Edit this page

Dataset release format

RiverBench's datasets can all be treated as RDF streams or just flat RDF files. RiverBench uses the RDF Stream Taxonomy (RDF-STaX) to organize the datasets – we recommend you have at least a brief look at it to understand the dataset structure.

Each dataset has three types of distributions:

Flat – flat distributions are simply compressed files with a sequence of RDF statements (this is a flat RDF stream, according to RDF-STaX).
Stream – grouped streaming distributions are compressed archives with a sequence of files. Each file corresponds to one stream element – an RDF graph or an RDF dataset (this is a grouped RDF stream, according to RDF-STaX).
Jelly – Jelly streaming distributions use the high-performance Jelly serialization format that natively supports RDF streams. The Jelly files can be used for both the grouped and flat streaming formulations.

There always a few size variants of each distribution to choose from, starting from 10K stream elements, with 10x increases up to the full dataset. The full distribution is the longest available distribution of the dataset.

Tip

Choose the size-limited distributions, if your work does not require the full dataset – they may be easier to work with. You can also want to choose them to have all datasets of the same size.

Tip

Jelly distributions are much faster to load and can support all benchmark tasks. They can be used instead of both flat and stream distributions.

Distribution types

Flat distributions

Flat distributions are serialized in the N-Triples or N-Quads format, depending on stream type. In case RDF-star is used in the dataset, the used formats are N-Triples-star or N-Quads-star. The files are compressed with gzip. This format is streamable, as you can decompress gzip-compressed files on-the-fly and read the statements line-by-line.

The flat distribution files are named flat_{size}.nt.gz or flat_{size}.nq.gz, where {size} is the number of stream elements in the file. For example, flat_10K.nt.gz is a flat distribution file of a triple stream with 10,000 stream elements. The full distribution is denoted by flat_full.nt.gz or flat_full.nq.gz.

Stream distributions

In streaming distributions each stream element is represented by a separate file. The files are compressed in a .tar.gz archive. The files are sequentially named starting from 0000000000.Y up to X.Y, where X + 1 is the number of stream elements in the dataset, and Y is the file extension. All numbers are zero-padded to exactly ten digits. The files are in nested directories with at most 1000 files per directory, to avoid issues with some file systems and file browsers. The number of levels of directories depends on the size of the distribution, with 10K–1M distributions having one level of directories and larger distributions having two.

The files are laid out in the archive sequentially, that is, physically the bytes of files X and X+1 are next to each other. This allows the package to be processed one element at a time without decompressing the entire archive. To do that, you will need a streaming decompressor / untar utility like the one in Pekko Streams (tarReader and gunzip), or Apache Commons Compress.

Each element is serialized in either the Turtle or the TriG format, depending on the stream type. In case RDF-star is used in the dataset, the used formats are Turtle-star or TriG-star.

The streaming distribution files are named stream_{size}.tar.gz, where {size} is the number of stream elements in the file. For example, stream_10K.tar.gz is a streaming distribution file with 10,000 stream elements. The full distribution is denoted by stream_full.tar.gz.

Example – RDF graph stream, `stream_10K.tar.gz`

- 000/
    - 0000000000.ttl
    - 0000000001.ttl
    - ...
    - 0000000999.ttl
...
- 009/
    - 0000009000.ttl
    - 0000009001.ttl
    - ...
    - 0000009999.ttl

Example – RDF dataset stream, `stream_10M.tar.gz`

- 000/
    - 000/
        - 0000000000.trig
        - 0000000001.trig
        - ...
        - 0000000999.trig
    - 001/
        - 0000001000.trig
        - 0000001001.trig
        - ...
        - 0000001999.trig
    - ...
    - 999/
        - 0000099000.trig
        - 0000099001.trig
        - ...
        - 0000099999.trig
...
- 999/
    - 000/
        - 0009990000.trig
        - 0009990001.trig
        - ...
        - 0009990999.trig
    - ...
    - 999/
        - 0009999000.trig
        - 0009999001.trig
        - ...
        - 0009999999.trig

Jelly distributions

Jelly is a high-performance binary serialization format for RDF. RiverBench's Jelly distributions simply use delimited RdfStreamFrames to denote the individual elements in the stream. The streams are either of TRIPLES physical type (for RDF graph streams) or QUADS for RDF dataset streams. The resulting file is gzip-compressed.

Parsing Jelly files can be 7 to 18 times faster than N-Triples/N-Quads and 10 to over 40 times faster than Turtle/TriG, depending on the dataset and your hardware. This can massively speed up your benchmark execution. Dataset sizes should be more-or-less the same when compressed, but when uncompressed, Jelly will be 3–4 times smaller.

Reading Jelly files is currently supported in Apache Jena and RDF4J, using the jelly-jvm library. Please refer to Jelly's website for usage examples and documentation.

Package stability guarantees

RiverBench takes a best-effort approach to keep the dataset distributions as stable as possible to help with reproducibility of certain benchmark tasks across dataset versions.

The following guarantees are provided:

Statement order:
- Since RiverBench version 2.2.0, the order of statements (triples or quads) in flat and Jelly distributions is guaranteed to be the same across different dataset versions with identical content. The statements are ordered by stream element, and then within a stream element using alphabetic sorting (by: graph, subject, predicate, object).
- Before RiverBench version 2.2.0, the statement order could vary within the scope of a single stream element (RDF graph or RDF dataset).
- In stream distributions (Turtle/TriG), the statement order across different dataset versions MAY be preserved, but it's not guaranteed.
Blank node identifiers:
- Since RiverBench version 2.0.0, blank node identifiers are guaranteed to be stable across different dataset versions with identical content.
Syntax:
- RiverBench tries to preserve the syntax of the RDF files across different dataset versions, but it's not guaranteed.
- Notably, in RiverBench version 2.2.0, the stream distributions were changed to use the PREFIX directive for namespaces, instead of the @prefix directive. This change was done due to a change in the default behavior in Apache Jena, and PREFIX being apparently the future default in RDF 1.2.
File checksums:
- RiverBench tries to keep the checksums of the dataset files the same across different dataset versions, but it's not guaranteed, as there are no strict guarantees on syntax stability.