cd /news/ai-infrastructure/jsonbench-is-back-and-clickhouse-sti… · home › topics › ai-infrastructure › article
[ARTICLE · art-170468] src=clickhouse.com ↗ pub= topic=ai-infrastructure verified=true sentiment=↑ positive

JSONBench is back, and ClickHouse still leads

ClickHouse's open-source JSONBench, which tests analytical queries over a billion Bluesky JSON documents on identical hardware, has expanded from five to 27 engines, and ClickHouse still leads the comparison nearly two years after the benchmark debuted in January 2025. On the event-type aggregation query, ClickHouse returned results in 661 ms on the cold run while MongoDB took 32 minutes and 26 seconds. The original five-system field included ClickHouse, MongoDB, Elasticsearch, DuckDB and PostgreSQL, each tested on the same Bluesky data with five analytical queries.

by read5 min views1 publishedNov 16, 2026
JSONBench is back, and ClickHouse still leads
Image: Clickhouse (auto-discovered)

TL;DR #

JSONBench has grown from five to 27 engines. Almost two years after its debut, ClickHouse still leads the billion-document comparison on identical hardware. You can run the open-source benchmark, reproduce the results, or add your system.

JSON is the lingua franca for handling semi-structured and unstructured data in modern systems, powering everything from logging and observability to real-time streaming, mobile app storage, and machine learning pipelines.

Document databases such as MongoDB make it easy to start working with JSON. Once the server is running, you can insert documents without declaring their fields or types upfront. The documents don’t have to share the same structure. Getting started feels refreshingly simple.

1# Connect to the local MongoDB server
2mongosh
3
4/* Create a collection */
5db.createCollection("bluesky");
6
7/* Insert JSON documents */
8db.bluesky.insertMany([
9  {"id":1,"author":"Tom","text":"I like pretzels 🥨"},
10  {"id":2,"author":"Mark","text":"I like coffee ☕️"}
11]);

ClickHouse can do this too. Create a table with a single JSON column, and you can start ingesting documents as they arrive, whatever their structure is, without defining the fields inside them upfront.

1# Connect to the local ClickHouse server
2clickhouse client
3
4/* Create a table with a JSON column */
5CREATE TABLE bluesky (data JSON);
6
7/* Insert JSON documents */
8INSERT INTO bluesky FORMAT JSONAsObject
9[
10  {"id":1,"author":"Tom","text":"I like pretzels 🥨"},
11  {"id":2,"author":"Mark","text":"I like coffee ☕️"}
12];

The first documents are easy. Now let’s run an aggregation over a billion documents in MongoDB and ClickHouse.

JSONBench: the billion-document test #

The query below comes from JSONBench, our open-source benchmark for analytics on JSON.

For this test, we use real Bluesky events with their original nested JSON structure, rather than the simplified posts above. Each document represents an event such as a post, like, or repost. We’ll count which event types occur most often.

The corresponding MongoDB query:

1db.bluesky.aggregate([
2  {$group: {_id: "$commit.collection", count: {$sum: 1}}},
3  {$sort: {count: -1}}
4]);

The ClickHouse query:

1SELECT data.commit.collection AS event, count() AS count
2FROM bluesky
3GROUP BY event
4ORDER BY count DESC;

The result is a list of event types and their counts, ordered from most common to least common.

1event                      count
2app.bsky.feed.like          …
3app.bsky.graph.follow       …
4app.bsky.feed.post          …
5…                           …

ClickHouse returns the result in 661 ms on the cold run. MongoDB takes 32 minutes and 26 seconds. Both run on the same machine type, using the published benchmark configurations. The hot runs are shown alongside them.

ClickHouse brings its native columnar storage, compression, and parallel query engine to JSON. Documents retain their flexibility, while queries benefit from reading and processing the fields they need.

JSONBench began with five systems #

We introduced JSONBench in January 2025 with five systems: ClickHouse, MongoDB, Elasticsearch, DuckDB and PostgreSQL. Each was tested on the same Bluesky data with five analytical queries.

In the original comparison, ClickHouse consistently outperformed all other tested JSON data stores in query performance and storage efficiency. Analytical query speedups reached thousands of times, while the stored JSON was even smaller than the compressed source files.

The charts use JSONBench’s combined score, which brings together hot and cold query runtimes, time and storage size. Lower is better. The historical chart applies today’s calculation to the original measurements, omitting time because it was not recorded then.

Almost two years later: 27 systems, ClickHouse still leads #

The competition has had almost two years to catch up. Alexey Milovidov, the inventor of ClickHouse, has now revived JSONBench, merged community contributions and added more systems. The original five have been retested with newer versions and updated configurations. ClickHouse still leads their billion-document comparison.

The field has grown from five to 27 database engines, including OpenSearch, Apache Doris, Databend, SingleStore, SQLite, and Umbra. The next chart shows 14 selected entries with billion-document scores on the same machine type. ClickHouse leads this broader comparison too.

The current charts show Combined scores relative to ClickHouse at 1.0. Lower is better. These are benchmark scores, not query time multipliers; unsuccessful queries and incomplete loads receive penalties.

We haven’t stood still #

Since the original JSONBench, we’ve shipped 178 changes to ClickHouse’s native JSON type, covering performance improvements, new features and fixes.

That includes faster reads and lower memory use, new indexes over JSON field names and values, direct JSON support in Parquet, and the option to add type hints without rewriting historical data.

The chart below shows JSON pull requests in yellow and work on the supporting Dynamic type in gray. Dynamic preserves the original types when a JSON field contains different kinds of values across documents.

And we’re not done. The JSON umbrella issue tracks ongoing work and welcomes feature requests.

JSONBench is now also part of our Versions Benchmark, where you can follow JSON query performance across releases.

Run JSONBench and add your system #

JSONBench is an open-source benchmark with public data, queries, configurations and results. You can reproduce the measurements, inspect how each system is tested, and contribute improvements.

If you know a system can do better, contribute an improved configuration or query. If your database is missing, add it. Vendors and community contributors are welcome.

JSON flexibility at ClickHouse speed #

JSON is the lingua franca for logs, events, applications, and machine learning pipelines. ClickHouse combines the flexibility to store those documents as they arrive with fast analytics and efficient storage at scale. That gives people and agents a common analytical engine for the applications they build, the systems they observe, and the security events they investigate.

Get started with ClickHouse Cloud today and receive $300 in credits. At the end of your 30-day trial, continue with a pay-as-you-go plan, or contact us to learn more about our volume-based discounts. Visit our pricing page for details.

── more in #ai-infrastructure 4 stories · sorted by recency
── more on @clickhouse 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/jsonbench-is-back-an…] indexed:0 read:5min 2026-11-16 · —