SLOconf 2023

With a Serverless mindset in place, there should no longer be a need to define resource requests or any other kinds of information. Just developed your app or function and run it.

But how can we achieve performance and make sure we run the system most efficiently? This is were we can now let user define what they truly care about - their SLOs. Based on these performance targets our Intent Driven Orchestration Planner will do the rest. No need to define resource requests and limit s on e.g. Kubernetes cluster anymore. The planner will set up the systems in such a way that your app/functions behave as expected, without the need to know anything about the underlying infrastructure – less knowledge needed, fewer errors made!

This is a shift in how we do orchestration and SLO management, that could be of interest to this community. Away from a monitoring & alerting way of looking SLOs, towards a way of using SLOs to manage the system. Furthermore this truly allows for ease-of-use of the user; rather than defining numbers an values based on domain & contextualized knowledge we let them define what they truly care about: their SLOs!

Present a real-life scenario of using SLOs to manage product requirements and be smarter about allocating the engineering team’s focus. The talk will discuss an example of how SLOs helped monitor a potential problem instead of assigning an engineering team to jump into solving a complex issue. It also helped focus on a specific area of a wider problem. In a broader sense, the talk will discuss the collaboration between the Product Manager and the Engineering Manager and how SLOs can lead to more productive conversations.

In this talk, we would share our experience in promoting Service Level practice across a large organization with over 900 engineers. Learn about the challenges we faced, the strategies we used to encourage adoption, and the valuable lessons we learned along the way.

Whether you're looking to implement SLs in your own organization or simply interested in how to drive adoption of new engineering practices in general.

In this talk, we explore the synergy between time series databases and OpenTelemetry standards, which together help organizations maintain their Service Level Objectives (SLOs) with precision and ease. Additionally, we will highlight the benefits of columnar-based data storage for high-performance queries and storage logs, traces, and metrics.

 Service Levels Objectives are often perceived as by SRE, for SRE, which limits the impact we can have on improving our systems because enforcing SLOs often collides with other priorities. Can we help others in the organization understand the value of improving the system? Can we apply SLOs to qualities of the system which other people already care about? The speed run introduction to SLOs as a commitment to improve using metrics on the work people do for building a coalition who will care about system reliability.

I will talk through the details of how SLOs at Zalando have evolved from the initial implementation ("SLOs for everything!") to the challenge of ensuring SLOs have the organizational power to drive changes in engineering priorities, to the current design of "critical business operations" and SLOs on those operations.

I'll discuss how to address the "fast burn" SLO problem by leveraging distributed tracing to identify regression in the customer experience automatically. When those regressions are identified, automatically identify and page the team best empowered to address them.

I'll discuss how to address the "slow burn" SLO problem through periodic operational review meetings, in which the SLOs are evaluated, and violations to the SLO (or slow burn issues) are allocated to an owner to investigate and address.

Lastly, I'll talk about challenges with the existing approach, including the difficulty of modelling event systems as a reliable flow, difficulty in rolling out more SLOs for non-customer-facing aspects of the organization and returning to service-specific SLOs.

For any large scale production system, the ability to effectively forecast potential capacity issues is crucial for the smooth functioning of the environment. With a reliable prediction, teams can proactively plan ahead, implement necessary scaling changes in a controlled manner and avoid unexpected availability issues that can cause stress and harm to the system.

Before implementing Tamland, the capacity planning process at GitLab.com was ad-hoc, and relied heavily on manual processes and intuition. Unfortunately, this approach often resulted in oversights, with issues going unnoticed until it was too late, sometimes only surfacing when site availability was impacted.

This talk delves into how GitLab leveraged the power of statistical analysis to greatly improve its capacity planning process. The session will be a practical demonstration of how we analyse long-term metrics data using the Meta’s Prophet library to build sophisticated forecast models.

Tamland, the capacity planning tool built by GitLab, is an open-source project and attendees will have access to the source code if they're interested in exploring the implementation in greater detail. This session is for anyone interested in learning about how forecasting libraries such as Prophet, Greykite, or NeuralProphet, and how they can be integrated into an observability system to provide greater insight into the health of a system.

