I think this post is a response to some new file format initiatives, based on the criticism that the Parquet file format is showing its age.
One of the arguments is that there is no standardized way to extend Parquet with new kinds of metadata (like statistical summaries, HyperLogLog etc.)
This post was written by the DataFusion folks, who have shown a clever way to do this without breaking backward compatibility with existing readers.
They have inserted arbitrary data between footer and data pages, which other readers will ignore. But query engines like DataFusion can exploit it. They embed a new index to the .parquet file, and use that to improve query performance.
In this specific instance, they add an index with all the distinct values of a column. Then they extend the DataFusion query engine to exploit that so that queries like `WHERE nation = 'Singapore'` can use that index to figure out whether the value exists in that .parquet file without having to scan the data pages (which is already optimized because there is a min-max filter to avoid scanning the entire dataset).
Also in general this is a really good deep dive into columnar data storage.
One question that the article does not cover: compaction. Adding custom indexes means you have to have knowledge of the indexes to compact Parquet files, since you'll want to reindex each time compaction occurs. Otherwise the indexes will at best be discarded. At worst they would even be corrupted.
So it looks as if adopting custom indexes mean you are adopting not just a particular engine for reading but also a particular engine for compaction. That in turn means you can't use generic mechanisms like the compaction mechanism in S3 table buckets. Am I missing something?
What are the new file format initiatives you're referencing here?
This solution seems clever overall, and finding a way to bolt on features of the latest-and-greatest new hotness without breaking backwards compatibility is a testament to the DataFusion team. Supporting legacy systems is crucial work, even if things need a ground-up rewrite periodically.
There are also a bunch of ideas coming out of academia, but I don't know how many of them have a sustained effort behind them and not just a couple of papers
I’ve been excited about lancedb and its ability to support vector indexes and efficient row level lookups. I wonder if this approach would work for their design goals and still allow broader backwards compatibility with the parquet ecosystem. Have been intrigued by Ducklake, and they’ve leaned into parquet. Perhaps this approach will allow more flexible indexing approaches with support for the broader parquet ecosystem which is significant.
Yeah I'm happy to see this, we have been curious as part of figuring out cloud native storage extensions to GFQL (graph dataframe-native query lang), and my intuition was parquet was pluggable here... And this is the first I'm seeing a cogent writeup.
Likewise, this means, afaict, it's likewise pretty straightforward to do novel indexing schemes within Iceberg as well just by reusing this.
The other aspect I've been curious about is the happy path pluggable types for custom columns. This shows one way, but I'm unclear if same thing.
Cool, but this is very specific to DataFusion, no? Is there any chance this would be standardized so other Parquet readers could leverage the same technique?
The technique can be applied by any engine, not just DataFusion. Each engine would have to know about the indexes in order to make use of them, but the fallback to parquet standard defaults means that the data is still readable by all.
But does data fusion publish a specification of how this metadata can be read, along with a test suite for verifying implementations? Because if they don't, this cannot be reliably used by any other impl
The technique described in the article, seems to use this key-value pair to store pointers to the additional metadata (in this case a distinct index) embedded in the file. Note that we can embed arbitrary binary data in the Parquet file between each data page. This is perfectly valid since all Parquet readers rely on the exact offsets to the data pages specified in the footer.
This means that DataFusion does not need to specify how the metadata is interpreted. It is already well specified as part of the Parquet file format itself. DataFusion is an independent project -- it is a query execution engine for OLAP / columnar data, which can take in SQL statements, build query plan, optimize them, and execute. It is an embeddable runtime with numerous ways to extend it by the host program. Parquet is a file format supported by DataFusion because it is one of the most popular ways of storing data in a columnar way in object storages like S3.
Note that the readers of Parquet need to be aware of any metadata to exploit it. But if not, nothing changes - as long as we're embedding only supplementary information like indices or bloom filters, a reader can still continue working with the columnar data in Parquet as it used to; it is just that it won't be able to take advantage of the additional metadata.
