[
https://issues.apache.org/jira/browse/CASSANDRA-7837?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14111408#comment-14111408
]
Jason Brown commented on CASSANDRA-7837:
----------------------------------------
I recently did some work on Gossiper to eliminate it as a singleton as a part
of a gossip simulation project I experimented with. The code wasn't too
heinous, but does introduce some layers of indirection. My first pass at it did
not buy us complete modularity of the gossip subsystem, but it actually became
possible to test our gossip in isolation.
Interested?
> Factor out static initialization / singletons
> ---------------------------------------------
>
> Key: CASSANDRA-7837
> URL: https://issues.apache.org/jira/browse/CASSANDRA-7837
> Project: Cassandra
> Issue Type: Improvement
> Reporter: Blake Eggleston
>
> I've been spending some time experimenting with ways to factor out static
> initialization and static singletons in Cassandra.
> This is a known issue, with obvious implications on testability and
> extensibility. It also leads to subtle changes in behavior of the database
> and tests, as access patterns to static methods reorder the initialization
> order of the static state. For instance, the test
> org.apache.cassandra.config.DatabaseDescriptorTest.testKSMetaDataSerialization
> is no longer testing anything, because Schema.getKeyspaceDefinitions() is
> now returning an empty array. Once the initialization order is changed, the
> test fails because system keyspaces aren't being filtered out, and fromThrift
> throws an exceptions when it sees LocalStrategy. Removing static state and
> singletons will make testing each component in isolation easier, leading to
> fewer regressions.
> With a large refactor like this, we'd want to avoid having the refactor
> branch, and the trunk branch diverging too far, or for too long. After a few
> failed attempts, I've worked out a way to perform the refactor in a way that
> can be performed with a series of (relatively) small patches, and avoid a
> big, scary merge. Each patch will build, the tests will pass, and can be
> merged into trunk daily/regularly as they are completed.
> The process is as follows:
> # Condense all static state into static singletons.
> (DatabaseDescriptor.getPartitioner() ->
> DatabaseDescriptor.instance.getPartitioner())
> ** In some cases, classes will need to be split into pairs of instance and
> factory classes, like in the case of Keyspace and ColumnFamilyStore
> # Identify all non-singleton usages of static state (grep for '.instance.'),
> and refactor to pass dependencies into their constructors from their
> instantiating services.
> # Work out and implement an initialization order for the singletons,
> preferring final members passed into their constructor, using volatile
> members assigned during startup where there are circular dependencies.
> ** We should consider splitting some of the singletons into more component
> specific classes before this step. There is definitely a case for making that
> another ticket, but with the non-singleton dependencies clearly defined at
> this point, we will have a clear idea of what can be split up. Doing that
> would reduce the number of circular dependencies between singletons, and make
> this stage a lot easier, with the additional benefit of further modularizing
> the code.
> # Remove static singleton instances. Initialize and start C* from a static
> method that takes a configuration as an argument.
> ** CassandraDaemon and utilities will need to be updated, and we'll need a
> helper method so tests can instantiate the database with a one or two liner.
> Obviously, this is a lot of work, but I think it's worth the effort. I've
> already gone through a few practice runs, and it's doable. I'm happy to do
> the work, with guidance and input from the C* team.
> Thoughts?
--
This message was sent by Atlassian JIRA
(v6.2#6252)
