If its mostly on disk, you're paying ms for block io; get more iops. If the index used are sensitive to low key diversity, fix that, and maybe shard. You migh

If its mostly on disk, you're paying ms for block io; get more iops. If the index used are sensitive to low key diversity, fix that, and maybe shard. You might prefer a vm type with nvme. If u can divide the data into more small ones, more indexes and cache fits in ram. Also, cluster data by what works, time, region, userclass, etc. Only decompress the blobs at point of use. Design your app to need only small fast data which cant be cached. Etc