I think the competition for the future is between DuckDB and Polars. Will we stick with the DataFrame model, made feasible by Polars's lazy execution, or will we go with in-process SQL a la DuckDB? Personally I've been using DuckDB because I already know SQL (and DuckDB provides persistence if I need it) and don't want to learn a new DataFrame DSL but I'd love to hear other the experience of other people.
I’d recommend using the polars SQL context manager if wanting to defer learning how to do everything through their API. The API is a big enough shift from pandas it took me a minute to figure out but I really enjoy having the choice to stay in dataframe methods or switch to SQL only transformations. It has global state too if that’s needed. I like that it isn’t a RDBMS but provides all of the SQL I use.
I really like the dataframe approach. I think it’s because I like REPL-driven-development where I can drop into the REPL and work through how to transform the data interactively.
To be fair, it can nearly always be done in SQL also (unless it’s ML or some Python-specific thing like that), but the SQL with nested queries and numerous CTEs is harder for me to wrap my brain around.
If I were betting, I’d pick DuckDB, because DuckDB seems more able to implement something Polars-like, than Polars is to implement something DuckDB-like.
I'm very split. There's a lot of interactive exploration and data transformations that SQL lends itself to poorly (try transposing in SQL - not fun!) but I really like the idea of data system that is language agnostic like DuckDB
I am just using duckdb on a 3TB dataset in a beefy ec2, and am pleasantly surprised at its performance on such a large table. I had to do some sharding to be sure but am able to match performance of snowflake or other cluster based systems using this single machine instance.
To clarify Clickhouse will likely match this performance as well, but doing things on a single machines look sexier to me than it ever did in decades.
Where does your data reside, is it on an attached EBS volume, or in S3, or somewhere else?
I had some spare time and tinkered with duckdb with a 70GB dataset, but just getting the 70GB on to the EC2 took hours. Would be pretty rocking if duckdb team could somehow set up a ~1TB sized demo that anyone can setup and try for themselves in, say, under an hour.
Local drives. DONT USE EBS! you’ll incur a huge IO charge. You have to choose instances with attached nvme storage which means one of the storage optimized instances.
Reading the data off s3 will mean you will be slower than offerings like snowflake. Snowflake has optimized the crap out of doing analytics in s3, so you can’t beat it with something as simple as duckdb.
Importantly you need the data in some distributed format like parquet or split csv. Otherwise duckdb can’t read it in parallel.
Hi – DuckDB Labs devrel here. It's great that you find DuckDB useful!
On the setup side, I agree that local (instance-attached) disks should be preferred but does EBS incur an IO fee? It incurs a significant latency for sure but it doesn't have a per-operation pricing:
> I/O is included in the price of the volumes, so you pay only for each GB of storage you provision.
Can’t remember anymore, but it’s either (a) the gp2 volumes were way too slow for the ops or (b) the IOPs charges made it bad. To be clear I didn’t do it on duckdb but hosted a Postgres. I moved to light sail instead and was happy with it (you don’t get attached SSD in ec2 until you go to instances that are super large).
I tried to spread large dataset into thousands of files on S3 and use StepFunctions Distributed Map to launch thousands of Lambda instances to process those files in parallel, using DuckDB (or other libs) in Lambda. The parallel loading and processing is way faster than doing this in a single big EC2 instance.
I’ve tried reading streamed parquet via PyArrow with Duck, and it’s been pretty promising. Depending on the query, you won’t need to download everything off HTTP.
we use partitioned parquet files in s3. we use a csv in the bucket root to track the files. i’m sure there’s a better way but for now the 2tb of data are stored cheaply and we get fast reads by only reading the partitions we need to read.
I'm curious how much simpler to build, manage, and run vs cost it would be to simply running a database on a large vultr/DO instance and paying for 2tb of storage?
I feel like you'd get away with the whole thing for around $500/mo depending on how much compute was needed?
well that's not the infrastructure we have. we are primarily an aws shop so we use the resources available to us in the context of our infrastructure decisions. it would be a hard sell to buy something outside of that ecosystem.
The test case of a simple aggregation is a good example of an important data science skill knowing when and here to use a given tool, and that there is no one right answer for all cases. Although it's worth noting that DuckDB and polars are comparable performance-wise for aggregation (DuckDB slightly faster: https://duckdblabs.github.io/db-benchmark/ ).
For my cases with polars and function piping, certain aspects of that workflow are hard to represent in SQL, and additionally it's easier for iteration/testing on a given aggregation to add/remove a given function pipe, and to relate to existing tables (e.g. filter a table to only IDs present in a different table, which is more algorithmically efficient than a join-then-filter). To do the ETL I tend to do for my data science workin pandas/polars in SQL/DuckDB, it would require chains of CTEs or other shenanigans, which eliminates similicity and efficincy.
The real winner is going to be a framework that, during dev, transparently materializes CTEs to temporary tables so you can iterate on them like you’re saying, while continuing to harness SQL for the end product.
> Note that DuckDB automatically figured out how to parse the date column.
It kinda did and it kinda didn't. Author got lucky that Transaction.csv contained a date where the day was after the 12th in a given month. Had there not been such a date, DuckDB would have gotten the dates wrong and read it as dd/mm/yyyy.
I think a warning from DuckDB would have been in order.
My biggest issue with DuckDB is its not willing to implement edits to blob storages which allow edits (Azure). Having common object/blob storages that can be interacted and operated by multiple process will make it much more amenable to many data science driven workflows.
It's not so much about size but about usage pattern.
If your workloads require fast writes and reads, SQLite will probably work fine.
If you're looking to run analytic, columnar queries (which tend to involve a lot of aggregation and joins on a few columns (say less than 50) at a time), then DuckDB is way more optimized.
Oversimplifying, Sqlite is more OLTP and DuckDB is more OLAP.
It's an expression of a personal experience, preferences, and thoughts on a personal blog, thrown for others that might care about DuckDb and Pandas/Polars (and many did, as it got in the HN's first page).
They didn't write it to be some novel research, some canonical tutorial about the tech, or to teach/amuse each and every random reader.
