https://github.com/juji-io/datalevin/blob/master/doc/query.md#direct-counting-for-cardinality-estimation-new > “the counting is capped by the current minimum, so the time spent in counting is minimized” This is brilliant. So simple. > “We found that the opportunities lie precisely in the “Achilles Heel” of RDBMS optimizer: cardinality estimation [6]. It is hard to have good cardinality estimation in RDBMS because the data are stored in rows, so it becomes rather expensive and complicated trying to unpack them to get attribute value counts or to sample by rows [4]. On the other hand, it is cheap and straightforward to count elements directly in the already unpacked indices of triple stores.” Thanks for writing documentation like this, it’s very insightful.