---

Tome is our internal platform for managing, reporting and alerting on SLOs. A design goal was to enable SLOs to be defined with Configuration as Code. This has become the primary way that SLOs are maintained within Atlassian. Working with SLOs this way has many benefits:

• Enforces consistency in how we organise, define and validate SLOs
• Changes are tracked and attributed through a version control system
• Updates can be deployed as part of existing continuous integration and delivery pipelines

We have written a custom Terraform plugin for provisioning SLOs using Terraform, which interfaces with Tome's backend API.

• Optimizes deployments by tracking deployment state and applying diffs only
• Performs custom validation on configurations

In this talk I'll run through our implementation of the plugin and associated API, and how it compares to other SLO configuration systems.

SLICK is our reliability tracking platform at Meta, pioneering an SLO-focused culture across the company. While we have been very successful in onboarding teams to SLICK, we started to notice that a significant amount of teams only got limited value out of their SLOs after the initial onboarding.

In order for SLOs to be useful, the whole team needs to adopt them, use them regularly and retrospect on them frequently. As an initial attempt to help socialize SLOs, we built various integrations into SLICK. This includes periodic reports in our internal work groups to increase the visibility of SLOs, and collaborative data annotations to enable retrospecting on the root causes of SLO violations.

Bringing all of the above together, we will present our brand new SLO review tooling. This provides a structured workflow to have meaningful discussions about SLOs and identify follow-ups, enabling teams to get the most value out of SLOs.

I'll cover Flexera's journey from our internally developed SLO solution, to partnering with NOBL9, and also include how we engaged teams to help develop SLO's. I'll also cover how SLO's are now part of our engineering goals for 2023.

By managing Service Level Objectives as code we can co-locate SLO definitions and ownership with the product code and team. This supports horizontal scaling of SLO ownership, while establishing a single source of truth, adding transparency, integrating with the code management process, and creating an audit trail for SLOs.

The talk will be based off this article published on Procore's Engineering blog: https://engineering.procore.com/implementing-slos-as-code-a-case-study-2/ 

Have you ever wondered if there's something behind the experiential knowledge that we hold as best practices? I've noticed that things that "feel" right can often be connected together, and the connection between SLOs and observability feels right, like there's something deeper underneath.

So, I've been digging into relationships between SLOs, observability, known knowns, unknown unknowns, cardinality, entropy, and more. There's still a lot of details to work out, but what I present here in this talk is a rough overview of where I'm at so far when it comes to motivating SLOs from a more interconnected perspective.

By managing Service Level Objectives as code we can co-locate SLO definitions and ownership with the product code and team. This supports horizontal scaling of SLO ownership, while establishing a single source of truth, adding transparency, integrating with the code management process, and creating an audit trail for SLOs.

The talk will be based off this article published on Procore's Engineering blog: https://engineering.procore.com/implementing-slos-as-code-a-case-study-2/ 

"Zero to SLO Hero" is a two-part talk that will guide you through the journey of implementing Service Level Objectives (SLOs) in your organization. Part 1 of the talk will cover the basics of SLOs, including what they are, why they are important, and how they can help you measure and improve the reliability of your services. In Part 2, we'll dive deeper into the practicalities of implementing SLOs in your organization with examples in Prometheus. We'll also cover some common pitfalls to avoid when implementing SLOs, and how to overcome them. By the end of this two-part talk, you'll have a solid understanding of what SLOs are, why they matter, and how to implement them in your organization. You'll be well on your way to becoming a SLO hero and improving the reliability of your services!"

SLOs govern org-wide expectations for how your software runs in production: but how do you know it's actually working at a functional level? Configuration, user analytics, and data flows are all highly engineered code, but they generally aren't treated as such. So how do you incorporate testing in the modern world of SRE?

