[BUG] on_change events of DataTable CDS fail when source is set after init

Software version info bokeh 2.0.0 chrome 80.0.3987.132 (Official Build) (64-bit) ubuntu 19.10

Description of expected behavior and the observed behavior Context:

• An on_change(‘data’, callback) is registered on the ColumnDataSource which is attached to a DataTable
• An on_change(‘indices’, callback) is registered on the “selected” field of a ColumnDataSource which is attached to a DataTable

Expect:

• The callback will be called while the content in the DataTable is edited
• The callback will be called while a row in the DataTable is selected

Observed:

• Both callbacks are not called

Complete, minimal, self-contained example code that reproduces the issue

from bokeh.io import curdoc
from bokeh.models import TableColumn, ColumnDataSource, DataTable

columns = [TableColumn(field='a', title='a'),
             TableColumn(field='b', title='b'),
             TableColumn(field='c', title='c')]

cds = ColumnDataSource(dict(a=[1,2,3],b=[2,3,4], c=[3,4,5]))

table = DataTable(editable=True, width=300, height=150)
table.source = cds
table.columns = columns
def on_data(attr, old, new):
    print('on_data')

def on_select(attr, old, new):
    print('on_select')


table.source.on_change('data', on_data)
table.source.selected.on_change('indices', on_select)

curdoc().add_root(table)

[bryevdv]

@maeglin89273 your code works for me if I change to pass source directly to DataTable:

table = DataTable(editable=True, width=300, height=150, source=cds)

It should work the way you have it too, and I am not sure why it does not, but hopefully this is an immediate solution for you.

[bryevdv]

@mattpap for reference this solves the issue:

diff --git a/bokeh/document/util.py b/bokeh/document/util.py
index 3583323b2..ab012ec80 100644
--- a/bokeh/document/util.py
+++ b/bokeh/document/util.py
@@ -125,6 +125,7 @@ def references_json(references):

     references_json = []
     for r in references:
+        r._finalize()
         struct = r.struct
         struct['attributes'] = r._to_json_like(include_defaults=False)
         references_json.append(struct)
diff --git a/bokeh/model.py b/bokeh/model.py
index c6594fb1f..e5f17077e 100644
--- a/bokeh/model.py
+++ b/bokeh/model.py
@@ -774,6 +774,16 @@ class Model(HasProps, PropertyCallbackManager, EventCallbackManager):

         return html

+    def _finalize(self):
+        """ This methods can be oeverriden if  a model needs to defer some
+        configuration until serialization time. E.g. if a model needs to set
+        up a relation bewteen two sub-components.
+
+        WARNING: no new models should ever be created inside this method
+
+        """
+        pass
+
 #-----------------------------------------------------------------------------
 # Private API
 #-----------------------------------------------------------------------------
diff --git a/bokeh/models/widgets/tables.py b/bokeh/models/widgets/tables.py
index 9d2c5aa38..aabf093cd 100644
--- a/bokeh/models/widgets/tables.py
+++ b/bokeh/models/widgets/tables.py
@@ -610,16 +610,15 @@ class TableWidget(Widget):
     The source of data for the widget.
     """)

-    view = Instance(CDSView, help="""
+    view = Instance(CDSView, default=lambda: CDSView(), help="""
     A view into the data source to use when rendering table rows. A default view
     of the entire data source is created if a view is not passed in during
     initialization.
     """)

-    def __init__(self, **kw):
-        super().__init__(**kw)
-        if "view" not in kw:
-            self.view = CDSView(source=self.source)
+    def _finalize(self):
+        if self.view.source is None:
+            self.view.source = self.source

 class DataTable(TableWidget):
     ''' Two dimensional grid for visualisation and editing large amounts

Would you be in favor of this approach to get rid of initializers on models more generally?