I do not think it is there just for testing. The original patch includes this commit message:
Fix sstable formats configuration - refactored sstable format configuration - sstable formats are discovered via ServiceLoader ... So the fact it is explicitly stated like this does not seem to be related just to testing only but it is a feature which is officially advertised. Here Jacek asks if we can use ServiceLoader (1) and here you answer to him (2) that you 100% support this. I have not dug deep but it seems that it is not related to tests. Here (3) Caleb says in 1) that it mimics what Lucene does. To your points about "what happens when it is not there". Yes, this is a problem. But I do not think that this is something tremendously serious. This is for users who know what they do and why. They would need to specifically implement an interface and put it on CP and it would be their responsibility to ensure it is everywhere. If they ever want to get rid of that implementation, I think they would need to: 1) remove these constraints in the first place so they are not present in any schema 2) is not there something like TCM checkpoint or similar which would make it all consolidated? What I mean by that is that it would not be replayed from the very first epoch to the last one as it would go through epochs the constraints are not in anymore (hence it would fail, similar to (4)) There is a similar problem to this in 4) (same for custom indexes etc) and I do not think that we will solve this here. (1) https://issues.apache.org/jira/browse/CASSANDRA-18441?focusedCommentId=17711283&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-17711283 (2) https://issues.apache.org/jira/browse/CASSANDRA-18441?focusedCommentId=17711500&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-17711500 (3) https://issues.apache.org/jira/browse/CASSANDRA-18441?focusedCommentId=17711539&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-17711539 (4) https://issues.apache.org/jira/browse/CASSANDRA-20287 On Mon, Aug 11, 2025 at 6:55 PM David Capwell <dcapw...@apple.com> wrote: > I have 0 issues with ServiceLoader, but I think the bigger question is > “what is a public api”. In #1 I see the first comment is > > > The `CustomConstraintsProvider` class will create custom constraints. > > > This implies to mean that this is a new public API that must be > maintained, and the current process for these is to reach out to dev@. > We don’t have any process to even know what is public API nor do we do > anything to make sure we don’t break compatibility with java APIs. > > > In the example you shown with SSTables, the format isn’t public API (at > least thats my memory), this was kinda mostly done to enable testing. AKA > if someone tries to use this and one of our patch fix versions makes a > breaking change, thats on you for using a private API. > > > 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. > > If they are not stored in TCM and side loaded into each node, what happens > when some of the nodes do not have them loaded? This had been a big pain > point with the memtable patch (issue is partially fixed, we no longer > crash, but we no longer respect user configs either). > > > > On Aug 11, 2025, at 8:31 AM, Štefan Miklošovič <smikloso...@apache.org> > wrote: > > > > 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 > >
