Production traces dataset

A trace is the step-by-step record of how one request moves through software, like a package scan at every stop. On July 18, 2026, Applied Compute posted a dataset of production traces from coding, quality-assurance, and office workflows for engineers to study. (x.com) Applied Compute said the traces come from real systems and are meant for failure-analysis experiments rather than toy demos. The post had about 1,400 views when it circulated, a small audience for a dataset aimed at infrastructure and evaluation engineers. (x.com) Distributed tracing is the standard way teams follow one job across many services, tools, or agents. OpenTelemetry, the main open-source framework in the field, says it captures traces, metrics, and logs across an application with shared context. (opentelemetry.io) Cloud vendors use the same model in production systems. Google describes Cloud Trace as a service that collects application latency data and shows it in near real time, which is the basic workflow teams use to find slow or broken steps. (docs.cloud.google.com) That makes real trace datasets useful beyond one company’s stack. Engineers can replay them to test observability tools, compare sampling and storage choices, and inspect where multi-step workflows split, retry, or fail under load. (opentelemetry.io; docs.cloud.google.com; qaskills.sh) The same idea is moving into artificial-intelligence systems, where one user request can trigger a chain of model calls, tools, and approvals. LangChain said teams are increasingly building test datasets from production traces to evaluate agent quality and speed root-cause analysis. (langchain.com) Applied Compute’s own pitch centers on that production-learning loop. Its website says the company turns historical data, standard operating procedures, and live execution data into training-ready environments inside a customer’s own systems. (appliedcompute.com) Public production-trace releases are still uncommon, especially outside large cloud or platform operators. One of the better-known precedents is Alibaba’s cluster trace program, which has published multiple versions of production data from roughly 1,300 machines to support research on real datacenter workloads. (github.com) The new post does not turn tracing into a solved problem, but it does put one more real-world dataset into a field that often relies on synthetic examples. For teams debugging complex agent and software workflows, the value is in seeing where actual systems bend before they break. (x.com; github.com; langchain.com)