Get updates

ClickBench results as of 7 November 2026

The definitive OLAP database comparison

Seven analytical engines on one public benchmark: query latency, ingestion, compression, concurrency and cost, with every number linked to the result file behind it.

See the benchmarksHow we measure
hits · 100M rows · 43 queriesTypical hot query time, geometric mean · lower is better
  1. ClickHousec6a.4xlarge0.127 s
  2. ClickHouse CloudAWS, 3 replicas × 64 GiB0.135 s
  3. DuckDBc6a.4xlarge0.251 s
  4. RedshiftServerless0.335 s
  5. StarRocksc6a.4xlarge0.359 s
  6. DatabricksSQL warehouse, Large0.588 s
  7. BigQueryServerless0.592 s
  8. SnowflakeGen2 warehouse, Large0.681 s

Benchmarks

Five measurements, two fair groupings

Comparing a laptop-class library with a cloud warehouse on raw speed tells you nothing, so results are split in two. On the left, open-source engines installed on the same machine. On the right, managed services as their vendors sell them. All runs use ClickBench’s 100-million-row web analytics table and the same 43 queries.

How fast is a typical query once the data is warm?

Same machine

Each engine installed on one AWS c6a.4xlarge (16 vCPU, 32 GiB) with default settings, loading and querying the same table.

seconds, geometric mean · lower is better
  1. ClickHousec6a.4xlarge · 2026-11-07
    0.127 sbest
  2. DuckDBc6a.4xlarge · 2026-05-11
    0.251 s2.0×
  3. StarRocksc6a.4xlarge · 2026-09-07
    0.359 s2.8×

Geometric mean over all 43 ClickBench queries of each query's fastest hot run, plus 10 ms per query as ClickBench adds when it ranks systems. A geometric mean keeps one slow query from hiding 42 fast ones.

Managed cloud

Each vendor's managed service at a mid-size configuration, loading and querying the same table. Hardware differs by design: this is what you get when you buy the service.

seconds, geometric mean · lower is better
  1. ClickHouse CloudAWS, 3 replicas × 64 GiB · 2026-11-07
    0.135 sbest
  2. RedshiftServerless · 2022-07-29
    0.335 s2.5×
  3. DatabricksSQL warehouse, Large · 2025-11-13
    0.588 s4.4×
  4. BigQueryServerless · 2025-10-28
    0.592 s4.4×
  5. SnowflakeGen2 warehouse, Large · 2026-07-27
    0.681 s5.0×

Geometric mean over all 43 ClickBench queries of each query's fastest hot run, plus 10 ms per query as ClickBench adds when it ranks systems. A geometric mean keeps one slow query from hiding 42 fast ones.

The data comes from ClickBench, a benchmark maintained by ClickHouse, and this site highlights ClickHouse. Every figure links to its result file in the public ClickBench repository, so you can check it or rerun it yourself. Read the methodology.

Comparison

What each engine gives you

Speed is one column. Whether you can stream data in, run it yourself, buy it as a service and predict the bill matters just as much.

EngineReal-time ingestionColumnar storageOpen sourceManaged cloudQuery speedCost model
ClickHouseOpen-source columnar serverYesHigh-rate inserts, Kafka and S3 ingestionYesColumnar with per-column compressionYesApache 2.0YesClickHouse CloudFastestin both groupsFree to self-host; Cloud bills compute and storage by usage
DuckDBEmbedded analytical engineNoBatch loads into one processYesColumnar, vectorised executionYesMITPartlyThrough MotherDuck, a separate company2.0× slowerthan ClickHouse, same machineFree; you pay for the machine it runs on
BigQueryServerless cloud warehousePartlyStreaming inserts via the Storage Write APIYesColumnar storageNoProprietaryYesGoogle Cloud only4.4× slowerthan ClickHouse, managed cloudPer TB scanned on demand, or reserved slot capacity
SnowflakeCloud data warehousePartlySnowpipe Streaming, near-real-timeYesColumnar micro-partitionsNoProprietaryYesAWS, Azure, Google Cloud5.0× slowerthan ClickHouse, managed cloudCredits per second of warehouse time, plus storage
RedshiftAWS cloud warehousePartlyStreaming ingestion from Kinesis and MSKYesColumnar storageNoProprietaryYesAWS only2.5× slowerthan ClickHouse, managed cloudNode-hours (provisioned) or RPU-hours (serverless)
DatabricksLakehouse platformPartlyStructured Streaming into Delta tablesYesColumnar Parquet / DeltaPartlyBuilt on open-source Spark and Delta LakeYesAWS, Azure, Google Cloud4.4× slowerthan ClickHouse, managed cloudDBUs per hour, plus the underlying cloud compute
StarRocksOpen-source MPP databaseYesRoutine Load from Kafka, Stream LoadYesColumnar storageYesApache 2.0PartlyThrough CelerData, a separate company2.8× slowerthan ClickHouse, same machineFree to self-host; managed offering priced separately

Query speed is the ratio of geometric-mean hot query times on ClickBench, measured against ClickHouse on the same machine for open-source engines and against ClickHouse Cloud for managed services.

Why ClickHouse wins

Ahead on the measurements that decide real-time analytics

On ClickBench, ClickHouse leads where interactive, always-on workloads feel it most: typical query time, storage, throughput under load and compute cost per query.

