary feature.
A second opinion on your shortlist
Bring your evaluation brief and your current cost baseline. We will test your shortlist against a real workload and your two disputed KPIs.
Or write to connect@logesys.com