As the year comes to a close, I decided to explore a fun yet somewhat impractical challenge: Can DuckDB run the TPC-DS benchmark using just 2 cores and 16 GB of RAM? The answer is yes, but with a caveat—it’s slow. Despite the limitations, it works!
Notice; I am using lakehouse mounted storage, for a background on the different access mode, you can read the previous blog
Data Generation Challenges
Initially, I encountered an out-of-memory error while generating the dataset. Upgrading to the development release of DuckDB resolved this issue. However, the development release currently lacks support for reading Delta tables, as Delta functionality is provided as an extension available only in the stable release.
Here are some workarounds:
Increase the available RAM.
Use the development release to generate the data, then switch back to version 1.1.3 for querying.
Wait for the upcoming version 1.2, which should resolve this limitation.
The data is stored as Delta tables in OneLake, it was exported as a parquet files by duckdb and converted to delta table using delta_rs (the conversion was very quick as it is a metadata only operation)
Query Performance
Running all 99 TPC-DS queries worked without errors, albeit very slowly( again using only 2 cores ).
I also experimented with different configurations:
4, 8, and 16 cores: Predictably, performance improved as more cores were utilized.
For comparison, I ran the same test on my laptop, which has 8 cores and reads my from local SSD storage, The Data was generated using the same notebook.
Results
Python notebook compute consumption is straightforward, 2 cores = 1 CUs, the cheapest option is the one that consume less capacity units, assuming speed of execution is not a priority.
Cheapest configuration: 8 cores offered a good balance between cost and performance.
Fastest configuration: 16 cores delivered the best performance.
Interestingly, the performance of a Fabric notebook with 8 cores reading from OneLake was comparable to my laptop with 8 cores and an SSD. This suggests that OneLake’s throughput is competitive with local SSDs.
Honestly, It’s About the Experience
At the end of the day, it’s not just about the numbers. There’s a certain joy in using a Python notebook—it just feels right. DuckDB paired with Python creates an intuitive, seamless experience that makes analytical work enjoyable. It’s simply a very good product.
Conclusion
While this experiment may not have practical applications, it highlights DuckDB’s robustness and adaptability. Running TPC-DS with such limited resources showcases its potential for lightweight analytical workloads.
This is not an official benchmark—just an exercise to experiment with the new Fabric Python notebook.
You can download the notebook and the results here
There is a growing belief that most structured data will eventually be stored in an open table format within object stores, with users leveraging various engines to query that data. The idea of data being tied to a specific data warehouse (DWH) may soon seem absurd, as everything becomes more open and interoperable.
While I can’t predict the future, 2024 will likely be remembered as the year when the lakehouse concept decoupled from Spark. It has become increasingly common for “traditional” DWHs or any Database for that matter to support open table formats out of the box. Fabric DWH, for instance, uses a native storage layer based on Parquet and publishes Delta tables for consumption by other engines. Snowflake now supports Iceberg, and BigQuery is slowly adding support as well.
I’m not particularly worried about those DWH engines—they have thousands of engineers and ample resources, they will be doing just fine.
My interest lies more in the state of open source Python engines, such as Polars and DataFusion, and how they behave with a limited resource environment.
Benchmarking Bias
Any test inherently involves bias, whether conscious or unconscious. For interactive queries, SQL is the right choice for me. I’m aware of the various DataFrame APIs, but I’m not inclined to learn a new API solely for testing. For OLAP-type queries, TPC-DS and TPC-H are the two main benchmarks. This time, I chose TPC-DS for reasons explained later.
Benchmark Setup
All data is stored in OneLake’s Melbourne region, approximately 1,400 km away from my location, the code will check if the data exists otherwise it will be generated, the whole thing is fully reproducible.
I ran each query only once, ensuring that the DuckDB cache, which is temporary, was cleared between sessions. This ensures a fair comparison.
I explicitly used the smallest available hardware since larger setups could mask bottlenecks. Additionally, I have a specific interest in the Fabric F2 SKU.
While any Python library can be used, as of this writing, only two libraries—DuckDB and DataFusion—support:
Running the 99 TPC-DS queries (DataFusion supports 95, which is sufficient for me).
Native Delta reads for abfss or at least local paths.
Python APIs, as they are required to run queries in a notebook.
Other libraries like ClickHouse, Databend, Daft, and Polars lack either mature Delta support or compatibility with complex SQL benchmarks like TPC-DS.
Why TPC-DS ?
TPC-DS presents a significantly greater challenge than TPC-H, with 99 queries compared to TPC-H’s 22. Its more complex schema, featuring multiple fact and dimension tables, provides a richer and more demanding testing environment.
Why 10GB?
The 10GB dataset reflects the type of data I encountered as a Power BI developer. My focus is more on scaling down than scaling up. For context:
The largest table contains 133 million rows.
The largest table by size is 1.1GB.
Admittedly, TPC-DS 10GB is overkill since my daily workload was around 1GB. However, running it on 2 cores and 16GB of RAM highlights DuckDB’s engineering capabilities.
btw, I did run the same test using 100GB and the python notebook with 16 GB did works just fine, but it took 45 minutes.