Join Marcus for a quick story about how testing can raise early warnings about SLOs that might be slipping--and how to preserve your error budget for only the highest-risk concerns.

Once you have implemented SLOs for your organization how do you move forward?
At Polar Signals, we have quarterly SLO reviews. We're first doing a retrospective and discussing where we did great and also could have improved. For the upcoming quarter, we discuss the OKRs and from those, we derive SLOs. Sometimes OKRs are easy to derive from and sometimes they need to be rephrased to make sense for SLOs.

Matthias will walk you through an example of an SLO that we implemented quarters ago and how it changed over time. The example will showcase the SLO tracked in the open-source Pyrra project which makes SLOs with Prometheus manageable, accessible, and easy to use for everyone.

I'm pretty bad at time management, so I was looking into ways to improve that part of my life.

I turned to a lo-fi way to use SLOs to manage my day by turning my tasks and other things I need to do during the day into SLOs.

It's increased my productivity and am interested in sharing it with others!

You've already sold your organization on SLOs, now it's time to sell them to your customers. But instead of pitching, this is a collaborative exercise to ensure that they understand your system and you understand their needs.

Measuring and having SLIs on latency could help, but what if your customers care about correctness?
In this talk, we'll discuss how to better meet your customers needs proactively since your SLOs will mirror what their expectations are and will reduce asking if there's an issue.

With this approach, you can reduce alert fatigue and improve the customer experience by making them happier and increasing trust in your system.

Present a real-life scenario of using SLOs to manage product requirements and be smarter about allocating the engineering team’s focus. The talk will discuss an example of how SLOs helped monitor a potential problem instead of assigning an engineering team to jump into solving a complex issue. It also helped focus on a specific area of a wider problem. In a broader sense, the talk will discuss the collaboration between the Product Manager and the Engineering Manager and how SLOs can lead to more productive conversations.

In today's digital world, it's important to ensure that our services are reliable, performant, and meet the needs of our users. In this video, we'll dive into the world of SLO and explore how we can apply it to edge networks.

Neil Pagaduan, Manager of Technology and Engineering at Cox Edge will begin by discussing the importance of SLO in Edge Networks, and how it can help to deliver a better user experience. Then, we'll explore the steps involved in achieving SLO, from defining objectives to measuring performance. Finally, Neil will discuss how to apply SLO to an edge network, including some key considerations to keep in mind.

SREs, as the name implies, care about service reliability. But, often, they struggle with having a way to define, measure and assess their services reliability. In practice, they lack a Reliability Framework.

How can SLOs help? They provided an opinionated way to do just that: define, measure and assess service reliability from the users perspective. They provide a common language to talk about reliability and prioritize work. They help fix the anti-pattern of trying to ensure service reliability without clearly defining what it means.

Recording rules is a clever concept introduced by Prometheus for storing results of query expressions in a form of a new time series. This concept is used for SLO calculations. But due to the nature of recording rules they have no retroactive effect. And since SLO objective usually captures a time window no less than 30d, recording rules produce incomplete results until the whole time window is captured.

The talk will cover how this can be fixed in VictoriaMetrics monitoring solution via retroactive rules evaluation on example of rules generated via https://github.com/slok/sloth framework.

While the direction and intent of SRE has been established and is becoming better-understood, the details on how to achieve "SRE" is still an exercise left for the reader.

Steve from Google and Sal from Nobl9 will present two independently developed methods for teaching SRE topics to customers, which we have discovered are actually quite similar. Huzzah!

Steve will present the "reliability map" and Sal will show Nobl9's SLODLC. Both are essentially sets of documentation presented in a way that allows customers to take a *guided* path through the dark and scary woods that is today's SRE.

The "reliability map" provides a detailed accounts of what development, infrastructure, operational, observability, and people/culture activities organizations can expect to be practicing to achieve different eras of reliability (99%, 99.9%, 99.99%, 99.999%). The "reliability map" takes it one step further and provides descriptions and references to additional content for all activities.

