I am not experienced enough in machine learning to say when it's not, but I can say it's extremely useful for quickly building classifiers of stuff that would be a nightmare to label by hand for training purposes.
I do a lot of text classification where I'm working with a set of semi-structured sample documents. I can usually hard-code rules to classify my sample set, but I know that any new sample document might well break my rules. Typically, I know a handful of things about what distinguishes some documents from others-- some I know will always be true but many I only know will sometimes be true. The naive way to handle this tends to be a score-based approach, where documents get "points" for meeting certain criteria. But unless it turns out that stuff you thought was sometimes true is always true and/or the things that are always true are enough to classify correctly 100% of the time, the naive approach is always going to regress (no pun intended) to a model because you have to identify a cutoff for your scoring. You can take statistical approaches to figuring out the best cutoff to achieve the result you want (e.g. a Principal Component Analysis optimized for a certain chi-square distribution), but it's still turning into a modeling task no matter what you wanna call it.
It's great to be able to dump all of my roughly understood features into Snorkel and let it build a model around them. It's also super fast because I've usually already written the hard-coded "rules", so I have very little new code to write.
Usually the classifier ends up being a supplement to what I'm already doing but in certain cases it could become an outright replacement.
We can only do classification (for now), and we need users to be able to provide enough labeling functions that are correlated well with the true labels. Text and some video tasks are examples of where it can work well, since it's possible to write heuristics over the primitives (words for text, presence of an off-the-shelf object detector firing).
There are other tasks where the right decision might be to go get a bunch of human-labeled examples -- an example I like to think of is how you would describe a chair, if you didn't have any object detectors lying around. That's a situation where a) it's hard for a human to describe what you're looking for, and b) you would potentially get more bang for your buck by just having an army of Turkers label a bunch of chairs for you, since everyone knows what a chair is.
If I read the claims correctly, it takes an unlabeled dataset + a classifier model + a handful of heuristic labeling functions, which may be both partial and inaccurate, and trains the model to be a good classifier for the dataset, without ever requiring per-example manual labeling.
