On Instability of Minimax Optimal Optimism-Based Bandit Algorithms