As someone that uses SQLite a bit for analytics, I feel like the elevator pitch on the github and the website ( https://www.duckdb.org/ ) is missing something.
I'm sure there's some great reason, but I don't see it: Why would I want to use this over SQLite? Is it considerably faster? Does it handle large scale better? I suppose I could try and throw it some use cases I could come up with, but I'd worry I'm missing the best use case for this tool.
Here's a YT video of one of the implementers giving a talk to CMU students about what DuckDB is, why they made it, and how it works: https://www.youtube.com/watch?v=PFUZlNQIndo
In a nutshell, it's a file-per-database embedded DB like SQLite, and it's installable as a Python or R package from PyPI or CRAN, but unlike SQLite, it's a column store with vectorized operations so it's more performant for analytical queries.
The video clarified my initial confusion. DuckDB is an embedded column store library and borrows the SQLite C API and some of the SQLite philosophy (single cpp/hpp amalgamation file, no external dependencies, and resource sharing) but it uses PostgreSQL syntax and parser, and a column oriented memory/storage engine based on their experience with MonetDB/VectorWise.
DuckDB borrows code from HyPer DB which looks like a in-memory hybrid OLTP/OLAP clone of SAP HANA. They did not discuss any hybrid OLTP/OLAP implementation details in the video so the borrowed HyPer code may only be the wrapper for the PostgreSQL parser.
Desktop column stores do have a precedent in the enterprise market. Microsoft's first attempt at a column store started with embedded integration with PowerPivot and eventually made its way into SQL Server. Sybase IQ, built on zero-admin SQL Anywhere, has been used as a local column store for Pivot Table applications (think BusinessObjects, Cognos, Tableau, QlikView, and Excel Pivot Tables). Pivot Table integrations are typically custom-built for each data warehouse vendor or MDX/XMLA driver.
DuckDB is focused on integrating with the desktop data science platforms R and Python (dbplyr and pandas). Analytics/OLAP in SQL means lightning fast GROUP BY queries and more recently SQL OLAP Windows. "SQLite for Analytics" confused me but that is probably due to my own historical mental clutter. The name DuckDB refers to a pet duck and not the duck typing system in SQLite.
The fact that it’s column oriented is a massive improvement for analytics over SQLite, I don’t understand why they don’t mention this more explicitly on their website and GitHub pages.
I usually just extend SQLite with virtual columns and try to do the filtering/aggregate before sending tuples to the engine. This does the trick most of the time and it's relatively straightforward to implement.
Consider someone who analyzes medium-sized volumes of (slowly-changing) data -- OLAP, not OLTP. People who need to do this primarily have 2 alternatives:
The latter can be slow and wasteful, because the data is stored in a form that allows very limited indexing. So imagine you want query speeds that require the former.
Traditional databases can be hugely wasteful for this usecase -- space overhead due to no compression, slow inserts due to transactions. The best analytics databases are closed-source and come with vendor lockin (there are very few good open-source column stores -- clickhouse is one, duckdb is another). Most solutions are multi-node, so they come with operational complexity. So DuckDB could fill a niche here -- data that's big enough to be unwieldy, but not big enough to need something like redshift. It's analogous to the niche SQLite fills in the transactional database world.
1. Is DuckDB similar to having indexes on each column? Because generally when something is slow, the solution is indexes. I have a 100 GB database which records real time data and is lightning fast because of some minor tuning.
2. The example of STDDEV not be available shows the author's unfamiliarity with SQLite which worries me.
ad 1) An index is an additional data structure that takes space and needs maintenance when the data changes. This is especially problematic for realtime data. DuckDBs main data structure and implementation is designed for efficient scanning without needing an additional index ("vectorised columnar").
ad 2) We support STDDEV_SAMP and STDDEV_POP. There is also an interface to define your own functions. We are very aware of how SQLite is doing things and even support their client API. But we also might eventually add their way of adding UDFs from Python.
Indexes don't solve the same problem as columnar because with an index you still have to dereference a pointer, read the page (performing I/O if on-disk), and locate each tuple you want within the page--the tuples you need to aggregate from may be scattered across many pages, and you have to read the whole page into memory even if you only want one column.
With a column store, the values of the same column are stored contiguously so if you're summing a column for example, you can read one or more pages that could consist entirely of values the column you need to sum, sequentially. It's much more efficient for aggregations due to cache locality--much fewer cache misses and I/O operations needed to summarize a column.
The major trade-off is that updates and deletes are more expensive in a column store. But this is usually fine for OLAP workloads.
Unsure if DuckDB implements this in its column store, but another common feature is compression.
A page of field values is a much better structure to compress than a page of tuple values. This is especially true of the sort of denormalized data that is typical in analytical workloads. E.g. dates may only be considered in a period of a few years - this will take much less space to encode than the full allocation for a date data type. Similarly we may have categorical fields with only a few values. Imagine a product category with only 8 values. This can take a set of numbers which may be otherwise stored as an integer, and encode it to 3 bits, plus some space for a dictionary.
Another easy bit of compression to apply to columns is run length encoding. Dates, again, serve as a good example. If the data is sorted by date, you can choose to store a data structure that captures (start index, date value, end index). If you had 1,000 records on the same date, you could encode this with only three likely-integer-sized values, rather than 1,000 repetitions of a single likely-integer-sized value.
