Petr Fedchenkov created HIVE-30212:
--------------------------------------
Summary: HMS ZooKeeper registration ignores the configured quorum
when started through main
Key: HIVE-30212
URL: https://issues.apache.org/jira/browse/HIVE-30212
Project: Hive
Issue Type: Bug
Components: Standalone Metastore
Affects Versions: 4.2.1, 4.2.0, 4.1.0
Reporter: Petr Fedchenkov
With metastore.service.discovery.mode=zookeeper, metastore.thrift.uris holds
the ZooKeeper quorum (HIVE-20794). HiveMetaStore.main clears that key so
in-process calls use the embedded metastore (hive.metastore.uris since
HIVE-20905, metastore.thrift.uris too since HIVE-29078 in 4.1.0), and only
afterwards startMetaStore builds the ZooKeeperHiveHelper from the configuration
(MetastoreConf.getZKConfig). The helper gets an empty quorum, which becomes
":2181":
{noformat}
ZooKeeperHiveHelper: Creating curator client with connectString: :2181 ...
java.net.ConnectException: Connection refused
HiveMetaStore: Error adding this metastore instance to ZooKeeper:
{noformat}
A metastore started through main (hive --service metastore, start-metastore)
exits unless a ZooKeeper server happens to listen on localhost:2181. In that
case a chroot in the quorum is silently dropped. The existing ZooKeeper tests
call startMetaStore directly and do not go through main, so they do not catch
this.
Reproduced with the official apache/hive:standalone-metastore-4.2.1 image on
Kubernetes. apache/hive:4.0.1 registers with metastore.thrift.uris but fails
the same way with hive.metastore.uris.
Reproduce locally: run ZooKeeper on localhost:2182 and start HMS with
metastore.service.discovery.mode=zookeeper and
metastore.thrift.uris=localhost:2182.
Fix: build the registration helper in main before the URIs are cleared and pass
it to startMetaStore. An empty quorum in zookeeper mode should fail fast
instead of connecting to localhost.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
