Use glue's ability to randomly sample values when computing histogram…

…s and image statistics, and remove code to handle NaN values that caused the whole array to be loaded into memory.

jdaviz/configs/default/plugins/plot_options/plot_options.py

@@ -36,6 +36,8 @@
 
 __all__ = ['PlotOptions']
 
+RANDOM_SUBSET_SIZE = 10_000
+
 
 def _register_random_cmap(
     cmap_name,
@@ -894,6 +896,7 @@ def _update_stretch_hist_sync(self, msg={}):
     @skip_if_no_updates_since_last_active()
     @with_spinner('stretch_hist_spinner')
     def _update_stretch_histogram(self, msg={}):
+
         if not hasattr(self, 'viewer'):  # pragma: no cover
             # plugin hasn't been fully initialized yet
             return
@@ -999,15 +1002,21 @@ def _update_stretch_histogram(self, msg={}):
             arr = comp.data
             sub_data = arr.ravel()
 
-        # filter out nans (or else bqplot will fail)
-        if np.any(np.isnan(sub_data)):
-            sub_data = sub_data[~np.isnan(sub_data)]
-
+        self.stretch_histogram.viewer.state.random_subset = RANDOM_SUBSET_SIZE
         self.stretch_histogram._update_data('histogram', x=sub_data)
 
         if len(sub_data) > 0:
-            interval = PercentileInterval(95)
-            hist_lims = interval.get_limits(sub_data)
+
+            # Use glue to compute the statistics since this allows us to use
+            # a random subset of the data to compute the histogram
+            glue_data = self.stretch_histogram.app.data_collection['histogram']
+            hist_lims = (
+                glue_data.compute_statistic('percentile', glue_data.id['x'],
+                                            percentile=2.5, random_subset=RANDOM_SUBSET_SIZE),
+                glue_data.compute_statistic('percentile', glue_data.id['x'],
+                                            percentile=97.5, random_subset=RANDOM_SUBSET_SIZE)
+            )
+
             # set the stepsize for vmin/vmax to be approximately 1% of the range of the
             # histogram (within the percentile interval), rounded to 1-2 significant digits
             # to avoid random step sizes.  This logic is somewhat arbitrary and can be safely