Cassandra and Consistency - A Simplified Explaination

Cassandra is architected for read-write (RW) anywhere, enabling any client to connect to any node in any datacenter (DC). The concept of a primary or single master node does not exist in Cassandra. This allows for drastically increased RW times as master node is not taking all the traffic. The flip side is that there is not a single source of truth (SSOT) as you would see in a RDBMS such as MySQL where you assume the master or primary is always correct. Cassandra has abilities to manage this though.

REPLICATION FACTOR (RF): This determines how many copies of data exists.
CONSISTENCY LEVEL (CL): This determines how many nodes must acknowledge a read or write before an acknowledgement of a write commit or the data is returned. Examples would be ALL, ONE, QUORUM, etc. Effectively a level of assurance that a write occurred or your read is against the freshest data.

Let’s look at two nodes for example.

WRITES: If the data is simply being split between the two nodes evenly and goes down, you have effectively lost half your data. To overcome this for writes, you would need to set the CL to ALL (or TWO) to ensure data was written to both nodes, or set it to ONE with an RF of TWO to ensure data was copied to the other node. At first this seems like plausible strategy, but there are a couple issues. If the CL is set to the total number of nodes or ALL and one goes down, you do not meet the CL to do an operation and it will fail, which leaves us with only the option of doing a CL of ONE with an RL of 2, which works for writes.

READS: The problem with the CL=1 and RL=2 strategy comes from the reads. With a CL of 1, when an object is updated it will replicate, but a read on that data may occur before the replication happens. If the update occurs on NODE A and the client checks NODE B, how does it know the data is fresh or if it is waiting on replication? The only way is to compare the timestamp to another node to ensure consistency. Once again the issue then becomes is you have a CL of ALL or TWO, a single node outage prohibits this analysis, and disables reads.

So effectively with just two nodes, if a single node goes down due to issue or maintenance, you either sacrifice redundancy or consistency.

QUORUM: A majority of nodes.

CL does not just support ALL or a hard number set, it also supports a QUORUM. This can be a local or all-encompassing majority when you look at regionally separated, fully redundant DCs for example. With a three node cluster, the majority, or quorum, would be two. This then allows for assured redundancy and consistency, even in the event of a single node outage.


As to why three, because it is the minimum number to ensure redundancy, consistency, and the capability to do a quorum. With Cassandra the more nodes the better however. If you have enough resources to stand up 5 Cassandra nodes in a DC, there is no downside that I know of. 

Comments

Post a Comment

Popular posts from this blog

API Security - The Next Generation with Elastic Beam

Image
The Prime Directive - Open more technology doors without increasing risk. I can distinctly recall being fresh in the Federally focused aspect of my career and having a conversation with a customer where he said to me "We don't measure failure in dollars, but in the loss of lives." It was a powerful statement that has stuck with me ever since that speaks to the sensitivity of the data and the critical nature of US Federal systems. The challenge becomes that this creates a situation where disruption can be far from ideal, and risk-adverse strategies with a strong focus on security are often paramount.  And yet, recent times have brought an explosion of new technologies and capabilities with tactical and practical value that cannot be ignored. Servicemen and women look to utilize their own devices to accessing PII focused data and services, or basic applications such as time sheets, without all the hassle of going through CAC-secured desktop systems. Officers in command wa...
Read more

Firewall, IDS, IDP, WAF, API Gateway: Choose Your Shield

To long a read? Feel free to jump to the bottom for the exciting conclusion! Often in US Federal we see different requirements and use cases arise than our commercial counterparts. Quite often I might get asked to provide insight into security, identity, and compliance queries in the same way I might ask for insight on team development or CICD requirements from my commercial focused brethren. It just so happens that this week I have provided some insights on three of four discussions around the role of our API Gateway in enterprise security. The role of an API Gateway in enterprise architecture in general always seems to bring up interesting discussion points (can it replace my load balancer?), but none more so than enterprise security. I have often heard of API Gateways jokingly referred to as a 'Swiss army knife' because you can really do so much with them. Often I find myself saying to both customers and colleagues that 'It's not what you can do with it, it is what...
Read more

REST API Best Practices: HTTP Status Codes and You, Part 2(xx) - Status 200; Success!

Ah, the HTTP Status Code 200 series. Request sent by client, received by server, understood and processed, and a response has been sent. He shoots he scores, GOALLLLLLLL! The status code 200 series is the "Pass GO, Collect 200 Dollars" of HTTP Responses, this is the result for which we strive. In this part of the series, I will cover some of the major HTTP 200 series response codes and how they can be used as part of an effective REST API strategy. Do note, that I will not cover them all here, but I want to touch on some of the ones you may already know, and some you may not know, but might easily start to use in API development process. So with that, let's start at the very beginning, a very good place to start! HTTP Status Code 200 - OK What it means: Success! Everything worked out. What content to expect returned: This largely depends on your request verb. For example, a GET request might return the requested resource, such as a JSON formatted response. A P...
Read more