Successfully Deploying AIOps, Part 3: The AIOps Apprenticeship

This is a copy of an original post on the AppDynamics blog here.



The Strategic Brief:

By augmenting operations teams, AIOps enables organizations to preemptively ensure that applications, architectures and infrastructures are ready for rapid digital transformation.



Part one of our series on deploying AIOPs identified how an anomaly breaks into two broad areas: problem time and solution time. Part two described the first deployment phase, which focuses on reducing problem time. With trust in the AIOps systems growing, we’re now ready for part three: taking on solution time by automating actions.

French Clock

© 2019 Marco Coulter

Applying AIOps to Mean Time to Fix (MTTFix)

The power of AIOps comes from continuous enhancement of machine learning powered by improved algorithms and training data, combined with the decreasing cost of processing power. A measured example was Googles project for accurately reading street address numbers from its street image systems—a necessity in countries where address numbers don’t run sequentially but rather are based on the age of the buildings. Humans examining photos of street numbers have an accuracy of 98%. Back in 2011, the available algorithms and training data produced a trained model with 91% accuracy. By 2013, improvements and retraining boosted this number to 97.5%. Not bad, though humans still had the edge. In 2015, the latest ML models passed human capability at 98.1%. This potential for continuous enhancement makes AIOps a significant benefit for operational response times.

You Already Trust AI/ML with Your Life

If you’ve flown commercially in the past decade, you’ve trusted the autopilot for part of that flight. At some major airports, even the landings are automated, though taxiing is still left to pilots. Despite already trusting AI/ML to this extent, enterprises need more time to trust AI/ML in newer fields such as AIOps. Let’s discuss how to build that trust.

Apprenticeships allow new employees to learn from experienced workers and avoid making dangerous mistakes. They’ve been used for ages in multiple professions; even police departments have a new academy graduate ride along with a veteran officer. In machine learning, ML frameworks need to see meaningful quantities of data in order to train themselves and create nested neural networks that form classification models. By treating automation in AIOps like an apprenticeship, you can build trust and gradually weave AIOps into a production environment.

By this stage, you should already be reducing problem time by deploying AIOps, which delivers significant benefits before adding automation to the mix. These advantages include the ability to train the model with live data, as well as observe the outcomes of baselining. This is the first step towards building trust in AIOps.

Stage One: AIOps-Guided Operations Response

With AIOps in place, operators can address anomalies immediately. At this stage, operations teams are still reviewing anomaly alerts to ensure their validity. Operations is also parsing the root cause(s) identified by AIOps to select the correct issue to address. While remediation is manual at this stage, you should already have a method of tracking common remediations.

In stage one, your operations teams oversee the AIOps system and simultaneously collect data to help determine where auto-remediation is acceptable and necessary.

Stage Two: Automate Low Risk

Automated computer operations began around 1964 with IBM’s OS/360 operating system allowing operators to combine multiple individual commands into a single script, thus automating multiple manual steps into a single command. Initially, the goal was to identify specific, recurring manual tasks and figure out how to automate them. While this approach delivered a short-term benefit, building isolated, automated processes incurred technical debt, both for future updates and eventual integration across multiple domains. Ultimately it became clear that a platform approach to automation could reduce potential tech debt.

Automation in the modern enterprise should be tackled like a microservices architecture: Use a single domain’s management tool to automate small actions, and make these services available to complex, cross-domain remediations. This approach allows your investment in automation to align with the lifespan of the single domain. If your infrastructure moves VMs to containers, the automated services you created for networking or storage are still valid.

You will not automate every single task. Selecting what to automate can be tricky, so when deciding whether to fully automate an anomaly resolution, use these five questions to identify the potential value:

  • Frequency: Does the anomaly resolution occur often enough to warrant automation?
  • Impact: Are you automating the solution to a major issue?
  • Coverage: What proportion of the real-world process can be automated?
  • Probability: Does the process always produce the desired result, or can it be impacted by environmentals?
  • Latency: Will automating the task achieve a faster resolution?

Existing standard operating procedures (SOPs) are a great place to start. With SOPs, you’ve already decided how you want a task performed, have documented the process, and likely have some form of automation (scripts, etc.) in place. Another early focus is to address resource constraints by adding front-end web servers when traffic is high, or by increasing network bandwidth. Growing available resources is low risk compared to restarting applications. While bandwidth expansion may impact your budget, it’s unlikely to break your apps. And by automating resource constraint remediations, you’re adding a rapid response capability to operations.

