I was facing similar issue; it turned out the issue was due to bot memory storage (IStorage) not being distributed across the scaled-out instances.