Alkesh Chauhan

If Apache Iceberg adoption is fast, the ecosystem around it is growing even faster and that's creating one of the cheapest lakehouse infrastructure options available today To support this ecosystem, we have ClickHouse with its recent updates and how it is shaping our stack of choice . The Cost-Effective Stack: OLake by Datazip + Apache Iceberg + ClickHouse Here's what I'm seeing teams build for maximum cost efficiency: OLake, handles the heavy lifting of real-time database replication - streaming from PostgreSQL, MySQL and MongoDB directly into Apache Iceberg with throughput of 46K+ records/second. It's open-source, requires minimal infrastructure (no Spark, no Flink, no Debezium), and supports all major Iceberg catalogs. Not to state the hero of the stack- Iceberg provides the open table format that eliminates vendor lock-in while delivering ACID transactions, schema evolution, and time travel capabilities. ClickHouse delivers the analytics layer with proven 5-15x cost advantages over traditional warehouses but what I want to highlight are it's recent updates . ClickHouse v25.8 dropped major updates for Iceberg support (full source in comments) : ✅ Native Write Support - Full CRUD operations, not just reads ✅ Production-Ready Catalogs - REST, Glue, Unity all promoted to beta ✅ Schema Evolution - Add/drop/modify columns seamlessly ✅ Better Deletes - Position deletes merged efficiently ✅ Near Real-time Streaming - Perfect for ingestion platforms like OLake Why This Stack Works: Real-time ingestion → Open storage → Fast analytics = Maximum performance at minimum cost. #ApacheIceberg #OpenTableFormats #ClickHouse #DataLakehouse

37

As we move past the initial excitement of how to build and experiment with agents, we are now facing the classic day-two operations problems at enterprise scale: how do we meaningfully observe systems whose failure modes are more semantic than syntactic?     Traditional metrics like latency, error rates, and token counts feel like trying to debug a complex distributed system with only “ping”. They tell you if a service is alive, but not if it’s lucid. Businesses need to answer higher-order questions about factual accuracy and coherence.     Today at .conf, Splunk and AGNTCY | A Linux Foundation Project announced their collaboration to:   — Drive agentic semantic conventions directly into the OpenTelemetry (OTel) schema to establish a common language  —Integrate the AGNTCY Metrics Compute Engine (MCE) that provides the tooling, using an LLM-as-a-judge approach to evaluate qualitative signals like coherence alongside the operational metrics we already know. Consider a banking AI agent handling loan applications. Your operational dashboards might be all green, but the MCE could flag a persistent dip in "factual accuracy" for queries about variable APRs, revealing a silent, compliance-critical failure mode. Now you are no longer just monitoring the health of the container; you're monitoring the integrity of its reasoning.  This foundation becomes even more critical as we move from single agents to multi-agent systems, where specialized agents collaborate to execute complex jobs. Get your hands on the code. Join the workgroup meetings. Provide feedback.   Start with the open-source AGNTCY Observe toolkit on GitHub and instrument an agent here https://lnkd.in/gCUqVsff Photo from .conf today on the main stage highlighting where AGNTCY is delivering with agent observability

109

5 Comments

Many AI teams are scaling LLM inference workloads on-prem for security, performance, and cost savings. ✅ Most will start with: 1. Infrastructure layer: Servers, GPUs, networking, storage, power, cooling, etc. 2. Container orchestration layer: K8s/OpenShift to manage and schedule workloads. ⚠️ What’s often missing? A 𝗽𝘂𝗿𝗽𝗼𝘀𝗲-𝗯𝘂𝗶𝗹𝘁 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺 𝗹𝗮𝘆𝗲𝗿 for LLM inference and agentic systems, the overlooked layer where things often fall short: - Slow time to market - Poor GPU utilization and cost visibility - Performance bottlenecks - Lack of LLM-specific observability - High operational overhead In our latest post, we break down the modern on-prem LLM stack and why the inference platform layer is a critical piece. 👉 Learn more: https://lnkd.in/gBFisHZJ

41