YouTube's Database “Procella”

I find lots of commonalities with other systems aimed at processing big datasets. The only thing I find different is their management of the metadata at a centralized place, which can bring lots of network IO savings since you don't have to query the stripe stats (like in ORC) in order to enable stripe skipping.

The risk I see with that is that you have to be extra careful when processing updates to the metastore, as sometimes there are competing processes writing metadata about the same partition and if not done properly you'll end up with a corruption that might cause problems (ie. you skip a whole file because the metadata was incorrect, causing a correctness problem down the road).