If you want to perform large scale imports. Then it’s probably best to use a staging table without any defaults, layouts, contexts or other logic. Just a plain full insert. 

The import performs an insert for each row separately (much like a cursor). This means that for each row also the default and context are triggered. I did a simple test. 100 rows = 1 second without logic. With a default that does nothing the runtime goes up to 8 seconds. You can imagine what this does to an import that has 8000 rows with a default that actual logic. 

I see a couple of opportunities to increase performance on this process, I will discuss this with the development team. 

So my advice: create staging table without logic. 
Run a process to check the data for the default/reference etc. 
Create a task to import the succesfull rows. And make some logging for the rows that cause an error. 

Thanks for your answer Erwin. I will see if I can do that.