Composite-model performance issue

[Kraftfood]

Hello, I am encountering a performance issue due to (thank GPT) composite model.

I have a PBI pro license and the workspace I use has a Premium Per Capacity license

Situation: I have downloaded a set of tables from an SQL database, structured them into a nice semantic model and built additional measures on top, the storage more is 'Import'. I published this new semantic model into PBI service.

I have built several new reports using this online PowerBI Semantic model, the storage mode is then 'DirectQuery'.

So far so good.

BUT, when I linked a new report built from this online semantic model, to a local excel, and linked both bases, oef, the performance...

I've discussed a lot with GPT and searched through this forum but so far I couldn't find a solution to my issue.

From my searches, I am facing an issue that does not have a perfect/quick fix. I will need to re-think my semantic model, do I build it in Power BI and then re-use it for other reports, or do I built 1 semantic model for each report that needs additional data, etc.

I'm sure others faced this issue and I am keen to hear about their solutions!

Additional information:

1. I extract the original data from an SQL database. This database (we call it our 'datamart'), extracts data from our system via API.
2. I am planning on sharing the semantic model I built with other users, filtering their view via RLS
3. The other users which will have access to my semantic model will, 100% sure of it, link it with local excels
4. This is only with 1 semantic model, what if I create several beautiful semantic models (for example 1 for IT, 1 for Finance, 1 for HR, etc). Then I face the same issue when linking them on a new Power BI desktop.
5. Most important feedbacks from GPT to help you understand my current issue: I have:

   1. 1. Excel tables = Import mode

      2. Semantic model = DirectQuery to AS / Fabric dataset

      3. You relate them (e.g., Date, ID, Case, Mapping)

      Result:

      ➡️ Every visual triggers a DAX query that requires joining IMPORT + DIRECTQUERY
      ➡️ Power BI must compute filters in the AS model and in the local model
      ➡️ Fabric returns a large intermediate result
      ➡️ Power BI filters it locally
      ➡️ The UI becomes slow, sometimes unusable

      This is inherent to composite modeling.

[Kraftfood]

Alright, not the answer I wanted to see but the answer I expected to see nonetheless. Thank you!

[rohit1991 · Accepted solution]

Hii @Kraftfood 

Joining a DirectQuery-to-AS semantic model with local Excel tables forces Power BI to run cross-engine queries, create large intermediate results, and compute filters in both places there is no fast path. The only real fix is to avoid the local Excel tables and keep everything inside one semantic model.