SLODLC provides a framework of documentation and templates to allow customers to iterate on improving their SLOs. This is especially necessary when customers have high reliability targets in which rely upon automation to execute predetermined playbooks. (rollback, exponential backoff, autoscale, etc) If SLOs are the indicator the performance of a user journey and decide when actions needs to be taken, then customers need a clear and useable way to explain:

• What they are attempting to measure (golden signals?, something else?)
• Why was it decided to measure X in such a way?
• How is X impactful for the targeted user journey?
• When was the last time the SLO objective, metric, time window, etc was changed?
• When the error budget is in danger of being breeched what actions should be take?

In short both projects are approaching the same topic of helping customers improve their reliability from different points of view and they synergize well together.

My background has mostly been in developing operating systems for data storage companies. In this environment, almost everything is controlled internally. For example, if the SLO of the operating system is five 9s, then the error budget is usually all consumed by software bugs owned internally by the company. As I switched to developing SaaS products in the cloud, this has drastically changed. Below are examples of lessons learned during different phases of working on a product, from development, to production, to support and operation. My goal is sharing these lessons so other developers can learn from my experience.

In my previous role, two main things I relied on while developing operating system products were the suite of testing as well as the release cadence. Shipping a new operating system every 6 months was considered fast. This gave developers a lot of time for their code to soak and be tested internally before being released to customers. Additionally, with a slower release cadence there was a lot of effort invested in creating different layers of testing. After all, a bug fix would take months to reach customers. Even if we released a patch quickly, which in this context means a few weeks, customers would still need to update their operating systems to apply that patch, and who knows how long a customer will wait before applying that patch. After moving to building SaaS products in the cloud, the release cadence became much faster. This required shifting my thought process. Instead of relying on soak time and various levels of testing, activities such as code review and in-build unit tests now take a front row seat. Metrics like code coverage from unit tests mattered more, while metrics like how long it's been since QA found a bug mattered less.

Another difference between my old role and new role is how developers access production and production metrics. In the old role, gathering production metrics was no easy feat.
Harder access to production metrics, implied that changing SLOs internally was harder, took longer, and a lot of developers didn't know how/why SLOs were changing. For a SaaS product running in the cloud, developers have access to overall system performance metrics at the click of a button, while maintaining compliance. This makes it easier for developers to know why/how SLOs are affected by SLAs and also gives them faster reaction times.

From the operation perspective, the old role and the new role were quite different. In the old role developers didn't go on-call. There was a customer support organization that would handle any customer issues first. Developers were brought in occasionally if customer support needed a bug fixed. Developers didn't need to wake up in the middle of the night because there's an outage. In the new role, developers are on-call, which means developers can and do occasionally get paged in the middle of the night. This shift caught me off-guard as I had to think much harder about how my service impacts the quality of life of my colleagues and myself. It made me think about how to change and measure SLOs to eventually avoid waking someone up in the middle of the night.

Another lesson that caught me off guard is not all SLAs can be trusted when working in the cloud. Yes, I knew this lesson theoretically, but learning it in practice is a different story. One example was an outage that was caused by a cloud provider breaching their SLAs for a managed service that we relied on for our product. This resulted in our product breaching its SLAs. When something like this happens, you think hard about how to update your SLOs to prepare for these issues, catch such issues, and react to them. I found that using SLOs that are tighter than SLAs was helpful in that regard.

In conclusion, even with years of professional experience, moving from developing On-Prem products to cloud SaaS offerings changed how I think about SLOs and it's truly not a one-size-fits-all.

At eBay we use a wrapped version of Prometheus as our centralized time series database. Data residing in Prometheus is used for mission critical functions like detecting issues on the site through anomaly detection, SLOs or simple threshold based alerts. To be able to enforce SLOs across the 1000s of micro services that are deployed inside of eBay, we would need to be able to aggregate commonly instrumented metrics at an application level.

Why are these aggregations necessary? Host level metrics when used for computation of things like burn rates over a 30 day window can become very very expensive. Using a single query across all micro services to generate such aggregations have several scale limitations. Manually onboarding each micro service into such aggregations can also be tedious.

