a Future makes sense though that would be a separate method right?
AutoCloseable.close() doesn't return Future.
Perhaps the new method could be called 'join' and it blocks until the
graph is fully closed.
I'm taking inspiration from embedded jetty.
On Tue, Oct 6, 2015 at 9:15 AM, Bryn Cooke <[email protected]> wrote:
Hi,
I'm not sure what the answer is here,
What I am sure of is that the goal of close should be a graceful shutdown.
Once a resource is closed then it should not accept new requests.
Descendant resources may or may not continue to be operational, for
instance you may allow ongoing transactions to complete but that is
implementation dependent.
The question is after calling close when how can you be sure that all
descendant resources are finished before returning to the user? Only by
keeping references to all spawned resources can you tell.
Perhaps Graph.close() should return a Future. That way you can block until
the graph is actually closed.
Bryn
On 05/10/15 20:56, Stephen Mallette wrote:
There are two tickets in JIRA that relate to the semantics of closing and
releasing resources in the Graph hierarchy:
+ Graph
+ TraversalSource
+ TraversalStrategy
+ GraphTraversal
+ Transaction
https://issues.apache.org/jira/browse/TINKERPOP3-789
https://issues.apache.org/jira/browse/TINKERPOP3-790
There's been some discussion on them already. I bring this to everyone's
attention as the change could have wide repercussions depending on the
direction it goes. The short of the matter is that currently:
1. Graph implements AutoCloseable, but we don't enforce the notion of
close() itself in the test suite.
2. TraversalSource should likely implement AutoCloseable as there are
sometimes resources that need to be released when a TraversalSource is no
longer in use.
3. Transaction implements AutoCloseable but it applies in the context of
how a Transaction will behave when Graph.close() is called.
Where is this all going? Matt Frantz summarized the thought points nicely
in this comment:
https://issues.apache.org/jira/browse/TINKERPOP3-790?focusedCommentId=14710389&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14710389
Basically, there's two options Matt identified:
1. Making it so that Graph.close() would release resources and effectively
close() all things it spawned is one approach, but I''m not sure how
straightforward/practical to implement that is. I suspect it will
complicate Graph implementations as well.
2. The alternative is that the things in this hierarchy can have their own
expensive resources which can be independently closed. I think that this
approach more closely fits with the code as we have it now and won't
unduly
burden implementers. TinkerPop Developers will have to know when to call
close() (e.g. if a TraversalStrategy has expensive resources and the user
assumes that a call to TraversalSource.close() will release those
resources, then they might be wrong - depends on the approach we decide
on).
When I first created 790, i had the second option in mind, but since Matt
brought wrote that comment, i figured it was worth thinking through as a
whole.
Once that is decided then we should figure out 789 which is how to enforce
the semantics of close() (e.g. what does a Graph do when it's closed and
someone calls graph.traversal()?).
Thoughts?
