ClickHouse is C . Adding Rust without triggering a full rewrite is harder than it sounds. Alexey wrote about how to do it incrementally, back in December. clickhou.se/3Sw4a6G

🚨 내일입니다! 내일 오후 2시(KST), ClickHouse Cloud의 ClickPipes와 Materialized View를 활용해 더 간단하고 비용 효율적인 실시간 ETL 구현 방법을 라이브 데모와 함께 소개합니다. 지금 등록하세요 → clickhou.se/4fkMSSV

Bangalore, July 11 📍 Join us for Pipes, Streams, and Queries: Engineering Fast Data at Scale - a joint meetup with GlassFlow and Bangalore Streams at Awfis Residency Square, Richmond Rd. On the agenda: • Making Postgres Migration Boring With ClickPipes - Amogh Bharadwaj, Senior Software Engineer @ ClickHouse • From Raw Telemetry to Actionable Insights: Processing OpenTelemetry Streams for ClickHouse - Sasi Teja, Open Source Community Catalyst • ClickHouse in Production: Understanding and Managing Resource Behaviour - Shri Veena M P, Data Platform Engineer @ Platformatory Doors open at 9:30 AM. Lunch and networking after the talks. Registration is via Luma only → clickhou.se/44Ndq9e

Cognitiv processes millions of ad bidding opportunities per second on ~2 petabytes of data, but their data science team was bottlenecked by queries that took hours or days to complete. clickhou.se/4fV8Ywb A while back, they explained how they moved ~2PB to ClickHouse Cloud to power their ML feature store, and what changed for the team building the models.

What are huge pages, and why do they matter for Postgres? The OS hands out memory in 4KB pages, and the CPU caches address translations in a small on-chip table, the TLB. A huge page is the same memory in 2MB or 1GB units, so one entry covers up to 262,144x more ground. Here's how ClickHouse Managed Postgres uses them. 🧵

56 new features, 79 perf optimizations, 366 bug fixes: ClickHouse 26.6 is our 10-year anniversary release. 🎂 clickhou.se/4v60LJn Highlights: hypothetical skip indexes, cascading materialized views, 3x lower latency on deeply nested queries, and a built-in CLI help system.

Vector search where you control precision at query time, not index time. QBit: 6.05 GB → 739 MB, ~2× faster on 29M vectors. We published this back in October. clickhou.se/43VoYHm

🤳 Great user experiences depend on real-time visibility, but retaining complete telemetry shouldn't mean blowing your observability budget. ClickStack makes that possible. Built on ClickHouse, it enables you to retain full-fidelity OpenTelemetry logs, metrics, traces, and session data without sampling, run lightning-fast queries, and reduce observability costs by up to 100x. The result? Better visibility, faster root cause analysis, and lower MTTR, without breaking your budget. 📚️ Join our free APJ-friendly virtual, hands-on Observability with ClickStack Workshop Series (Levels 1–3) on July 8–10 to learn how to deploy, optimize, and scale ClickStack for production. 👉 Register here: clickhou.se/4gjDKPE

Most teams keep operational data in Postgres and analytics in ClickHouse, then stitch the two together by hand. clickhou.se/3QA1CnH ClickHouse Agents now queries both directly, in plain English, in the same request.

ClickStack inherited its schema from the OpenTelemetry Collector. At petabyte scale that became a problem. The team rebuilt it, resulting in 5× faster queries, validated at 1.5 PB/month. clickhou.se/4y1kcFS

Our strategy for unifying OLTP and OLAP isn’t to unify storage just because it sounds “cool”. Each use case is fundamentally different and has its own characteristics. That’s exactly why row and columnar storage formats exist. Unifying storage sounds “flashy” on paper, but whether it works well in the real world is a big question. Enterprise-grade AI applications can be brutal, demanding low latency under high concurrency, everytime, where every trade-off matters. CDC, when done right and with the care like what we did with PeerDB/ClickPipes, can make the integration between OLTP and OLAP feel magical while still leveraging best-of-breed databases, as they should be. Today, there are thousands of customers running multi-terabyte Postgres and ClickHouse workloads using CDC. TL;DR Our approach for unifying OLTP and OLAP is to build the best-of-breed data stack: embrace Postgres and ClickHouse as the leading open source OLTP and OLAP databases, and consistently innovating to make the integration between them as native and seamless as possible for our users. Enterprise-grade CDC is just the beginning. Stay tuned for more (native) innovations that deepen the integration of Postgres and ClickHouse while continuing to let each do what it does best. @ClickHouseDB

PgBouncer uses one CPU core per process, so on a big box the pooler caps out before Postgres does. clickhou.se/4aQiXj9 In ClickHouse Managed Postgres we run a fleet of them on a single port: so_reuseport spreads connections across the processes, and peering makes cancellation work when it lands on the wrong one. Following Kaushik Iska's write-up earlier this week, here's the design measured on real hardware.

