// time series · AI agents · data engineering · infrastructure
Hi, I'm Sven.
I work on time-series data, AI agents, data engineering, and infrastructure. I write about what I build and measure—what works, what breaks, and what it costs to run.
- time-series
- ai-agents
- data-engineering
- infrastructure
# writing 9 posts
-
How fast can cheap LLMs query ClickHouse? Comparing cheap models on SQL accuracy, response time and cost, then using one to query wildlife and ship data. -
Can LLMs Write ClickHouse SQL? A Zero-Shot Benchmark of 22 Models 22 LLMs generate ClickHouse SQL for 60 questions across six difficulty tiers, zero-shot. Strict and adjusted accuracy, cost per correct query, and the failure modes that actually matter. -
ClickHouse Projections vs Materialized Views: A Practical Benchmark Projections and materialized views benchmarked on 200M rows of web analytics data in ClickHouse: query speedup, storage and ingest cost, and the unmerged-parts case where both are slower than the base table. -
ClickHouse Compression Codecs for Time-Series Data: A Benchmark Five ClickHouse codec configurations benchmarked on 100M rows of time-series data: storage, query latency and ingest throughput, with a decision tree for codec selection. -
Kubernetes Trouble-Cheat-Sheet with 8 most important best practices (PDF, direct download) Download the Kubernetes Trouble-Cheat-Sheet PDF with 8 essential best practices to improve your workflow and prevent future issues -
Benchmarking Hetzner's Storage Classes for Database Workloads on Kubernetes Hetzner Storage Classes Comparison on Kubernetes -
2.5x Performance: PostgreSQL to Clickhouse without Kafka with MaterializedPostgreSQL -
Bun+ElysiaJS vs Node+Express.js Using a Single Real-World Request Dive into my benchmark comparison of Bun+ElysiaJS and Node+Express.js, featuring a real-world request scenario. -
Decoding Plot Data with ChatGPT: A Journey into Chart Analysis Experiments with ChatGPT reading thermostat charts and time-series plots: where prompting helped, and where visual interpretation remained unreliable.