Technology

ClickHouse

A column-oriented DBMS designed for real-time analytical queries, with native support for reading from and writing to S3.

9 connections 4 resources 5 posts

Summary

What it is

A column-oriented DBMS designed for real-time analytical queries, with native support for reading from and writing to S3.

Where it fits

ClickHouse occupies the performance tier above pure lakehouse queries. It can use S3 as a storage backend (S3-backed MergeTree) while maintaining its own columnar indexes for sub-second query performance — bridging the gap between S3 data lakes and dedicated analytics databases.

Misconceptions / Traps
  • ClickHouse with S3 storage is not the same as querying S3 directly. ClickHouse maintains local indexes and metadata for performance; it uses S3 for durability and cost.
  • The S3 table function (for ad-hoc S3 reads) and the S3-backed MergeTree engine (for persistent tables) are different features with different performance characteristics.
Key Connections
  • depends_on Apache Parquet — reads/writes Parquet for S3 interop
  • implements Separation of Storage and Compute — S3-backed storage with independent compute
  • scoped_to S3, Lakehouse

Definition

What it is

A column-oriented database management system designed for real-time analytical queries, with native support for reading from and writing to S3. The 25.x series (early 2026) added **bidirectional Iceberg read/write**, **Delta Lake INSERT** support, and native **Apache Paimon** compatibility — converting ClickHouse from a hot-tier accelerator into a first-class lakehouse engine. ClickHouse Inc. acquired **Langfuse** (LLM observability) in March 2026, planting an LLM-trace-store flag adjacent to the analytics engine.

Why it exists

Some analytical workloads require sub-second query performance on recent data, which pure S3-backed query engines cannot consistently deliver. ClickHouse uses S3 as a storage backend while maintaining its own columnar indexes for speed; with Iceberg/Delta/Paimon integration it can also serve as the query layer over an open-format lakehouse without rewriting source-of-truth data.

Primary use cases

Real-time analytics dashboards backed by S3 storage, log analytics with S3 archival, hybrid hot/cold query patterns, LLM observability stores (Langfuse), bidirectional read/write against open table formats.

Recent developments

Latest signals

  • Latest release: 26.7.4 (current stable, August 2026); 26.3.x is the active LTS line. ClickHouse ships frequent calendar-versioned releases with LTS designations twice a year. Per ClickHouse releases.

  • 26.7 (July 22, 2026) ships a backward-incompatible S3 security change — audit your pipelines before upgrading. S3 access originating from user SQL no longer resolves the server's own cloud credentials (environment variables, IMDS/IRSA, instance profiles, AWS config files, role_arn-based STS, GCP OAuth metadata) by default — such requests must now use explicit credentials or NOSIGN, and named collections default use_environment_credentials to 0. The change closes an SSRF-style exfiltration path where any user query could borrow the server's identity to read cloud storage. Previous behaviour is restorable per-request with use_environment_credentials = 1 plus the new s3_allow_server_credentials_in_user_queries server setting (disabled by default); persistent S3 tables whose definitions relied on server credentials load anonymously on startup instead of aborting. The same release lands QBit Int8 quantization for faster vector search and lets the right side of a JOIN prune left-side granules via min/max primary-key analysis — up to 313× speedups in ordered aggregations. Per the ClickHouse 2026 changelog and ClickHouse Release 26.7.

  • Vector Search GA + 9,000× faster JSON than PostgreSQL JSONB. Per ClickHouse's 2025 roundup, the v25.8 release brought vector search with binary quantization to general availability — ClickHouse joins the analytic-database-with-vector-search shape that Snowflake and Databricks reached earlier, but at OSS-engine speed. The real-time analytics database guide for 2026 reports native JSON throughput 9,000× faster than PostgreSQL JSONB on JSONBench — a real workload pattern for telemetry, event-store, and LLM-trace ingestion (the Langfuse acquisition is downstream of this number).

  • Automatic Global Join Reordering — TPC-H SF100 wins by ~1,450×. Per the same engineering guide, v25.09's Automatic Global Join Reordering posted a ~1,450× speedup on TPC-H SF100 vs the prior planner. v25.10 added runtime bloom filters for an additional 2.1× speedup on selective joins. The cumulative effect: ClickHouse closes the "complex JOIN performance" gap that historically pushed teams toward Snowflake or Databricks for multi-fact-table analytics.

