How Fubo uses AI to run on-call for the biggest live sports events in streaming

600+ Total on-call alerts triaged by Cleric since December
9 out of 10 On-call alerts during the FIFA World Cup 2026™ that Cleric handled without pinging a human
4.57 Average SRE user feedback rating (out of 5) for Cleric’s responses
With Cleric, we were able to reduce at least 70% of the noise from infrastructure alerts. Previously, a developer who was on call would spend so much time attending to and looking at these alerts, and now they have that time and focus back.
Ravisankar Velmurugan
Ravisankar Velmurugan Vice President of Engineering, Fubo

In the past six months, the on-call experience for Fubo’s engineering team has completely changed. “Before, if you were somewhere like a coffee shop and you got pinged, you’d have to get on a computer to see what happened,” explained Ravisankar (Ravi) Velmurugan, VP of Engineering. “Now you can just handle everything from your phone, with a side chat with Cleric, to see if it’s a serious issue or not and what actions you might need to take.”

Fubo is one of the leading live TV streaming platforms in the US. Live sports are its pride, from the FIFA World Cup to the NBA Finals and more than 80,000 live sporting events streamed on its platform every year. Within Fubo, engineering is a globally distributed team that protects those critical live moments.

With Cleric as part of their agentic stack, the team is able to resolve production issues quickly (often automatically) and get time back to focus on high-impact engineering work. Fubo’s engineering team runs on-call for the biggest live sports events and 9 out of 10 alerts are automated thanks to Cleric.

The on-call challenge for live event scaling

Live events present unique scaling challenges for streaming platforms. A channel might have modest viewership, and then a sports event starts, and suddenly millions of users jump onto the stream at the same time. These extreme scale-up scenarios can be hard to predict as well: a channel might go from a few thousand viewers to millions the moment a playoff game kicks off.

A live game can drive 10× more user traffic within minutes

Conceptual live-game scenario, not measured Fubo traffic

User traffic

8:00

8:15

8:30

8:45

Time

Scale-up scenarios can create a lot of alerts, not all urgent, but all requiring a quick assessment. “Alert triage was consuming a significant share of engineering time,” Ravi explained. “It’s completely distracting. You have to come out of the day-to-day engineering task you’re concentrating on, do some work, and then switch back.”

On-call engineers would often receive pings late at night, which is appropriate when alerts are urgent, but not when they’re noisy. Complicating matters, the team’s SREs are based in India, while engineers are in the US.

Why production triage needed a purpose-built agent

Fubo’s engineering team uses AI extensively for development, but production triage requires reasoning that’s grounded in live system state instead of general programming knowledge. They quickly ruled out building a solution in-house.

“We knew the reasoning would be the hard part,” Ravi explained. “And with the landscape changing so quickly, we didn’t want to allocate resources to an approach that might become obsolete in six months.”

As they explored solutions, Cleric stood out for several reasons:

  • Fast time-to-results. “We could use it right away, from day one, which was pretty interesting, because most of the companies we talked to wanted more time,” Ravi said.
  • Ease of integration. Unlike walled-garden enterprise tools, Cleric didn’t come with extra integration costs, and it worked with every component in Fubo’s stack. “With other tools, you have to buy certain things to get certain things. You can’t just use your data, you have to use their whole observability or telemetry system. With Cleric, that’s not an issue, and that makes it easy for us to expand it to multiple teams and pick and choose features, which helps on the cost side as well.”
  • Does the work of an engineer. “Most competitors just had reporting tools that would show you the number of issues. Nobody had an interactive tool like Cleric that could proactively fix issues in production.”

How Cleric works

The Fubo SRE team first tested Cleric until they felt they could trust the solution, then exposed it to the entire team as the default on-call agent.

When an alert comes in on Slack, Cleric auto-triggers an investigation. For false positives and alerts where the conditions are no longer valid, Cleric typically resolves without involving an on-call engineer. For the few issues that require human intervention, Cleric provides the likely cause and reasoning. Engineers can chat with it to understand the issue in more detail, and can push Cleric’s recommendation straight to Claude Code to make a fix.

Same alerts, far fewer interruptions, and a head start on the rest

Cleric changes what is escalated to an engineer, not just who performs the investigation.

Before Cleric Engineer paged Investigate from scratch With Cleric Resolved by Cleric No one paged Engineer paged Likely cause Proposed remediation Cleric

Cleric is integrated with Fubo’s entire engineering stack, including containers, APM systems, logging systems, code repositories and knowledge bases. It required minimal setup: “I didn’t give any indication as to which workload is coming from which repo; it basically understands what each application or service is doing based on the code base and the README files,” explained Karthik Rudrayya, the team’s lead SRE.

These integrations allow Cleric to understand the production system deeply and triage the kinds of issues that can be challenging for on-call engineers, like the main server being unable to read from GKE, or a pod failing to start before the GKE metadata service. “Engineers understand the need for more memory or CPU, but not necessarily the nuances specific to our cloud infrastructure,” Ravi said. “That reasoning is what makes Cleric so valuable.”

For example, the team had an out-of-memory restart, and Cleric saw that it happened during a routine graph download and shared the pattern that it saw on one particular pod. “It was even able to isolate the exact file that caused the out-of-memory error and say ‘there’s something happening with this, it doesn’t look normal,’” Karthik said.

