LAKEHOUSE ENGINE
BENCHMARK MATRIX
Independent stress-tests across five production-grade lakehouse engines. Cyan cells mark category leaders. Every number is reproducible — methodology and raw Parquet files linked in the full report.
| BENCHMARK DIMENSION | Databricks3 wins | Iceberg4 wins | Delta Lake1 wins | Hudi0 wins | Dremio4 wins |
|---|---|---|---|---|---|
Query Latency p50unit: ms | 142▲ BEST | 198 | 211 | 287 | 163 |
Query Latency p99unit: ms | 412▲ BEST | 573 | 604 | 891 | 488 |
Time-Travel Query Costunit: $/query | $0.0052 | $0.0031▲ BEST | $0.0037 | $0.0089 | $0.0041 |
Concurrent Write Throughputunit: MB/s | 1840 | 1620 | 1710 | 980 | 2210▲ BEST |
Schema Evolution Supportunit: score/100 | 94 | 98▲ BEST | 91 | 87 | 82 |
Partition Pruning Efficiencyunit: % | 97.2% | 98.8%▲ BEST | 96.1% | 91.4% | 95.7% |
Small-File Compactionunit: sec/GB | 8.4 | 11.2 | 9.7 | 14.6 | 7.1▲ BEST |
ACID Transaction Overheadunit: ms/txn | 23 | 31 | 18▲ BEST | 44 | 38 |
Cloud Storage Egressunit: $/TB | $0.87 | $0.64 | $0.71 | $1.12 | $0.58▲ BEST |
Spark Compatibilityunit: score/100 | 99▲ BEST | 97 | 98 | 94 | 88 |
Flink Compatibilityunit: score/100 | 91 | 96▲ BEST | 88 | 89 | 83 |
Trino/Presto Read Perfunit: GB/s | 3.2 | 4.1 | 3.8 | 2.9 | 4.7▲ BEST |
METHODOLOGY &
ENVIRONMENT SPECS
We disclose every variable. If a number surprises you, click through to the raw Parquet files and run the assertion yourself.
Instance Type
r6i.4xlarge (128 GB RAM, 16 vCPU)
Cloud Region
AWS us-east-1 (N. Virginia)
Dataset Size
10 TB TPC-DS (SF=10000)
File Format
Apache Parquet · Snappy compression
Spark Version
Apache Spark 3.5.1
Flink Version
Apache Flink 1.19.0
Test Duration
72 hours per engine · 847 total runs
Isolation
Dedicated VPC · no shared workloads
Reproducible by Default
Every test script, Terraform config, and seed dataset is published to GitHub. Any team with an AWS account can replicate the full suite in under 4 hours.
Vendor-Blind Configuration
Each engine is configured using its own published best-practices guide. No deliberate handicapping. If Databricks recommends Delta caching, we enable it.
Quarterly Re-Tests
Engines ship fast. We re-run the full benchmark suite every 90 days and publish diffs. Subscribers get email alerts when their primary engine changes rank.
git clone github.com/lakehouse-lab/benchmarks · ./run_suite.sh --engine all --dataset tpcds-10tb
# Full suite completes in ~4 hours on r6i.4xlarge. Raw results output to /results/*.parquet
DIMENSION-LEVEL
ANALYSIS
Each section isolates one benchmark dimension — methodology, raw numbers, and a one-sentence verdict you can paste into your RFP.
Databricks leads p50 by 28% over nearest rival
We ran 99 TPC-DS queries 5× each engine at steady-state (warm cache, no cold starts). p50 reflects median interactive query performance — the number your BI users feel every day. Databricks Photon engine's vectorized execution delivers 142ms median versus Iceberg's 198ms on the same r6i.4xlarge hardware.
Test Environment
r6i.4xlarge · 10TB TPC-DS · Warm cache · 5 runs per query
Editorial Verdict
Databricks wins p50 latency by 28%. For interactive BI workloads, this gap is user-perceptible. Iceberg on Trino closes to within 15% at p99 — making it the better choice for batch-heavy pipelines where tail latency matters more than median.