OneLake Access Modes
You can query OneLake using either abfss or mounted storage. I prefer the latter, as it simulates a local path and libraries don’t require authentication or knowledge of abfss. Moreover, it caches data on runtime SSDs, which is an order of magnitude faster than reading from remote storage. Transactions are also included in the base capacity unit consumption, eliminating extra OneLake costs.
It’s worth noting that disk storage in Fabric notebook is volatile and only available during the session, while OneLake provides permanent storage.
You can read more about how to laverage DuckDB native storage format as a cache layer here
Onelake Open internet throughput
My internet connection is not too bad but not great either, I managed to get a peak of 113 Mbps, notice here the extra compute of my laptop will not help much as the bottleneck is network access.
Results
The table below summarizes the results across different modes, running both in Fabric notebooks and on my laptop.
DuckDB Disk caching yielded the shortest durations but the worst individual query performance, as copying large tables to disk takes time.
Delta_rs SQL performance was somewhat erratic.
Performance on my laptop was significantly slower, influenced by my internet connection speed.
Mounted storage offered the best overall experience, caching only the Parquet files needed for queries.
And here is the geomean
Key Takeaways
For optimal read performance, use mounted storage.
For write operations, use the abfss path.
Having a data center next to your laptop is probably a very good idea 🙂
Due to network traffic, Querying inside the same region will be faster than Querying from the web (I know, it is a pretty obvious observation)
but is Onelake throughput good ?
I guess that’s the core question, to answer that I changed the Python notebook to use 8 cores, and run the test from my laptop using the same data stored in my SSD Disk, no call to onelake, and the results are just weird
Reading from Onelake using mounted storage in Fabric Notebook is faster than reading the same data from my Laptop !!!!
Looking Ahead to 2025
2024 has been an incredible year for Python engines, evolving from curiosities to tools supported by major vendors. However, as of today, no single Python library supports disk caching for remote storage queries. This remains a gap, and I hope it’s addressed in 2025.
For Polars and Daft, seriously works on better SQL support
Use SQGLot to parse the SQL query and extract the list of Delta tables that need to be scanned from OneLake.
Track Table Metadata
Capture the Delta table version and ID to ensure consistency.
Selective Copy
Copy only the necessary tables locally to satisfy the query.
Reuse Cached Data
For subsequent queries, check if the Delta table has changed:
If unchanged, read data from the local SSD.
If new tables are required, repeat the caching process for those tables.
Why It Works
This approach effectively creates a temporary, ad hoc disk cache in the notebook. The cache:
Persists only for the session’s duration.
Evicts automatically when the session ends.
Ensures consistency by validating whether the Delta table in OneLake has changed before reusing cached data.
Thanks to the Delta format, this validation is a relatively cheap operation.
Leverages the user-level isolation in Fabric notebooks to eliminate risks of data inconsistency.
Despite its simplicity, this method has proven to be highly effective for query acceleration! 🚀
Limitations
Yes, I know—the cache is rather naïve since it loads the entire table. Other systems go further by:
Copying only the columns needed for the query.
Fetching just the row groups relevant to the query.
However, these optimizations would need to be implemented natively by the engine itself.
Industry Gap
Although virtually all Python engines (e.g., Polars, DataFusion, etc.) support reading formats like Delta and Iceberg, almost none offer built-in disk or RAM caching. This lack of caching support limits performance optimization opportunities.
Hopefully, this will change in the future, enabling more efficient workflows out of the box.
Btw, this is really fast !!! just a hint, this is faster than the results obtained by a state of the art DWH in 2022 !!!
I have a notebook that processes hot data every 5 minutes. Meanwhile, another pipeline processes historical data, and I want to create a summary table that uses the hot data incrementally but refreshes entirely when the historical data changes.
Problem
Checking for changes in historical data every 5 minutes is inefficient, slows down the hot data pipeline, and increases costs. There are many potential solutions for this use case, but one approach I used has been working well.
Solution
Using Delta Table Version
Delta tables provide a variety of functions to access metadata without reading actual data files. For instance, you can retrieve the latest table version, which is highly efficient and typically takes less than a second.
dt = try_get_deltatable(f'/lakehouse/default/Tables/{schema}/scada', storage_options=storage_options)
if dt is None:
current_version = -1
else:
current_version = dt.version()
Storing Custom Metadata
You can store arbitrary metadata, such as a Python dictionary, when writing a Delta table. This metadata storage does not modify Parquet files and can contain information like who wrote the table or any custom data. In my case, I store the version of the historical table used in creating my summary table.
The hot data pipeline incrementally adds data and checks the version of the historical table, storing it in the summary table. If the stored version differs from the latest version, this indicates a change, triggering a full refresh of the summary table.
Example Scenarios
When the Historical Table Has Not Changed
When a Change is Detected in the Historical Table
Key Takeaway
The Python Delta package is a versatile tool that can solve complex data engineering challenges efficiently.
Small Data And self service 