I had one truly good math class during my secondary education, and the way the teacher conducted things sounds a lot like this. Small groups, sitting around a table, talking about the subject. The overall class size was much larger, of course. So what we had was a class subdivided into smaller groups, with the teacher circulating among the tables, spending maybe 10 minutes at a time at each one.
I suspect the real problem with this approach, at least from an American perspective, is that it's not easy enough to instrument for data collection. You can't measure "mastery", so instead you measure performance on multiple-choice questions. And then you tie funding to goals related to those tests. . . at which point, no matter what you _want_ the teacher's job to be, what they're really being paid to do is train kids to regurgitate information on multiple-choice tests.
I'm a professional data scientist, and, ironically, being one has resulted in me becoming a deep skeptic of the movement toward data-driven everything. "Data" is understood to be quantitative data, and not everything can be studied quantitatively, so it leads to people habitually deciding, intentionally or not, for better or for worse, to re-frame all their activities in ways that make them easier to quantify and micro-quantify.
To take another example, there's a whole lot of well-established research out there indicating what the best way to pick up a second language is. And you'll never see any of this knowledge being applied to language learning classes in American public schools, because its implications about how we should teach second languages are almost universally incompatible with the teachers' mandate to always be quantifying.