Research Engineer - Benchmarks

ClickHouse and Hud connect the dots between AI code, production data and engineering decisions. New York, USA, August 13th, 2026, FinanceWire The promise of AI-powered development has largely centered on speed. Coding agents can generate changes quickly, allowing engineering teams to increase the amount of software they produce. But faster code generation also creates a new pressure point: teams need to understand those changes quickly enough to decide whether they belong in production. Hud says AI now generates or assists with 42% of the code developers ship, with that figure expected to reach 65% by 2027. The company's new integration with ClickHouse is built around the idea that production behavior can provide some of the context needed to manage that acceleration. The integration connects ClickStack, ClickHouse's open-source observability stack, with Hud's Runtime Code Sensor. Rather than treating observability as a separate troubleshooting function, the companies are connecting it with workflows for assessing, releasing, and fixing AI-generated software. From Infrastructure Signals to Code-Level Context The distinction between the two platforms is central to the integration. ClickStack provides broad visibility across applications and infrastructure, helping teams determine the service, deployment, or endpoint associated with an issue. Hud operates at the code layer, following function-level behavior and connecting production activity to the changes behind it. Shared trace IDs allow a coding agent to connect the two systems. An engineering team investigating an issue in ClickStack can move directly toward the corresponding code-level context in Hud. "Our users already trust ClickHouse to store and query their Open Telemetry data at scale," said Mike Shi, Head of Observability of ClickHouse. "The shift now underway is from simply watching systems or investigating issues to using that data to make day-to-day engineering decisions. Hud connects observability data to the code and changes behind it, making the entire stack more useful for teams building and shipping software with AI." That approach expands the role of production data. Instead of becoming useful only after a deployment goes wrong, runtime information can help inform decisions about changes before they are released. Making the Release Process More Dynamic The integration supports pre-deployment risk assessment for code changes. Teams can evaluate changes against real-time information about how affected code behaves in production, with higher-risk changes held for additional review and deeper context while safer changes can move faster or be automatically merged. The workflow continues after deployment. ClickHouse and Hud support release verification, automatic reversion upon regression, automated detection and investigation, and agentic workflows that can create pull requests to fix underlying issues. Rollback and remediation workflows are also part of the integration. That means the same connection can support several decisions around a single change: whether it should ship, whether it is behaving as expected after deployment, and what should happen if production behavior diverges from expectations. "AI is accelerating how quickly teams can generate code, but safely shipping it with high confidence requires production context," said May Walter, CTO of Hud. "Hud and ClickHouse bring real production behavior at an unparalleled breadth and depth, so teams can build a production-aware AI SDLC: gating changes before they ship, proactively verifying them once deployed, and fixing issues as they arise - all using real runtime truth. Together with ClickHouse, we are bringing that intelligence across the entire AI SDLC." The Case for a Production-Aware AI SDLC The practical use case becomes clearer when a seemingly small change produces an unexpected production result. A query can slow down, a function can behave differently under a particular workload, or a code path can consume more resources. Hud is designed to detect these issues at the function level and provide forensic context into their causes. ClickStack contributes the wider operational view, helping teams connect the code-level behavior to what is happening across the application. "Like every modern engineering organization, a growing share of our code is now written with AI," said Rom Kadria, Senior Software Engineer, monday.com. "We write code much faster, but the challenge has shifted to shipping just as quickly while maintaining confidence that new code won't cause harm. ClickHouse gives us the wide operational picture at scale, while Hud gives us the runtime intelligence and next-level introspection needed to evaluate and ship AI-generated code confidently. When issues do arise, combining ClickHouse and Hud allows us to triage and resolve them quickly. For a company building with AI, that combination is the obvious choice." Engineering teams can get started by installing the Hud SDK and connecting it to their ClickStack service. Hud's runtime intelligence then flows alongside the OpenTelemetry data teams already collect. The broader proposition is a shift in where production information sits within software development. As coding agents generate more of the code, ClickHouse and Hud are attempting to make real-world application behavior part of the feedback loop that determines how that code is reviewed, released, monitored, and repaired. Contact. Editorial team [email protected]

Hud, an Israeli startup co-founded by Roee Adler, May Walter and Lemonade CEO Shai Wininger, has emerged with $21 million in early-stage funding. The company has developed what it calls the first platform offering real-time, function-level visibility into software behaviour in production environments, designed for both engineers and AI coding agents. The startup's Runtime Code Sensor installs in under one minute without configuration, recording live function-level behaviour, performance, errors and dependencies whilst correlating system issues with root causes in real time. The technology runs with minimal overhead and is already deployed across millions of services. Crucially, the data is structured for AI coding tools like Cursor, Copilot and Claude, enabling them to generate code based on actual production dynamics rather than guesswork.