Measure the full ETL workload.
Query speed is one part of pipeline performance. Our benchmarks also measure incremental loads and table maintenance, where storage-aware execution makes a difference.
TPC-DS · 10 TB
Measure scans, joins, aggregations, and shuffles across 99 queries.
Explore query results Load & update 4.1× fasterLakeLoader · 1 TB
Measure the cost of merging changes into existing dimension tables.
Explore load results Maintain tables ~4× fasterIceberg / Hudi compaction
Measure the work needed to consolidate table data after writes.
Explore maintenance resultsQuanton vs. OSS Apache Spark. Separate workloads and configurations; these speedups are not a combined ETL score.
An open approach to
benchmarking ETL.
We propose measuring ETL as a set of workloads, with open tools and an explicit account of what each test covers.
Read the methodology- 01
Separate queries from loads.
Use TPC-DS for reads and transformations. Its 99 queries do not measure incremental writes.
- 02
Replay realistic changes.
Use LakeLoader for DIM, FACT, and EVENT patterns. Vary update distribution and repeat loads until performance stabilizes.
- 03
Account for table maintenance
Include compaction and other deferred work when comparing storage models.
- 04
Apply your workload mix.
Compare costs within each workload, then weight them by your pipeline’s compute-time split. One query speedup is not an ETL cost estimate.
Where the engine stands.
Current measurements cover selected workloads in this methodology. They do not yet cover every load pattern, deletion workload, or concurrency scenario.
TPC-DS
99 analytical queries exercise scans, joins, aggregations, and shuffles across a 10 TB dataset. This measures reads and transformations, without incremental loads.
Quanton completes the suite in 2,034 seconds: 6× faster than OSS Spark and 25% faster than Photon on the same hardware.
Total runtime · lower is better
11 × m8gd.4xlargeComet and Gluten results use tuned configurations. Out-of-the-box runs encountered out-of-memory failures on Q67 and Q93.
TPCx-BB
A mix of Python transformations and SQL exercises the engine beyond SQL-only analytics.
Quanton runs the 1 TB workload 2.0× faster than OSS Spark, with native execution and plan optimizations that reduce intermediate work.
Iceberg 1.5.2 · Spark 3.5 · 1 r8g.4xlarge driver + 10 r8g.4xlarge executors
TPC-DI
Data integration combines transformations with historical and incremental loads into target tables.
Quanton runs the 1 TB workload 2.05× faster than OSS Spark, extending acceleration to the work of loading and updating tables.
Iceberg 1.5.2 · Spark 3.5 · 1 m8g.4xlarge driver + 1 m8g.4xlarge executor
LakeLoader
LakeLoader replays changes into existing tables to measure MERGE INTO and CDC workloads; this result uses dimension tables.
Quanton runs the 1 TB DIM workload 4.1× faster than OSS Spark, using storage-aware joins and columnar writes to reduce merge work.
Explore LakeLoader on GitHubIceberg 1.5.2 · Spark 3.5 · 1 m8g.4xlarge driver + 5 m8g.xlarge executors
Iceberg / Hudi compaction
Compaction consolidates table data after writes, so repeated updates do not leave an ever-growing read and maintenance cost.
Quanton’s native columnar compaction takes about 75% less time than the OSS Spark path in the reported tests.
Native columnar compaction vs. the OSS row-wise rewrite path. Hardware and data scale are not specified in the source result.
