Agents-OSS
EN
Submit

ClickHouse vs Langfuse

Both are catalogued under Observability & Evals. The figures come from the GitHub API; the assessments are ours.

ClickHouseColumnar SQL database that agent tracing stacks are built onOfficialApache-2.0LangfuseTracing, prompt management and evaluation for LLM applicationsOfficialMIT

At a glance

At a glanceClickHouseLangfuse
LicenseApache-2.0MIT
LanguagesC++, PythonTypeScript, Python
DeploymentSelf-hosted / Runs locally / Managed cloudSelf-hosted / Managed cloud
MaturityEstablishedEstablished
Stars49.3k33.2k
Star growth over the last 7 days—+7 ★
Forks8.8k3.6k
Open issues6.9k781
Last commit16 Aug 202615 Aug 2026
ActivityActiveActive

What each one does

ClickHouse

A column-oriented engine for append-heavy, high-cardinality data, which is the shape agent telemetry takes: one row per model call, filtered later by model, cost or error. Langfuse moved its traces here from Postgres in December 2024 and ClickHouse acquired the project in January 2026, so a self-hosted observability stack increasingly means running this underneath. It is the wrong choice for anything transactional — updates and deletes rewrite whole column parts rather than edit rows — so treat it as where agent runs land, not as the database your application writes state to.

Full entry →

Langfuse

Records what an agent actually did — every call, tool invocation and cost — and turns those traces into evaluation datasets. The managed service runs the same codebase you can self-host with Docker Compose, so migrating in either direction stays cheap. Some enterprise features sit outside the MIT core; check which tier you need before committing.

Full entry →

What you can do

ClickHouse

  • Store one row per model call — Columnar storage means a query filtering observations by model, cost or error status reads only those columns while multi-megabyte prompt and completion payloads stay on disk, and the engine is built for large batch inserts merged in the background rather than single-row writes — which is how a queued trace ingestion pipeline wants to write.
  • Inherit the storage layer under Langfuse — Langfuse moved tracing data out of Postgres into ClickHouse in December 2024, and ClickHouse acquired Langfuse on 16 January 2026, saying the core stays MIT-licensed and self-hostable at production scale. ClickStack, its OpenTelemetry-native observability stack, runs on the same engine.
  • Search embeddings without a second database — A vector_similarity index builds an HNSW graph over Array(Float32), Array(Float64) or Array(BFloat16) columns with L2Distance, cosineDistance or dotProduct, available from version 25.8. The index has to be loaded from disk into memory in full to serve a search, and building it slows inserts and merges.
  • Expect exact scans once you filter — Pre-filtering by metadata is, in the documentation's own words, an unsolved problem — ClickHouse falls back to exact nearest-neighbour search — and post-filtering can return fewer rows than the LIMIT asked for when candidates fail the WHERE clause. If narrow metadata filters over a large collection are your main access pattern, a dedicated vector store handles it better.
  • Let an agent query it over MCP — ClickHouse publishes mcp-clickhouse in a separate repository, exposing list_databases, list_tables and run_query plus a chDB tool for embedded queries. It runs read-only unless CLICKHOUSE_ALLOW_WRITE_ACCESS=true is set, with DROP and TRUNCATE behind a second flag.

Langfuse

  • Instrument an app function by function — Importing from langfuse.openai import openai captures each OpenAI call with its model parameters, and every function you wrap in @observe() is recorded as its own nested step — the README's main() and story() pair — so retrieval and agent code appears in the trace only once it is decorated too.
  • Version prompts outside your code — Prompt Management holds prompts centrally with version control, and caching on both server and client means fetching the current version adds no latency to your application.
  • Turn traces into test sets — Datasets build benchmarks for pre-deployment testing, scored with LLM-as-a-judge, code evaluators, manual labelling or collected user feedback.
  • Retry failures in the playground — From a bad result in a trace you can jump straight into the LLM Playground and re-run it against a different prompt or model configuration.
  • Self-host on your own infrastructure — docker compose up for a local instance, a Helm chart as the preferred production path on Kubernetes, and Terraform templates for AWS, Azure and GCP.

Other comparisons