Skip to content

OPA to Kyverno: The business justification

Kyverno, the CNCF-graduated policy engine Nirmata created, writes Kubernetes policy as Kubernetes YAML. Gatekeeper writes it in Rego. That one difference sets who can review a policy, how fast you can ship one and what it costs to run. At Wayfair, it meant a policy in under a day instead of a week, and 75% lower admission controller cost.

For platform and security leaders weighing a move from OPA Gatekeeper to Kyverno.

Policy writing shouldn’t require paying the Rego tax

  • A deploy is blocked and the one engineer who reads Rego is out this week.
  • An exemption for one workload lands in a cluster-wide policy, and every team inherits it.
  • An auditor asks which policies carry exceptions. The answer is spread across templates and tickets.

The question isn’t which engine can express more. It is who can read, review and change a policy a year from now, and what each change costs.

TakeawayPolicy cost is a people cost: who can review it, not what it can express.

AI agents raise the stakes

  • Agents write changes faster than people review them. Coding agents now draft manifests, Helm values and even policies. Admission control is the last check before a change lands. If an agent writes a Rego constraint that only one engineer can read, you aren’t reviewing your guardrails. You’re trusting them.
  • Agents in your clusters need guardrails too. AI agents run with real credentials, and most clusters don’t check whether an action fits the task. Kyverno policies are Kubernetes YAML and CEL, the same expression language Kubernetes uses for ValidatingAdmissionPolicy, so platform, security and app teams can all review them. Read how to govern AI agents in Kubernetes.
  • Who stands behind your engine? In August 2025, the founders and core team behind Styra, the company that created OPA, announced they were joining Apple. OPA and Gatekeeper remain open source. Before your next audit, ask who backports your CVE fixes and who answers a 2 a.m. call.

TakeawayWhen agents write the changes, the policy layer has to be something your whole team can read, and someone has to stand behind it.

One UK bank, same team, both engines

OPA Gatekeeper 6 months
5 policies
Kyverno under 3 months
30 policies in production

Six times the policies in half the time, more than 10x the monthly rate.

Customer experience shared with Nirmata, anonymized. Policy complexity varies, so treat it as one team’s result, not a benchmark.

Wayfair: faster policies, smaller bill

  • 75%cost reduction from lower admission controller resource use after moving from Gatekeeper to Kyverno
  • 1 week → <1 dayAn experienced engineer spent a week on a simple Gatekeeper constraint. The same engineer shipped a Kyverno policy in under a day.
  • 2in-house webhooks retired and replaced with Kyverno policies

The savings didn’t stop at the policy engine. Wayfair uses Kyverno mutate policies to change pod scheduler selectors for better bin packing and to improve resource requests in a widely used sub-chart, across about 2,000 engineers and 15,000 production deploys a month, without asking developers to change a line.

Source: Wayfair’s “Migration to Kyverno” and “Mutating Policy Examples” slides from its KubeCon NA 2023 talk, shown in Why and how Wayfair migrated from OPA to Kyverno. Figures are as Wayfair reported them, not a Nirmata benchmark.

Where the cost sits

Cost driverWith Gatekeeper todayWith Kyverno
Writing and testingAuthoring and review hoursRego, plus a ConstraintTemplate and Constraint per policy. Separate test CLI (gator). Exceptions need custom logic.Kubernetes YAML, with CEL where logic needs it. kyverno test is built in. Anyone who reads manifests can review. PolicyException is reviewed through Git and RBAC.
PerformanceCompute and admission latencyRego evaluation. One team reported Gatekeeper’s in-memory data sync causing bottlenecks.Wayfair cut admission controller resource cost by 75% after the move. CEL is designed to evaluate in nanoseconds to microseconds (cel.dev) and uses less CPU and memory, per Kyverno’s guide. Confirm end-to-end latency in your pilot.
CVEs and supportEmergency upgrades, exposureBackports and 24/7 support come from your platform vendor or a third party, where you have one.Nirmata Enterprise for Kyverno: backported fixes for every supported release, 18 months of patch support, 24/7 SLAs.
Extra toolsTo buy, run and secureNo built-in generation. Image verification and cleanup come from extensions or separate tools.Generate, cleanup and verifyImages (Cosign and Notary) are built in. Wayfair replaced two in-house webhooks with Kyverno policies.
Fleet visibilityAudit and reportingFleet views and audit evidence come from tooling you assemble.Nirmata Control Hub enforces policies across clusters and produces compliance evidence.

