For work along the same lines you should check out the HBase metastore
work in Hive 2.0. It still uses the thrift server and RawStore but puts
HBase behind it instead of an RDBMS. We did this because we found that
most of the inefficiencies of Hive's metadata access had to do with the
layout of the RDBMS and the way it was accessed. In the same work I
built short-circuit options in to avoid using thrift and enable sharing
of objects across HiveMetaStore and HiveMetaStoreClient.
On the backwards incompatibilities, yes IMetaStoreClient evolves in lock
step with the thrift interface. My point was we often add calls, add
new fields to structs, etc. Your code would still compile in these
cases, new features just wouldn't work. Given that a couple major
Hadoop support vendors now support rolling upgrades they are devs
interested in making sure that client version x works properly with
server version x+1.
Still, we don't test for the use case you are proposing so we could end
up breaking your code without knowing it.
When I said it wasn't external, I meant we did not expect end users to
write code against it (like say the UDF interface). Yes it's external
to the metastore package as you point out.
Alan.
Austin Lee <mailto:austin.t....@gmail.com>
December 15, 2015 at 10:46
Yes, a more efficient implementation is what I am trying to achieve.
I also want to retain the ability to talk to a remote metastore that
is not necessarily thrift.
To be more precise, what I would like is a more efficient metastore.
In looking at the current architecture, I came to a conclusion that
there are three logical boundaries where I can inject an improved
implementation or alternative to what Hive offers in the metastore space.
1) RawStore
I think the existing mechanism that Hive offers users to choose from
major RDBMSes works fine. I suppose there's still room for
improvement here, but the impact of those improvements would be
limited to the storage aspects of metadata.
2) Thrift server
An alternative HiveMetaStore that talks Hive Metastore Thrift. It's
almost a coin toss between this and #3, but I think for the reasons I
will state below, #3 is preferable.
3) IMetaStoreClient
I feel this gives me the most freedom since I can be embedded or
remote. I am not tied to the Thrift interface or the RawStore
interface, if I choose to roll my own.
One thing that does concern me is your statement about
IMetaStoreClient being an internal interface, which is true. Do the
changes to this interface really happen ad-hoc? Doesn't it evolve in
lock step with the Thrift interface? If so, wouldn't backward
compatibility guarantees for Thrift translate to backward
compatibility guarantees for this interface as well? From the way it
is used by Query Planning, I think it could be made an "external"
interface that belongs in hive-metastore.
Alan Gates <mailto:alanfga...@gmail.com>
December 15, 2015 at 10:14
I don't see an issue with this, it seems fine. One caveat though is
we see this as an internal interface and we change it all the time. I
wouldn't want to be pushed into making backwards compatibility
guarantees for IMetaStoreClient. Which means that if you develop a
different implementation of it outside Hive it will likely break on
every upgrade.
I don't understand your example use case. You can run Hive now
without the thrift server, so I'm guessing that's not what you're
really trying to do. Are you just interested in building a more
efficient implementation or do you have another use case in mind?
Alan.
Austin Lee <mailto:austin.t....@gmail.com>
December 14, 2015 at 20:48
Hi,
I would like to propose a change that would make it possible for users to
choose an implementation of IMetaStoreClient via HiveConf, i.e.
hive-site.xml. Currently, in Hive the choice is hard coded to be
SessionHiveMetaStoreClient in org.apache.hadoop.hive.ql.metadata.Hive.
There is no other direct reference to SessionHiveMetaStoreClient other
than
the hard coded class name in Hive.java and the QL component operates only
on the IMetaStoreClient interface so the change would be minimal and it
would be quite similar to how an implementation of RawStore is specified
and loaded in hive-metastore. One use case this change would serve would
be one where a user wishes to use an implementation of this interface
without the dependency on the Thrift server. I would appreciate the
community's input and feedback on this proposal.
Thank you,
Austin