Benchmark · 19 November 2026

One job, six tools, measured

We built the same job in six tools, the way a normal user of each would, with default settings, and ran it five times on the same machine. All six produced the same 144 rows. Where another tool wins, the page says so.

The job: Read 5,000,000 orders and 200,000 customers from CSV, keep the paid orders, look up each customer’s country, group by country and month (count and sum of the amount), sort, and write the 144 result rows to CSV.

Median of 5 runs. Not all numbers measure exactly the same thing: Odara and NiFi run on a server that is already started; Talend, Hop and dbt include the start of their runtime; KNIME is timed in its desktop application. Each card says what was measured.

What the numbers say

  • Odara was the fastest of the visual tools: 2.0 s, with one native binary and no JVM.
  • Talend Open Studio 8 came close at 4.5 s, its JVM start included: its generated Java is efficient. It no longer gets updates or security fixes.
  • dbt with DuckDB has the fastest engine: about 0.6 s for the query itself, 3.0 s with dbt’s start-up. It has no visual editor and only transforms data.
  • KNIME (26.8 s), NiFi (35.5 s) and Apache Hop (39.9 s) were much slower on this job with default settings. Each card says where the time went.

Odara and Talend Open Studio

Full comparison: Odara and Talend Open Studio

Odara ETL 0.4.4

2.0 s median of 5 runs

Measured
Run on a running Odara server, through the API (server start: 2.3 s, not included)
Memory
0.44 GB (whole server process)
Download
139 MB
Needs
Nothing (no JVM)
Licence
Proprietary, free for any use, continuity pledge

Talend Open Studio 8.0.1 (the last free release)

4.5 s median of 5 runs

Measured
Standalone job built in the Studio (Build Job), run with Java 11; includes JVM start
Memory
0.26 GB (job JVM, default -Xmx1024M)
Download
831 MB, plus a Java runtime
Needs
Java 11 (Studio and jobs)
Licence
Apache 2.0; retired by Qlik on 31 January 2024
  • Talend’s generated Java is fast: 4.5 s including the JVM start, the closest result to Odara among the visual tools.
  • Open Studio has no command-line build: each job is built from the Studio. The first start also asks to download 48 required libraries.
  • No more updates or security fixes since January 2024.

Choose Talend Open Studio if

  • your jobs already run well and you can live without updates or security fixes
The benchmark job in Talend Open Studio 8.0.1 after a run: tFileInputDelimited, tMap with a lookup, tAggregateRow, tSortRow and tFileOutputDelimited, with row counts on each link.
Talend Open Studio 8.0.1 (the last free release): the benchmark job, as built and run for this test.

Odara and Apache Hop

Full comparison: Odara and Apache Hop

Odara ETL 0.4.4

2.0 s median of 5 runs

Measured
Run on a running Odara server, through the API (server start: 2.3 s, not included)
Memory
0.44 GB (whole server process)
Download
139 MB
Needs
Nothing (no JVM)
Licence
Proprietary, free for any use, continuity pledge

Apache Hop 2.9.0

39.9 s median of 5 runs

Measured
hop-run with the local engine, Java 17; includes JVM start (the pipeline itself: 36.5 s)
Memory
0.68 GB
Download
883 MB, plus a Java runtime
Needs
Java 17
Licence
Apache 2.0, open source
  • Built with standard transforms and default settings. The Filter rows transform took about 32 of the 36.5 s: reading the 5 million rows alone took 2.3 s.
  • We did not find a setting that changed it (lazy conversion off, a non-compressed lookup). If you know a faster way to build this pipeline in Hop, send it to [email protected] and we will publish the new number.

Choose Apache Hop if

  • an open-source licence is a requirement today
  • you need macOS or ARM64, or Apache Beam (Spark, Flink, Dataflow)
  • your team knows Kettle / Pentaho
The benchmark pipeline in Apache Hop 2.9.0: two CSV file inputs, Filter rows, Stream lookup, Strings cut, Memory group by, Sort rows and Text file output.
Apache Hop 2.9.0: the benchmark job, as built and run for this test.

Odara and KNIME Analytics Platform

Full comparison: Odara and KNIME Analytics Platform

Odara ETL 0.4.4

2.0 s median of 5 runs

Measured
Run on a running Odara server, through the API (server start: 2.3 s, not included)
Memory
0.44 GB (whole server process)
Download
139 MB
Needs
Nothing (no JVM)
Licence
Proprietary, free for any use, continuity pledge

KNIME Analytics Platform 5.12.0

26.8 s median of 5 runs

Measured
Execute all in the desktop application (the free version has no headless run since 5.9)
Memory
Not comparable (desktop application, about 10 GB in total)
Download
849 MB (Java included)
Needs
Its bundled Java 21
Licence
GPL v3 with extra permissions, open source
  • Slowest steps: GroupBy 8 s, CSV Reader 6 s, Joiner and String Manipulation 5–6 s each.
  • Running workflows from the command line or on a schedule now needs a paid KNIME product (KNIME forum, April 2026).

