Andy Seaborne created JENA-2049:
-----------------------------------
Summary: Binding Improvements
Key: JENA-2049
URL: https://issues.apache.org/jira/browse/JENA-2049
Project: Apache Jena
Issue Type: Improvement
Components: ARQ
Reporter: Andy Seaborne
Assignee: Andy Seaborne
Fix For: Jena 4.0.0
{{Binding}} is a map from variables to nodes used in query processing. It is
central to execution of SPARQL.
h3. Context
One feature is that binding objects form a tree. To add variables to an
existing binding, whether a join or assignment (BIND), a new binding object is
created with the original as parent. This avoid copy costs. The parent can be
shared across solutions.
This design leads to large numbers of {{Binding}} objects, the majority of
which are quite small (1 or 2 variables).
h3. Current Situation
For general building, the binding implementation is the general {{BindingMap}}.
This is used a lot because the number of variables is not easy to determine at
the time the {{Binding}} object is allocated.
{{Binding1}} is the special case of one var/node pair. This "one slot" binding
requires the application to know there will be only one slot binding so it has
limited use.
{{BindingMap}} is not immutable, although as a {{Binding}} the mutable
operations need to be accessed with a cast.
{{BindingMap}} is implemented using a Map ({{HashMap}}) with a special case of a
binding of one slot.
Maps are several java object including a array. It is quite large and Java
objects require initialization. For streaming queries, the total footprint is
quite low; the GC will reclaim space as results rows are finished with. For
queries that accumulate results (e.g general sort), bindings are held for a
period of time.
h3. Proposal
* A builder pattern to accumulate "(var,node)" pairs during a constriction
phase before a build step that can then choose the most appropriate immutable
{{Binding}} implementation.
This takes the work of selecting the right binding implementation off the
rest of the code.
* Expand the number fixed length bindings: Special cases for 0, 1, 2, 3, and 4
pairs.
* The builder can be reused (same thread) so the cost of object creation for
the builder
can be amortized.
* Truly immutable bindings for long term stability.
h3. Compatibility
Retain, marked deprecated, the {{BindingFactory}} methods to create a
{{BindingMap}} so existing code continues as before, with deprecation to
indicate it would be a good idea to switch to the new way.
h3. Non-goal
This is not a speed up (nor a slow down).
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
