Some high-level work that we expect to be achieved ont he time-scale of months. This list is not exhaustive, but rather aims to whet the appetite for what Intake can be in the future.
Since Intake aims to be a community of data-oriented pythoneers, nothing written here is laid in stone, and users and devs are encouraged to make their opinions known!
Broaden the coverage of formats¶
Data-type drivers are easy to write, but still require some effort, and therefore reasonable impetus to get the work done. Conversations over the coming months can help determine the drivers that should be created by the Intake team, and those that might be contributed by the community.
Similarly, there are many third-party data services, some of which are paid and prorietary. Currently, SQLCatalog is the only example, and therefore reference, of this, but Intake will become more universally useful when it can act as a bridge to several other systems.
Integration with Apache Spark¶
The spark ecosystems and Intake will co-operate nicely! Firstly, Spark sources (i.e., named tables) will become standard data sources, so that the data can be streamed from Spark to a python process, and the data-sets referenced in a catalog as usual. These data-sets will necessarily be data-frame type, although an RDD-to-sequential method may also be possible
Later, automatic streaming of data into Spark should be possible also, with a to_spark() method appearing on data-frame (and maybe sequence, later) type sources.
Often, we can conceive of a data-type as being a modified version of another data-type. For example: the “csv” plugin produced data-frames from a set of files in the CSV format, while another plugin takes data-frames with a particular set of fields as input, and produces new data-frames based on some model predictions.
Rather than allow a general pipeline with arbitrary code specified in catalogues, we aim to allow the creation of arbitrary plugins, where the inputs are the outputs of other data-sources. This way, the logic stays in the code of the plugin, which can be distributed as python/conda packages as usual, but a path is in place to generate “second-order” data products. Naturally, such derived plugins ought to be thorough about describing the process in the metadata of the resultant data-source. Preliminary work: PR.