This would be great cleanup, big +1 from me for all three of these adjustments, including the promotion of pe, ltt, and friends out of the test scope.
I believe that we included hbase test jars because we used to freely mix classes needed for minicluster between runtime and test jars, which in turn relied on Hadoop minicluster capabilities. The big cleanup around HBaseTestingUtil/it addressed much (or all) of these issues on branch-3. I believe that we include a Hadoop distribution in our assembly because that makes it easy for a new user to download our release bin.tgz and get started immediately with learning. I guess it’s high time that we work out the with- and without-Hadoop variants. Thanks, Nick On Tue, 5 Mar 2024 at 09:14, Istvan Toth <st...@apache.org> wrote: > DISCLAIMER: I don't have a patch ready, or even an elegant way mapped out > to achieve this, this is about discussing whether we even want to make > these changes. > These are also substantial changes, but they could be targeted for HBase > 3.0. > > One issue I have noticed is that we ship test jars and test dependencies in > the assembly. > I can't see anyone using those, but it bloats the assembly and classpath, > and adds unnecessary JARs with possible CVE issues. (for example Kerby > which is a Hadoop minicluster dependency) > > My proposal is to exclude the test jars and the test scope dependencies > from the assembly. > > The advantages would be: > * Smaller distro size > * Faster startup (this is marginal) > * Less CVE-prone JARs in the binary assemblies > > The other issue is that the assembly includes much of the Hadoop > distribution. > The basic assumption in all scripts and instructions is that the node has a > fully configured Hadoop installation, and we include it in the classpath of > HBase. > > If that is true, then there is no reason to include Hadoop in the assembly, > HBase and its direct dependencies should be enough. > > One could argue that it would simplify the client side, which is true to > some extent (though 95% of the client distro use cases are served better by > simply using hbase-shaded-client). > > We could either remove the Hadoop libraries from either or both of the > assemblies unconditionally, or provide two variants for either or both > assemblies, one with Hadoop included, and one without it. > Spark already does this, it has binary distributions both with and without > Hadoop. > > The advantages would be: > * Smaller distro size > * Faster startup (this is marginal) > * Less chance of conflicts with the Hadoop jars > * Less CVE-prone JARs in the binary assemblies > > > Thirdly, we could consider excluding the > full-fat org.apache.hbase:hbase-shaded-client JAR from the Hadoop-less > binary assemblies. It is not used by the assembly, and AFAIK it is not > included in any of the 'hbase classpath' command variants. > > This would make sure that no Hadoop libraries are included (even in shaded > form) and would make the HBase distribution fully insulated from Hadoop's > CVE issues. > > (The full-fat hbase-shaded-client works best as direct build-time > dependency anyway) > > best regards > Istvan >
