Kubernetes SLA Credits: What EKS, AKS, GKE, and DOKS Owe You When the Control Plane Goes Down
August 17, 2026
Every managed Kubernetes service has an SLA that covers exactly one thing: the control plane. When that breaks, you are owed a credit, and the four providers handle it four different ways, with commitments ranging from a genuine 99.99% down to a 99.5% that is barely a promise at all. On March 1, 2026, objects struck an AWS data center in the UAE, sparks ignited a fire, and the emergency power cut took 38 services in that region down, EKS among them, before rippling into Bahrain. On August 16, the AWS Health Dashboard logged another multiple-services operational issue in us-east-1, this one tied to DNS resolution of the DynamoDB API endpoint. Control planes fail, and they fail more often than your worker nodes do. Here is what each of the big four owes you when it happens, and how to collect it.
The commitments at a glance
Each provider commits to a monthly uptime percentage for the control plane, then pays a sliding credit if it misses. The tables below use the current SLA documents, including the EKS agreement AWS rewrote on March 20, 2026.
| Provider and mode | Covered | Commitment | Measured in | Credit tiers |
|---|---|---|---|---|
| EKS Standard | Kubernetes endpoint | 99.95% | 5-minute intervals | 10% / 25% / 100% |
| EKS Provisioned | Kubernetes endpoint | 99.99% | 1-minute intervals | 10% / 25% / 100% |
| AKS Standard or Premium, with zones | API server | 99.95% | billing month | 10% / 25% / 100% |
| AKS Standard or Premium, no zones | API server | 99.9% | billing month | 10% / 25% / 100% |
| GKE regional or Autopilot | control plane | 99.95% | 5+ consecutive minutes | 10% / 25% / 50% |
| GKE zonal | control plane | 99.5% | 5+ consecutive minutes | 10% / 25% / 50% |
| DOKS with HA enabled | control plane | 99.95% | 5-minute intervals | 10% / 25% / 100% |
Three gaps stand out. AKS on the Free tier has no financially backed SLA at all. DOKS has no SLA unless you pay for the HA control plane add-on. GKE caps every credit at 50% instead, where the other three will go to 100%. Note also what is not in the table: worker nodes. EKS nodes fall under the EC2 Compute SLA, AKS nodes under the Virtual Machines SLA, GKE nodes under the Compute Engine SLA, and DOKS nodes under the Droplets SLA. A node outage never triggers the Kubernetes SLA.
What the credit is actually calculated on
Here is the number that surprises everyone. The credit is a percentage of what you paid for the control plane, not what you paid for nodes. EKS charges $0.10 per hour per cluster, about $73 a month. GKE charges a $0.10 per hour management fee, about $74.40 a month, and offsets one cluster's fee per billing account with a free credit. AKS charges a similar management fee on the Standard tier. DOKS gives the control plane away and charges $40 a month for HA, the feature that creates the SLA in the first place.
| Provider | Credit applies to | Typical monthly fee | 10% tier | Maximum |
|---|---|---|---|---|
| EKS Standard | control plane fee | $73 | $7.30 | $73 |
| EKS Provisioned (XL) | control plane fee | about $1,280 | about $128 | about $1,280 |
| AKS Standard | cluster management fee | about $73 | about $7.30 | about $73 |
| GKE any mode | cluster management fee | $74.40 | $7.44 | $37.20 (50% cap) |
| DOKS with HA | HA add-on fee | $40 | $4 | $40 |
Your Kubernetes SLA covers the control plane, and the credit is a percentage of the control plane fee you pay, not your node bill. A month-long breach at the 10% tier on a $73 EKS cluster is worth $7.30. Claim it anyway: the evidence packet is the same one your own customer SLAs run on.
EKS: two agreements, and the new one is strict
The standard EKS control plane must hold 99.95% monthly uptime, measured in 5-minute intervals. An interval only counts as unavailable if every connection request to the cluster's Kubernetes endpoint fails during it. One successful request and the interval does not count, so a 5-minute window with a single lucky call can drop out of the math entirely.
