Operations | Monitoring | ITSM | DevOps | Cloud

Qovery now sends your organization ID as the sts:ExternalId when it assumes the IAM role in your AWS account, and the role's trust policy only accepts it on behalf of your organization. Your role is already updated. If you manage it with the qovery-role-creation CloudFormation stack, update the stack to the latest template. Romain is a staff engineer at Qovery working on the control plane and AI integrations. He writes about distributed systems, incident response, and MCP-based tooling.

A code sandbox like Daytona or E2B isolates a process. A sandbox environment isolates a whole stack: services, databases, queues and configuration, cloned for one agent task and destroyed when it is done. Here is how sandbox environments work, what they are good for, and how to run them on Qovery. Romaric founded Qovery to make Kubernetes accessible to every engineering team. He writes about platform strategy, developer experience, and the future of cloud infrastructure.

Railway's disclosure of leaked customer environment variables is a reminder that deployment platforms hold the keys to everything. Here is how secrets leak from platforms, the practices that prevent it or limit the damage, and exactly how Qovery stores secrets across its control plane and your data plane, with AWS Secrets Manager, Parameter Store, GCP Secret Manager, Doppler and Infisical. Romaric founded Qovery to make Kubernetes accessible to every engineering team.

When it comes to multi-tenancy isolation, most platforms only support one approach, and you have to make it work for every customer. Usually that means either all tenants share the same infrastructure and rely on the application to keep them apart, or every customer gets a separate stack that your ops team has to deploy, patch, and monitor on its own. Both approaches work. The trouble is that they're rigid.

The rise of large-language models (LLMs) that can understand and create text, and reason about problems has opened the door for software agents that use the reasoning power of LLMs to interact with the wider world and solve problems. At Kubex, our bread and butter is compute and GPU resource optimization, so it’s natural for us to ask how agents could help our customers more effectively optimize their resources.

If you've spent any time in an engineering Slack lately, you've probably noticed that AI comes up in just about every conversation. Someone's using an assistant to scaffold a new service. Someone else has an agent triaging alerts overnight. And at least one person is asking whether all of this means the team can finally skip the "platform stuff.".

Google, CoreWeave and Nebius all say they are short of GPUs. Here is where capacity still exists in 2026, which GPU clouds offer managed Kubernetes, how to connect a second cloud to your first one, and how to run GPU services on all of them from one platform. Romaric founded Qovery to make Kubernetes accessible to every engineering team. He writes about platform strategy, developer experience, and the future of cloud infrastructure.

Google says it is compute constrained, and GKE teams now hit ZONE_RESOURCE_POOL_EXHAUSTED on CPU machine families, not only GPUs. Here is what the error means, what Google recommends, what AWS can and cannot offer, and how to offload workloads from GCP to AWS with one platform instead of two. Romaric founded Qovery to make Kubernetes accessible to every engineering team. He writes about platform strategy, developer experience, and the future of cloud infrastructure.

TL;DR: Google Cloud’s post of 22nd July 2026 on why AI apps fail in production describes how a YouTube team gets AI prototypes in front of user research in weeks instead of quarters: a prototyping stack that hands each prototype read-only access to live production data through a proxy, keeps its code out of the production binary, and throws the prototype away once the idea is proven.