It was definitely MinIO related, I probably should have made that clearer. We noticed that with zero fault tolerance, MinIO objects would randomly become corrupted, which MinIO would present as "you're making too many requests, please slow down". We were certainly not making too many requests.
We're looking at alternatives, I've made some previous comments on that front. Sadly MinIO was the only option with sufficient performance for this particular situation. Thankfully we're not using any MinIO-specific features, so at least the migration path away is clear.
Did you figure out what caused corruption? Was minio losing blocks or was juicefs corrupted even though minio was consistent?