It starts really early on. When I first used MongoDB I thought to myself "what a fresh breath, I can finally ignore normalization!" The documents I inserted would contain all the information I needed. One retrieval and I got what I needed. Wow!! But then I started doing sorting, searching and I had to do most of the work on the client side (my backend). At that point, I found myself in trouble because in my other tables I also dup my information. Data has to be updated in multiple places. So I thought "let's take out the dup data." Then I found myself not knowing how to structure my document data anymore... back to some sort of normalization. At the time the searching and ranking in MongoDB were also poor, so I was forced to doing the entire thing on the client side regardless.

I went back to PostgreSQL since. I probably needed a good one-on-one expert training with MongoDB, but I just found myself happier with RDBMS. You don't have to be strict normalization in RDBMS, just enough to make sense for your use cases.

One thing I really did like about MongoDB back then was storing blob (files). It was the best solution available without setting up S3. At the time there was a limitation with 4GB (??) but MongoDB worked for my use case anyway. That being said, please don't store files in any databases today. Use DB to store references to a real object storage like S3. When the DB crashed, you better hope no corruption.