Such compression offers several benefits. First, it can reduce data volumes, and allow more query workload to come from RAM, either through DB or filesystem caching, or by an in-memory architecture. Additionally, the smaller the data size is, the faster a scan can be completed. Analytic workloads tend toward queries that require a full table scan (though with a columnstore, it's more a full column scan). If the data size is smaller, then you can stream the entire thing through whatever your query is faster.
Yep, compressibility is another big win made possible by columnar layout, and it looks like DuckDB implements at least some form of column compression.
Yes, it's geared toward reporting / analytics / data science workloads where the most common read operation is calculating aggregate metrics like count/min/mean/max and histograms. Statistics stuff.
I'd love to see how it compares to MonetDB. I've used MonetDB successfully to look at smallish data set (few hundred millions rows) that neither postgres nor sqlite could complete queries at a useful speed on my laptop (at the time an x220 with a spinning disk).
The primary difference for analytic workloads is the processing model. MonetDB uses a column-at-a-time processing model which causes large intermediates to be materialized in memory. This significantly increases memory usage, and causes poor performance when data exceeds memory. DuckDB uses a vectorized execution engine, which materializes only small subsets of the columns. This increases cache-locality, decreases memory usage and also improves parallel execution capabilities (although parallel execution is still WIP currently).
DuckDB is developed by the same database research group at CWI that developed MonetDB, so they have applied a lot of the lessons they learned from MonetDB.
I bookmarked DuckDB some time ago intending to revisit, and when I did, I had forgotten why I bookmarked it and the website's contents unfortunately didn't help for reminding me.
If the documentation had said something like "a columnar database based on sqlite, optimized for analytics", I think that would pique the interest of many. Columnar database = column compression = fast filtering/aggregations = high performance analytics compared to pure row stores.
There are very few free columnar databases available that are production grade (or built on production grade technology), so a columnar database built on top of sqlite is big news.
Apologies, a website rework is currently in progress :)
DuckDB is indeed a free columnar database system, but it is not entirely built on top of SQLite. It exposes the same front-end and uses components of SQLite (the shell and testing infrastructure), but the execution engine/storage code is new.
Just came here to say the same thing. I read the readme and it's not clear what this does exactly as compared to SQLite, particularly since it used SQL. What kinds of analytics is it optimized for? Does it have different data structures for streaming data? I can tell a lot of work went into this project and I think it could benefit from a more thorough introduction.
It's a columnar database and the query engine is inspired from MonetDB which processes columns in vectors - "Execution engine: The vectorized execution engine is inspired by the paper "MonetDB/X100: Hyper-Pipelining Query Execution" by Peter Boncz, Marcin Zukowski and Niels Nes.".
I wouldn't call that "buried in the acknowledgements section." I wish more software projects would present themselves in this way actually. Inevitably everything is built upon prior art, so it makes sense to say where the ideas came from or which parts are novel (if any - sometimes packaging together a bunch of existing ideas is valuable in itself).
It sure beats "[project] is a new key-value store that is even faster than [other project]".
Agreed. I was trying to understand it earlier, and there doesn't seem to be a description of what its features are and what it's good for. It jumps straight to "how to use" without "why to use".
My first instinct based on a cursory glance (ie. specific language compatibilities) is that it's a portable embedded database for data analytics of the cybersecurity domain, particularly mobile penetration testing possibly.
If you want to dig into the differences more, I highly recommend the book Designing Data Intensive Applications. Analytics systems have different requirements from normal transaction processing systems, so it naturally follows that DuckDB could specialize to fill those requirements better than SQLite. Many people bring up row stores (SQLite) vs column stores, but there are many other interesting differences and optimizations to be made, so I can see how SQLite may be leaving some niche unfilled.
Adding to the pile of people agreeing. SQLite works okay for analytics for now, and I would be super-interested in something embeddable more tailor-made for analytics, but don't understand yet in what specific ways DuckDB is an improvement.
It might be dating me, but seem like a larger version of the difference between C and FORTRAN arrays, C being relational like and FORTRAN being column store like.
Well, seems to me the difference between rowstores and columnstores might be a little more profound than just row and column major orders in C and FORTRAN. (Which I remember from back in the day)
In columnstores, it’s not just the ordering that’s advantageous for analytics. It’s also the possibility of scanning only the columns required (to the exclusion of others), and the possibility of column compression using RLE compression, which speeds up search when there are many repeated elements (also compression makes searches cache efficient). This means if you workload only involves a few columns, and your primary operations are filter and aggregate — analytics workloads — then the performance gains are tremendous.
On the other hand, if you want to optimize for row by row writes (OLTP), and if your workloads involve operations on many columns, then a rowstore is advantageous.
Row and column major layouts are more about exploiting sequentiality, but in practice both C and FORTRAN can provide good linear algebra performance.
But for analytics workloads on row vs columnstores, the difference in performance can be in the orders of magnitude. The fastest analytics databases in the world (Kdb, Vertica, Exasol, some of the newer GPU based databases) are almost all columnar.
I'm sure there's some great reason, but I don't see it: Why would I want to use this over SQLite? Is it considerably faster? Does it handle large scale better? I suppose I could try and throw it some use cases I could come up with, but I'd worry I'm missing the best use case for this tool.