3.0

[bmschmidt]

Major refactor.

3.0.0 includes a number of pent-up breaking changes. The underlying motivation for many of these is to allow library to now fully pass typescript compilation tests, with all the stability benefits that provides.

Breaking changes:

1. The library is now structured as named exports,
   rather than a single default export. Instead of

   import Scatterplot from 'deepscatter';

   a typical first line will be

   import { Scatterplot } from 'deepscatter';

   This allows the export of several useful types for advanced functions in scatterplots we've found useful at Nomic. The initial set of exported items are {Dataset, Bitmask, Scatterplot}.

2. The distinction between QuadTile and ArrowTile
   has been eliminated in favor of Tile, and with it the need to provide
   generics around them through the system. Similarly, QuadTileDataset and ArrowDataset are both removed in favor of Dataset.
   Instead, the TileProxy object is used to provide a wrapper than can turn anything into a
   dataset. Although datasets are presumed to be quadtiles right now, formally they can be any
   any collection of arrow batches structured as a tree. (This is increasingly how I've come to think of the data parts of deepscatter: as a system for navigating dataframes that consist of trees rather than of linear lists of points.)

3. Deepscatter no longer accepts strings as direct
   arguments to Scatterplot.plotAPI in places where they were previously cast to functions
   as lambdas, because linters rightfully get crazy mad about the unsafe use of eval. If
   you want to use deepscatter in scrollytelling
   contexts where definining functions as strings inside json is convenient (I still will do this myself in static sites) you must turn them
   into functions before passing them into deepscatter.

4. Shortcuts for passing position and position0 rather than naming the x and y dimensions explicitly have been removed.

5. The behavior of categorical scales in certain circumstances has been tightened; it is possible, as a result, that places where it previously possible to treat categorical scales as numbers (referring to the underlying ints) will no longer work. I am not aware of specific such issues at the moment, and will act responsively to address any issues.

6. Dataset and Tile objects can now be instantiated with a manifest that allows listing all the tiles in a dataset. When passed, this allows a dataset to instantiate all tiles at creation time without actually loading any data. This represents a major change for any code that access the Tile.record_batch attributes, because they may now error on well-formed tiles since the presence of data is no longer necessary for something to be a Tile. Additionally, the Tile.ready promise has been retired; instead, to check if necessary data exists for a dataset, you must explicitly check if Tile.hasLoadedColumn('foo').

7. (Another way of expressing this change is that where previously there was a 'primary record batch' and 'sidecar batches', in version 3.0 of deepscatter this distinction is much less important; it is possible, for example, to draw a scatterplot without loading the x and y columns if other columns are passed to encoding.x and encoding.y.)

8. The tile prioritization rules which previously applied to core tiles in a dataset now apply to all sidecars as well.

9. It is possible to aggressively load any columns to any depth in the dataset without loading other data using 'Dataset.spawnDownloads()' and Dataset.runDownloads().

10. The Dataset object is now more independent of the scatterplot, to the point that it can independently run in a non-browser environment like Node. See the unit tests for an example of this.