DNS cold start performance and new AWS VPC connectivity for AI tools

Internet Systems Consortium engineer Ondřej Surý presented DNS cold start behavior at RIPE 92 in May 2026, showing how even a minimal query count for a simple domain can balloon past 200 queries for names relying heavily on CNAME records. The presentation focused on the tension between resolver resilience and speed during initial cache population.

A cold start recursive resolver running Bind 9.20 needs four queries to resolve a three-label name like www.potaroo.net, using query name minimization. But the same resolver took 94 queries to resolve teams.microsoft.com, a name that uses a chain of four CNAME records. Surý noted Bind 9.21 has reduced that count to 29 queries through improved caching logic.

CNAME chains and CDN fragmentation

Content Distribution Networks rely on CNAME records to decouple domain labels from zones, but each CNAME forces the resolver to restart resolution from the root for the target name. For teams.microsoft.com, the resolver walks through teams.office.com, tmc-g2.tm-4.office.com, a dual-s-msedge.net subdomain, and finally s-0005.dual-s-msedge.net. Each hop adds latency proportional to the number of zone cuts.

Key takeaways from the analysis include:

• Four CNAME records in sequence cause five independent name resolutions from potentially uncached zones.
• A cold resolver must fetch root, .net, .com, and each subdomain's NS records before reaching the final A record.
• Bind 9.21's query path optimizations cut queries by more than two thirds compared to 9.16.
• Operators with high cache hit rates rarely see cold start penalties, but first user requests after restart can add seconds of latency.

AWS expands VPC and monitoring for AI endpoints

Amazon Quick, the company's AI assistant, now supports VPC connectivity for Model Context Protocol servers. Previously limited to public internet endpoints, enterprises can host MCP servers inside their own VPC and connect them to Quick without exposing traffic to the public internet. The change allows organizations to use proprietary data sources and internal tools with the assistant.

Amazon Bedrock, meanwhile, added CloudWatch metrics for its bedrock-mantle endpoint, which supports the OpenAI Responses API, OpenAI Chat Completions API, and the Anthropic Messages API. Customers can now monitor inference traffic the same way they monitor the bedrock-runtime endpoint. Both updates reflect broader demand for private networking and observability in enterprise AI deployments.