Sources: Nirmata Kyverno vs. OPA/Gatekeeper posts; Kyverno evaluation guide; Adevinta tech blog; cel.dev; nirmata.com product pages.

Engineering time

OPA GatekeeperKyverno
Write one policy (with Rego expertise)8 to 12 hrs4 hrs
Maintenance, troubleshooting and tooling, per month40 to 80 hrs10 to 20 hrs
Migrate one existing policy, one time (convert, unit test, live test with Chainsaw)n/a1 to 2 hrs
Annual savings, conservative$72,000360 hours a year
Annual savings, high$216,0001,080 hours a year
Writing policies120 to 240 hrs a year
Maintenance and tooling240 to 840 hrs a year
Payback on migration0.3 to 2.0 monthsSooner for teams that also save audit hours.

Conservative pairs the low OPA estimate with the high Kyverno estimate. High pairs the high OPA estimate with the low Kyverno estimate. Excludes learning time and tools retired, so real savings may be higher. Starting values are examples. Replace them with yours.

Audit and risk

Audit prep savings$40,000 to $80,000Total with engineering: $112,000 to $296,000 a year
Incident-cost equivalent0.03 incidents a yearavoiding this many incidents a year is worth the low total

The default incident cost is the global average in IBM's 2025 Cost of a Data Breach report ($4.44M). Financial services runs higher. We don't estimate how often incidents happen. Enter your own cost to see how many incidents a year the savings are worth. We don't know your incident rate, so this is an equivalent, not a break-even.

A migration you can run beside the old engine

  1. Inventory and mapMatch each constraint to the Kyverno policy library (kyverno.io/policies/gatekeeper/).
  2. Convert the restThe Nirmata policy-as-code AI agent converts custom Rego and generates test cases.
  3. Audit beside GatekeeperRun Kyverno in Audit mode while Gatekeeper enforces, and compare reports.
  4. Switch to EnforceWhen results agree. Gatekeeper stays until then, so there is no cutover gap.

TakeawayYou run both engines until the results match. There is no cutover day.

What to measure in a pilot

  • Authoring time for someone outside the policy team
  • Latency, CPU and memory of the admission controller at your scale (Wayfair cut cost by 75%)
  • Days from CVE disclosure to a patched build
  • Tools you can retire
  • How long “which clusters violate this policy?” takes to answer

Teams that made the move

Wayfair

About 2,000 engineers on GKE, no central policy team. Rego complexity and documentation gaps drove the change. They retooled constraints, ran both engines, moved from audit to enforce, then retired Gatekeeper. Today: about 56 validate rules and 20 mutate policies, with 75% lower admission controller cost.

Source: Nirmata blog and the KubeCon NA 2023 talk.

Adevinta

Gatekeeper’s Assign couldn’t express a toleration mutation. YAML let more people write policies. They froze new Gatekeeper use, migrated gradually and tested with kyverno test and Chainsaw. Watch webhook latency and policy report buildup.

Source: Adevinta tech blog.

Kyverno runs in production at Bloomberg, Coinbase, Deutsche Telekom, LinkedIn, Spotify, Vodafone and Wayfair. More accounts are listed in the Kyverno evaluation guide.

Bring your constraint templates.
We’ll map them to Kyverno.

We’ll tell you what converts cleanly and what to test in a pilot.