Choose KNIME Analytics Platform if

  • you do analytics and data science as much as ETL
  • you want the largest catalogue of nodes and a big community
  • you need macOS
The benchmark workflow in KNIME 5.12: two CSV Readers, Row Filter, Joiner, String Manipulation, GroupBy, Number Rounder, Sorter and CSV Writer.
KNIME Analytics Platform 5.12.0: the benchmark job, as built and run for this test.

Odara and Apache NiFi

Full comparison: Odara and Apache NiFi

Odara ETL 0.4.4

2.0 s median of 5 runs

Measured
Run on a running Odara server, through the API (server start: 2.3 s, not included)
Memory
0.44 GB (whole server process)
Download
139 MB
Needs
Nothing (no JVM)
Licence
Proprietary, free for any use, continuity pledge

Apache NiFi 2.12.0

35.5 s median of 5 runs

Measured
Flow already running on a NiFi server, from the file landing to the result written (server start: 16 s)
Memory
1.66 GB (server, default 1 GB heap)
Download
847 MB, plus a Java runtime
Needs
Java 21
Licence
Apache 2.0, open source
  • NiFi moves and routes data continuously; a batch join and aggregation is not what it is built for. Each step reads and writes the whole file again.
  • Default settings, no parallel tasks or file splitting.

Choose Apache NiFi if

  • you move data continuously from many sources and need back-pressure, guaranteed delivery and a record of where every record went
The benchmark flow in Apache NiFi 2.12: GetFile, QueryRecord, LookupRecord, QueryRecord, UpdateAttribute and PutFile, with their counters.
Apache NiFi 2.12.0: the benchmark job, as built and run for this test.

Odara and dbt Core + DuckDB

Full comparison: Odara and dbt Core + DuckDB

Odara ETL 0.4.4

2.0 s median of 5 runs

Measured
Run on a running Odara server, through the API (server start: 2.3 s, not included)
Memory
0.44 GB (whole server process)
Download
139 MB
Needs
Nothing (no JVM)
Licence
Proprietary, free for any use, continuity pledge

dbt Core + DuckDB dbt-core 1.12.7, dbt-duckdb 1.11.1, DuckDB 2.0.0

3.0 s median of 5 runs

Measured
dbt run, end to end; includes about 2 s of dbt start-up (the query itself: about 0.6 s)
Memory
0.46 GB
Download
391 MB (Python virtual environment)
Needs
Python
Licence
Apache 2.0 (dbt) and MIT (DuckDB), open source
  • DuckDB’s engine is the fastest in this test: the query takes about 0.6 s, against about 2 s for Odara’s run.
  • dbt only transforms: no connectors to move data between systems, no visual editor, and no scheduler in dbt Core (dbt Cloud is paid).

Choose dbt Core + DuckDB if

  • your data is already in one warehouse or database and your team works in SQL and Git
The dbt docs lineage graph: two CSV sources feeding one model.
dbt Core + DuckDB dbt-core 1.12.7, dbt-duckdb 1.11.1, DuckDB 2.0.0: the benchmark job, as built and run for this test.
The benchmark pipeline in the Odara editor: two CSV sources, one SQL step with two inputs, one CSV target; completed in 2179 ms.
Odara ETL 0.4.4: the benchmark job, as built and run for this test.

Method

  • Machine: AMD Ryzen 5 3600 (6 cores, 12 threads), 62 GB RAM, NVMe disk, Linux 7.0. One tool at a time: a lock made sure no two timed runs overlapped.
  • Data: synthetic, generated with a fixed seed: 5,000,000 orders (189 MB) and 200,000 customers (6 MB).
  • Runs: five per tool, median reported; memory is the peak resident size of the process.
  • Builds: standard components and default settings, as a normal user would build it. No tuning for any tool, including Odara.
  • Downloads: official sources with checksums where published. Talend Open Studio 8.0.1 is no longer offered by Qlik; we used the archived copy of the official file (TOS_DI-20211109_1610-V8.0.1.zip) on archive.org.
  • Not tested: Pentaho Data Integration (its free download needs a registration we did not do), and paid products (Informatica, Matillion, SSIS, Fivetran).

Repeat it yourself. The data generator and the job in each tool:

The NiFi flow was built through its REST API with GetFile, QueryRecord, LookupRecord (SimpleCsvFileLookupService), QueryRecord, UpdateAttribute and PutFile. Paths in the files point to our test folder: change them to yours.

Found a faster way to build the job in one of these tools, or a mistake? Write to [email protected] and we will run it and publish the new number.