prometheus

To address redundant data scanning and high query latency caused by overlapping windows in time-series monitoring systems, this paper proposes PromSketch—the first “approximation-first” intermediate caching framework. PromSketch innovatively integrates Count-Min Sketch with sliding-window approximate aggregation to construct a pluggable caching layer, augmented by a rule-aware cache eviction policy and a lightweight metadata index. It achieves efficient coverage for 70% of prevalent time-series aggregation queries, with average approximation error ≤5%. Experimental evaluation demonstrates that, compared to Prometheus, PromSketch reduces query latency and computational cost by two orders of magnitude; relative to VictoriaMetrics, it cuts cost by over 4×. This work establishes the first systematic framework enabling low-overhead, high-coverage approximate query processing for windowed time-series monitoring workloads.

Manually authoring PromQL queries for Prometheus monitoring poses high barriers for engineers, requiring both domain expertise and programming proficiency. Method: This work formally defines the text-to-PromQL task—automatically translating natural language questions into executable PromQL queries—and introduces the first collaborative reasoning framework integrating a domain-specific knowledge graph with large language models (LLMs) to support context-aware domain-specific language (DSL) generation. The framework incorporates prompt engineering, semantic parsing, and query validation to ensure correctness and robustness. Contribution/Results: Evaluated on a novel, manually curated benchmark of 280 real-world questions, the approach significantly outperforms existing baselines. It demonstrates the feasibility and effectiveness of natural language–driven metric analysis in observability, establishing a new paradigm for intelligent, accessible monitoring systems.

This work addresses the challenge developers face in comprehensively evaluating the cost, code quality, and behavioral patterns of AI-assisted programming. The paper presents the first AI observability system grounded in real API invocations, integrating precise token tracking, a multi-provider LLM gateway, a pricing database covering 24 models, a response validation pipeline, and LLM-driven code review analysis. A unified dashboard enables joint insights into cost and quality, while Prometheus-based metrics ensure systematic monitoring. In six months of real-world deployment, the system maintained per-review cost estimation errors within 2% and improved the efficiency of analyzing AI usage patterns by an order of magnitude.

Identifying critical health indicators and configuring effective alerts in microservice systems remain challenging due to heavy reliance on manual expertise and the absence of scalable, automated approaches. Method: This paper proposes a lightweight, fully automated method for critical indicator identification that leverages only historical monitoring metrics and minimal trace data. We introduce an indicator–trace coupling modeling framework that jointly integrates information-theoretic measures (mutual information, conditional entropy) with a topology-aware graph neural network—enabling reliability-sensitive indicator ranking without logs or full multimodal telemetry. Results: Evaluated on mainstream benchmarks including DeathStarBench and OnlineBoutique, our approach significantly improves alert coverage and fault detection rates while reducing SRE effort for manual indicator filtering by over 70%. It establishes a scalable, low-overhead paradigm for intelligent alert definition in cloud-native systems.

Distributed astronomical systems (e.g., the ASTRI Mini-Array) suffer from delayed quality control, reactive anomaly response, and fragmented multi-source telemetry and scientific data. Method: We propose a scalable intelligent monitoring pipeline integrating edge computing, cloud-native technologies (Prometheus, Kafka, Cassandra, InfluxDB), and time-series analytics. We introduce the first real-time application of Isolation Forest for anomaly detection on Cassandra performance metrics and pioneer cross-modal correlation analysis between telemetry and scientific observation data. Contribution/Results: The system enhances operational resilience, enabling millisecond-scale stream processing and visualization; detects early-stage performance degradation; reduces unplanned downtime by >35%; and enables AI-driven proactive maintenance. It establishes a reusable intelligent operations paradigm for next-generation large-scale astronomical facilities.

This work addresses the limitations of traditional real-time analytics systems, which rely on manually defined queries and struggle to proactively uncover the vast array of potential insights within complex, dynamic data streams. To overcome this, the authors propose a multi-agent architecture that establishes a continuous closed-loop process for autonomous insight discovery, encompassing hypothesis generation, compilation of executable analyses, result validation, and visualization. A key innovation is the introduction of a contract-driven design based on typed intermediate artifacts, which ensures modularity, observability, lineage tracking, and secure execution of dynamic analyses. The system leverages Kafka as its event coordination backbone and Flink for stream processing, integrating large language models to power specialized agents. Empirical evaluations in retail, financial, and public data scenarios demonstrate an effective paradigm shift from query-driven to proactive discovery-driven analytics.

Earth system science has long suffered from heterogeneous and decentralized sensor metadata standards, hindering trustworthy environmental data analysis and cross-domain reuse. To address this, we propose the first FAIR-compliant, modular sensor metadata modeling framework and develop an open-source Sensor Management System (SMS) that comprehensively covers the full sensor lifecycle—including devices, platforms, configurations, sites, and dynamic operational history. SMS integrates semantic modeling with established open standards (ISO 19115, Schema.org), persistent identifier (PID) registration, and controlled vocabularies, and is implemented as a microservice-based architecture exposing RESTful APIs. Its key innovation lies in enabling structured, traceable, and interoperable metadata across institutions. Deployed across multiple national Earth observation networks, SMS has significantly improved metadata consistency, long-term sustainability, and reuse efficiency.

This study addresses the lack of systematic evaluation of data quality tools with respect to their measurement capabilities and integration with large language models (LLMs). It presents the first multidimensional assessment framework grounded in real-world enterprise use cases, systematically evaluating six prominent tools—including open-source solutions such as Great Expectations and Deequ, as well as commercial platforms like Informatica and Experian—across dimensions including rule definition, duplicate detection, metric aggregation, and uncertainty handling, along with their LLM integration mechanisms. The findings reveal that commercial tools offer more comprehensive functionality and初步 support for LLM-assisted rule generation, whereas open-source tools provide greater flexibility at the cost of higher implementation effort. Notably, none of the evaluated tools currently enable direct LLM-based data validation. This work provides empirical guidance for selecting data quality tools and advancing their integration with LLMs.

This work addresses the unreliability of developer productivity dashboards, which often stems from ad hoc scripts that introduce undetected silent data gaps, eroding organizational trust. To resolve this, we propose a robust ELT pipeline grounded in DAG-based orchestration and the Medallion architecture, decoupling data extraction from transformation to preserve the immutability of raw data. Our approach introduces a state-driven dependency scheduling mechanism and, for the first time, treats metric pipelines as production-grade distributed systems. We emphasize the critical role of immutable raw history in enabling reliable metric redefinition. This methodology significantly enhances data reliability and freshness while effectively eliminating silent failures, thereby restoring organizational confidence in DevOps metrics.

This work proposes an automated log aggregation and analysis framework based on large language models to address the growing challenge of log analysis in increasingly complex systems, where engineers traditionally rely on domain expertise to manually craft intricate LogQL queries. The framework enables end-to-end generation of LogQL queries from natural language instructions by integrating a hierarchical log knowledge base, natural language understanding, knowledge retrieval, and tool invocation mechanisms. Evaluated on four real-world log datasets, the approach achieves an average accuracy of 76.8%, significantly outperforming existing baselines and demonstrating its effectiveness and practicality for log analysis tasks.