Save Bokeh Plot as Bokeh Plot

[mrocklin]

I would like to save a snapshot of Bokeh plot created by a Bokeh server application in all of the following ways:

1. Download the contents of the ColumnDataSource as JSON/msgpack/whatever
2. Download a static HTML file of the Bokeh plot
3. Publish that bokeh plot as a gist

Dask's use case

For example with Dask.distributed, the current best way to have a discussion about performance is by referring to our task-stream diagnostic Bokeh plot. This occurs when I write blogposts discussing algorithms and when users have performance questions. If there was a Bokeh Tool that let users publish a static view of their task-stream plot as a gist from within the Dask Dashboard then it would elevate the level of conversation significantly (and produce a lot of cool looking Bokeh images). As a pleasant side effect, I would start including Bokeh plots way more often when writing blogposts and documentation.

How to do this?

When I briefly mentioned this to @birdsarah she recommended building a custom tool like the current SaveTool for Dask which would dump to one of the above forms rather than to png. However, looking at this a bit more I don't think that this functionality necessarily needs to be Dask-specific. This may be of general utility and would, I think, encourage people to embed Bokeh plots more often within broadcast publications. For example if they built a Bokeh plot in a notebook but wanted to include it in a blogpost then this might help.

[gryBox]

This is something we are tried/trying to jerry rig. Some of the issues we ran into were (1) legend placement (2) combining multiple widgets with widgets that were not on a canvas (3) data tables.

Download the contents of the ColumnDataSource as JSON/msgpack/whatever

That was trivial, since ColumnDataSource (for us) originates as a dataframe. Personally not a fan of msgpack for this user case. We ran into datetime datatype issues when we wanted to read the file back for debugging.

Publish that bokeh plot as a gist

Yes!

I for one would love to have all three functions proposed by @mrocklin

[mattpap]

Download the contents of the ColumnDataSource as JSON/msgpack/whatever

As part of "whatever" we could consider CSV as proposed in PR #4228. Allowing JSON could be useful as well.

[gryBox]

Using the Bokeh stock example as a simple use case to illustrate some problems we encountered.

Problem(s)

There are plenty of times where the user may want to save only certain widgets/figures.
Example:

1. A debug pretext window under all the plots - Rarely needs to get saved
2. In the example above, maybe only the figures need to be saved.
3. Data tables:

• Saving bokeh pretext and data tables is not currently supported out of the box. Is this a separate PR? My hope was to eventually make legends out of data tables -
• They can be 'to large to save.' I have seen comments on the bokeh board where people are loading copious amounts of data into tables. Imagine the pretext description in the doc above. What if that was a table with ~100,000 records. Does it get saved as part of the HTML or only what's 'being viewed'?

Since every Bokeh doc is unique it becomes hard to generalize what the user wants saved or not.

Proposed Solution

Creating a tools for a bokeh doc. It's first widget would be the SaveTool with the behavior of the HoverTool checkbox. Where each checkbox is mapped to the grid in `layout'.

# set up layout
widgets = column(ticker1, ticker2, stats)
main_row = row(corr, widgets)
series = column(ts1, ts2)
layout = column(main_row, series)

# List of checkboxes that magically appear in bokeh SaveTool
SaveTool(... ,components=[layout, main_row, series, corr, widgets, ticker1, ticker2, stats])

# Or a dictionary that can accept default values
SaveTool(... ,components= {layout : False, main_row: True ...})

[gryBox]

Download the contents of the ColumnDataSource as JSON/msgpack/whatever

@mrocklin Would odo work?

[mrocklin]

@mrocklin Would odo work?

Probably not, no

[bryevdv]

All the current tools apply only to the plot they are on, and I don't think that should change. That is, an export tool that was on a particular plot should save information only for recreating that particular plot. If the intent is to save an entire layout or document, I think that should be handled differently. For example, a button, outside of any particular plot, could be configured with what to save. Although, I think no-configuration would be easier for users and developing alike: a button callback that just exports the current document as-is in its current state would be my definite preference.

[bryevdv]

There are two possible routes to go that I can think of:

• export directly from BokehJS

  pros: works for any document, server or standalone
  cons: a fair amount of BokehJS code to write, no doc->json code exists currently in BokehJS (only the converse, json->doc exist)

• export from a python server callback

  pros: probably easiest to get something up and running for the "whole-document" case, Bokeh python side already knows how to export docs to JSON
  cons: only works for server apps

[mrocklin]

Personally I care more about saving single plots than entire documents.
I'm also obviously biased towards bokeh server apps.

[bryevdv]

Personally I care more about saving single plots than entire documents.

That's probably actually a fairly harder case. Basically individual models can't be meaningfully or faithfully serialized outside the context of the entire document they inhabit. What if the plot has a CustomJS callback and the callback is configured with models or data sources for/from other plots? Things can't really be taken in isolation like that. The more I think about it, for a general tool, "whole-document" export is probably the only realistic option.

@mrocklin this is another reason I think you might be better off with a custom tool. A custom tool can bake in simplifying assumptions (in your case about the structure of the Dask dashboard app, and what is the exact and "minimum" set of things that could be surgically exported) that a general tool can't make.

[mrocklin]

OK, sounds like I should pursue this through a custom tool. There is really just one major plot that I care about. I'm ok to close this.

[bryevdv]

Here is a reference implementation of whole document export:

import numpy as np
from numpy import pi

from bokeh.core.properties import Int, String
from bokeh.driving import cosine
from bokeh.embed import file_html
from bokeh.io import curdoc
from bokeh.models import Tool
from bokeh.plotting import figure
from bokeh.resources import CDN


JS_CODE = """
p = require "core/properties"
ActionTool = require "models/tools/actions/action_tool"


class ExportToolView extends ActionTool.View

  initialize: (options) ->
    super(options)
    @listenTo(@model, 'change:content', @export)

  do: () ->
    # This is just to trigger an event a python callback can respond to
    @model.event = @model.event + 1

  export: () ->
    if @model.content?
      x = window.open()
      x.document.open()
      x.document.write(@model.content)
      x.document.close()

class ExportTool extends ActionTool.Model
  default_view: ExportToolView
  type: "ExportTool"
  tool_name: "Export"
  icon: "bk-tool-icon-save"

  @define {
    event:   [ p.Int,   0 ]
    content: [ p.String   ]
  }

module.exports =
  Model: ExportTool
  View: ExportToolView
"""


class ExportTool(Tool):
    __implementation__ = JS_CODE
    event   = Int(default=0)
    content = String()

x = np.linspace(0, 4*pi, 80)
y = np.sin(x)

def export_callback(attr, old, new):
  # really, export the doc as JSON
  export_tool.content = None
  html = file_html(p, CDN, "exported plot")
  export_tool.content = html

export_tool = ExportTool()
export_tool.on_change('event', export_callback)

p = figure(tools=[export_tool])
r1 = p.line([0, 4*pi], [-1, 1], color="firebrick")
r2 = p.line(x, y, color="navy", line_width=4)

@cosine(w=0.03)
def update(step):
    r2.data_source.data["y"] = y * step
    r2.glyph.line_alpha = 1 - 0.8 * abs(step)

curdoc().add_root(p)
curdoc().add_periodic_callback(update, 50)