I highly recommend, 'The theory that would not die' as a highly entertaining account of the history of Bayesian statistics.
It is not a technical text on the subject, but it provides some of the high level reasoning as well as several of the most important stories of success over many years.
My impression as a student, however, is that among statisticians, Bayesian statistics is viewed as a part of the mainstream way of doing statistics. So it has perfectly okay reputation in that sense. But when it comes to applications, the extent that Bayesian methods are used varies between fields quite much. In machine learning and computational statistics community, it seems to be very popular. On the other hand, ome other fields (esp. the published papers) appear still to be mostly about hypothesis testing and p-values. (For example, last time I checked, the "Introduction to Statistics for Social Scientists" first undergrad level course in our uni was mostly about the traditional frequentist stuff.)
It's freaking rare so they don't have any opinion on it if they don't know it.
If I ever need to model I would do it in Bayesian hierarchical modeling.
But seeing that I'mma do data science many companies don't really care about such intricate modeling especially when it take forever to run MCMC is sooo slow (haven't tried Hamiltonian MC tho).
If it's so mainstream then it wouldn't be so hard to freaking google and find implementation in Stan for chinese buffet or indian buffet.
There were some pretty savage fights between Bayesians and Frequentists over the 20th century (see e.g. The Theory That Would Not Die), but by now there is peace. The main thing that's changed is computers, algorithms, and software have gotten a lot faster. So you can actually do bayesian inference on nontrivial models and datasets.
The math behind Bayesian methods is pretty solid and indisputable. The main controversy is how to specify priors. And you can always use uninformative priors if you have to.
>The math behind Bayesian methods is pretty solid and indisputable. The main controversy is how to specify priors.
But as the Bayesians always point out, specifying which sampling procedures, test statistics, and inferences to run under frequentist statistics also amounts to inserting subjective judgements into your statistical analyses, soooo... you might as well just pick the one that reviewers in your field like, or that works for your problem.
Increasing in recent years. For a long time, one had to use mathematical shortcuts to make the techniques computationally feasible (for example conjugate priors). Now, with better sampling algorithms (like NUTS) you don't have to make those compromises anymore. This means as a practitioner you get all the benefits (like built in, rich error distributions and good small sample properties) without the compromises.
It is not a technical text on the subject, but it provides some of the high level reasoning as well as several of the most important stories of success over many years.