In stage two, you augment your operations teams with automated tasks that can be triggered in response to AIOps-identified anomalies.

Stage Three: Connect Visibility to Action (Trust!)

As you start to use automated root cause analysis (RCA), it’s critical to understand the probability concept of machine learning. Surprisingly, for a classical computer technology, ML does not output a binary, 0 or 1 result, but rather produces statistical likelihoods or probabilities of the outcome. The reason this outcome sometimes looks definitive is that a coder or “builder” (the latter if you’re AWS’s Andy Jassy) has decided an acceptable probability will be chosen as the definitive result. But under the covers of ML, there is always a percentage likelihood. The nature of ML means that RCA sometimes will result in a selection of a few probable causes. Over time, the system will train itself on more data and probabilities and grow more accurate, leading to single outcomes where the root cause is clear.

Once trust in RCA is established (stage one), and remediation actions are automated (stage two), it’s time to remove the manual operator from the middle. The low-risk remediations identified in stage two can now be connected to the specific root cause as a fully automated action.

The benefits of automated operations are often listed as cost reduction, productivity, availability, reliability and performance. While all of these apply, there’s also the significant benefit of expertise time. “The main upshot of automation is more free time to spend on improving other parts of the infrastructure,” according to Google’s SRE project. The less time your experts spend in MTTR steps, the more time they can spend on preemption rather than reaction.

Similar to DevOps, AIOps will require a new mindset. After a successful AIOps deployment, your team will be ready to transition from its existing siloed capabilities. Each team member’s current specialization(s) will need to be accompanied with broader skills in other operational silos.

AIOps augments each operations team, including ITOps, DevOps and SRE. By giving each team ample time to move into preemptive mode, AIOps ensures that applications, architectures and infrastructures are ready for the rapid transformations required by today’s business.

Successfully Deploying AIOps, Part 2: Automating Problem Time

This is a copy of an original post on the AppDynamics blog here.



The Strategic Brief:

Built-in AI/ML—such as in AppDynamics APM—delivers value by activating the cognitive engine of AIOps to address anomalies.



Asian Clock 1

© 2017 Marco Coulter

In part one of our Successfully Deploying AIOps series, we identified how an anomaly breaks into two broad areas: problem time and solution time. The first phase in deploying AIOps focuses on reducing problem time, with some benefit in solution time as well. This simply requires turning on machine learning within an AIOps-powered APM solution. Existing operations processes will still be defining, selecting and implementing anomaly rectifications. When you automate problem time, solution time commences much sooner, significantly reducing an anomaly’s impact.

AIOps: Not Just for Production

Anomalies in test and quality assurance (QA) environments cost the enterprise time and resources. AIOps can deliver significant benefits here. Applying the anomaly resolution processes seen in production will assist developers navigating the deployment cycle.

Test and QA environments are expected to identify problems before production deployment. Agile and DevOps approaches have introduced rapid, automated building and testing of applications. Though mean time to resolution (MTTR) is commonly not measured in test and QA environments (which aren’t as critical as those supporting customers), the benefits to time and resources still pay off.

Beginning your deployment in test and QA environments allows a lower-risk, yet still valuable, introduction to AIOps. These pre-production environments have less business impact, as they are not visited by customers. Understanding performance changes between application updates is critical to successful deployment. Remember, as the test and QA environments will not have the production workload available, it’s best to recreate simulated workloads through synthetics testing.

With trust in AIOps built from first applying AIOps to mean time to detect (MTTD), mean time to know (MTTK) and mean time to verify (MTTV) in your test and QA environments, your next step will be to apply these benefits to production. Let’s analyze where you’ll find these initial benefits.

Apply AI/ML to Detection (MTTD)

An anomaly deviates from what is expected or normal. Detecting an anomaly requires a definition of “normal” and a monitoring of live, streaming metrics to see when they become abnormal. A crashing application is clearly an anomaly, as is one that responds poorly or inconsistently after an update.

With legacy monitoring tools, defining “normal” was no easy task. Manually setting thresholds required operations or SRE professionals to guesstimate thresholds for all metrics measured by applications, frameworks, containers, databases, operating systems, virtual machines, hypervisors and underlying storage.

AIOps removes the stress of threshold-setting by letting machine learning baseline your environment. AI/ML applies mathematical algorithms to different data features seeking correlations. With AppDynamics, for example, you simply run APM for a week. AppDynamics observes your application over time and creates baselines, with ML observing existing behavioral metrics and defining a range of normal behavior with time-based and contextual correlation. Time-based correlation removes alerts related to the normal flow of business—for example, the login spike that occurs each morning as the workday begins; or the Black Friday or Guanggun Jie traffic spikes driven by cultural events. Contextual correlation pairs metrics that track together, enabling anomaly identification and alerts later when the metrics don’t track together.