The talk discusses our commonly instrumented metrics, our metrics store architecture, scale challenges seen with SLI/burn rate computation and how we came up with concepts like templated rule groups that automatically generate Prometheus rule groups to be able to do aggregations for every application that is emitting metrics.

SLICK is our reliability tracking platform at Meta, pioneering an SLO-focused culture across the company. While we have been very successful in onboarding teams to SLICK, we started to notice that a significant amount of teams only got limited value out of their SLOs after the initial onboarding.

In order for SLOs to be useful, the whole team needs to adopt them, use them regularly and retrospect on them frequently. As an initial attempt to help socialize SLOs, we built various integrations into SLICK. This includes periodic reports in our internal work groups to increase the visibility of SLOs, and collaborative data annotations to enable retrospecting on the root causes of SLO violations.

Bringing all of the above together, we will present our brand new SLO review tooling. This provides a structured workflow to have meaningful discussions about SLOs and identify follow-ups, enabling teams to get the most value out of SLOs.

Adopting an SLO culture involves identifying the metrics that matter without drowning in noise and alert fatigue. The black box concept lets us aggregate granular metrics into SLIs that focus on the user experience as an indicator of system reliability.

The talk will be based off this article published on Procore's Engineering blog: https://careers.procore.com/blogs/engineering-at-procore/black-box-slis.

By managing Service Level Objectives as code we can co-locate SLO definitions and ownership with the product code and team. This supports horizontal scaling of SLO ownership, while establishing a single source of truth, adding transparency, integrating with the code management process, and creating an audit trail for SLOs.

The talk will be based off this article published on Procore's Engineering blog: https://engineering.procore.com/implementing-slos-as-code-a-case-study-2/ 

In my experience, many teams encounter SLOs as something they've been told to "do" and the flip side is it's been something I've been asked to help them with. Naturally this is not ideal - as engineers we prefer things to be self-evident.

I have a number of strategies to use when communicating the process and the value of creating and using SLOs. I'll give away the thing I say the most right here: "SLOs are for decision-making". They're not magic or even a single thing. They're a tool that helps us do our jobs.

Audience will come away either better-able to articulate the pragmatic usefulness of an SLO mindset, or with one or two real motivations why they should consider developing and using SLOs for their own systems.

While the direction and intent of SRE has been established and is becoming better-understood, the details on how to achieve "SRE" is still an exercise left for the reader.

Steve from Google and Sal from Nobl9 will present two independently developed methods for teaching SRE topics to customers, which we have discovered are actually quite similar. Huzzah!

Steve will present the "reliability map" and Sal will show Nobl9's SLODLC. Both are essentially sets of documentation presented in a way that allows customers to take a *guided* path through the dark and scary woods that is today's SRE.

The "reliability map" provides a detailed accounts of what development, infrastructure, operational, observability, and people/culture activities organizations can expect to be practicing to achieve different eras of reliability (99%, 99.9%, 99.99%, 99.999%). The "reliability map" takes it one step further and provides descriptions and references to additional content for all activities.

SLODLC provides a framework of documentation and templates to allow customers to iterate on improving their SLOs. This is especially necessary when customers have high reliability targets in which rely upon automation to execute predetermined playbooks. (rollback, exponential backoff, autoscale, etc) If SLOs are the indicator the performance of a user journey and decide when actions needs to be taken, then customers need a clear and useable way to explain:

• What they are attempting to measure (golden signals?, something else?)
• Why was it decided to measure X in such a way?
• How is X impactful for the targeted user journey?
• When was the last time the SLO objective, metric, time window, etc was changed?
• When the error budget is in danger of being breeched what actions should be take?

In short both projects are approaching the same topic of helping customers improve their reliability from different points of view and they synergize well together.

The talk tells the story of how conversations around SLOs can be a great trigger to start a shift to a Product Thinking mindset, with practical examples. We'll also take a light dip into constraint mapping from an SLO perspective.

