Home › Data engineering › Metric views › 3 · Limits
Data engineering · BI and analytics · free guide · 3 · Limits
Limits of each, seen from eight seats
Every product has limits. What matters is which seat you sit in. A table that is fine for an author may be painful for a security reviewer. Below are the three products seen from eight seats, then the hard limits of metric views, then verdicts for common setups.
Eight seats
Limits seen from each seat
| Seat | Metric view | dbt Semantic Layer | Power BI semantic model |
|---|---|---|---|
| Author (who writes it) | YAML plus SQL expressions; ALTER VIEW replaces the whole definition; only the owner edits; window measures need care with ordering and gaps. | YAML plus Jinja filters; entity and time-spine concepts to learn; spec changed in 2026 (old and new formats in the wild). | DAX is powerful but hard to review; models are binary files unless you use PBIP/TMDL folders. |
| Business user | Sees friendly names, formats and synonyms in dashboards and Genie; needs a Databricks login. | Reaches metrics only through a connected tool (Sheets, Excel, Tableau and others on the integrations list). | Best experience: drag and drop in a tool they know; Excel via Analyze in Excel. |
| BI tool outside the platform | Tools must send MEASURE() in custom SQL or read wrapper views; publish-to-Power-BI is Private Preview; partner native support is arriving. | Partners integrate through the Semantic Layer APIs; the list is finite. | Outside Microsoft tools, reach is through XMLA, which few non-Microsoft tools use well. |
| AI and agents | Genie reads metric views natively; agent metadata (synonyms, display names) helps; agents call them through SQL tools or Genie API. | dbt MCP server gives agents governed metrics (paid platform). | Copilot in Power BI uses it; agents outside Microsoft reach it with effort. |
| Security | Unity Catalog grants, row filters and masks on source tables apply as the user; no materialization while those filters exist. | No RLS of its own; depends on the warehouse identity; a shared service account sees everything. | RLS and OLS roles in the model; Import copies data out, so rules live twice if Databricks has them too. |
| Performance at scale | Compiles to SQL on gold; speed comes from table layout, the warehouse and optional materialization; no query-time joins with other tables (use a CTE). | Adds a hop (API → compile → warehouse); declarative caching and exports help. | Import is very fast until model size, refresh time and capacity memory bite; DirectQuery depends on the source. |
| Cost | No licence; you pay warehouse compute per query and pipeline compute for materialization. | MetricFlow free; serving needs a dbt platform plan; plus warehouse compute. | Per-user licences or Fabric capacity; refresh compute; Import duplicates storage. |
| Portability and change | Databricks-only object (sharing across metastores is Beta); YAML features depend on runtime version; Apache Ossie may help later. | Most portable: same YAML runs on several warehouses. | Locked to Power BI/Fabric; TMDL and Git integration make change control possible. |
Metric views
Hard limits of metric views today, and what they mean
| Limit (from the Databricks docs, Nov 2026) | What it means for you |
|---|---|
| No SELECT *; every measure needs MEASURE() (or AGG on runtime 18.1+) | BI tools that generate their own SQL can't use metric views without custom SQL or a compatibility mode. |
| No joins with other tables at query time | Wrap the metric view query in a CTE, then join the result. |
| Materialization refused when the view or its sources use row filters, column masks or ABAC policies | Secured data must be sped up some other way (layout, your own secured aggregate tables). |
| Parameterized metric views can't be materialized | Pick: flexibility or pre-computation. |
| at_most_one_match is a promise, not a check | If the join is not truly many-to-one, sums are silently wrong. |
| Relationships between metric views are Beta, metadata only, not enforced, SQL only | Tools choose join paths; treat them as hints. |
| ALTER VIEW replaces the whole YAML; DROP is permanent; owner-only edits | Treat the YAML file in git as the source; deploy, don't hand-edit. |
| Wildcards expand at create time | New columns in the source don't appear until you recreate the view. |
| Features are gated by runtime and YAML 1.1 | A warehouse or cluster on an older runtime may reject a view that works elsewhere. |
| No data profiling on metric views | Profile the gold tables underneath instead. |
Choosing
Common setups with a verdict
| Situation | Setup | Verdict |
|---|---|---|
| Data on Databricks; dashboards, Genie and agents on Databricks | Metric views as the only home for definitions | Pursue |
| Same, plus Power BI for formal reports | Metric views, plus a thin Power BI model whose measures match them, with a number test | Pursue |
| dbt builds gold; several warehouses or tools outside Databricks | dbt Semantic Layer as the home; Databricks reads gold as normal tables | Pursue |
| Power BI is the only consumer, data fits in Import, no AI on top | Power BI semantic model alone | Pursue (small scope) |
| Definitions written in both metric views and DAX by hand | Two homes that drift | Drop |
| Metric views and dbt metrics both defined for the same numbers | Two compilers, two truths | Drop |
| No semantic layer; each dashboard and Genie writes its own SQL on silver | Fastest to start, slowest to trust | Drop |
← 2 · Three products · Next: 4 · Parts →
FAQ
Questions people ask
What is the biggest limit of metric views?
Reach outside Databricks: tools must send MEASURE() through custom SQL until their connectors support metric views natively. Inside Databricks, the biggest design limit is that materialization is refused when source tables use row filters or masks.
What is the biggest limit of the dbt Semantic Layer?
It has no access rules of its own and its serving layer (APIs, caching, MCP) is a paid dbt platform feature. The open-source MetricFlow only compiles SQL.
What is the biggest limit of a Power BI semantic model?
Its logic is locked inside Power BI and Fabric. Genie, agents and non-Microsoft tools can't easily reuse DAX, and Import mode copies data, so security rules must be applied twice.
Sources
Checked against (11 Nov 2026)
- Databricks: Unity Catalog metric views and the YAML reference
- Materialization · Advanced techniques · Relationships · BI tools · Feature availability
- Row filters and column masks · ABAC evaluation through views (KB)
- dbt semantic models · MetricFlow open-sourced · Apache Ossie
Metric views change every few months. Confirm runtime minimums and preview status before a design depends on them.