Five years in, the modern data stack works — for the companies that committed to it. The half-measures left a graveyard.
Around 2020 a category called the "modern data stack" coalesced. Fivetran for ingestion, Snowflake for storage, dbt for transformation, Looker or Mode for visualization, and Census or Hightouch for reverse ETL. Five years later, the stack has matured but the lessons are not what the original pitch promised.
What worked
The decoupling of ingestion from transformation from visualization. Each layer is now best-in-class and independently replaceable. Teams that adopted the stack have fundamentally cleaner pipelines than the previous generation.
What disappointed
Cost. The combined Fivetran + Snowflake + Looker bill for a mid-size company is shockingly high. Teams have learned to push back on every layer's pricing model. Snowflake credit consumption optimization is its own job function now.
What changed
- ClickHouse arrived as the warehouse of choice for sub-second analytics on event data. Cheaper than Snowflake by an order of magnitude for the right workloads.
- DuckDB made local analytical queries fast enough that you do not always need a warehouse.
- The lakehouse pattern (Iceberg + Delta) is replacing pure warehouse architectures for large enterprises.
- Reverse ETL went from category to feature — both Fivetran and Stitch now ship it built-in.
The current best stack (2026)
- Ingestion: Airbyte (open source) or Fivetran (managed) for SaaS sources, custom CDC for high-volume operational data.
- Storage: Snowflake or BigQuery for general analytics, ClickHouse for event analytics, Iceberg on S3 for the lakehouse pattern.
- Transformation: dbt remains the standard.
- Orchestration: Dagster has overtaken Airflow for new projects.
- Visualization: Metabase for self-serve, Mode or Hex for analyst-driven reports, Looker if you have the budget.
- Reverse ETL: Census or Hightouch.
The philosophical lesson
The modern data stack is a kit, not a solution. Companies that bought the whole kit and treated it like a turnkey product have been disappointed. Companies that picked the layers they actually needed and skipped the rest are happy.