From 72c0eb6dfe0cbab149e40797b87630812b1a01b6 Mon Sep 17 00:00:00 2001 From: Catherine Chambers Date: Tue, 10 Nov 2026 14:54:56 -0500 Subject: [PATCH 1/2] docs: split workload isolation into Workload QoS and Full Workload Isolation The workload isolation page said every deployment gets hardware only you use, and listed Cloud. Cloud SpiceDB pods run on a node pool shared with other Cloud customers; they get reserved CPU and memory, not their own machines. Dedicated runs in a cloud account and cluster dedicated to the customer, and production permission systems get node pools of their own. - Rename the page to "Workload Isolation and QoS" (same URL) and describe both levels, with anchors for each. - Split the Picking a Product matrix row into Workload QoS (Cloud and Dedicated) and Full Workload Isolation (Dedicated only). - Correct the Deployments page, which said Cloud runs on hardware dedicated to your organization. Co-Authored-By: Claude Opus 5.5 Signed-off-by: Catherine Chambers --- app/authzed/concepts/deployments/page.mdx | 7 +-- .../concepts/workload-isolation/page.mdx | 44 ++++++++++++++++--- app/authzed/guides/picking-a-product/page.mdx | 6 ++- 3 files changed, 45 insertions(+), 12 deletions(-) diff --git a/app/authzed/concepts/deployments/page.mdx b/app/authzed/concepts/deployments/page.mdx index 9eb1a8de..a01ceb21 100644 --- a/app/authzed/concepts/deployments/page.mdx +++ b/app/authzed/concepts/deployments/page.mdx @@ -45,7 +45,8 @@ There is no global endpoint that automatically routes requests to the nearest De ## Workload isolation -Deployments on AuthZed Dedicated and Cloud run on hardware dedicated to your organization, separate from the shared control plane services (such as the Management Dashboard and observability tooling). -See [Workload Isolation] for details. +Deployments on AuthZed Cloud and Dedicated run on nodes separate from the shared control plane services (such as the Management Dashboard and observability tooling), with CPU and memory reserved for each SpiceDB pod. +On AuthZed Dedicated, the cloud account and cluster are also dedicated to your organization; on AuthZed Cloud, SpiceDB nodes are shared with other Cloud customers. +See [Workload Isolation and QoS] for details. -[Workload Isolation]: workload-isolation +[Workload Isolation and QoS]: workload-isolation diff --git a/app/authzed/concepts/workload-isolation/page.mdx b/app/authzed/concepts/workload-isolation/page.mdx index f5c1bc92..94c6ade8 100644 --- a/app/authzed/concepts/workload-isolation/page.mdx +++ b/app/authzed/concepts/workload-isolation/page.mdx @@ -1,17 +1,47 @@ --- -title: "Workload Isolation" -description: "AuthZed functionality that gives SpiceDB deployments dedicated hardware to guarantee performance and prevent noisy neighbor problems." +title: "Workload Isolation and QoS" +description: "How AuthZed Cloud and Dedicated protect SpiceDB deployments from noisy neighbors, with reserved resources on Cloud and fully isolated infrastructure on Dedicated." --- import YouTube from "@/components/youtube-wrapper"; -# Workload Isolation +# Workload Isolation and QoS -Workload Isolation gives your SpiceDB deployments access to hardware that is only used by you, to guarantee performance and prevent "noisy neighbor" problems. +Other workloads running on the same machines as SpiceDB can slow down your permission checks, especially at the tail of your latency distribution (p99). +This is often called the "noisy neighbor" problem. + +AuthZed protects your SpiceDB deployments from noisy neighbors at two levels: + +| Level | [Cloud] | [Dedicated] | +| --------------------------------------------------- | :--------: | :---------: | +| [Workload QoS](#workload-qos) | | | +| [Full Workload Isolation](#full-workload-isolation) | | | + +[Cloud]: ../guides/picking-a-product#cloud +[Dedicated]: ../guides/picking-a-product#dedicated + +## Workload QoS + +Every SpiceDB pod gets a fixed amount of CPU and memory that is reserved for it. +Other workloads can't use that share, and your pods are the last to be evicted if a node runs low on memory. + +SpiceDB also runs on nodes that are separate from AuthZed's own platform services (see [Control Plane](#control-plane)), so