Both are vectorised columnar engines and both are absurdly fast on one machine. The split is the deployment model. DuckDB is a library: analytics inside your process, against files, on a laptop or a Lambda, zero servers. ClickHouse is a service: continuous ingestion, materialized rollups, and thousands of concurrent dashboard queries against a cluster that is always on. Pick the shape, not the speed.
The DuckDB vs ClickHouse comparison is unusual because the engines agree on almost everything (columns, vectorised execution, SQL) and the products share almost nothing. One is an embedded dependency you version-bump; the other is infrastructure you run, scale, and page someone about. Most teams that frame this as a performance question are facing an architecture question.
A library. pip install duckdb, import, query: the engine runs inside your Python, Node, Java, or Wasm process, even in the browser. There is no server, no port, no auth layer, because there is no separate thing to connect to.
A server (or a fleet). Clients connect over HTTP or the native protocol; clusters add Keeper for coordination, replicas, and distributed tables. ClickHouse Cloud removes the hosting but not the client-server shape.
A single-file columnar format with compression, plus first-class external reads: Parquet, CSV, JSON, Iceberg, Delta, over local disk, HTTP, or S3. For many users DuckDB never "stores" anything; it is a query engine over files they already have.
MergeTree: inserts land as sorted immutable parts, background merges consolidate, a sparse primary index skips data at scan time. Specialised codecs, TTLs, and projections layer on. One of the most refined storage engines for analytics, and you tune it accordingly.
One process, one machine. Vectorised, multi-threaded, and spills to disk past memory, so a big instance carries you further than people expect, but there is no clustering and none is planned. The ceiling is real.
Sharding and replication are native; production clusters run to petabytes and trillions of rows. When data outgrows a node, ClickHouse keeps going and DuckDB simply stops being the right tool.
Batch-shaped: load files, append from a pipeline step, rebuild. A single writer per database and no streaming-ingest machinery: fine for hourly jobs, wrong for firehoses.
Built to drink from the firehose: millions of rows per second per node, a Kafka table engine, async inserts for many small writers, and reads stay fast while writes pour in. This is the workload ClickHouse was invented for at Yandex.
Concurrency is per-process: your app can run parallel reads, but DuckDB is not a shared server many clients hammer at once. Embedding one engine per worker works; pretending it is a warehouse endpoint does not.
Designed to serve: thousands of concurrent queries with workload isolation via quotas and settings profiles. Customer-facing analytics, every user slicing their own dashboard, is a flagship use case.
Standard views only; no incremental materialized views. Rollups are pipeline steps you re-run: idiomatic in a batch world, a gap if you wanted always-fresh aggregates.
Incremental materialized views compute aggregates at insert time, so the rollup is ready before anyone queries it. With TTL-based downsampling, raw data ages out while summaries stay. A killer feature for metrics and clickstream.
The friendliest SQL in analytics: GROUP BY ALL, SELECT * EXCLUDE, reading a CSV by passing its filename, zero configuration, instant startup. Tight pandas/Polars/Arrow interop makes it the default engine of the Python data stack.
Rich SQL with hundreds of analytical functions, but also a server to configure and MergeTree decisions to get right (ORDER BY keys, partitioning, merges). clickhouse-local offers a taste of the file-querying workflow without the server.
MIT-licensed with a foundation holding the IP; an extension ecosystem (httpfs, spatial, Iceberg, community extensions) and DuckLake, DuckDB Labs’ SQL-metadata lakehouse format. Adoption rode the Python data wave: it is in everything now.
Apache 2.0 core with a well-funded company driving fast releases; integrations across Kafka, Grafana, dbt, and every BI tool; chDB embeds ClickHouse in-process: the company’s explicit answer to DuckDB. Both projects are healthy; neither is a risky bet.
Each project has reached into the other’s territory, which is why the comparison stays interesting.
ClickHouse can play the embedded game: clickhouse-local queries Parquet and CSV files from the command line with no server, and chDB, now developed inside ClickHouse Inc., packages the whole engine as an in-process Python library, a direct answer to DuckDB’s pip-install ergonomics. From the other side, DuckDB keeps growing upward: larger-than-memory execution, Iceberg and Delta readers, and DuckLake, a lakehouse format that keeps data in Parquet while putting the metadata in a real SQL database. The honest reading: ClickHouse versus DuckDB on a single node is closer to a coin flip on raw speed than either fan base admits, and ClickBench, ClickHouse’s own published benchmark suite, shows both in the top tier depending on query and hardware.
So the tiebreakers in the overlap zone are not speed. They are: who maintains it (a server needs an owner; a library needs a version bump), what the query concurrency looks like (one analyst vs a thousand dashboard sessions), and whether data arrives in batches or as a stream. A single beefy ClickHouse node serving Grafana panels around the clock is doing a job DuckDB structurally cannot; a DuckDB process inside a Lambda crunching one tenant’s Parquet on demand is a job where a ClickHouse server would be pure overhead.
The least-regret pattern we see: DuckDB for exploration, local pipelines, and anything embedded; ClickHouse when the same queries graduate to always-on, multi-user serving. The SQL dialects are close enough that the migration is mostly DDL, which is also why getting this choice wrong early is survivable.
Data team querying a Parquet lake for internal analysis
Query the lake in place from notebooks and dbt. No cluster bill, no ingestion pipeline, no ops.
Product feature: per-customer analytics dashboards
Concurrent serving with materialized rollups is the core competency. Embedding DuckDB per request reloads data and wastes it.
Observability: logs, traces, metrics at high volume
Streaming ingest, brutal compression, TTL downsampling. Half the observability vendors run on ClickHouse for a reason.
Analytics inside a CLI, desktop app, or serverless function
In-process is the requirement, and DuckDB is the best in-process analytical engine: with Wasm if it has to run in the browser.
Single-node, always-on analytics service in the overlap zone
If it serves many clients continuously, the server shape wins even on one box. Revisit DuckDB only if concurrency stays trivial.
You want both: local dev speed, server-grade prod
DuckDB in development and pipelines, ClickHouse in serving. The pairing is common and the SQL gap is small.
Published Updated