Kevin Wilfong created HIVE-3826:
-----------------------------------
Summary: Rollbacks and retries of drops cause
org.datanucleus.exceptions.NucleusObjectNotFoundException: No such database row)
Key: HIVE-3826
URL: https://issues.apache.org/jira/browse/HIVE-3826
Project: Hive
Issue Type: Bug
Components: Metastore
Affects Versions: 0.11
Reporter: Kevin Wilfong
Assignee: Kevin Wilfong
I'm not sure if this is the only cause of the exception
"org.datanucleus.exceptions.NucleusObjectNotFoundException: No such database
row)" from the metastore, but one cause seems to be related to a drop command
failing, and being retried by the client.
Based on focusing on a single thread in the metastore with DEBUG level logging,
I was seeing the objects that were intended to be dropped remaining in the
PersistenceManager cache even after a rollback. The steps seemed to be as
follows:
1) First attempt to drop the table, the table is pulled into the
PersistenceManager cache for the purposes of dropping
2) The drop fails, e.g. due to a lock wait timeout on the SQL backend, this
causes a rollback of the transaction
3) The drop is retried using a different thread on the metastore Thrift server
or a different server and succeeds
4) Back on the original thread of the original Thrift server someone tries to
perform some write operation which produces a commit. This causes those
detached objects related to the dropped table to attempt to reattach, causing
JDO to query the SQL backend for those objects which it can't find. This
causes the exception.
I was able to reproduce this regularly using the following sequence of commands:
Hive client 1 (Hive1): connected to a metastore Thrift server running a single
thread, I hard coded a RuntimeException into the code to drop a table in the
ObjectStore, specifically right before the commit in preDropStorageDescriptor,
to induce a rollback
Hive client 2 (Hive2): connected to a separate metastore Thrift server running
with standard configs and code
1: On Hive1, CREATE TABLE t1 (c STRING);
2: On Hive1, DROP TABLE t1; // This failed due to the hard coded exception
3: On Hive2, DROP TABLE t1; // Succeeds
4: On Hive1, CREATE DATABASE d1; // This database already existed, I'm not sure
why this was necessary, but it didn't work without it, it seemed to have an
affect on the order objects were committed in the next step
5: On Hive1, CREATE DATABASE d2; // This database didn't exist, it would fail
with the NucleusObjectNotFoundException
The object that would cause the exception varied, I saw the MTable, the
MSerDeInfo, and MTablePrivilege from the table that attempted to be dropped.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