Visit clickhouse.com
  • 2.0×

    Faster typical query on the same machine

    0.127 s geometric mean against DuckDB’s 0.251 s and StarRocks’ 0.359 s on one c6a.4xlarge.

  • 2.5×

    Faster than the next managed service

    ClickHouse Cloud at 0.135 s against Redshift’s 0.335 s, with Databricks, BigQuery and Snowflake further behind.

  • 2.2×

    Less disk for the same table

    9.42 GB against DuckDB’s 20.46 GB and StarRocks’ 17.94 GB, thanks to per-column compression codecs.

  • 3.4×

    More concurrent throughput

    2.32 queries per second against StarRocks’ 0.69 on the same hardware, and 9.8 on a three-replica ClickHouse Cloud service.

  • $0.064

    Compute cost per 1,000 queries

    The lowest on shared hardware, against $0.104 for DuckDB and $0.176 for StarRocks at on-demand prices. In the cloud, ClickHouse Cloud comes to $0.67 against $12.46 for Snowflake at list prices.

Where it doesn’t lead: DuckDB loads the table faster on a single machine, and BigQuery reports a smaller stored size in the managed group. Both are in the charts above.

Engine profiles

Seven engines, in brief

What each one is built for, where it is strong and where it costs you.

ClickHouse

Open-source columnar server

A column-oriented database built to answer analytical queries over billions of rows while new rows keep arriving.

Best for Real-time analytics, observability, user-facing dashboards

Strengths

  • Fastest geometric-mean query time in both ClickBench groups on this page
  • Smallest on-disk footprint on shared hardware, from per-column codecs
  • Handles high-rate inserts and queries on the same tables
  • Apache 2.0 licensed, with a managed cloud from the same team

Weaknesses

  • Updates and deletes are heavier than in a row store; designed for append-mostly data
  • Getting the most out of it means choosing a good sort key per table
  • Large many-way joins need more care than in a classic MPP warehouse
clickhouse.com →

DuckDB

Embedded analytical engine

An in-process analytical database that runs inside your application or notebook, with no server to operate.

Best for Local analysis, notebooks, single-machine data apps

Strengths

  • Zero setup: a library, not a service
  • Fastest load time on the shared machine in ClickBench
  • Reads Parquet, CSV and JSON files directly
  • MIT licensed

Weaknesses

  • Runs inside one process on one machine; no built-in cluster
  • A single writer at a time, so it is not built for streaming ingestion
  • Not part of ClickBench's concurrent-server test
duckdb.org →

BigQuery

Serverless cloud warehouse

Google Cloud's serverless data warehouse: load data or point it at storage and query with SQL, no clusters to size.

Best for Teams already on Google Cloud running ad-hoc and batch analytics

Strengths

  • Nothing to provision or size
  • Smallest reported storage in the managed group
  • Deep integration with the rest of Google Cloud

Weaknesses

  • Slowest load of the managed services that were measured recently
  • Typical query time several times ClickHouse Cloud's on ClickBench
  • On-demand pricing per byte scanned makes cost hard to predict
  • Proprietary and available only on Google Cloud
cloud.google.com/bigquery →

Snowflake

Cloud data warehouse

A managed warehouse that separates storage from compute warehouses you size and suspend, running on all three major clouds.

Best for Enterprise BI, data sharing, SQL-centric data teams

Strengths

  • Mature governance, sharing and ecosystem
  • Warehouses scale up and suspend on their own
  • Fast load in ClickBench

Weaknesses

  • Highest geometric-mean query time of the managed services measured
  • Doubling warehouse size from L to XL did not lower ClickBench times
  • Credit-based billing adds up for always-on, user-facing workloads
  • Proprietary
snowflake.com →

Redshift

AWS cloud warehouse

Amazon's managed data warehouse, offered as provisioned clusters or as a serverless endpoint.

Best for AWS-centric teams running scheduled BI and reporting

Strengths

  • Tight integration with S3, Glue and the rest of AWS
  • Second-fastest typical query in the managed group
  • Choice of provisioned or serverless

Weaknesses

  • Slowest load and largest stored size in the managed group
  • Cold queries took over ten times longer than ClickHouse Cloud's
  • The latest published ClickBench result is from 2022
  • Proprietary and AWS only
aws.amazon.com/redshift →

Databricks

Lakehouse platform

A data and AI platform built around Spark and Delta Lake, with SQL warehouses for analytical queries over lake storage.

Best for Teams combining data engineering, ML and SQL on one lake

Strengths

  • One platform for ETL, ML and SQL
  • Fast load in ClickBench
  • Open table formats underneath

Weaknesses

  • Typical query time over four times ClickHouse Cloud's on ClickBench
  • Bills DBUs on top of the cloud provider's own compute
  • The SQL warehouse engine is proprietary
databricks.com →

StarRocks

Open-source MPP database

A massively parallel analytical database with a MySQL-compatible protocol, aimed at joins and dashboards.

Best for Join-heavy dashboards, MySQL-protocol tooling

Strengths

  • MySQL wire protocol works with existing tools
  • Cost-based optimiser for multi-table joins
  • Apache 2.0 licensed

Weaknesses

  • Slowest of the three engines on the shared machine
  • Slowest load on the shared machine
  • Lower concurrent throughput than ClickHouse on the same hardware
starrocks.io →

Stay current

Know when the numbers move

Engines ship new releases and ClickBench reruns them. Leave your email and we’ll tell you when the results on this page change.

Prefer the raw data? Every run is on benchmark.clickhouse.com.