Last month, Artemis came out of stealth with $70M in funding to build the future of AI-powered cybersecurity. clickhou.se/4wogIfe See how they're using @ClickHouseDB to run detection queries 69x faster and keep investigative lookups <2 seconds across many TBs of log data a day 👇

Today, @AnthropicAI released @claudeai Sonnet 5, and we’re excited to support it from day one in ClickHouse Agents 🚀 At ClickHouse, we build for real-time agentic analytics: agents that explore live data and produce insights on the fly. As in everything we do, we obsess over performance, so we put the new Sonnet through our internal benchmarks before launch. Against our evals, the new Sonnet delivered the same performance we rely on from Sonnet 4.6 while using fewer tokens overall, with up to 30% fewer tool-call tokens. The reasoning steps are also tighter, which gets users to insights faster. Claude Sonnet has been a sweet spot for this workload for a long time, and we run it extensively in production. For a performance-focused workload like ours, faster and leaner is always better. Congratulations to the @Anthropic team on the launch! We’re excited to put Claude Sonnet 5 in front of our users. clickhouse.com/ai

"Time really matters," says Jua engineering lead Mark Frey. "If you have the data before others do, then you have an edge." clickhou.se/4f7RpYR See how Jua, the AI company behind the world's first universal physics simulation engine, delivers forecasts 3x faster with @ClickHouseDB 👇

Part three of our LogHouse series: 19 PiB -> 431 PiB across 1.59 quadrillion rows. clickhou.se/4vI346C Two years ago, LogHouse stored 19 PiB. Today it stores 431 PiB across 1.59 quadrillion rows, ingesting peaks of 80 GiB/s and 190 million rows/sec.

Can you render ray tracing inside your database? If you use ClickHouse, you can! This video was rendered entirely inside ClickHouse: clickhou.se/4bjNiqe A ray tracer computes lighting math for every pixel, millions of times over, and none of those calculations depend on each other. The query generates one row per pixel, ClickHouse spreads the work across every CPU core at once, then a GROUP BY collapses it into the final image. It is the same pattern as scanning a billion rows and aggregating them, which is what an analytical database does all day. The only difference is that the output is a picture instead of a number. So why do it? Partly because it's fun. But also because it's hard. ClickHouse users push the database hard every day, and projects like this give us novel ways to push it even harder. Building this, Alexey Milovidov found real improvements to make, like better recursive CTEs, optimization-pass fixes for complex queries, and loads of ideas for improving JIT compilation to speed up heavy math. Stuff that improves the core engine for real, everyday queries. It also forced us to finish a long-standing wish: rasterising images straight from ClickHouse with FORMAT PNG. ClickHouse is often used to wrangle the data behind visualisations, and most of the time you reach for a tool like Grafana, Hex or just matplotlib to do the plotting. But sometimes you just want to drop a quick heatmap in Slack, and now you can do that with ClickHouse pushing the pixels itself. Is ClickHouse pivoting to replace Blender? No. But building the fastest database in the world doesn't have to be boring!

We're building Click-a-thon with some incredible companies. @InMobi. @atlys. @SonyLIV. These partners are bringing real-world engineering problems to the table, and 150 developers will spend 24 hours building open-source solutions around them. That's what co-innovation looks like. 📍 Bengaluru | 01–02 August 2026 🏆 ₹10,00,000 prize pool ⏰ Applications close tonight at 11:59 PM IST Register now: clickhou.se/4erivJ9

Databricks claimed ClickHouse “crashed.” We tried to reproduce it. clickhou.se/4gLIRIn It didn't crash. They also provided no hardware, cost, config, settings, or reproducible methodology either. If it can’t be reproduced, it shouldn’t be trusted.