When to use Dataflow vs Dataset?

[PowerrrBrrr]

Hi, So our organization has been using Postgre connector as source for building reports and have been working without any issue. Recently, we had issues with some of the users where data size has been large and have been using good amount of power queries to do the transformation and it takes lot of time to load data each time in power Bi service. We see dataflows do provide an answer to our problem and we are able to handle most of our use case with dataflows. But then we have the same thing with Power Bi datasets where user can create a dataset and use it to build the report and have different timestamp to refresh dataset and Power Bi report as we do now with dataflow. So my question is why dataflow and on what scenario will dataflow will actually be useful with Power Bi rather than power bi datasets? I did some research and agree tha with dataflows the data is hosted over a Azure instance and that is well integrated with PowerBi then data being hosted some where else. but is that the only reason why we should go for dataflow?
Would really like this to be answered by some expert who has been using both services

[Burningsuit · Accepted solution]

HI @PowerrrBrrr 

I don't think I could be considered an "expert" but I am using both Datasets and Dataflows, and as I'm teaching this topic today, I can share these slides I use ....

Why Use Dataflows ?

• Not a replacement for a data warehouse, but useful when:
  –There is no data warehouse in your organisation
  –The data warehouse does not contain the data you need

• Reduces overall data refresh time:
  –Extracting once and re-using multiple times means you only pay the performance price for the initial slow extract once
  –Reading data from a dataflow is fast, probably much faster than extracting data from the original source

• Reduces load on/number of calls to source system
  –Eg when refresh could affect the performance of a line-of-business database
  –Eg when there is a limit on the number of calls to an API

• More consistency between datasets – less chance that different users will make different decisions when preparing data
• Share complex M queries that some users would not be able to write
• Share tables that have no source, eg Date dimensions generated in M

These come from the excellent Matthew Roche's blog Dataflows – BI Polar (ssbipolar.com) Matthew is pretty much the expert on Dataflows and his blog is well worth a read.

Hope this helps

Stuart