Dora Noda
Software Engineer
Dora Noda is a Software Engineer writing about Web3 infrastructure, blockchain networks, and self-hosting.
3404 posts · View all authors
Pinterest Ran 66,000 MCP Tool Calls a Month — Here's the Gate Your Deploy Agent Needs
Pinterest's production MCP fleet handled 66,000 tool calls a month across 844 users. Here's the registry-and-gate architecture behind those numbers, and the concrete rate-limit and audit-retention policy a deploy-from-chat PaaS should ship on day one.
Your Deploy Agent's MCP Server Is a Trust Boundary — Here's the Threat Model
A malicious MCP server already backdoored a mail pipeline in the wild. Here's the threat model for what happens when the same protocol holds your deploy and rollback credentials — and the scoping, gating, and logging practices that actually bound the damage.
Fly.io's 'Global Edge' Has a 6x Egress Cliff — and Flat Hetzner Pricing Wins the Workload It's Marketed For
Fly.io bills egress at 6x more to Africa and India than to North America and Europe — exactly the traffic its 'deploy close to your users' pitch is supposed to serve. Here's the worked cost comparison against Hetzner's flat, destination-blind bandwidth.
Kubelet Config Drop-in Directories Just Went GA: What Kubernetes 1.35 Actually Buys a Cluster API Fleet
Kubernetes 1.35 took kubelet's --config-dir flag to GA, letting drop-in KubeletConfiguration snippets replace one monolithic file. Here's the exact Cluster API YAML that uses it, and the node drift it still doesn't fix.
The Kubernetes Descheduler Can Cut Compute Spend 30-50% — Here's the Math on a Real Hetzner Fleet
A worked example turning the abstract 'bin-packing cuts compute 30-50%' claim into real node counts and euros on a Hetzner fleet, plus the PodDisruptionBudget and eviction-cadence config a multi-tenant PaaS needs before enabling it.
Hetzner Raised Prices Twice in 2026: What It Actually Does to the Self-Hosting Cost Pitch
Hetzner's shared-vCPU lines rose 30-38% in 2026; its dedicated-vCPU lines rose up to 173%. Here's the redone self-host-vs-PaaS break-even math by node class, and the rescale trap that quietly reprices a fleet you thought was grandfathered.
Cluster API's Docker Resources Are Deprecated, Not Removed: The Real CAPD→CAPDev Timeline and What to Migrate Before v1.15
CAPD's DockerCluster/DockerMachine types are deprecated as of Cluster API v1.13.0, not v1.12, and removal is targeted for v1.15 with no committed date. Here's the real timeline, the concrete manifest migration, and what a CAPH-based fleet actually needs to touch.
Cursor and Windsurf Can Now Run Your Code in the Cloud. Neither One Can Ship It.
Cursor's Cloud Agents and Windsurf's Devin handoff both run your code in a sandboxed cloud VM and hand back a PR. Here's the documented, vendor-confirmed line where that stops and deploying starts — and what actually closes it.
Depot's Cache Meter vs. Your Own NVMe: What $0.20/GB Actually Costs a Real Cache Footprint
Depot bills its remote build cache at $0.20/GB/month past your plan's allotment. Here's the worked math on what that costs a real dependency cache footprint against a self-hosted fleet's own NVMe, which charges nothing extra for storage it already owns.