projectsveltos

This thread describes how sveltos handles long running tasks in #Kubernetes reconciliation loop. #Controllers are the core of Kubernetes. It’s a controller’s job to ensure that, for any given object, the actual state matches the desired state defined in the object spec. (1)

Release v1.12.0 ift.tt/0Bfkv1Q 🚀 New Features Supply Chain Verification for Helm Charts Sveltos can now verify the integrity and origin of a Helm chart before deploying it, with two mechanisms targeting different chart sources. For charts pulled from OC…

Release v1.11.1 ift.tt/wflGrpB 🚀 New Features Health Checks: Metric-Based Validation ValidateHealth now supports querying a Prometheus-compatible endpoint as an additional data source. Named scalar values are exposed to the Lua evaluate() function via a…

K8s cleaner is a controller that identifies, removes, or updates stale/orphaned or unhealthy resources in a Kubernetes cluster ➤ ku.bz/0m2jdQzWx

Local Platform Engineering on Your Laptop | vind, Sveltos and ArgoCD - by Markof #Kubernetes #PlatformEngineering #ArgoCD #DevOps #Gitops #CloudNative #Sveltos itnext.io/local-platform…

Local Platform Engineering on Your Laptop | vind, Sveltos and ArgoCD by Markof medium.com/p/local-platfo…

Release v1.10.0 ift.tt/EGHZwMt 🚀 New Features OIDC Authentication in Dashboard Users can now log into the Sveltos dashboard using OpenID Connect (Authorization Code Flow with a public client), as an alternative to manual token authentication. PR: dashb…

@foliotrail Loved this "Beach-head services pattern (ksm + Sveltos) — clean place to layer NVIDIA GPU Operator, Network Operator (DOCA/BlueField), ... consistently across every child cluster." and just FYI, Sveltos is used but not owned by Mirantis

$IREN investors research. I just run an AI SWOT analysis on k0rdent public github code (the main product from Mirantis). I still don't know what else do they have internally (I guess not everything is opensource) so take this SWOT analysis with a pinch of salt. Here it goes: k0rdent SWOT — for GPU Cloud at Scale (Blackwell / Vera Rubin) Caveat: k0rdent is a Kubernetes multi-cluster control plane (CAPI-based "super control plane" + observability/FinOps via kof). It is GPU-agnostic by design — it doesn't ship Blackwell/Rubin-specific operators. The analysis is about its fit as a substrate for a GPU cloud. Strengths - CAPI-native, declarative multi-cluster lifecycle — fits the "fleet of GPU clusters across regions/DCs/colos" pattern naturally; same YAML across AWS/Azure/bare-metal/edge. - Hybrid + bare-metal first-class — GPU economics push toward owned/colo bare metal (Blackwell HGX, GB200 NVL72, future Rubin racks); k0rdent's provider model handles this without forcing public cloud. - Beach-head services pattern (ksm + Sveltos) — clean place to layer NVIDIA GPU Operator, Network Operator (DOCA/BlueField), MIG/MPS config, NCCL tuning, Run:ai/KAI scheduler, Kueue, etc., consistently across every child cluster. - kof for observability + FinOps — GPU-hour cost attribution and utilization telemetry are the #1 ops problem at scale; having this in-tree is differentiating. - CNCF/Apache 2.0, Mirantis-backed, active community — neutral governance is important for sovereign/regulated GPU buyers (logsight/SPRIND case is exactly this). - Sovereign-AI proof point already — geo-distributed heterogeneous GPU orchestration is a validated use case in ADOPTERS.md, not a hypothetical. Weaknesses - No native GPU/accelerator abstractions — no first-class CRDs for GPU SKUs, NVLink/NVSwitch topology, fabric manager state, MIG profiles, or InfiniBand/Spectrum-X fabric. Operators must compose this themselves. - No scheduler opinion — gang scheduling, topology-aware placement (NVL72 domain awareness), and fractional GPU sharing require bolt-ons (Volcano/Kueue/Run:ai). Out-of-the-box K8s scheduler is wrong for training. - Young project — kcm is pre-1.0; operators running 10k-GPU production may hit sharp edges in upgrade/rollback paths. - CAPI bare-metal story is uneven — Tinkerbell/Metal3 work but provisioning DGX/HGX/NVL72-class hardware (BMC quirks, NVSwitch firmware, fabric bring-up) is still bespoke. - Limited inference-serving primitives — no built-in equivalent of KServe/NIM/Triton autoscaling patterns; again, layering required. Opportunities - Blackwell/Rubin rack-scale shifts the unit of compute from "node" to "rack domain" (NVL72, future Rubin Ultra NVL576). A super-control-plane that treats a rack as a cluster fits this better than traditional node-centric K8s management. - Sovereign AI / neocloud tailwind — CoreWeave/Lambda/Crusoe-style buildouts and EU/ME sovereign clouds need exactly k0rdent's "fleet of clusters across jurisdictions with policy" shape. - Define GPU-aware templates — shipping opinionated ClusterTemplates for Blackwell HGX, GB200 NVL72, and Rubin reference designs would be a strong moat. - FinOps for GPU-hours is an underserved category; kof could become the reason people pick k0rdent. - Partner with NVIDIA NIM / DGX Cloud Lepton / Mission Control — interoperability story rather than reinvention. Threats - Direct competition is well-funded: Run:ai (NVIDIA-acquired), NVIDIA Mission Control / Base Command, Rafay, Spectro Cloud Palette, CoreWeave's internal stack, OpenShift AI. Several have GPU-specific features k0rdent lacks. - NVIDIA vertical integration risk — Mission Control + BCM + Run:ai may become the default "rack OS" for Blackwell/Rubin, squeezing third-party control planes into a thinner layer. - Hardware churn cadence — Blackwell → Blackwell Ultra → Rubin → Rubin Ultra in ~12-month steps. Any platform that doesn't track NVIDIA's reference architectures lags hard. - Networking is the real battleground — Spectrum-X / Quantum-X / NVLink fabric management is increasingly proprietary; a generic CAPI control plane has limited leverage there. - Bare-metal Day-2 (firmware, fabric, GPU RMA workflows) — if k0rdent stays K8s-layer-only, big GPU operators will keep their own out-of-band tooling and treat it as "just another K8s manager." Bottom line Good substrate, not yet a GPU cloud platform. k0rdent's multi-cluster + bare-metal + FinOps trio is a credible foundation for a Blackwell/Rubin fleet, especially for sovereign and neocloud builders who don't want NVIDIA's full vertical stack. To win the GPU-at-scale narrative it needs: (1) opinionated Blackwell/NVL72/Rubin templates, (2) a default topology-aware scheduling story, (3) deeper bare-metal Day-2, and (4) a clear interop position vs. NVIDIA Mission Control rather than overlap.