AIOps will define “normal” by letting built-in ML watch the application and automatically create a baseline. So again, install APM and let it run. If you have specific KPIs, you can add these on top of the automatic baselines as health rules. With baselines defining normal, AIOps will watch metric streams in real time, with the model tuned to identify anomalies in real time, too.

Apply AI/ML to Root Cause Analysis (MTTK)

The first step to legacy root cause analysis (RCA) is to recreate the timeline: When did the anomaly begin, and what significant events occurred afterward? You could search manually through error logs to uncover the time of the first error. This can be misleading, however, as sometimes the first error is an outcome, not a cause (e.g., a crash caused by a memory overrun is the result of a memory leak running for a period of time before the crash).

In the midst of an anomaly, multiple signifiers often will indicate fault. Logs will show screeds of errors caused by stress introduced by the fault, but fail to identify the underlying defect. The operational challenge is unpacking the layers of resultant faults to identify root cause. By pinpointing this cause, we can move onto identifying the required fix or reconfiguration to resolve the issue.

AIOps creates this anomaly timeline automatically. It observes data streams in real time and uses historical and contextual correlation to identify the anomaly’s origin, as well as any important state changes during the anomaly. Even with a complete timeline, it’s still a challenge to reduce the overall noise level. AIOps addresses this by correlating across domains to filter out symptoms from possible causes.

There’s a good reason why AIOps’ RCA output may not always identify a single cause. Trained AI/ML models do not always produce a zero or one outcome, but rather work in a world of probabilities or likelihoods. The output of a self-taught ML algorithm will be a percentage likelihood that the resulting classification is accurate. As more data is fed to the algorithm, these outcome percentages may change if new data makes a specific output classification more likely. Early snapshots may indicate a priority list of probable causes that later refine down to a single cause, as more data runs through the ML models.

RCA is one area where AI/ML delivers the most value, and the time spent on RCA is the mean time to know (MTTK). While operations is working on RCA, the anomaly is still impacting customers. The pressure to conclude RCA quickly is why war rooms get filled with every possible I-shaped professional (a deep expert in a particular silo of skills) in order to eliminate the noise and get to the signal.

Apply AI/ML to Verification

Mean time to verify (MTTV) is the remaining MTTR portion automated in phase one of an AIOps rollout. An anomaly concludes when the environment returns to normal, or even to a new normal. The same ML mechanisms used for detection will minimize MTTV, as baselines already provide the definition of normal you’re seeking to regain. ML models monitoring live ETL streams of metrics from all sources provide rapid identification when the status returns to normal and the anomaly is over.

Later in your rollout when AIOps is powering fully automated responses, this rapid observation and response is critical, as anomalies are resolved without human intervention.  Part three of this series will discuss connecting this visibility and insight to action.

Successfully Deploying AIOps, Part 1: Deconstructing MTTR

This is a copy of an original post on the AppDynamics blog here.



The Strategic Brief:

Quantifying the value of successful AIOps deployment requires tracking subsidiary metrics within the industry default of mean time to resolution (MTTR). This post breaks out the metrics that form MTTR and divides them into two categories: problem and solution.



Somewhere between waking up today and reading this blog post, AI/ML has done something for you. Maybe Netflix suggested a show, or DuckDuckGo recommended a website. Perhaps it was your photos application asking you to confirm the tag of a specific friend in your latest photo. In short, AI/ML is already embedded into our lives.

The quantity of metrics in development, operations and infrastructure makes development and operations a perfect partner for machine learning. With this general acceptance of AI/ML, it is surprising that organizations are lagging in implementing machine learning in operations automation, according to Gartner.

The level of responsibility you will assign to AIOps and automation comes from two factors:

  • The level of business risk in the automated action
  • The observed success of AI/ML matching real world experiences

The good news is this is not new territory; there is a tried-and-true path for automating operations that can easily be adjusted for AIOps.

It Feels Like Operations is the Last to Know

The primary goal of the operations team is to keep business applications functional for enterprise customers or users. They design, “rack and stack,” monitor performance, and support infrastructure, operating systems, cloud providers and more. But their ability to focus on this prime directive is undermined by application anomalies that consume time and resources, reducing team bandwidth for preemptive work.