  • v26.4 lands JSON skip indexes + NATURAL JOIN + parameterized Web UI queries (April 30, 2026). Per the Changelog 2026, v26.4 ships MergeTree skip index support for JSON columns (the missing piece for using native JSON as a queryable column type at scale), NATURAL JOIN syntax for terser SQL, commit_order projection index, and parameterized queries in the Web UI. chDB — the embedded ClickHouse runtime — picked up the v25.8 kernel for a reported 61× performance improvement, keeping the in-process analytical-database story competitive with DuckDB.

  • v26.5 (May 21, 2026) — text index GA + quantized-vector (QBit) GA + cluster-distributed vector search. Two search primitives graduated to GA in one release: the full-text (inverted) index and the QBit data type — quantized, bit-packed vector storage for approximate-nearest-neighbor that trades a controlled recall hit for large memory savings on billion-vector sets. v26.5 also lets vector search distribute across cluster replicas, so ANN scales horizontally rather than being pinned to one node. This is the query-engine convergence in miniature: an OLAP engine absorbing full-text + vector on the same Parquet/Arrow-on-S3 substrate that Doris and DuckDB are also racing to own. Per ClickHouse release 26.5 (Changelog).

  • Open House 2026 (May 27) — ClickHouse goes agent-native: Claude-powered Agents plus a JSON-emitting CLI. ClickHouse crossed $250M ARR / 4,000 customers and launched ClickHouse Agents into public beta — a fully managed agentic-analytics service on ClickHouse Cloud, powered by Claude, with a no-code agent builder, sandboxed code interpreter, memory, multi-agent workflows, and native MCP connectivity (built on the LibreChat platform it acquired in November 2025). Alongside it, clickhousectl (beta, April 9, 2026) makes the database operable through a single deterministic, JSON-emitting CLI and ships an agent-skills installer that pushes official ClickHouse skills into 15+ coding agents (Claude, Codex, Cursor, Windsurf). The stated thesis: a database that can't be driven end-to-end by a JSON-emitting CLI risks becoming invisible to the agentic development loop. Per ClickHouse — Open House 2026 Day 1, ClickHouse — $250M ARR + Claude-powered Agents, and Futurum — ClickHouse Builds a CLI to Make its Databases Agent-Native.

  • ClickBench (May 2026) puts ClickHouse ahead of every major cloud warehouse. Self-hosted ClickHouse on c6a.4xlarge takes 32 category wins at 148ms median query latency; ClickHouse Cloud (2 nodes) takes 38 wins at 109ms median. Snowflake, Redshift, BigQuery, and Databricks all trail 3-5x on median latency despite larger configurations, and the cost gap widens with scale — 7-13x cheaper at 10B rows, 23-32x at 100B rows. Per The fastest OLAP databases in 2026 (ranked by ClickBench).

  • ClickHouse now markets itself as "the leading database for AI," citing Claude 4. ClickHouse's homepage states Anthropic used ClickHouse in developing and shipping Claude 4, alongside Tesla, Lyft, and Cisco as named users — consistent with the agentic-analytics push (ClickHouse Agents, clickhousectl) already noted above. Per ClickHouse homepage.

  • Native JSON performance goes deeper than the headline 9,000x number. A PR-by-PR analysis of the JSON type's path to GA (v25.3) finds v25.8's shared-data serialization delivers 58x faster reads and 3,300x less memory over the earlier column implementation, with JSONBench also showing ClickHouse 2,500x faster than MongoDB and 10x faster than Elasticsearch on JSON aggregations. Per ClickHouse Native JSON Support in 2026: A PR-by-PR Analysis (secondary).

  • ClickHouse Cloud's managed PostgreSQL service reached GA in January 2026. The OLTP companion to the analytics engine runs on NVMe-backed storage (up to 10x faster than EBS-backed setups) with failover-safe replication slots that preserve CDC state across HA failovers — pairing a transactional front end with ClickHouse so change data flows into the analytical store without third-party CDC plumbing. Per Unifying OLTP and OLAP (ClickHouse engineering).

