The behavior you observe is as designed. Whenever all the roads completely surrounding a region are included within the Service Area result, that entire region is included in the Service Area Polygon. This is so that the polygon isn't generated with an excessive number of "holes". For example, for a Service Area analysis in a standard city grid, you would expect your polygon to come back solid, as opposed to looking like a fish net.
To achieve what you are seeking, perform the Service Area analysis to generate Lines as output, and then run the Buffer tool on the resulting lines to generate polygons that only cover the areas that are within the specified buffer distance (say, 250 meters) from the road.
I am trying to understand in more detail how service area polygon geometry is generated.
The enclosed an image with few spikes highlighted. How or why are these generated?
I see the same results with a more detailed network i.e. Esri streetmap data.
Let me know if you need more information.
Recall that service area polygons are only meant to be accurate at the points where its boundary crosses the roads; what the boundary does in the space away from the roads is not formally defined and is mostly an "art", if you will. What Network Analyst tries to do to define the boundary in the open space is to take into account the impedance level (the travel time/distance away from the facility) of all the roads that surround the open spaces -- roads both inside and outside of the polygon are considered.
What makes things even more obscure is that polygons are meant to be a smooth and continuous two-dimensional surface. Unfortuantely the road network is often three-dimensional (with overpasses and underpasses) and directional (the two parallel one-way sleeves of a divided highway are separated by a very narrow median of space, and one-way streets that can only be accessed by "going around" from the other side), which means that the impedance level for a road (i.e., the travel time/distance that road is from the facility) can be drastically different for two roads that are in such close XY proximity. These usually are the culprit of most of the sharper spikes you see in the resulting polygon, as the polygon generator does its best to try to account for such drastically differing values in such a small space.
Hopefully this somewhat explains why the polygons are generated the way they are. Keep in mind that you can always export the polygon result to a feature class and smooth it out yourself to the desired visual appeal to effectively communicate the area that is "covered" by the facility in the particular scenario you're presenting.