I vote AVL!, I have some different perspectives on this issue. From an aesthetic standpoint, AVL trees are more visually appealing, and their balancing principles are more straightforward. Red-black trees, on the other hand, introduce the concept of color, using color changes to replace some rotation operations. However, rotation operations are not significantly more complex than color changes, especially during deletion, where red-black trees also require extensive case discrimination, which detracts from the elegance of the code.

The mainstream view is that red-black trees perform better than AVL trees in terms of insertion and deletion. However, few people have implemented both red-black trees and AVL trees in the same language, with similar styles and approaches, and conducted systematic testing and comparison.

I implemented both red-black trees and AVL trees using numpy and numba, and tested them by sequentially inserting and then deleting 1e7 random float numbers. The results showed that AVL trees had a slight advantage.

For comparison, I also included the STL Set with O3 optimization and Java's TreeMap. The comparison results are as follows:

1e7 random float push/pop one by one

I believe this issue requires more implementations and testing, as factors beyond theoretical computational complexity, such as cache-friendliness, can significantly impact performance. Even when considering theoretical performance, it is essential to record in the code the number of branch selections, node operations, and color changes, rather than making assumptions.

And I had post another question: Does red-black tree really have advantages over avl tree?