Refresh duration of a semantic model plugged on a dataflow

[BrunoM]

Dear all,

I would like to know if this situation is normal :

i am working in a premium capacity.

I load in a semantic model 2 tables from a dataflow Gen 1. 

In the semantic model, there is no transformation in Power Query (the purpose is only to create visuals) no calculated Table / Columns and no Measures.

The size of the semantic model is about 1,6 Go. The 2 tables loaded have each 20 million records.

The refresh online of the semantic model take 1 hour. My feeling is that it is very long has there are absolutly no transformation, only the transfer of the data from the dataflow.

In advance thank you for your comments.

[anilgavhane · Accepted solution]

@BrunoM  Why the Refresh Might Be Slow

Here are some key considerations:

1. Dataflow Gen 1 Performance

• Gen 1 dataflows store data in CSV format on ADLS Gen2, which is less optimized for high-speed ingestion compared to Gen 2 (which supports Lakehouse, Warehouse, etc.).
• Reading large CSVs (20M rows per table) can be I/O intensive, especially if partitioning isn't used.

2. Semantic Model Import Mode

• Even without transformations, Power BI still loads data into VertiPaq memory engine, compresses it, and builds internal structures.
• A 1.6 GB model with 40M rows total can take time to process, especially if the dataflow refresh and semantic model refresh are not well synchronized.

3. No Incremental Refresh

• If you're doing a full refresh every time, Power BI re-ingests all 40M rows, which is costly.
• Consider enabling incremental refresh if your data supports it (e.g., has a date column).

4. Network and Capacity Load

• Even in Premium, performance can vary based on concurrent workloads, network latency, and backend throttling.
• Check the refresh history for signs of bottlenecks or retry attempts.
  
  

What You Can Do

• Switch to Dataflow Gen 2: It’s more performant and integrates better with Fabric’s Lakehouse architecture.
• Enable Incremental Refresh: This can drastically reduce refresh time if your data is time-partitioned.
• Use Data Pipelines: For better control over refresh sequencing and dependencies.
• Monitor Refresh Logs: Use refresh history and CSV logs to pinpoint slow phases (e.g., data read vs. model load).

[BrunoM]

Thank you @anilgavhane 
I will try with a dataflow Gen 2 and also Pipeline (i do not know anything on this techno)