As an Enterprise Architect for Intel IT, I worked with IT Engineering and our Software and Services group on the elastic scaling of the APIs that power the Intel AppUp® center. Our goal was to scale our APIs to at least 10x our baseline capacity (measured in transactions per second) by moving them to our private cloud, and ultimately to be able to connect to a public cloud provider for additional availability and scalability. Here’s a quick set of practices we used to achieve our goal:
- Virtualize everything. This may seem obvious and is probably a no-op for new APIs, but in our case we were using a bare-metal installs at our gateway and database layers (the API servers themselves were already running as VMs). While our gateway hardware appliance had very good scalability, we knew we were ultimately targeting the public cloud and that our need for dynamic scaling could exceed our ability to add new physical servers. Using a gateway that scales in pure software virtual machines without the need for special purpose-built hardware helped us achieve our goal here.
- Instrument everything. We needed to be able to correlate leading indicators like transactions per second to system load at each layer so we could begin to identify bottlenecks. We also needed to characterize our workload for testing – understanding a real-world sequence of API methods and mix/ordering of reads and writes. This allowed us to create a viable set of load tests.
- Identify bottlenecks. We used Apache jmeter to generate load and identify points where latency became an issue, correlating that against system loads to find out where we had reached saturation and needed to scale.
- Define a scaling unit. In our case, we were using dedicated DB instances rather than database-as-a-service, so we decided to scale all three layers together. We identified how many API servers would saturate the DB layer, and how many gateways we would need to manage the traffic. We then defined a collection of VMs that would provision all of these VMs together. We might have scaled each layer independently had our API been architected differently, or if we were building from scratch on database-as-a-service.
- Repeat. The above let us scale from 1x to about 5x or 6x without any problem. However, when we hit 6x scaling we discovered that a new bottleneck: the overhead of replicating commits across the database instances. We went back to the drawing board and redesigned the back end for eventual consistency so we could reduce database load.
- Automate everything. We use Nagios and Puppetto monitor and respond to health changes. A new scaling unit is provisioned when we hit predefined performance thresholds.
- Don’t forget to test scaling down. If you set a threshold for removing capacity, it’s important to make sure that your workflow allows for a graceful shutdown and doesn’t impact calls that are in progress.
The above approach got us to 10x our initial capacity in a single data center. Because of some of our architecture decisions (coarse-grained scaling units and eventual consistency) we were then able to add a GLB and scale out to multiple data centers – first to another internal private cloud and then to a public cloud provider.