The most surprising finding to me (as already said in the article) is that even at BB's scale, the sample size (and time period) for lot of data categories are

The most surprising finding to me (as already said in the article) is that even at BB's scale, the sample size (and time period) for lot of data categories are simply not statistically large enough to conclude much, at least not with this quarterly data. You can see it clearly when the failure rate fluctuates heavily QoQ and has lots of outliers when count of units is (relatively) low. You still can get some trends, but for instance if you want to find say the "best" model, this probably isn't too helpful. I guess it just shows that the failure of an HDD is a very random thing.