I want to ask what people think about ServiceLoader's (Java SPI) in the context of Cassandra.
There is already a precedent with ServiceLoader<SSTableFormat.Factory> loader = ServiceLoader.load(SSTableFormat.Factory.class, DatabaseDescriptor.class.getClassLoader()); in DatabaseDescriptor.applySSTableFormats I am asking because I would like to do (1, 2) and I find SPI to be the most suitable for it. Just drop a JAR among Cassandra jars and be done with it. This means that there is no change in the configuration of cassandra.yaml and similar. On the other hand, we kind of lose the visibility of what we are actually running. What I mean by that is that when it is configured in cassandra.yaml and done as "ParameterizedClass", it is pretty explicit what configuration we use for what. On the other hand, by placing a JAR to Cassandra class path, while it is super simple to do, the only visible way to tell a user if there is some custom stuff loaded is to log it as Cassandra starts. What do you think about the SPI approach in general? There might also be other advantages as providing more implementations and ordering them by priority etc. Is SPI's way of registering / loading custom implementations just fine and can be considered as something developers can choose to do if they think it fits the use-case enough? For constraints, I think it would be very welcome if people are allowed to code their own constraints and just put it on CP and it would be picked up automatically. It would speed up the development, they would not need to wait for us to merge them and release them and they could have proprietary / in-house constraints just for their use case while they do not need to patch Cassandra itself. Hence they can still use the official release but they could extend it as they want. Regards (1) https://github.com/apache/cassandra/pull/4297/files#diff-d6e8754d8326be6063d9d59265e882416d74007e47f07f809267b07f9860e26d (2) https://issues.apache.org/jira/browse/CASSANDRA-20824
