Choosing a Cloud for Data — Azure vs AWS vs GCP
Map equivalent data-platform services across Microsoft Azure, AWS, and Google Cloud, then choose based on workload requirements rather than brand claims.
Prerequisites and learning objective
Before this lesson, understand files and object storage, batch versus streaming, SQL analytics, and the basic purpose of identity and networking. You do not need hands-on experience with all three providers.
By the end, you can translate a data workload into cloud capabilities, identify the closest service families across Azure, AWS, and Google Cloud, and defend a selection using operational constraints.
Think in capabilities before product names
A cloud data platform commonly needs durable storage, ingestion, transformation, orchestration, analytics, governance, identity, and networking. Products overlap but are not interchangeable: a warehouse, a Spark service, and an event broker solve different jobs even when all appear in one vendor's data catalog.
Purpose Azure AWS Google Cloud
Object / lake storage ADLS Gen2 / Blob Amazon S3 Cloud Storage (GCS)
Managed Spark Azure Databricks EMR / Glue Dataproc / Databricks
ETL / data integration Data Factory Glue / Step Functions Dataflow / Composer
Warehouse / SQL Synapse / Fabric Redshift / Athena BigQuery
Event ingestion Event Hubs Kinesis Pub/Sub
Secrets Key Vault Secrets Manager Secret Manager
Identity Entra ID + Azure RBAC IAM Cloud IAMTranslate the workload into a shortlist
Suppose a delivery marketplace receives order files each day and payment-status events throughout the day. A sensible design identifies the required capabilities first: land immutable raw data, process events, run batch transforms, expose governed analytics, and let an operations team observe failures.
# Translate a workload into capabilities before selecting products.
workload = {
"source": "daily order files and payment events",
"storage": ["durable object storage", "retention controls", "open formats"],
"processing": ["scheduled batch", "incremental streaming", "Spark"],
"serving": ["SQL analytics", "BI access", "data sharing"],
"controls": ["private connectivity", "audit trail", "regional residency"],
}
# Compare provider implementations against these needs, not product names.
for capability, requirements in workload.items():
print(capability, requirements)For each candidate cloud, score the fit of each capability, including people and existing contracts. A team already operating Azure may prefer ADLS, Data Factory, and Databricks because identity and networking are familiar. Another team may favor S3 and Glue, or GCS and BigQuery. The right answer is conditional, not universal.
Compare the full operating model
| Decision area | Questions to answer |
|---|---|
| Workload fit | Does the platform support the needed batch, streaming, Spark, SQL, and sharing patterns? |
| People and ecosystem | Which cloud skills, operational tooling, partner integrations, and hiring pool already exist? |
| Governance and residency | Which regions, data classifications, audit controls, and retention rules apply? |
| Economics | What are storage, compute, network egress, idle capacity, and support costs for this usage pattern? |
| Reliability | What are the recovery objectives, quotas, service dependencies, and failure-handling expectations? |
A practical decision workflow
- Write down data sources, formats, volume shape, freshness, consumers, and recovery needs.
- Mark mandatory constraints: region, identity integration, private connectivity, compliance, and existing agreements.
- Build one small representative slice in each plausible service combination; test failure recovery and access boundaries.
- Compare total operating effort and cost, not only a single service's advertised rate.
- Record why the choice was made and which conditions would trigger a review.
Common mistakes
- Choosing by popularity: it ignores team skills, residency, and integration constraints. Start with requirements and evidence.
- Assuming equivalents are identical: managed Spark, serverless SQL, and workflow services expose different controls. Compare capabilities and limits.
- Comparing list prices only: forgotten egress, orchestration, observability, backups, and support distort the estimate. Model end-to-end usage.
- Designing for a single provider everywhere: portability has a real engineering cost. Keep open formats and interfaces where useful, but do not add multi-cloud complexity without a requirement.
Practice
- Map object storage, event ingestion, a Spark workload, and SQL serving for all three clouds.
- Choose a shortlist for a company already using Microsoft identity and explain what evidence could change your choice.
- List five cost line items beyond compute for a daily batch pipeline.
- Explain one service pair that looks equivalent but differs in how engineers operate it.
Hint: Compare one concrete workload and include access, network paths, scheduling, and monitoring in your diagram.
🎯 Key Takeaways
- ✓Start with workload capabilities and constraints, then map them to provider services.
- ✓Azure, AWS, and Google Cloud offer overlapping storage, processing, orchestration, streaming, and analytics families, not exact copies.
- ✓A useful selection includes people, governance, reliability, and total operating cost.
- ✓Test a representative slice and measure end-to-end behavior before committing.
- ✓Use open formats and clear interfaces where they help; avoid unnecessary multi-cloud complexity.
HELP US IMPROVE
How was this lesson?
Your note is shared with the Chaduvuko team for review. Please don’t include personal or sensitive information.
Spotted a mistake or something unclear? Suggest a fix on GitHub or email us. Select the text first and it is quoted for you.
Discussion
0Have a better approach? Found something outdated? Share it — your knowledge helps everyone learning here.
