[
https://issues.apache.org/jira/browse/IMPALA-10082?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17182913#comment-17182913
]
Robbie Zhang commented on IMPALA-10082:
---------------------------------------
Not sure how LocalCatalog is affected yet. But I believe I found a bug in
CatalogServiceCatalog.java.
Let's take
[removeTable|https://github.com/apache/impala/blob/master/fe/src/main/java/org/apache/impala/catalog/CatalogServiceCatalog.java#L2076-L2078]
for example.
{code:java}
public Table removeTable(String dbName, String tblName) {
Db parentDb = getDb(dbName);
if (parentDb == null) return null;
versionLock_.writeLock().lock();
try {
Table removedTable = parentDb.removeTable(tblName);
{code}
The method getDb is called before the lock is occupied. If the thread which is
processing a global invalidate metadata occupies the lock at that time,
parentDb will be stale. In other words, the table is removed from HMS but not
removed from the latest database object in catalogd. Then we can't create a
table with the same name until we run invalidate metadata again.
The method addFunction has the same issue.
> Concurrent invalidate metadata and create/drop table cause discrepancy in
> metadata
> ----------------------------------------------------------------------------------
>
> Key: IMPALA-10082
> URL: https://issues.apache.org/jira/browse/IMPALA-10082
> Project: IMPALA
> Issue Type: Bug
> Affects Versions: Impala 4.0
> Reporter: Robbie Zhang
> Priority: Major
>
> The symptom is similar to IMPALA-7093 but is a different issue. Here is how I
> reproduce it:
> 1) Ran the first script to keep running create/insert/drop queries
> {code:java}
> #!/bin/bash
> while [ 1 ]
> do
> shell/impala-shell -q "create table if not exists test(i int); insert into
> test(i) values(1); drop table test;" 2>&1| tee test.output
> n=`egrep "Exception" test.output | wc -l`
> if [ $n -lt 0 ]; then
> rm -f /tmp/testing
> exit
> fi
> done
> {code}
> 2) Ran the second script to keep running global invalidate metadata
> {code:java}
> #!/bin/bash
> while [ 1 ]
> do
> shell/impala-shell -q "invalidate metadata"
> done
> {code}
> Sometime later, the first scrip ended with "Table default.test does not
> exist":
> {code:java}
> Starting Impala Shell with no authentication using Python 2.7.12
> Warning: live_progress only applies to interactive shell sessions, and is
> being skipped for now.
> Opened TCP connection to localhost:21000
> Connected to localhost:21000
> Server version: impalad version 4.0.0-SNAPSHOT DEBUG (build
> f95f7940e4a290d75ee85fd78e85bc26795f0f9f)
> Query: create table if not exists test(i int)
> Fetched 1 row(s) in 0.01s
> Query: insert into test(i) values(1)
> Query submitted at: 2020-08-13 22:57:51 (Coordinator: http://impala34:25000)
> ERROR: AnalysisException: Table does not exist: default.test
> Could not execute command: insert into test(i) values(1){code}
> Even after I change to local catalog mode, this issue still exists:
> {code:java}
> Starting Impala Shell with no authentication using Python 2.7.12
> Warning: live_progress only applies to interactive shell sessions, and is
> being skipped for now.
> Opened TCP connection to localhost:21000
> Connected to localhost:21000
> Server version: impalad version 4.0.0-SNAPSHOT DEBUG (build
> f95f7940e4a290d75ee85fd78e85bc26795f0f9f)
> Query: create table if not exists test(i int)
> Fetched 1 row(s) in 0.07s
> Query: insert into test(i) values(1)
> Query submitted at: 2020-08-13 22:10:16 (Coordinator: http://impala34:25000)
> ERROR: AnalysisException: org.apache.impala.catalog.TableLoadingException:
> Could not load table default.test from catalog
> CAUSED BY: TableLoadingException: Could not load table default.test from
> catalog
> CAUSED BY: TException:
> TGetPartialCatalogObjectResponse(status:TStatus(status_code:GENERAL,
> error_msgs:[TableLoadingException: Table default.test no longer exists in the
> Hive MetaStore. Run 'invalidate metadata default.test' to update the Impala
> catalog.]), lookup_status:OK)
> Could not execute command: insert into test(i) values(1)
> {code}
> And in local catalog mode, the newly created table was lost but it's still
> visible in the coordinator. After running 'invalidate metadata default.test',
> the table disappeared at all.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org
For additional commands, e-mail: issues-all-h...@impala.apache.org