Join Surya Bhagvat, Director of Site Reliability Engineering at Harness, to discuss how his team used business objectives to create SLOs that positively impacted customer experience and his engineering team. Surya will share lessons learned in choosing and implementing the SLIs and SLOs aligned with customer expectations and Harness’ desired business outcomes.

With a Serverless mindset in place, there should no longer be a need to define resource requests or any other kinds of information. Just developed your app or function and run it.

But how can we achieve performance and make sure we run the system most efficiently? This is were we can now let user define what they truly care about - their SLOs. Based on these performance targets our Intent Driven Orchestration Planner will do the rest. No need to define resource requests and limit s on e.g. Kubernetes cluster anymore. The planner will set up the systems in such a way that your app/functions behave as expected, without the need to know anything about the underlying infrastructure – less knowledge needed, fewer errors made!

This is a shift in how we do orchestration and SLO management, that could be of interest to this community. Away from a monitoring & alerting way of looking SLOs, towards a way of using SLOs to manage the system. Furthermore this truly allows for ease-of-use of the user; rather than defining numbers an values based on domain & contextualized knowledge we let them define what they truly care about: their SLOs!

There's been a lot of (really good!) discussion in the last several years about how to think about, monitor, and alert on long-tailed distributed quantities, such as latency. However, I'm worried that in our desire to describe the long tail we may be too eager to abandon tools that are still useful.

In this talk I'll (re)introduce everyone's favorite summary statistic, the average. I'll talk about the difference between a sample average and a random variable's expectation, and how the two are uniquely linked by the law of large numbers. I'll also talk about how the central limit theorem allows us to treat sample averages as draws from a Gaussian distribution, irrespective of the distribution the samples come from, and then I'll talk about the exceptions. We'll finish up by looking at how the properties of expected values can give us insight into the behavior of systems, even at the tail.

Through all of this we'll be looking at examples drawn from real-world latency data, and comparing the insights gleaned from this versus other common summary statistics.

Resiliency is a critical piece to building reliable systems. It allows us to feel safe knowing failure is inevitable. After all, as noted in the OG SRE book, 100% is terrible target for basically everything.

We spend a lot of resources to add in layers of resiliency from redundant multi-region compute stacks to backups on our backups. How do we know this resilience achieved our ultimate outcome of reliability for our customers? We'll discuss the ways to observe your SLOs and error budgets during resiliency events.

The various events we'll look at include, regional failover, chaos experiments (such as latency injection), database recovery, and more! After failing a region, do we know if your customer's experienced a disturbance? When you're running a resiliency test or game day, how do you measure success?

Observing error budgets before, during, and after an event paint a picture of our customer's experience and can ultimately be part of the success criteria. It is critical that we know how our architecture and system changes unfold. What if new resiliency introduces latency that negatively impacts your customer? For example, if there's complexity introduced that makes your release engineering more convoluted, we may see a longer error budget burn while we remediate. On the flip side the partnership of adding resiliency and observing your SLOs can also lead to improving the objectives with newly matured levels of resiliency; raising the bar for performance.

At eBay we use a wrapped version of Prometheus as our centralized time series database. Data residing in Prometheus is used for mission critical functions like detecting issues on the site through anomaly detection, SLOs or simple threshold based alerts. To be able to enforce SLOs across the 1000s of micro services that are deployed inside of eBay, we would need to be able to aggregate commonly instrumented metrics at an application level.

Why are these aggregations necessary? Host level metrics when used for computation of things like burn rates over a 30 day window can become very very expensive. Using a single query across all micro services to generate such aggregations have several scale limitations. Manually onboarding each micro service into such aggregations can also be tedious.

The talk discusses our commonly instrumented metrics, our metrics store architecture, scale challenges seen with SLI/burn rate computation and how we came up with concepts like templated rule groups that automatically generate Prometheus rule groups to be able to do aggregations for every application that is emitting metrics.