This repository contains a collection of source code and scripts that can be used to implement load balancing for a Scylla Alternator cluster.
All code in this repository is open source, licensed under the Apache License 2.0.
Scylla is an open-source distributed database. Alternator is a Scylla feature which adds Amazon DynamoDB™ compatibility to Scylla. With Alternator, Scylla is fully (or almost fully) compatible with DynamoDB's HTTP and JSON based API. Unmodified applications written with any of Amazon's SDK libraries can connect to a Scylla Alternator cluster instead of to Amazon's DynamoDB.
However, there is still one fundamental difference between how DynamoDB and a Scylla cluster appear to an application:
- The entire DynamoDB service is presented to the application as a
single endpoint, for example
http://dynamodb.us-east-1.amazonaws.com. - Scylla is not a single endpoint - it is a distributed database - a cluster of many nodes.
If we configure the application to use just one of the Scylla nodes as the single endpoint, this specific node will become a performance bottleneck as it gets more work than the other nodes. Moreover, this node will become a single point of failure - if it fails, the entire service is unavailable.
So what Alternator users need now is a way for a DynamoDB application - which was written with just a single endpoint in mind - to send requests to all of Alternator's nodes, not just to one. The mechanisms we are looking for should equally load all of Alternator's nodes (load balancing) and ensure that the service continues normally even if some of these nodes go down (high availability).
In our blog post Load Balancing in Scylla Alternator we explained in more detail the need for load balancing in Alternator and the various server-side and client-side options that are available.
The goal of this repository is to offer Alternator users with such load balancing mechanisms, in the form of code examples, libraries, and documentation.
The most obvious load-balancing solution is a load balancer, a machine or a virtual service which sits in front of the Alternator cluster and forwards the HTTP requests that it gets to the different Alternator nodes. This is a good option for some setups, but a costly one because all the request traffic needs to flow through the load balancer.
In this document we surveyed some additional server-side load-balancing mechanisms besides the TCP or HTTP load balancer. These including DNS, virtual IP addresses, and coordinator-only nodes. In the dns subdirectory in this repository we demonstrate a simple proof-of-concept of the DNS mechanism.
But the bulk of this repository is devoted to client-side load balancing. In client-side load balancing, the client is modified to connect to all Alternator nodes instead of just one. Client-side load balancing simplifies server deployment and lowers server costs - as we do not need to deploy additional server-side nodes or services.
Of course, our goal is to require as little as possible changes to the client. Ideally, all that would need to be changed in an application is to have it load an additional library, or initialize the existing library a bit differently; From there on, the usual unmodified AWS SDK functions will automatically use all of Alternator's nodes instead of just one.
We currently provide libraries to do exactly that in five programming languages: go (AWS SDK v1), go (AWS SDK v2), java, javascript (node.js), python and C++. Each of these directories includes a README file explaining how to use this library in an application. These libraries are not complete DynamoDB drivers - the application continues to use Amazon's SDKs (e.g., boto3 in Python). Rather, what our libraries do is to automatically retrieve the list of nodes in an Alternator cluster, and configure or trick the Amazon SDK into sending requests to many different nodes instead of always to the same one.
