21:40. He mentions in the video that they had a one-producer, one-consumer input queue that they originally implemented with clojure.lang.PersistentQueue. They engaged Cognitect for some code review/consulting, and Cognitect's advice was to use a regular Java ConcurrentLinkedQueue for this instead.
I think this particular example supports this "Cognitect snake oil" theory very poorly. I don't think PersistentQueue can be properly considered a first-class data structure in Clojure, as it doesn't even have a constructor function in clojure.core. The Clojure documentation, and the other writings of Rich Hickey and Cognitect don't encourage people to use this data structure. It exists, I think, for completeness and because there are some situations where it's appropriate, but unlike the core Clojure persistent data structures - maps, vectors, seqs, etc. - it's not designed as a general abstraction that covers the vast majority of general use cases.
Clojure is meant to be a practical language. It's hosted on the JVM and other platforms, and designed to make it is easy to use what those platforms provide. Where it's sensible to do so, Cognitect and other Clojure developers would certainly advocate using Java libraries to solve problems.
The other thing mentioned in the GP post was the idea of not using STM. This falls under the same umbrella. There are problems for which STM is the best solution, but there's a big space where something simpler is the right answer. In my observation, overuse of Clojure's STM is far more often the result of a developer impulse of "wow this is a cool looking feature, I want to try it" than any advocacy by Rich or Cognitect.
I also have some firsthand knowledge here, having been involved in a Cognitect engagement. I have some concerns about Cognitect as a result, but certainly nothing like this suggestion that Clojure is a real manifestation of the old fake Bjarne Stroustrup interview. [1]
