Tracing an AI request through Cloudflare and its origin
Follow a slow AI request from Cloudflare to an instrumented origin, configure sampling, and verify that the trace preserves the evidence needed to investigate.
Articles
Guides, news and explanations about AI and reliable systems.
Not sure where to begin? Choose a short reading path, or browse at your own pace.
Measuring a latency SLO with Prometheus histograms
Build a latency SLI from Prometheus histogram buckets, include failed requests, and compare classic and native queries with a worked search-service example.
Why ninfer-ext’s smaller EXL3 file can take longer to answer
Match the .ninfer artifact to its loader and hardware, then weigh saved memory against decoding time on short queries and long inputs.
DeepSeek Harness connects a desktop agent to your models and tools
The desktop preview packages a configurable coding agent. Its plugins determine which model answers, which files it can read, and where commands run.
OpenTelemetry tail sampling through a checkout failure
Tail sampling selects recorded traces by error, latency and other evidence. A checkout timeout shows how routing, waiting time and memory affect the trace a responder can retrieve.
Recover the work left behind when OpenShell stops an agent
Containment can stop the next action while an earlier write remains unresolved. Follow the destination, current permission and service outcome before resuming.
When a Collector queue recovers, check the age of what arrives
A persistent queue can save telemetry through an outage. Check delivery, catch-up capacity and event age before letting the recovered data drive an action.
Recovering an automation write after a timeout
A missing response can leave a worker already created. Durable operation IDs and the destination’s idempotency contract let a replacement executor recover the original request.
Disruption budgets and the capacity left behind
Plan an inference-service node drain with a scoped disruption budget, replacement-capacity checks and a practical maintenance worksheet.
NVIDIA introduces NodeWright for Kubernetes host updates
NodeWright can coordinate host updates across a Kubernetes fleet. One labeled worker shows why the package, machine selection and observed result need to stay together.
Prompt caching: measure the cost of useful work
What prompt-cache hits save once retries and failed work are included in task cost, with a local calculator and ledger.
How to evaluate incident-response agents
Build repeatable cases and check recovery independently, with a runnable Python grader.