LinkedIn Scales GitOps Beyond Argo CD with Custom Engine, Highlighting Enterprise Challenges
LinkedIn has announced that it has replaced Argo CD with its own custom-built GitOps engine for deploying applications across its extensive Kubernetes infrastructure. This new system manages deployments across more than 250 Kubernetes clusters, supporting over 4,000 services and handling upwards of 100,000 deployments each week. The migration, which took place from July 2025 to April 2026, resulted in a significant reduction in deployment infrastructure failures, from 925 to 63, despite a 3.7x increase in deployment volume. The median time to detect and resolve failed deployments also drastically improved, dropping from 16.1 minutes to 2.45 minutes.
This development is highly significant for DevOps and cloud practitioners, especially those managing large-scale Kubernetes deployments. It signals that while tools like Argo CD are powerful and effective for most use cases, they may encounter architectural limitations when pushed to extreme scales, necessitating bespoke solutions. For organizations approaching similar scales, this serves as a case study in evaluating whether existing tools can meet future demands or if a custom approach is warranted. It also highlights that even mature open-source projects might not perfectly align with the unique operational models and performance requirements of hyperscalers.
This trend aligns with the broader industry movement towards highly optimized, domain-specific platforms as organizations mature their cloud-native operations. While the initial phase of cloud adoption often involves leveraging general-purpose tools, companies operating at LinkedIn's scale frequently find themselves building custom abstractions and engines to achieve further efficiencies and reliability. This is not a rejection of GitOps principles, but rather an evolution of their implementation to meet specific enterprise needs. Uber's Up platform, which assigns each deployment to a dedicated controller, is another example of this trend, indicating a move towards more isolated and specialized deployment logic rather than a shared, general-purpose controller.
In practice, this means that while Argo CD remains an excellent choice for the vast majority of organizations, those anticipating massive growth in cluster count, service diversity, and deployment frequency should carefully consider its scaling characteristics. Practitioners should pay attention to potential bottlenecks such as `managedFields` truncation, noisy-neighbor sync slowdowns, and the overhead of managing numerous Argo CD instances. LinkedIn's insights suggest that per-deployment isolation, rather than just controller tuning, can be key to scaling with compute. While LinkedIn's engine is internal, their findings encourage other large-scale users to investigate issues like field manager counts on long-lived objects and sync latency in relation to total pod count on shared clusters, even without a full rewrite.
Read original source