[VELODB.IO]
DATANOMIX.PRO // BLOG // BENCHMARKS

Apache Doris Up to 34x Faster Than ClickHouse in Real-Time Updates

SSB and ClickBench analysis: why update mechanics define modern OLAP performance.

Prepared by:
Datanomix.pro
Reading time:
~10 min

Key Facts from the Benchmark

  • SSB: Apache Doris shows up to 34x speedup vs ClickHouse ReplacingMergeTree in update-heavy workloads.
  • ClickBench: Apache Doris is 1.7x-4.6x faster depending on update ratio and resource profile.
  • Within ClickHouse, MergeTree consistently outperforms ReplacingMergeTree, exposing the cost of update semantics.
  • Business impact: more predictable latency under continuous updates for BI, fraud, and risk workloads.

1. What was tested and why

The VeloDB team benchmarked Apache Doris against ClickHouse in update-heavy analytics scenarios: new events, corrected transactions, customer profile changes, and operational data refreshes.

The objective was straightforward: identify which engine better sustains low query latency while continuously ingesting updates.

2. Test environment

  • VeloDB Cloud (Apache Doris): 1 node, 16 vCPU, 128 GB RAM
  • ClickHouse Cloud: 2 nodes × (16 vCPU, 64 GB RAM each)
  • ClickHouse Cloud: 2 nodes × (8 vCPU, 32 GB RAM each)
  • Benchmarks: SSB and ClickBench; update ratios: 25% and 100%

The benchmark included both CPU-aligned and memory-aligned comparisons to reduce configuration bias.

3. Update mechanism: Doris vs ClickHouse

Apache Doris uses Unique Key + Delete Bitmap. Older record versions are marked at write time, so query-time deduplication overhead is minimized.

ClickHouse relies on ReplacingMergeTree for update-like behavior. New versions are appended, while consolidation depends on background merges and often requires FINAL semantics.

In practice, this means update-heavy workloads increase read cost in ClickHouse faster than in Doris.

Apache Doris Unique Key and Delete Bitmap update mechanism
Unique Key + Delete Bitmap update mechanism in Apache Doris (from the original article).

4. SSB results (real-time updates)

  • Doris vs ClickHouse ReplacingMergeTree (32c/128GB): 18x faster at 100% updates, 14x faster at 25% updates.
  • Doris vs ClickHouse ReplacingMergeTree (16c/64GB): 34x faster at 100% updates, 25x faster at 25% updates.
  • Within ClickHouse itself, MergeTree outperforms ReplacingMergeTree by 1.7-2.5x.
SSB benchmark baseline comparison chart
SSB-A: baseline comparison without real-time update semantics (MergeTree vs Duplicate Key).
SSB benchmark MergeTree versus ReplacingMergeTree chart
SSB-B: performance drop inside ClickHouse when switching to ReplacingMergeTree.
SSB benchmark ClickHouse ReplacingMergeTree versus Doris Unique Key chart
SSB-C: direct real-time update comparison: ClickHouse ReplacingMergeTree vs Doris Unique Key.

5. ClickBench results (real-time updates)

  • Doris vs ClickHouse ReplacingMergeTree (32c/128GB): 2.5x faster at 100% updates, 1.7x faster at 25% updates.
  • Doris vs ClickHouse ReplacingMergeTree (16c/64GB): 4.6x faster at 100% updates, 3.1x faster at 25% updates.
  • Within ClickHouse, MergeTree is 2.7-3.9x faster than ReplacingMergeTree.
ClickBench MergeTree versus ReplacingMergeTree benchmark chart
ClickBench-A: MergeTree vs ReplacingMergeTree (cost of update semantics in ClickHouse).
ClickBench ClickHouse versus Apache Doris benchmark chart
ClickBench-B: real-time update comparison between ClickHouse and Apache Doris.

6. Why this matters for modern OLAP

  • Real-time analytics requires both low-latency ingestion and low-latency ad-hoc queries.
  • If updates degrade query latency, dashboards and alerting drift behind reality.
  • This directly impacts fraud detection, risk systems, and SLA-sensitive operations.
  • That is why update handling is a first-class OLAP selection criterion.

8. Benchmark Limits and Fairness Conditions

To interpret the benchmark responsibly, keep these boundary conditions in mind:

  • The test was executed on specific managed cloud configurations (VeloDB Cloud and ClickHouse Cloud).
  • The scenario focuses on update-intensive workloads; pure append/scan workloads may produce different outcomes.
  • Results are sensitive to table model, merge/background compaction settings, and query profile.
  • SSB and ClickBench are representative, but final architecture decisions should rely on your own workload POC.

Recommendation: run a short workload-driven POC with your SLA, schema design, and real query patterns before migration.

7. Apache Doris customer stories

NetEase Cloud Music

Migrated from ClickHouse to Doris for log analytics at trillion-event scale, with peak 6 GB/s ingest and improved query concurrency.

Lakala

Consolidated Elasticsearch/Hive/HBase/TiDB/Oracle into Doris with up to 15x faster queries and 52% fewer servers.

Kwai

Built a unified lakehouse on Doris, handling nearly one billion queries per day with high concurrency and lower complexity.

Original source

VeloDB Engineering Team, "Apache Doris Up to 34x Faster Than ClickHouse in Real-Time Updates", 2025-10-01.

https://www.velodb.io/blog/apache-doris-34x-faster-clickhouse-realtime-updates

FAQ

Was the benchmark fair in terms of resources?

Yes. The study used both CPU-aligned and memory-aligned configurations to compare systems under comparable resource conditions.

Why not compare only with MergeTree?

Because the benchmark focuses on real-time updates. MergeTree is faster, but it does not provide equivalent update semantics to ReplacingMergeTree.

Where can I review SQL and benchmark details?

The SQL update workload is published at github.com/dataroaring/ClickBench/tree/main/clickhouse-cloud-update.

Want to validate this on your own workload?

./REQUEST_TECHNICAL_POC.sh
© 2026 DATANOMIX.PRO — EXCLUSIVE VELODB PARTNER IN CENTRAL ASIA
VeloDB — Real-Time Analytics /HOME