At New Relic, we believe that programmatically tracked service level indicators (SLIs) are foundational to our site reliability engineering practice. When we have programmatic SLIs in place, we lessen the need to manually track performance and incident data. We’re able to reduce that manual toil because our DevOps teams define the capabilities and metrics that define their SLI data, which they in turn collect automatically—hence “programmatic.”
Programmatic SLIs have three key characteristics:
- Current—they reflect the state of our system right now
- Automated—they’re reported by instrumentation, not by humans
- Useful—they’re selected based on what a system’s use care about
In this post, I’ll explain how New Relic site reliability engineers (SREs) help their teams develop and create programmatic SLIs.
A key part of creating programmatic SLIs is identifying the capability of the system or service for which you’re creating the SLI. At New Relic we use the following definitions:
- A system is a group of services and infrastructure components that exposes one or more capabilities to external customers (either end users or other internal teams).
- A service is a runtime process (or a horizontally scaled tier of processes) that make up a portion of a system.
- A capability is a particular aspect of functionality exposed by a service to its users, phrased in plain-language terms.
You can create SLIs at any layer, but this post will focus primarily on system-level SLIs.
SLIs—indicators and objectives
But first, we need some more definitions. An indicator is something you can measure about a system that acts as a proxy for the customer experience. An objective is the goal for a specific indicator that you’re committed to achieving.
X should be true… Y portion of the time.
Configuring indicators and objectives is the easy part. The hard part is thinking through what measurable system behavior serves as a proxy for customer experience. When setting system-level SLIs, think about the key performance indicators (KPIs) for those systems; for example:
- User-facing system KPIs most often include availability, latency, and throughput.
- Storage system KPIs often emphasize latency, availability, and durability.
- Big data systems, such as data processing pipelines, typically use KPIs such as throughput and end-to-end latency.
Your indicators and objectives should provide an accurate snapshot of the impact of your system to your customers.
A more precise description of the indicator and objective relationship is to say that SLIs are expressed in relation to service level objectives (SLOs). When you think about the availability of a system, for example, SLIs are the key measurements of the availability of the system while SLOs are the goals you set for how much availability you expect out of that system; (and service level agreements—SLAs—explain the results of breaking the SLO commitments).
SLI SLO SLA
X should be true… Y portion of the time or else.
So, as an example, for the routing capability of a data-ingest tier, a plain-language definition for the data routing capability might look like, “Incoming messages are available for other systems to consume off the message bus without delay.” With that definition then, we might establish the SLI and SLO as, “Incoming messages are available for other systems to consume off of our message bus within 500 milliseconds 99.xx% of the time.”
Creating programmatic SLIs
You should write your programmatic SLIs in collaboration with your product managers, engineering managers, and individual contributors who work on a system. To define your programmatic SLIs (and SLOs), apply these steps:
- Identify the system and its services.
- Identify the customer-facing capabilities of the system or services.
- Articulate a plain-language definition of what it means for each capability to be available.
- Define one or more SLIs for that definition.
- Measure the system to get a baseline.
- Define an SLO for each capability, and track how you perform against it.
- Iterate and refine our system, and fine tune the SLOs over time.
Here are two example capabilities and definitions for an imaginary team that manages an imaginary dashboard service:
Capability: Dashboards overview
Availability Definition: Customers are able to select the dashboard launcher, and see a list of all dashboards available to them.
Capability: Dashboards detail view
Availability Definition: Customers can view a dashboard, and widgets render accurately and timely manner.
To express these availability definitions as programmatic SLIs (with SLOs to measure them), you’d state these service capabilities as:
- Requests for the full list of available dashboards returns within 100 milliseconds 99.9% of the time.
- Requests to open the dashboard launcher complete without error 99.9% of the time.
- Requests for an individual dashboard returns within 100 milliseconds 99.9% of the time.
- Requests to open an individual dashboard complete without error 99.9% of the time.
Tracking programmatic SLIs in New Relic
New Relic provides a bounty of resources for tracking and measuring your SLIs—an essential part of creating programmatic SLIs. You’ll need to identify existing instrumentation (if any), and deploy instrumentation in your systems and services where is doesn’t already exist. Note that tracking any of your business logic in New Relic will likely require some kind of custom instrumentation, which allows you to track any interactions that may not already be captured by New Relic’s automatic instrumentation.
The key steps in New Relic will be:
- Gather metrics and events (custom or automatic) through instrumentation, and, if your capability involves APIs, run New Relic Synthetic API tests to ensure they behave as expected.
- Create alert conditions that will create a violation if your SLIs exceed their objectives.
- Create New Relic query language (NRQL) queries and New Relic dashboards that reveal when your services miss their indicators.
Here are three examples of programmatic SLIs set for three systems in a highly simplified version of a system like New Relic. Each contains a capability definition, a measurement of the SLI (with an SLO), and a NRQL query for creating dashboard widgets and alert conditions.
1. Ingest system: Data is sent to an API endpoint, serviced by a load-balancer, to a service that aggregates data by account and publishes to a Kafka topic with an account-based partition scheme.
Capability: Ingest is able to receive data.
Measurement: 99.96% of data is consumed within 1 minute.
SELECT percentile(duration, 99.96) FROM ingest_consumer WHERE appName = 'Ingest Production' AND action = ‘processed’
Alert condition: Query result is > 60 at least once in 1 minute.
2. Storage system: Data from the Kafka topic is consumed by a service that writes the data to a distributed database.
Capability: Storage writes the incoming data to disk in a timely manner.
Measurement: 99.96% of data is written to disk within 10ms.
SELECT percentile(duration, 99.96) FROM storage_writer WHERE appName = ‘Storage Production' AND action = ‘write_success’
Alert condition: Query result is > .01 at least once in 1 minute.
3. User interface: Customers view graphs of their data via a scaled backend that reads data from a distributed datastore.
Capability: Customers are able to login to the user interface and see their data.
Measurement: 99.96% of transactions complete
SELECT count(*) from SyntheticCheck where result = 'FAILED' and monitorName = 'UI Status Check'
Alert condition: Query result is > 1 at least once in 1 minute.
One size doesn’t fit all
New Relic SREs spend a great deal of time working with our teams to define their SLIs, but it’s not a one size fits all or a one-stop process. Modern software systems evolve and change rapidly, so SLIs aren’t something we can set once and forget about.
Once you’ve settled on your SLIs, they should be reasonably stable, but they’ll need to be revisited regularly because systems evolve. It’s a good idea to revisit them quarterly, or whenever you make changes to your services, traffic volume, and upstream and downstream dependencies.