Abstrakt
There is a multitude of tools for preparation of Linked Data from data sources such as CSV and XML files. These tools usually perform as expected when processing examples, or smaller real world data.
However, a majority of these tools become hard to use when faced with a larger dataset such as hundreds of megabytes large CSV file. Tools which load the entire resulting RDF dataset into memory usually have memory requirements unsatisfiable by commodity hardware.
This is the case of RDF-based ETL tools. Their limits can be avoided by running them on powerful and expensive hardware, which is, however, not an option for majority of data publishers.
Tools which process the data in a streamed way tend to have limited transformation options. This is the case of text-based transformations, such as XSLT, or per-item SPARQL transformations such as the streamed version of TARQL.
In this paper, we show how the power and transformation options of RDF-based ETL tools can be combined with the possibility to transform large datasets on common consumer hardware for so called chunkable data - data which can be split in a certain way. We demonstrate our approach in our RDF-based ETL tool, LinkedPipes ETL.
We include experiments on selected real world datasets and a comparison of performance and memory consumption of available tools.