AWS re:Invent 2015 | (BDT323) Amazon EBS & Cassandra: 1 Million Writes Per Second on 60 Nodes
Why Stack Overflow recommends it
Stack Overflow users link it when answering whether EBS is a viable storage choice for Cassandra on AWS, citing it as evidence that modern EBS performs well enough.
Recommended by 2 people across 2 different months, 2015 to 2016.
What it is
A 42-minute AWS re:Invent 2015 session on running Apache Cassandra on Amazon EBS GP2 volumes instead of ephemeral instance storage. It walks through a multi-petabyte production case, covering instance choices, configuration and low-level tuning that reached 1.3 million writes per second at replication factor 3 on 60 nodes.
What people said
Recommends Milen Kovachev · 2016-05 · Stack Overflow · on “Recommended AWS storage type for Cassandra?”
I came upon this presentation, which clearly answers the question with a very interesting use case: [this video] To summarize: EBS has changed a lot since 2011 when major companies like Netflix had problems with it. EBS and GP2 are now the recommended storage for Cassandra and you should not expect any bottlenecks there. Datastax have recently updated their documentation to also recommend EBS: [LINK]
Recommends piotrwest · 2016-06 · Stack Overflow · on “Datastax Cassandra - Amazon EC2 instance - Cluster with three node spanning across Amazon region”
It's not true that you get Datastax AMI only with EC2 ephemeral storage. Starting from version 2.5 they claim you can choose EBS as well: Introducing the DataStax Auto-Clustering AMI 2.5 [LINK] . That's an relatively easy way of getting started which I've personally chosen. Should you choose EBS or EC2 ephemeral storage? The answer is: it depends... The past (~2012-2013): EC2 instances with ephemeral storage were a better choice. There were detailed performance benchmarks over the years which indicated that EBS is getting better, but still, attached physical drives were better. The past (~2014): EC2 choice is still better. Datastax wrote a nice post about pricing, network and failure resilie
Cites Jeff Jirsa · 2015-12 · Stack Overflow · on “Large Write Performance Questions”
250 nodes with 6TB of disk (4TB in use, 2TB for compaction/etc), 8-16 cores, 48GB of RAM. We (my employer, Crowdstrike) gave a talk about this at AWS Reinvent - [this video] . Same concept.