Actually, we started out using mongo to focus on the analysis instead of the schema, but we quickly ran into performance issues as the datasets grew. We were simply using mongoengine for Python, so we didn't spend a significant amount of time trying to optimize our schema or implement things like sharding.

Our performance issues with mongo largely stemmed from our poor use of indexes - we defined a lot of indexes because how we needed to query was a very organic and undefined process as we got new analysis requirements. Because we would have to frequently go back and compute new feature vectors across the whole (or large parts of) the dataset, we weren't able to implement a lot of the aggregation capabilities you'll see implemented in many other time series schemas.