  • Lightweight Updates ("patch parts") — up to 1,700x faster than classic mutations. UPDATE-heavy workloads no longer pay the historical rewrite-the-whole-part penalty: updates land as small patch parts and deletes write small mask files for instant consistency with minimal overhead, benchmarking up to 1,700x faster than Classic Mutations and up to 4,000x faster than PostgreSQL on bulk cold updates. Per How to choose a database for real-time analytics in 2026 (ClickHouse engineering). Sources: ClickHouse 2025 Roundup (Vector Search GA) · How to choose a database for real-time analytics in 2026 (ClickHouse engineering) · Unifying OLTP and OLAP (ClickHouse engineering) · ClickHouse Changelog 2026 · chDB Kernel Upgrade to ClickHouse v25.8 · ClickHouse — Open House 2026 Day 1 · ClickHouse — $250M ARR + Claude-powered Agents · Futurum — ClickHouse Builds a CLI to Make its Databases Agent-Native

Connections 9

Outbound 6
scoped_to2
S3Lakehouse
depends_on1
Apache Parquet
implements2
Separation of Storage and ComputeIceberg Table Spec
augments1
Lakehouse Architecture
Inbound 3
depends_on1
chDB
alternative_to1
StarTree Cloud
used_by1
Apache Parquet

Resources 4

DocsHigh
clickhouse.com/docs

Official ClickHouse documentation covering the column-oriented OLAP database engine, SQL dialect, and all table engines.

GitHubHigh
github.com/ClickHouse/ClickHouse

The primary ClickHouse repository — one of the most active C++ database projects, with the full analytical engine source.

DocsHigh
clickhouse.com/docs/en/integrations/s3

ClickHouse's dedicated S3 integration page documents the S3 table function, S3Queue engine, S3-backed MergeTree, and S3 disk configuration.

ChangelogHigh
clickhouse.com/docs/en/whats-new/changelog

Detailed ClickHouse changelog tracks every release including S3 engine improvements and storage backend changes.

Featured in

2026-08-24 The Storage Tier Is the New Context Window A vendor published the break-even arithmetic for persisting LLM inference state to object storage: one cache reuse every 94 hours justifies the spend. Between that number, NVMe object tiers priced like commodities, and streaming engines that write payloads straight to S3, the quiet reclassification of object storage — from cold archive to active memory tier — now has receipts. 2026-07-05 The Frontier Moved Again — and the Floor Compounded In June we argued the AI stack had split into a closed, expensive frontier and an open, cheap floor. Six weeks later the frontier proved the point the hard way: Claude Fable 5 was pulled offline for 19 days by an export-control order, then restored with a retention mandate that quietly bars the most sensitive data from ever touching it. Meanwhile the floor didn't just hold — it compounded. Object storage got RDMA-fast, the query engines converged on one substrate, and the data plane grew a control plane built for agents instead of humans. This is the field report. 2026-06-16 The Whole Stack Went Open — Weights, Storage, and Sovereignty In the same quarter that a 1.6-trillion-parameter open-weights model landed on top of last year's closed frontier — and Anthropic's Claude Fable 5 promptly sprinted ahead again — the storage layer underneath it went open too: DeepSeek open-sourced its file system, Europe stood up sovereign S3, and the post-MinIO self-hosted stack matured. The frontier and the floor are pulling apart, and the pattern underneath all of it is the oldest pain point this index maps: vendor lock-in. 2026-06-10 The Catalog Became the Control Plane — for Humans and Agents Alike Within one quarter, Apache Polaris, Unity Catalog, and Apache Gravitino all converged on the same role: the catalog is no longer where you look up where a table lives — it's where access is granted, credentials are minted, query planning happens, and AI agents are let in or kept out. The companion shift to 'the catalog is becoming a database': the catalog is also becoming the control plane. 2026-05-01 What Survived MinIO: The Post-Archive S3 Server Landscape, May 2026 On April 25, 2026, MinIO archived its main GitHub repository. This index has tracked Vendor Lock-In as a first-class pain point since launch. Here is what the post-MinIO Apache-2.0 server set actually looks like — Alarik, RustFS, SeaweedFS, Garage — and what the fragmentation tells us about what we believed open-source meant.