An anomaly deviates from what is expected or normal. A crashing application is clearly an anomaly, yet so too is one that was updated and now responds poorly or inconsistently. Detecting an anomaly requires a definition of “normal,” accompanied with monitoring of live streaming metrics to spot when the environment exhibits abnormal behaviour.

The majority of enterprises are alerted to an anomaly by users or non-IT teams before IT detects the problem, according to a recent AppDynamics survey of 6,000 global IT leaders. This disappointing outcome can be traced to three trends:

  • Exponential growth of uncorrelated log and metric data triggered by DevOps and Continuous Integration and Continuous Delivery (CI/CD) in the process of automating the build and deployment of applications.
  • Exploding application architecture complexity with service architectures, multi-cloud, serverless, isolation of system logic and system state—all adding dynamic qualities defying static or human visualization.
  • Siloed IT operations and operational data within infrastructure teams.

Complexity and data growth overload development, operations and SRE professionals with data rather than insight, while siloed data prevents each team from seeing the full application anomaly picture.

Enterprises adopted agile development methods in the early 2000s to wash away the time and expense of waterfall approaches. This focus on speed came with technical debt and lower reliability. In the mid-2000s manual builds and testing were identified as the impediment leading to DevOps, and later to CI/CD.

DevOps allowed development to survive agile and extreme approaches by transforming development—and particularly by automating testing and deployment—while leaving production operations basically unchanged. The operator’s role in maintaining highly available and consistent applications still consisted of waiting for someone or something to tell them a problem existed, after which they would manually push through a solution. Standard operating procedures (SOPs) were introduced to prevent the operator from accidentally making a situation worse for recurring repairs. There were pockets of successful automation (e.g., tuning the network) but mostly the entire response was still reactive. AIOps is now stepping up to allow operations to survive in this complex environment, as DevOps did for the agile transformation.

Reacting to Anomalies

DevOps automation removed a portion of production issues. But in the real world there’s always the unpredictable SQL query, API call, or even the forklift driving through the network cable. The good news is that the lean manufacturing approach that inspired DevOps can be applied to incident management.

To understand how to deploy AIOps, we need to break down the “assembly line” used to address an anomaly. The time spent reacting to an anomaly can be broken into two key areas: problem time and solution time.

Problem time: The period when the anomaly has not yet being addressed.

Anomaly management begins with time spent detecting a problem. The AppDynamics survey found that 58% of enterprises still find out about performance issues or full outages from their users. Calls arrive and service tickets get created, triggering professionals to examine whether there really is a problem or just user error. Once an anomaly is accepted as real, the next step generally is to create a war room (physical or Slack channel), enabling all the stakeholders to begin root cause analysis (RCA). This analysis requires visibility into the current and historical system to answer questions like:

  • How do we recreate the timeline?
  • When did things last work normally or when did the anomaly began?
  • How are the application and underlying systems currently structured?
  • What has changed since then?
  • Are all the errors in the logs the result of one or multiple problems?
  • What can we correlate?
  • Who is impacted?
  • Which change is most likely to have caused this event?

Answering these questions leads to the root cause. During this investigative work, the anomaly is still active and users are still impacted. While the war room is working tirelessly, no action to actually rectify the anomaly has begun.

Solution time: The time spent resolving the issues and verifying return-to-normal state.

With the root cause and impact identified, incident management finally crosses over to spending time on the actual solution. The questions in this phase are:

  • What will fix the issue?
  • Where are these changes to be made?
  • Who will make them?
  • How will we record them?
  • What side effects could there be?
  • When will we do this?
  • How will we know it is fixed?
  • Was it fixed?

Solution time is where we solve the incident rather than merely understanding it. Mean time to resolution (MTTR) is the key metric we use to measure the operational response to application anomalies. After deploying the fix and verifying return-to-normal state, we get to go home and sleep.

Deconstructing MTTR

MTTR originated in the hardware world as “mean time to repair”— the full time from error detection to hardware replacement and reinstatement into full service (e.g., swapping out a hard drive and rebuilding the data stored on it). In the software world, MTTR is the time from software running abnormally (an anomaly) to the time when the software has been verified as functioning normally.

Measuring the value of AIOps requires breaking MTTR into subset components. Different phases in deploying AIOps will improve different portions of MTTR. Tracking these subdivisions before and after deployment allows the value of AIOps to be justified throughout.

With this understanding and measurement of existing processes, the strategic adoption of AIOps can begin, which we discuss in part two of this series.