> Note that the readers of Parquet need to be aware of any metadata to exploit it. But if not, nothing changes
The one downside of this approach, which is likely obvious, but I haven't seen mentioned is that the resulting parquet files are larger than they would be otherwise, and the increased size only benefits engines that know how to interpret the new index
There is no spec. Personally I hope that the existing indexes (bloom filters, zone maps) get re-designed to fit into a paradigm where parquet itself has more first class support for multiple levels of indexes embedded in the file and conventions for how those common types. That is, start with Wild West and define specs as needed
> That is, start with Wild West and define specs as needed
Yes this is my personal hope as well -- if there are new index types that are widespread, they can be incorporated formally into the spec
However, changing the spec is a non trivial process and requires significant consensus and engineering
Thus the methods used in the blog can be used to use indexes prior to any spec change and potentially as a way to prototype / prove out new potential indexes
The story here isn't that they've invented a new format for user defined indexes (the one proposed here is sort of contrived and I probably wouldn't recommend in production) but rather demonstrating how the user defined metadata space of the parquet format can be used for application specific purposes.
I work on a database engine that uses parquet as our on-storage file format and we make liberal use of the custom metadata area for things specific to our product that any other parquet readers would just ignore.
The Arrow/Parquet community is already discussing standardization via the Parquet format GitHub - this approach intentionally uses existing extension points in the format specification to remain compatible while the standardization discussions progress.
I think this post is a response to some new file format initiatives, based on the criticism that the Parquet file format is showing its age.
One of the arguments is that there is no standardized way to extend Parquet with new kinds of metadata (like statistical summaries, HyperLogLog etc.)
This post was written by the DataFusion folks, who have shown a clever way to do this without breaking backward compatibility with existing readers.
They have inserted arbitrary data between footer and data pages, which other readers will ignore. But query engines like DataFusion can exploit it. They embed a new index to the .parquet file, and use that to improve query performance.
In this specific instance, they add an index with all the distinct values of a column. Then they extend the DataFusion query engine to exploit that so that queries like `WHERE nation = 'Singapore'` can use that index to figure out whether the value exists in that .parquet file without having to scan the data pages (which is already optimized because there is a min-max filter to avoid scanning the entire dataset).
Also in general this is a really good deep dive into columnar data storage.
One question that the article does not cover: compaction. Adding custom indexes means you have to have knowledge of the indexes to compact Parquet files, since you'll want to reindex each time compaction occurs. Otherwise the indexes will at best be discarded. At worst they would even be corrupted.
So it looks as if adopting custom indexes mean you are adopting not just a particular engine for reading but also a particular engine for compaction. That in turn means you can't use generic mechanisms like the compaction mechanism in S3 table buckets. Am I missing something?
What are the new file format initiatives you're referencing here?
This solution seems clever overall, and finding a way to bolt on features of the latest-and-greatest new hotness without breaking backwards compatibility is a testament to the DataFusion team. Supporting legacy systems is crucial work, even if things need a ground-up rewrite periodically.
Off the top of my head:
- Vortex https://github.com/vortex-data/vortex
- Lance https://github.com/lancedb/lance
- Nimble https://github.com/facebookincubator/nimble
There are also a bunch of ideas coming out of academia, but I don't know how many of them have a sustained effort behind them and not just a couple of papers
Lance (from LanceDB folks), Nimble (from Meta folks, formerly known as Alpha); I think there are a few others
https://github.com/lancedb/lance
https://github.com/facebookincubator/nimble
I’ve been excited about lancedb and its ability to support vector indexes and efficient row level lookups. I wonder if this approach would work for their design goals and still allow broader backwards compatibility with the parquet ecosystem. Have been intrigued by Ducklake, and they’ve leaned into parquet. Perhaps this approach will allow more flexible indexing approaches with support for the broader parquet ecosystem which is significant.
Yeah I'm happy to see this, we have been curious as part of figuring out cloud native storage extensions to GFQL (graph dataframe-native query lang), and my intuition was parquet was pluggable here... And this is the first I'm seeing a cogent writeup.
Likewise, this means, afaict, it's likewise pretty straightforward to do novel indexing schemes within Iceberg as well just by reusing this.
