The problem is that the default/naive mode for an ETL in most ACID SQL implementations is "lock every table you're touching and give zero feedback as to how long it will be locked".
Eventually you end up writing more and more cryptic code to work around the fact that this is not something supported out-of-the-box in any mainstream RDBMS.
Now, should it be better? Would I expect a vendor-supported solution like Microsoft SQL server to provide a simple tool that lets me write a Merge statement with the understanding that I want it to be eventually-consistent rather than expecting me to spend weeks learning yet another tool with it's own DSL to handle this case? Absolutely.
Eventually you end up writing more and more cryptic code to work around the fact that this is not something supported out-of-the-box in any mainstream RDBMS.
Now, should it be better? Would I expect a vendor-supported solution like Microsoft SQL server to provide a simple tool that lets me write a Merge statement with the understanding that I want it to be eventually-consistent rather than expecting me to spend weeks learning yet another tool with it's own DSL to handle this case? Absolutely.