The advantage of the annotations is that Tom White already did the work for jdiff to ignore non-public classes over in Hadoop land. We could leverage that work, whether we re-use the o.a.h.classification annotations or add our own copies in org.apache.hbase.*.
-Todd On Mon, Apr 8, 2013 at 3:08 PM, lars hofhansl <la...@apache.org> wrote: > It seems we could just generally document that: > - no RPC incompatibilities > - no API breaking changes to any user facing classes (now we'll pay better > attention to this) > - best effort to keep coprocessor API changes backward compatible > > If - on the other hand - we wanted to automate API checks then we'd need > tagging (either in form of an annotation or Javadoc) > > +1 on the javadoc tagging if you're willing to take than on. As other have > said -1 on pulling Interface Audience in. > Your set of classes looks good. > > -- Lars > > > > ________________________________ > From: Elliott Clark <ecl...@apache.org> > To: "dev@hbase.apache.org" <dev@hbase.apache.org> > Sent: Monday, April 8, 2013 1:49 PM > Subject: Re: Declaring HBase Public API in 0.94 > > Please don't pull in @InterfaceAudience. Keeping 0.2x compatibility was > something that was hard won in 0.94, it would be a real shame to loose that > now. > > > On Mon, Apr 8, 2013 at 11:07 AM, Aleksandr Shulman <al...@cloudera.com > >wrote: > > > Hi everyone, > > > > In light of all the conversation on compatibility, I wanted to float the > > idea of documenting which Java packages, classes, and methods we want to > > declare as being API compatible in 0.94.x. I'd like your input on: > > 1. JavaDoc vs. using AudienceInterface > > 2. What the javadoc notation should look like > > 3. Which pieces of code should be tagged > > > > What do I mean by documenting API compatibility? That means that we > suggest > > the anyone building applications use specific methods because they would > > continue to be both binary and RPC-compatible going forward. Any > > application written, either running on a node of a cluster or on a remote > > machine, would continue to work properly without recompile for all > versions > > of 0.94.x running on the cluster. > > > > *Benefits:* > > It would prevent developers from using calls that are subject to change. > > This would give developers more confidence in using the platform, which > > will encourage more development on our platform. > > 0.94 will still be with us for some time and I think the > > better-late-than-never approach will save us pain down the road. Finally, > > it would allow us to more easily verify that we are in fact API > compatible. > > > > *Can we use AudienceInterface?* > > HBase 0.94 can be compiled against both hadoop 0.2x, 1.x, and 2.0.x. In > the > > case of 0.2x, the AudienceInterface classes were not bundled. Therefore, > we > > cannot expect HBase 0.94 to support it. For that reason, I think JavaDoc > > might be better. > > On the other hand, perhaps we might just want to bundle AudienceInterface > > with 0.94 going forward? Then we can have consistent annotations in 0.94, > > 0.95, and 0.96 without worrying about the hadoop version. > > > > Please correct me if I'm wrong about any of the above. > > > > *Clarification of RPC compatibility:* > > We care about RPC compatibility when we create clients that bundle their > > dependency jars with them. These jars are used to form a request that is > > executed on a remote machine (i.e. the cluster). If the cluster is > upgraded > > and no longer recognizes the command, then this will break RPC > > compatibility. > > > > *Clarification of Binary compatibility:* > > We care about binary compatibility when a client is created and compiled, > > and the jars on which is depends change. It should still be able to form > > requests using those jars. If the cluster is upgraded and the compiled > > client code cannot find a method it was depending on to be there, we > break > > binary compatibility. A recent example is in 0.94.2, where the return > type > > of HColumnDescriptor.setMaximumVersions was changed and those who > upgraded > > received this error: > > > > java.lang.NoSuchMethodError: org.apache.hadoop.hbase.** > > HColumnDescriptor.**setMaxVersions(I)V > > > > *What we currently have:* > > We have an @audience annotation set up in 0.95/0.96. In 0.94, I suggest > > either adding JavaDoc or pulling in the AudienceInterface annotation. > > > > *Suggested Javadoc language:* > > @custom.94_api > > > > *Granularity:* > > Just to the class level. The native java access level (e.g. public, > > protected, etc.) should indicate what should be kept compatible. > > > > *Suggested classes:* > > Here is a first-cut of things that should be declared and documented as > > public APIs. This list was obtained from looking at some MapReduce over > > HBase example code. > > > > *JAVA API:* > > > > *org.apache.hadoop.hbase (some selected classes, see below) > > org.apache.hadoop.hbase.client.* > > org.apache.hadoop.hbase.filter.* > > org.apache.hadoop.hbase.io.hfile.Compression.Algorithm > > org.apache.hadoop.hbase.util.* > > org.apache.hadoop.hbase.mapreduce.** > > > > *REST API: > > org.apache.hadoop.hbase.rest.client.** > > > > *Thrift API: > > All methods defined in: > > /hbase/src/main/resources/org/apache/hadoop/hbase/thrift/Hbase.thrift* > > > > *Selected classes in org.apache.hadoop.hbase:* > > > > *import org.apache.hadoop.hbase.ClusterStatus; > > import org.apache.hadoop.hbase.HBaseConfiguration; > > import org.apache.hadoop.hbase.HColumnDescriptor; > > import org.apache.hadoop.hbase.HRegionInfo; > > import org.apache.hadoop.hbase.HRegionLocation; > > import org.apache.hadoop.hbase.HServerAddress; > > import org.apache.hadoop.hbase.HTableDescriptor; > > import org.apache.hadoop.hbase.KeyValue;* > > > > -- > > Best Regards, > > > > Aleks Shulman > > 847.814.5804 > > Cloudera > > > -- Todd Lipcon Software Engineer, Cloudera
