Nick Dimiduk wrote:
> Well put, Nick.
>
> With Sean's point about the Hadoop shaded client, it seems to me that we
> have things which could be pursued in parallel:
>
> 1) Roadmap to Hadoop3 (and shaded hdfs client).
> 2) Identify components which we use from Hadoop, for each component:
> 2a) Work with Hadoop to isolate that component from other cruft (best
> example is the Configuration class -- you get something like 8MB of
> "jar" just to parse an xml file).
> 2b) Pull the implementation into HBase, removing dependency from
> Hadoop entirely.
>
> I think that both of these can/should be done in parallel to the
> isolation of the dependencies which HBase requires (isolating ourselves
> from upstream, and isolating downstream from us).
Hang on, these are two different concerns.
Isolating ourselves from Hadoop follows the line of thought around Hadoop's
shaded client jars. If we must have this for HBase2.0/Hadoop2.8, we can
probably backport their efforts as modules in our own build. See my early
comment about this being error prone for folks who re-package us. Either
way, it's time we firm up the boundaries between us and Hadoop.
Isolating our clients from our deps is best served by our shaded modules.
What do you think about turning things on their head: for 2.0 the
hbase-client jar is the shaded artifact by default, not the other way
around? We have cleanup to get our deps out of our public interfaces in
order to make this work.
+1 Worst case, people bloat their applications with dependencies (worst
case 2x the size), but it removes the runtime breakages due to multiple
versions of a class (because of HBase). I'd gladly pay size cost any day.
This proposal of an external shaded dependencies module sounds like an
attempt to solve both concerns at once. It would isolate ourselves from
Hadoop's deps, and it would isolate our clients from our deps. However, it
doesn't isolate our clients from Hadoop's deps, so our users don't really
gain anything from it. I also argue that it creates an unreasonable release
engineering burden on our project. I'm also not clear on the implications
to downstreamers who extend us with coprocessors.
I thought I had a reason as to why reducing our reliance on Hadoop
classes was important (even with a HDFS shaded client), but maybe my
only consideration was long-term cleanliness (avoid the double-packaging
of the same classes). I can't come up with an example anymore.
Constructing some exemplars for this work would likely be the best kind
of "acceptance test". Maybe we can pull common use-cases and just create
sample stub-projects as to how they'd work with whatever we come up
with. Hopefully, this would help us minimize downstream burden.
re CPs: I have also not considered them..