management and monitoring work never competes with your permission checks. + +On AuthZed Cloud, SpiceDB nodes are shared with other Cloud customers' deployments. +Your CPU and memory are reserved, but other resources on a node, such as CPU caches, memory bandwidth and network bandwidth, are shared. + +## Full Workload Isolation + +AuthZed Dedicated runs in a cloud account and Kubernetes cluster dedicated to your organization, so no other customer's workloads run on your machines. + +Within that cluster, permission systems marked as production run on node pools used only by that permission system. +Development permission systems share a node pool with your other development permission systems. + +Full Workload Isolation includes Workload QoS: SpiceDB pods on Dedicated also get reserved CPU and memory. ## Control Plane -Services that are secondary to serving real-time requests run one shared workload called the control plane. +Services that are secondary to serving real-time requests run as one shared workload called the control plane, on nodes separate from SpiceDB. These services include: @@ -19,10 +49,10 @@ These services include: - Observability services for the SRE team (e.g. Prometheus, Grafana, Jaeger) - Compliance infrastructure -## Workloads - ## Example Architecture Diagram +This example shows an AuthZed Dedicated cluster, with SpiceDB pods and control plane services running on separate nodes. + ![Workload Isolation Diagram](/images/workload-isolation-diagram.png) ## Low-latency Workloads on Kubernetes diff --git a/app/authzed/guides/picking-a-product/page.mdx b/app/authzed/guides/picking-a-product/page.mdx index 4273e316..46be28f0 100644 --- a/app/authzed/guides/picking-a-product/page.mdx +++ b/app/authzed/guides/picking-a-product/page.mdx @@ -24,7 +24,8 @@ This document is designed to give a high-level overview of the features supporte | ↳ [Event Streams (Early Access)] | | | | | | [Management Dashboard] | | | | | | [Private Networking] | DIY | | | DIY | -| [Workload Isolation] | DIY | | | DIY | +| [Workload QoS] | DIY | | | DIY | +| [Full Workload Isolation] | DIY | | | DIY | | [Automated Updates] | DIY | | | DIY | | [SOC2 Compliance] | DIY | | | DIY | | [Audit Logging] | | | | | @@ -51,7 +52,8 @@ This document is designed to give a high-level overview of the features supporte [Security Embargo]: ../concepts/security-embargo [Expedited Support]: https://authzed.com/pricing#support-packages [Private Networking]: ../concepts/private-networking -[Workload Isolation]: ../concepts/workload-isolation +[Workload QoS]: ../concepts/workload-isolation#workload-qos +[Full Workload Isolation]: ../concepts/workload-isolation#full-workload-isolation [Restricted API Access]: ../concepts/restricted-api-access [Rate Limiting]: ../concepts/rate-limiting From 1ef6f6b26d5ace62c0ba65905b7de46dcd77c2f4 Mon Sep 17 00:00:00 2001 From: Catherine Chambers Date: Mon, 16 Nov 2026 11:14:29 -0500 Subject: [PATCH 2/2] docs: say Cloud SpiceDB pods get pinned CPU cores Co-Authored-By: Claude Opus 5.5 Signed-off-by: Catherine Chambers --- app/authzed/concepts/workload-isolation/page.mdx | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/app/authzed/concepts/workload-isolation/page.mdx b/app/authzed/concepts/workload-isolation/page.mdx index 94c6ade8..938f215f 100644 --- a/app/authzed/concepts/workload-isolation/page.mdx +++ b/app/authzed/concepts/workload-isolation/page.mdx @@ -28,7 +28,8 @@ Other workloads can't use that share, and your pods are the last to be evicted i SpiceDB also runs on nodes that are separate from AuthZed's own platform services (see [Control Plane](#control-plane)), so management and monitoring work never competes with your permission checks. On AuthZed Cloud, SpiceDB nodes are shared with other Cloud customers' deployments. -Your CPU and memory are reserved, but other resources on a node, such as CPU caches, memory bandwidth and network bandwidth, are shared. +Each SpiceDB pod is pinned to its own whole CPU cores, which no other workload on the node can run on, and its memory is reserved. +Other resources on a node, such as CPU caches, memory bandwidth and network bandwidth, are still shared. ## Full Workload Isolation