I'm running batch 5311 and they're about 45 - 50 minutes long.

Batch 3518 are about the same, less than an hour.

I woke up this morning to very short MCM WUs also (5235, 5334, 5337). They seem OK since they're finishing as valid.

CJSL

Gotta keep crunching... there's a world to save !!!

Well, one reason you may have had to wait is the feeder was probably getting pounded with the short units. Maybe more demand than capacity to deliver.
I would not blame either WCG or BOINC for inability to adequately predict WU length. that should be placed with the researchers in my opinion. When these very short units first appeared the researchers stated something to the effect they were looking at a more efficient
algorithm and would need to work on how exactly to package it. i would agree with you that with an intermittent connection these short units will cause you fits.
Cheers

Reference kremmen's credit per unit comment, on june 14 the units were shortest at 1.26 hours average fetching 21.33 average per hour, then it rose to 4.04 hours p/u few days ago fetching 31.73 p/h at peak and now we're back at 1.37 hrs p/u and 23.74 credit p/h. This credit system so draws vacuum. Any brainy person can explain why a near 50%+ difference when doing the longer units? It just should not matter when doing a 10 minute task or a 10 hour task. A node should get a constant per unit of time, second/hour/day. Set that condition and the subject will semi permanently go away, now sucking energy on a near constant basis from forum posters and readers. An endless distraction if not annoyance.

Actually, for clarity, those numbers were computed using the mcm global daily numbers. Individual devices see roller-coaster effect, one on the rise, the other going down, the question being how many carts are doing the same up/down at the same time. Conversely, somehow fa@h runtimes have stabilized last few weeks now at about 3 hours. Credit has been fairly constant at about 24, but you see the same idiocy, average times extend and the credit/hour goes up, but there is always that lowest common denominator gravitation. Let it sit at constant time and credit will creep down. Once the statistics for a device have stabilized and then a change is made extending or shortening runtime and the same happens again. Whoever build this method, they might want to check up the 'autocorrellation' effects in statistics, which this program is apparently drawing on.