This category provides a set of conventions for working with tables in row-based, columnar or chunked/batched columnar formats.
|Metadata of the table, maps name of each column to its type.|
|Data of the table, see table types|
|Number of rows|
loaders.gl deals with (and offers utilities to convert between) three different types of tables:
data consists of an
Object instances, each representing a row.
Columnar tables are stored as one array per column. Columns that are numeric can be loaded as typed arrays which are stored in contigous memory.
data is an
Object that maps column names to an array or typed array.
Contiguous memory has tremendous benefits:
A problem with columnar tables is that column arrays they can get very long, causing issues with streaming, memory allication etc. A powerful solution is to worked with chunked columnar tables, where columns is are broken into matching sequences of typed arrays.
The down-side is that complexity can increase quickly. Data Frames are optimized to minimize the amount of copying/moving/reallocation of data during common operations such e.g. loading and transformations, and support zero-cost filtering through smart iterators etc.
Using the Arrow API it is possible to work extremely efficiently with very large (multi-gigabyte) datasets.