Apache Doris vs ClickHouse: Why Companies Are Migrating (And What Changed in 2025)
Both are fast. Both handle big data. So why are companies like Tencent Music switching from ClickHouse to Apache Doris? The answer is more nuanced than you’d expect.
This article breaks down what’s really driving migration decisions in 2025. Not marketing claims. Real architectural differences that matter when you’re running production workloads.
The Single Table vs Multi Table Problem
ClickHouse dominates single table queries. When you need extreme performance on flat tables with straightforward aggregations, it’s hard to beat. But complex multi table joins tell a different story. Apache Doris handles these scenarios significantly better, with performance advantages of 2 to 10 times in real world tests.
Companies deploying both systems found that while Doris and ClickHouse perform similarly on single table queries, Doris maintains consistent query performance under high concurrency where ClickHouse struggles.
This matters because business requirements evolve. Your flat table design works great until someone needs to join customer data with transaction history across multiple dimensions. At that point, ClickHouse’s limitations become operational bottlenecks.
Architecture: Simple vs Flexible
Doris uses a frontend and backend separation architecture. The FE handles metadata and query planning. Multiple FE nodes maintain consistency through Paxos. Adding new BE nodes is straightforward. The system automatically handles data distribution.
ClickHouse started as a single machine system. Building a cluster means configuring distributed tables manually and managing ZooKeeper coordination. This design works well for small deployments but adds complexity at scale.
The operational difference is stark. With Doris, cluster management and scaling are simpler. With ClickHouse, you manage local tables, distributed tables, and ZooKeeper configurations.
Data Consistency: Async vs Sync
ClickHouse handles updates and deletes asynchronously. When you delete a record, it doesn’t immediately disappear from query results. The system waits for background merge operations to complete. This approach optimizes write performance but sacrifices data consistency.
Doris supports synchronous updates and deletes. Data changes are immediately visible. Its UniqueKey model achieves primary key deduplication through Merge on Write, with performance reported as 10 times better than ClickHouse for these operations.
For real time scenarios like user tagging or live dashboards, asynchronous behavior creates problems. Deleted users still appearing in query results impacts business operations.
What Changed in 2025
Both systems shipped major improvements this year. Understanding these changes helps explain current migration patterns.
Apache Doris 3.0 Series
Recent Doris releases focused heavily on lakehouse capabilities, compute storage separation, and query optimizer enhancements. Version 3.0 introduced cloud native architecture that decouples computation and storage layers, enabling physical isolation between different workload types.
The 2025 roadmap emphasizes lakehouse and semi structured data analysis. New features include Iceberg update and delete support, Paimon data write back, and multi Kerberos environment support. Materialized view capabilities expanded with automatic management and improved observability.
ClickHouse 25.x Series
ClickHouse released multiple versions throughout 2025. Version 25.6 brought consistent snapshots across queries, enhanced projection filtering, and JSON in Parquet. Version 25.8 added a faster Parquet reader, Hive style partitioning for writes, and initial PromQL support.
ClickHouse Cloud evolved significantly with compute compute separation for better workload isolation. New CDC connectors for Postgres and MySQL reached public beta. The platform expanded to 25 regions across all major cloud providers.
Iceberg support matured with position deletes, equality deletes, and write capabilities for REST and Glue catalogs. AI powered SQL generation was added to the client, supporting OpenAI and Anthropic providers.
The Migration Pattern
Real migration cases reveal consistent patterns. Tencent Music improved data timeliness and reduced maintenance costs after moving to Doris. The flexible ingestion methods and robust consistency protocol ensured high availability.
Another company solved critical concurrency problems during their migration. ClickHouse would shut down under just 10 concurrent queries. After switching to Doris, they handled 1000 plus concurrent queries with millisecond response times and maintained stable QPS under load.
One insurance company replaced ClickHouse, MySQL, Presto, and HBase with a unified Doris platform. The consolidated architecture simplified maintenance and eliminated the need to join real time and offline data in application code.
Cost Considerations
Benchmark comparisons show Apache Doris requires 10 to 20 percent of the cost compared to ClickHouse for equivalent OLAP workloads. Performance tests across CoffeeBench, TPC H, and TPC DS consistently favor Doris.
Beyond compute costs, operational expenses matter. ClickHouse cloud services run expensive, with high dependency on components like ZooKeeper. Frequent interaction between ClickHouse and ZooKeeper during data ingestion creates stability pressure.
Cost isn’t just about infrastructure spend. Simpler architecture means less engineering time spent on maintenance and troubleshooting.
When ClickHouse Still Makes Sense
ClickHouse isn’t wrong for every use case. For workloads dominated by single table queries requiring extreme performance, ClickHouse delivers. Its columnar storage and vectorized execution excel in these scenarios.
Small scale deployments benefit from ClickHouse’s flexibility. The single machine design works well when cluster management complexity isn’t a concern.
If your queries rarely involve joins and you can tolerate eventual consistency for updates, ClickHouse’s performance advantages matter more than Doris’s architectural benefits.
Decision Framework
Choose Apache Doris when you need:
- Better multi table join performance
- High concurrency support above 100 simultaneous queries
- Synchronous data updates and deletes
- Simpler cluster management and scaling
- Unified handling of real time and batch workloads
Choose ClickHouse when:
- Your workload consists primarily of single table queries
- You need absolute maximum performance on flat table aggregations
- Cluster size remains small and manageable
- Eventual consistency for updates works for your use case
- You have existing expertise in ClickHouse operations
Looking Forward
Both systems continue evolving. Doris is pushing deeper into lakehouse territory with expanded Iceberg, Paimon, and Hudi support. Automatic materialized views and recursive CTE support are coming. AI integration scenarios include vector search and training data management.
ClickHouse focuses on streaming queries, smarter join reordering with statistics, and promoting JSON data type to production status. HTTP event stream support will enable real time data streaming via HTTP.
The competition between these systems benefits everyone. Each pushes the other to improve. Migration decisions depend less on which database is “better” and more on which architectural tradeoffs match your specific requirements.
Key Takeaways
Single table performance matters less than you think once your use cases expand beyond simple aggregations.
Architecture complexity creates hidden costs that exceed raw query performance differences.
Real time data consistency requirements often outweigh marginal performance gains.
Both systems shipped significant improvements in 2025, narrowing some historical gaps.
Migration success depends on matching database strengths to actual workload patterns, not theoretical benchmarks.
The shift toward Apache Doris reflects changing requirements in modern data platforms. Companies need systems that handle diverse workloads without architectural gymnastics. Sometimes the fastest single query performance matters less than consistent performance across query types.
Leave a Reply