The concept of architecture effecting costs is a key element when you are architecting for Cloud solutions. One of the reasons this is key, is that your ongoing costs have a greater potential to be variable. This is a good thing, one of the benefits of moving to Cloud models is the elastic nature and you want your applications to be right sized and running efficiently both technically but also financially.
There are two considerations when thinking about costs for Cloud.
First there is the selection cost. As an example, if you were using Amazon S3 there are two durability options available, one is lower durability and also lower cost. If you have scratch data or data that can be regenerated you might simply choose to use the less expensive storage.
Second is then understanding your elastic cost, based on your predicted demand. Based on your forecast demand, what will your peak and average expenditure be over time? What size of cheque are you signing up to if you implement your selected solution?
An example of this crossed my mind this last week as a read a blog post from VMware on their vFabric blog. It was entitled "ROBLOX: RabbitMQ, Hybrid Clouds, and 1 Billion Page Views/Month".
Don't be distracted by the 1 billion number. Its an interesting article on how the application architecture for this company needed to be changed as they started to scale out. The example was an AWS customer and in order to speed up their service they introduced a message queue to decouple the synchronous relationship between the web services and the back-end services (as you do). Here is a pertinent bit of text.
To dive deeper, ROBLOX implemented RabbitMQ to help deal with the volume of database requests and slow response times. The queue is managing 15 million requests a day. The example scenario is when there is an update to the ROBLOX catalog content, which needs to update the search index. Before RabbitMQ, the web server would have to wait for an update to complete. Now, the loosely coupled architecture allows the web server to update the message queue, not the database, and move on (technically, the update is enqueued and picked up by the consumers). At some point, the message is picked up by a service that updates their search platform index.Wonderful stuff. But but it hit me, why bother implementing RabbitMQ? It does not sound like they were using a lot of its sophisticated functions, surely they could have used the AWS Simple Queue Service (SQS). Thats when the "architect for cost" mantra kicked in. 15 million requests a day, that does not seam like much but lets see what that would cost for SQS. SQS is $0.01 per 10,000 requests and there are no data charges if the transfer is within a single region. Thats $15 a day which is reasonable. Thats about $465 a month. Thats about the same prices as a full license over a year potentially for RabbitMQ (or you could use the free version) but that is now a fixed cost and you have to factor in the cost for running the server to execute it along with the effort to maintain it (no cheating on the hidden costs). Looks like others have pondered this SQS vs RabbitMQ question as well (1, 2). However this is just an example of the point which jumped out this week.
So when architecting for Cloud, don't forget "architecture effects costs". With a Cloud there is not only the selection cost but also consider your elastic cost.
Rodos
No comments:
Post a Comment