Still slow on the scale of things these days. I just provisioned a few servers with over 500,000 IOP/s / 2000MB/s read and write each, 100% SSD with 3-10 year warranties and they use bugger all power. Very low running cost and maintenance overhead and cost less than 8k a unit (1u chassis, redundant power, 32GB RAM, 2x 6 core Xeon v3) and I can guarantee the performance is consistent and there when We need it.<p>I'm all for outsourcing hardware hosting ('cloud') to save costs and to allow for quick provisioning of new instances - but went you need raw power and in cases where it's inefficient to horizontally scale - the latest generation of PCIe NVMe SSDs are really very impressive and in a recent evaluation we performed of our storage - it was actually going to work out significantly cheaper to A) host our high speed storage ourselves and B) buy SSDs and do away with rotational drives.
At OktaWave[1] 20k IOPS with 2GB bandwidth is only the second tier and you can have up to 200k IOPS with 2GB/s bandwidth. Only size of single volume is max 300GB, but you can raid0 them.<p>[1] <a href="https://www.oktawave.com/pricing.html" rel="nofollow">https://www.oktawave.com/pricing.html</a>
Did the IOPS (and better yet, enhanced IOPS) help any big user? DataStax and Eleasticsearch Inc recommend to stripe instance stores (doing raid0), which in the cloud is not the worst idea as long as you have multiple nodes and proper ring designed. That being said, now I lose the ability create snapshot because the volumes are not EBS. From my experience the IOPS isn't very stable.