Release v1.9.0 ift.tt/cMz6OPU 🚀 New Features Remote URL Support in PolicyRefs Reference YAML content directly from HTTP/HTTPS endpoints: Previously, PolicyRef was limited to ConfigMaps and Secrets, which imposed an ~1 MB size cap. You can now define a r…

Release v1.8.0 ift.tt/mcfG4gu 🚀 New Features GitOps Friendly Kubeconfig Renewal Support for In-place Secret Updates: Previously, Sveltos rotated kubeconfigs by creating a new key (re-kubeconfig) and updating the SveltosCluster spec. This caused drift in…

Release v1.7.0 ift.tt/c4VKPNS ✨ Key Highlights Reconciliation Stability (#1657): Fixed "reconciliation storms" in production environments by implementing a robust NextReconcileTime guard and an in-memory cooldown map to prevent tight loops and high CPU u…

Release v1.6.1 ift.tt/7wSsKql 🐞 Bug Fixes #1641: Hardened Helm Lifecycle: Resolved issues causing Helm releases to stall or fail during upgrades. Improved handling of missing sub-charts within the dependency management flow. Added metadata sanitization…

Release v1.6.0 ift.tt/Ms5WqOo 🚀 Release Notes: Performance & Stability Update This release focuses heavily on infrastructure efficiency and core stability. We have significantly optimized the resource footprint of our edge components and addressed severa…

Release v1.5.1 ift.tt/xh9Wfb1 ✨ New Features & Enhancements Pre and Post-Delete Validation Checks (#1590): Introduces the ability to define health checks that must pass before and after a resource is deleted. This ensures that addons are only removed whe…

Release v1.4.0 ift.tt/NuPSUzO This release introduces significant enhancements to the Sveltos Dashboard, providing deeper visibility into event-driven workflows, and brings critical performance optimizations to EventReports and Pull-Mode cluster management…

Release v1.3.0 ift.tt/7TMdjtF ✨ New Features and Enhancements 🚀 Promotion Enhancements: Pre-Health Check Deployments Introduced the optional PreHealthCheckDeployment field to the ClusterPromotion resource. This allows users to deploy and reconcile spec…

Release v1.2.1 ift.tt/PnrobQD ✨ New Features and Enhancements Dashboard Enhancements Filter deployed resources by namespace, name, kind Filter deployed Helm charts by release namespace and name Filter Profiles by namespace, name Display when cluster …

This tool lets you define add-ons (Helm charts, YAML, kustomize, etc.) in a management cluster and automatically deploy them across multiple target clusters with dependency ordering and drift detection ➜ ku.bz/bfhT9vf44

Release v1.2.0 ift.tt/tJwhLDb ✨ New Features and Enhancements Progressive Rollouts with Cluster Promotion A new Custom Resource Definition (ClusterPromotion) is introduced to enable declarative, multi-stage, progressive rollout of a single configuration…

前職は内製していたけど、こういうツールあったのか👀👀👀 github.com/projectsveltos #CNDW2025

Release v1.1.0 (v1.1.1) ift.tt/BCwrzMQ 🚀 Features Sveltos MCP Server: This release introduces the new Sveltos MCP Server, which provides the ability to trigger debug sessions. This new component enhances the management capabilities of the Sveltos platfor…