The March 20, 2026 update added a second agreement. Clusters on the Provisioned control plane get 99.99%, measured in 1-minute intervals, with a new 8XL scaling tier. That drops the monthly allowance from about 21.6 minutes of downtime to about 4.3 minutes. It is opt-in and costs $1.65 to $13.90 per hour extra depending on tier, which is the one case where a credit can actually mean something: a 10% credit on an XL-tier cluster is over a hundred dollars.
Both agreements pay 10%, 25%, or 100% of the control plane charges, credits under $1 are not issued, and the credit lands against future EKS payments. To claim, open an AWS Support case with "SLA Credit Request" in the subject, include the billing cycle, region, dates and times, the affected cluster IDs, and request logs showing the errors. The deadline is the end of the second billing cycle after the incident, so an outage in March must be claimed by the end of May. Watch the exclusions: restarts, cluster version updates, VPC misconfiguration, bugs in open source Kubernetes itself, and "not following the best practices guide" all void the SLA.
AKS: the tier decides whether you are covered
Azure's Kubernetes SLA only exists on the paid tiers. Free tier clusters run best-effort with no financially backed agreement, which is fine for test clusters and quietly dangerous for anything that grew into production without a tier change. Standard and Premium tiers guarantee the API server at 99.95% when the cluster uses availability zones, and 99.9% without. AKS Automatic clusters come on Standard by default.
The credit schedule is the familiar 10/25/100, calculated against the cluster management fee. Claims go through an Azure support request within two months of the end of the billing month. One detail worth checking in your own environment: if the cluster was built without zones, the commitment is 99.9%, which allows about 43 minutes of API server downtime a month, double the allowance of a zoned cluster.
GKE: the weakest promise, and the version fine print
Google's numbers are the most uneven of the four. A zonal cluster gets 99.5%, about 3.6 hours of allowed downtime per month, which is the loosest commitment in this article. Regional and Autopilot clusters get 99.95%, and Autopilot pods spread across zones get 99.9%. In the Mexico and Stockholm regions everything drops another notch, with regional clusters at 99.9%. Downtime also has to last five consecutive minutes to count; a string of two-minute blips never accumulates.
Credits run 10%, 25%, and 50%, and the monthly total is capped at 50% of the covered service charges, so even a fully dead API server refunds half the management fee at most: $37.20. You must notify Google technical support within 30 days of becoming eligible, and you must attach server log files showing the loss of connectivity. There is also a version trap: the SLA applies only while the cluster runs a minor version offered in the Stable or Regular channels, or in Extended support. A Rapid-channel version has no SLA, so a cluster can fail, breach, and still pay full price.
DOKS: no HA, no SLA
DigitalOcean's Kubernetes SLA is conditional in a way people miss. It applies only to the control plane with high availability enabled. A standard single-replica control plane has no financially backed SLA, full stop. HA costs $40 a month per cluster, and that $40 is what the credit is a percentage of.
The commitment is 99.95% measured in 5-minute intervals, unavailability defined as all requests failing for more than five minutes. Credits are 10/25/100 of the HA charges and can be used on any cluster. Claims go through a support ticket within three months of the end of the billing cycle, the longest window of the four. Exclusions cover maintenance windows, automated upgrades, and anything you initiated yourself.
When the control plane goes dark
The claim workflow is the same shape everywhere, and the failure points are the same too: missed windows and missing logs.
- Record the outage window in UTC the moment it starts. Provider math runs on intervals and consecutive minutes, so your own timestamps decide which tier you land on.
- Capture evidence as it happens: failed API calls, kubectl errors against the endpoint, client logs with timestamps, and the status page incident ID.
- Confirm the cluster qualifies before filing: AKS tier, DOKS HA add-on, GKE version channel, EKS mode.
- Compute the monthly uptime percentage for the cluster and find your tier.
- File inside the window: EKS by the end of the second billing cycle, AKS within two months, GKE within 30 days, DOKS within three months.
The one concrete action that changes the outcome: the next time the control plane blips, screenshot the status page incident, note the UTC timestamps, and put the provider's filing deadline in your calendar before you go back to restoring workloads. The claim itself is ten minutes of copy-paste. The window is the only part that closes.