I'm not seeing the delays you talk of for MSP-SFO. Seven-day average shows all but one flight averaging only about 15-minute delays. Do you have contradictory data you can provide a link to?

In general, flow sets an arrival rate at an airport. Say there's bad weather at SFO and the arrival rate is set at 30. Then flow sets the scope of the delay program, which for SFO is typically 1000 nm, 1600 nm or all continental US and Canada. Flow looks at how many flights and what the avg delay is on a given scope and selects what looks to make the most sense. For example, if 1000nm scope is going to effect 40 flights with avg of 4-hour delays and a 1600nm scope will give 120 planes an avg of 70-minute delay, flow will go with the latter.

The decision is a collaborative one made on a telcon with the users included. Once the decision is made airlines are free to swap assigned delays. The algorithm for assigning individual delays is extremely complex, taking into account previous week's data, aircraft type, etc. so two ORD-SFO flights scheduled to leave at the same time will likely have different delays.

The airlines' processes for selecting delayed flights takes much into account, including how many passengers would miss an intl cnx, size of a/c (it costs much more to have a heavy sitting on the ground than a regional jet), and crew and a/c scheduling. So if there's a United Express flight going CEC-SFO and flow gives it a 20-minute delay, you can bet UA's going to take the short delay and give it to a 747 while the other, with low-paid crew and few passengers, sits on the ground.