The other aspect I've been curious about is the happy path pluggable types for custom columns. This shows one way, but I'm unclear if same thing.
We are actively working on supporting extension types. The mechanism is likely to be using the Arrow extension type mechanism (a logical annotation on top of existing Arrow types https://arrow.apache.org/docs/format/Columnar.html#format-me...)
I expect this to be used to support Variant https://github.com/apache/datafusion/issues/16116 and geometry types
(note I am an author)
I'm not sure if this is what you're looking for, but there is a proposal in DataFusion to allow user defined types. https://github.com/apache/datafusion/issues/12644
Thank you, looking forward to reading!
nice summary!
Note that there are "Puffin files" associated with Iceberg which have some overlap with this functionality: https://iceberg.apache.org/puffin-spec/#file-structure
Speaking of Puffin files, Apache Parquet always makes me think of this 1978 SNL intro with Bill Murray and SNL bass player Buddy Williams:
https://snltranscripts.jt.org/77/77sparaquat.phtml
Cool, but this is very specific to DataFusion, no? Is there any chance this would be standardized so other Parquet readers could leverage the same technique?
The technique can be applied by any engine, not just DataFusion. Each engine would have to know about the indexes in order to make use of them, but the fallback to parquet standard defaults means that the data is still readable by all.
But does data fusion publish a specification of how this metadata can be read, along with a test suite for verifying implementations? Because if they don't, this cannot be reliably used by any other impl
Parquet files include a field called key_value_metadata in the FileMetadata structure; it sits in the footer of the file. See: https://github.com/apache/parquet-format/blob/master/src/mai...
The technique described in the article, seems to use this key-value pair to store pointers to the additional metadata (in this case a distinct index) embedded in the file. Note that we can embed arbitrary binary data in the Parquet file between each data page. This is perfectly valid since all Parquet readers rely on the exact offsets to the data pages specified in the footer.
This means that DataFusion does not need to specify how the metadata is interpreted. It is already well specified as part of the Parquet file format itself. DataFusion is an independent project -- it is a query execution engine for OLAP / columnar data, which can take in SQL statements, build query plan, optimize them, and execute. It is an embeddable runtime with numerous ways to extend it by the host program. Parquet is a file format supported by DataFusion because it is one of the most popular ways of storing data in a columnar way in object storages like S3.
Note that the readers of Parquet need to be aware of any metadata to exploit it. But if not, nothing changes - as long as we're embedding only supplementary information like indices or bloom filters, a reader can still continue working with the columnar data in Parquet as it used to; it is just that it won't be able to take advantage of the additional metadata.
> Note that the readers of Parquet need to be aware of any metadata to exploit it. But if not, nothing changes
The one downside of this approach, which is likely obvious, but I haven't seen mentioned is that the resulting parquet files are larger than they would be otherwise, and the increased size only benefits engines that know how to interpret the new index
(I am an author)
So, can we take that as a "no"?
There is no spec. Personally I hope that the existing indexes (bloom filters, zone maps) get re-designed to fit into a paradigm where parquet itself has more first class support for multiple levels of indexes embedded in the file and conventions for how those common types. That is, start with Wild West and define specs as needed
> That is, start with Wild West and define specs as needed
Yes this is my personal hope as well -- if there are new index types that are widespread, they can be incorporated formally into the spec
However, changing the spec is a non trivial process and requires significant consensus and engineering
Thus the methods used in the blog can be used to use indexes prior to any spec change and potentially as a way to prototype / prove out new potential indexes
(note I am an author)
The story here isn't that they've invented a new format for user defined indexes (the one proposed here is sort of contrived and I probably wouldn't recommend in production) but rather demonstrating how the user defined metadata space of the parquet format can be used for application specific purposes.
I work on a database engine that uses parquet as our on-storage file format and we make liberal use of the custom metadata area for things specific to our product that any other parquet readers would just ignore.
The Arrow/Parquet community is already discussing standardization via the Parquet format GitHub - this approach intentionally uses existing extension points in the format specification to remain compatible while the standardization discussions progress.