SRE, DevOps and Platform Operations Insights | OpsWerks

Who Operates Data Pipelines and AI Infrastructure as a Managed Service? | OpsWerks

Written by OpsWerks | Nov 19, 2026, 4:12:47 PM
 
Operating Models

Who Operates Data Pipelines and AI Infrastructure as a Managed Service?

What a managed service covers for data and AI platforms, where those platforms break, and how to judge who should run them.

01:40An Airflow scheduler stalls, and the overnight jobs queue behind it.
03:15A Kafka consumer group falls behind. Nobody is paged, because the brokers are healthy.
07:00The morning training job runs on schedule, on yesterday’s data.

The models were fine; the platform under them wasn’t. The people who spend the morning finding out why are your data and ML engineers, and the models they were meant to ship wait another day.

So who should operate this layer, and what does handing it to a managed service actually cover? Here’s how we see it. We’ve been supporting enterprise engineering teams since 2015.

What the service actually covers

A managed service for data pipelines and AI infrastructure runs the platform your data and ML teams build on. It doesn’t build your models, and it doesn’t own your data.

OrchestrationAirflow schedulers, workers, and the job runs that move data on time.

StreamingKafka brokers and consumer groups, including lag that never trips a broker alert.

StorageObject storage and data services such as Cassandra, kept available and maintained.

ComputeThe Kubernetes clusters underneath it all, which run to hundreds at enterprise scale.

ObservabilityMetrics, dashboards, and alerts that show a pipeline is late, not just that a server is up.

Upgrades and patchingVersion upgrades and fixes across every layer, before support deadlines arrive.

Operating it means24/7 monitoring and response, upgrades, runbooks, and environment support for the agreed scope.
It doesn’t meanChoosing your architecture, owning your data policy, or training your models.

AI plans are moving faster than the ability to run them

13%
“Only 13% describe themselves as very or extremely confident in their ability to monitor AI reliability.”
The SRE Report 2026 (Catchpoint / LogicMonitor)
50%
of practitioners count Kubernetes and container orchestration among the areas taking the most engineering effort
The most recent OpsWerks State of SRE Survey

The SRE Report 2026 found that “more than half plan to deploy agentic AI systems in production within the next 12 months.” Confidence in watching those systems hasn’t kept up. Underneath them sit Kubernetes clusters that already take a large share of engineering effort before any new AI workload arrives.

Sources: The SRE Report 2026 (Catchpoint / LogicMonitor), January 2026; Catchpoint, “SRE Report: AI optimism and the economics of effort,” February 2026; the most recent OpsWerks State of SRE Survey

Where data platforms break in day-to-day operations

When a data platform struggles, the cause is often in how it’s run, not how it’s designed. Four places to look.

The upgrade treadmill

Kubernetes and each component on top of it go out of support on their maintainers’ schedules, not yours. Every deadline means testing, sequencing, and rolling an upgrade across every cluster.

When an upgrade slips, the compliance and security fixes that depend on it slip too.

Failures that don’t page anyone

A pipeline can be up while its data is late, partial, or stale. The infrastructure alerts stay green, because nothing on the infrastructure has failed.

The first person to notice is often an analyst downstream, or a model trained on the wrong day’s data.

Two stacks at once

Modernizing a data platform usually means running the legacy stack and the new one side by side, sometimes for years. The same engineers support users on both while trying to finish the move.

Support wins the day to day, so the convergence stalls.

Scale turns chores into programs

One upgrade on one cluster is a task. The same upgrade across hundreds of clusters, each with its own workloads and owners, needs planning, coordination, and people to run it.

Without them, the chore becomes the backlog that never clears.

What you keep, and what a managed service runs

Draw the line so your teams keep every decision, and the service runs the work that follows from them.

Your teams keep

Direction, policy, and risk. These stay with you whoever runs the platform.

  • Model development, training, and deployment decisions
  • Data architecture and platform tooling choices
  • Data quality, retention, and governance policy
  • Security, privacy, and access approvals
  • Workload priorities and the AI roadmap

A managed service runs

Repeatable operations, inside the scope you agree.

  • Day-to-day operations for the agreed platform scope
  • Pipeline monitoring and first response
  • Airflow, Kafka, storage, and data service maintenance
  • Version upgrades and patching
  • Runbooks, documentation, and recurring automation
  • Environment support within your approved standards

A quick test: if a task repeats and someone could do it from a written runbook, it can move. If it sets direction or accepts risk, it stays.

Four ways to get the work done

Each option can be the right one. What changes is who owns the result and who manages the people.

Four ways to operate data pipelines and AI infrastructure, compared
Criterion Managed services Staff augmentation Hire in-house Platform tooling alone
Who owns the outcome The provider, against defined service levels You You You
Who manages the people The provider You You You, with no new people added
How cost scales Agreed up front, tied to scope and service levels Per head, per hour Salaries, hiring, and retention Licenses and usage, plus your team’s time
Time to coverage Weeks, through a structured onboarding Weeks, once contractors are found and onboarded Months to hire, then ramp-up As fast as you adopt it, but your team still runs it
What happens to knowledge The provider documents it and keeps it with the team It leaves at the end of each contract It stays, until people move on It sits in the tool’s configuration and your team’s heads
Where it breaks Small teams, or platforms with nothing yet in production You still manage the people and own the result Hiring lags scope, and it takes 10+ FTEs to run sustainable 24/7 coverage in-house A tool flags the problem; someone still has to fix it

Managed services suit teams running production data platforms at scale. If nothing is in production yet, or your team is small, hiring or better tooling may serve you better.

What it looks like at scale

A Fortune 100 technology company had to converge a legacy data stack with a new one, while its developers and SREs supported users on both.

Hundreds to a thousand Kubernetes clusters across major public and hybrid clouds

Hundreds of Kubernetes clusters upgraded in a single month

More than two years of managed Kubernetes upgrades, version 1.24 through 1.32

A month from onboarding to running operations under light supervision

The customer’s developers and SREs stayed focused on converging the two stacks, and raising deprecations early kept their planning on schedule.

“One of the first big things I got to witness was the EKS 1.32 upgrade effort; and honestly, the patience and coordination you all showed was impressive. Upgrading 21 Airflow clusters in 2 days while keeping both teams in sync is genuinely not an easy thing to pull off.”
Senior SRE Manager, Major Technology Company

Due to strict enterprise confidentiality requirements, customer names and organizations are anonymized.

Questions to ask a provider

In the most recent OpsWerks State of SRE Survey, technical depth in their specific stack was the most-cited factor in choosing an operations partner, named by 66% of respondents. These questions test for it.

01

Who runs our pipelines day to day, and will it be the same people next quarter?Ask to meet them. Continuity is how knowledge of your platform builds up instead of leaving with each rotation.

02

How do you know a pipeline is late, not just down?Look for monitoring of data freshness and job outcomes, not only host and broker health.

03

How would you handle a version deadline across all of our clusters?A good answer covers testing, sequencing, and rollback, and when they’ll raise deprecations so you have time to plan.

04

What are you measured and paid on?Look for service levels and outcomes, not hours or headcount. Hours reward long fixes; outcomes reward permanent ones.

05

Where do the runbooks live, and could we take the work back?In your own repository or wiki, where your engineers can read and edit them. If you couldn’t take the work back, find out why before you sign.

AI-Ready Infrastructure: the data platform operations we run, and the decisions your teams keep.

See what we run

Want your data team back on models?

Bring a list of your pipelines, clusters, and upcoming version deadlines. We’ll show you where the operational load sits and what could move first.

You define the outcomes. We own the delivery.