← Lower is better
Dremio auto-compaction 40% faster than Hudi on 1M-file datasets
The small-file problem is the silent killer of lakehouse performance. We generated 1 million 512KB Parquet files (simulating 6 months of CDC ingestion) and measured time-to-compact to target 128MB file size. Dremio's Automatic Reflection Refresh triggered compaction in 7.1s/GB; Hudi's async compaction service required 14.6s/GB even with maxParallelism=32.
Test Environment
r6i.4xlarge · 1M × 512KB files · 128MB target file size · 32 parallel threads
Editorial Verdict
Dremio wins small-file compaction at 7.1s/GB. If your pipeline generates >100K files/day from streaming CDC, Dremio or Databricks are the only defensible choices. Hudi's compaction lag creates a 2–3× query penalty window that compounds under high-ingestion workloads.
← Lower is better
Delta Lake ACID overhead 21% lower than Iceberg at 1K concurrent writers
We measured the lock acquisition + commit latency overhead of ACID transactions under concurrent write pressure: 1,000 simultaneous upsert transactions on a 500GB table. Delta Lake's optimistic concurrency control with log-based conflict detection adds only 18ms overhead per transaction. Hudi's timeline-based locking adds 44ms — a 2.4× penalty that compounds at scale.
Test Environment
r6i.4xlarge · 500GB table · 1,000 concurrent upserts · optimistic concurrency
Editorial Verdict
Delta Lake wins ACID overhead at 18ms/txn. For high-frequency upsert workloads (CDC replication, event sourcing), Delta Lake's optimistic concurrency model outperforms Hudi's timeline locking by 2.4×. Iceberg's 31ms is a reasonable middle ground for teams already invested in the Iceberg ecosystem.
← Lower is better
Showing 3 of 12 benchmark dimensions.
Access All 12 Dimensions in Full ReportENGINEERS WHO
USED THE DATA
"We were 6 weeks into a Delta Lake POC when we found Lakehouse's ACID overhead benchmark. The 18ms vs 44ms gap on concurrent upserts matched exactly what we were seeing in production CDC. We migrated the evaluation criteria overnight and saved a $2.4M infrastructure mistake."
Priya Raghunathan
Staff Data Engineer · Meridian Financial
"The partition pruning efficiency numbers were the one artifact I needed for the board deck. 98.8% on Iceberg vs 91.4% on Hudi — that's a $180K/year egress difference at our query volume. The benchmark methodology section answered every question our security team raised about data provenance."
Marcus Okafor
Platform Architect · Axiom Logistics
"I sent the one-page benchmark summary to our CTO at 11 PM on a Tuesday. By Wednesday morning we had budget approval for the Dremio migration. Three months of internal debates ended because the numbers were specific, the methodology was clean, and there was nothing to argue with."
Tomás Herrera
Principal Engineer · Vektor Analytics
UPGRADE YOUR
ACCESS LEVEL
80% of the benchmark data is visible on this page — free, ungated, exportable. The full report adds depth: cost modeling, migration playbooks, and quarterly diffs.
- ✓This comparison table (all 12 dimensions, 5 engines)
- ✓Methodology overview and environment specs
- ✓Three deep-dive dimension analyses
- ✓Editorial verdicts for query latency, compaction, and ACID overhead
- ✦Full 12-dimension benchmark results with statistical confidence intervals
- ✦Cost modeling spreadsheet: TCO calculator for 1TB → 1PB scale
- ✦Quarterly re-test diffs: see how each engine changed over 4 quarters
- ✦Migration playbook: Hive → your chosen lakehouse in 14 steps
- ✦Raw Parquet result files for independent verification
- ✦Private Slack channel: ask methodology questions directly
$ ./download_report.sh --format pdf+xlsx --include raw-parquet
→ Enter credentials to authenticate...