APPJul 22nd at 8:11 PM
Memory for is approaching the critical threshold (currently above 75%).
@
avg(last_10m):(avg:.memory.usage{:, environment:prod, cluster_name:} by {pod_name} / avg:.memory.limits{:, environment:prod, cluster_name:} by {pod_name}) * 100 > 85
Notified
@
0 20 40 60 80 100 02:15 02:30 02:45 03:00 UTC > 85 < 85 last 10m avg: 75.42
3 replies
ClericAPPJul 22nd at 8:22 PM
Summary
The alert fired for pod at 75.4% memory usage, but this pod had already been terminated by HPA scale-down 9 minutes before the alert fired [1] [2]. The termination was routine autoscaling: the HPA explicitly stated “memory resource utilization (percentage of request) below target” as the reason [3].
Key Findings
  • No OOMKill: Zero OOMKilled events found across events, events, and logs
  • HPA scale-down sequence: deployment scaled from 16 → 11 replicas over ~20 minutes as traffic decreased
  • Fleet-wide memory spike: 10+ pods spiked to 72-85% memory at 03:00-03:02Z, but this resolved naturally [1]
  • Current state: All 11 pods are healthy at 35-48% memory utilization
  • Pod was at lower end: The alerted pod (72.45%) was actually at the LOWER end of fleet memory vs peers (73-85%)
Diagnosis
This was a timing artifact: the monitor captured a point-in-time memory reading for a pod that was subsequently terminated by routine HPA autoscaling. The HPA terminated the pod because overall memory utilization was BELOW target, not because of memory pressure. No action is required.
Unresolved
The cause of the fleet-wide memory spike (03:00-03:02Z) could not be determined; there’s a known pattern where large cache queries can cause OOM, but logging has 99% sampling configured which prevented verification.
Check memory trendReview HPA config
Open in Cleric Rate this answer ···
Karthik RudrayyaJul 23rd at 12:09 AM
@Cleric, check if the memory spike is at a time when KG graph is downloaded?
ClericAPPJul 23rd at 12:13 AM
Yes, the memory spike correlates exactly with a scheduled KG/cache refresh at 03:00 UTC.
Evidence
The metric shows cache refresh activity exclusively between 03:00:20Z - 03:07:20Z, this is a daily scheduled operation (confirmed by same pattern on 2026-07-22).
During this refresh:
  • Cache unique_keys spiked to 52,766 (19.5x the normal ~2,707 baseline) at 03:00:40Z
  • Cache rebuild duration peaked at 3,231ms at 03:00:40Z (vs ~2,100ms baseline)
  • Multiple hosts affected simultaneously as pods hold both old and new cache data in memory
Root Cause Confirmed
The fleet-wide memory spike was caused by the daily RTU (Real-Time Update) cache refresh at 03:00 UTC. During this refresh, pods temporarily hold both old and new cache data in memory, causing the ~19.5x spike in unique_keys and corresponding memory pressure. This is expected behavior for this scheduled operation, not an anomaly.
Open in Cleric Rate this answer ···
Cleric’s investigation of a memory alert in Slack, including investigation refinement from the team

The team uses Claude Code to resolve issues that Cleric identifies, but they haven’t seen an off-the-shelf model catch up to Cleric’s level of production understanding. Karthik explained, “We had some connectivity issues with a caching layer, and Cleric was able to quickly identify the root cause with conviction, which is what really matters in an incident.”

Results

Since the team adopted Cleric, it has redirected over 600 issues from Fubo’s on-call engineers. With the launch of Cleric’s verification engine in May, it’s also possible to measure how many issues were fixed successfully in production.

During June and July, when Fubo was streaming the FIFA World Cup 2026™ and experiencing extremely high traffic loads, Cleric handled 94% of the on-call alert volume without requiring human intervention. Of the few alerts that did need to be surfaced to engineers, all were fixed successfully:

During the FIFA World Cup 2026™, 94% of alerts never needed a human

June to July 2026, typical month

47 3 50 alerts 2 1 3 Escalated to an engineer Resolved by Cleric Escalated to an engineer Fixed with Cleric’s proposed remediation Fixed by other means

Engineers fixed all 3 escalated to them, 2 with Cleric’s proposed remediation

But it’s also had an impact beyond the quantitative outcomes. “Cleric has completely changed the on-call experience,” Ravi said. There’s less need for cross-timezone coordination, as alerts can often be resolved without reliance on the team in India. The noise has drastically reduced, and engineers have more flexibility and less stress when they’re on call.

Central to Cleric’s success is the team’s trust in the tool’s reasoning. “I would say it has at least an 80% rate in terms of a good answer,” said Rohit Nambiar, a software engineer at Fubo. On a five-point scale for in-product feedback, the average rating during the SRE team’s initial 30-day evaluation was 4.57.

Strong early trust from Fubo’s SRE team

Average in-product feedback rating, initial 30-day evaluation

4.57 out of 5, average rating

“I would say it has at least an 80% rate in terms of a good answer.”

Rohit Nambiar, Software Engineer at Fubo

Looking forward

With less context switching from on-call alerts, Fubo’s engineering team has freed up more of their time and focus for higher-impact infrastructure work, like multi-region scaling and backend migrations.

The team is already partnering with Cleric on new features that can impact the entire SDLC. “One thing we’d like is if Cleric could look at our PRs and flag if any of the changes would cause infrastructure issues,” Ravi said. “It would be great if it could even suggest ways to improve our infrastructure to see better results.”

This vision is aligned with the future that we see at Cleric, where our tool becomes production reliability for AI-built software: prevention where possible, fast response where not. Cleric detects and prevents failures in production, so teams keep shipping fast.

For other teams looking to adopt a similar tool, Ravi’s advice is to get direct experience. “You don’t really know the impact until you actually use something in the real world. We appreciated that Cleric was so easy to integrate and try out. We had amazing support from their team, they were very adaptive to our needs, and we were able to achieve results much more quickly than we anticipated.”

Give your on-call a head start.

Start for free, or talk to us about a plan built for your team’s scale and security needs.