This paper defines the "triangular dilemma of set size estimation" (the trade-off between generality, accuracy, and updatability) that hinders the practical application of learning-based cardinality estimation. To address this issue, we present DistJoin, an efficient, distribution-based join set size estimator utilizing a multiple autoregressive model. DistJoin estimates the join set size by separating the probability distributions of individual tables and develops a high-throughput distribution estimation model, Adaptive Neural Predicate Modulation (ANPM), to ensure efficiency. We formally address the variance accumulation problem of existing similar approaches through variance analysis and effectively reduce variance through a selectivity-based approach. DistJoin is the first data-driven method to support both equi- and non-equi-joins, demonstrating high accuracy, robust data updates, generality, and fast, flexible update and inference speeds. Experimental results show that DistJoin achieves the highest accuracy, robustness, and generality compared to existing methods, while delivering comparable speed.
