Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
2.1.2, 2.5.2
-
None
-
None
-
Operating system: Linux
Application server: Websphere v7
RDBMS: Oracle 11g
Jackrabbit: V2.1.2 (built into Liferay 6.0 EE SP2)
Description
-
-
- Experienced phenomenon ***
-
Our customer reported an issue regarding Liferay’s document library: while documents are being retrieved, the following exception occurs accompanied by temporary tablespace shortage.
[9/24/12 8:00:55:973 CEST] 00000023 SystemErr R ERROR [org.apache.jackrabbit.core.util.db.ConnectionHelper:454] Failed to execute SQL (stacktrace on DEBUG log level)
java.sql.SQLException: ORA-01652: unable to extend temp segment by 128 in tablespace TEMP
at oracle.jdbc.driver.T4CTTIoer.processError(T4CTTIoer.java:440)
at oracle.jdbc.driver.T4CTTIoer.processError(T4CTTIoer.java:396)
…
at oracle.jdbc.driver.OraclePreparedStatementWrapper.execute(OraclePreparedStatementWrapper.java:1374)
at com.ibm.ws.rsadapter.jdbc.WSJdbcPreparedStatement.pmiExecute(WSJdbcPreparedStatement.java:928)
at com.ibm.ws.rsadapter.jdbc.WSJdbcPreparedStatement.execute(WSJdbcPreparedStatement.java:614)
…
at org.apache.jackrabbit.core.util.db.ConnectionHelper.exec(ConnectionHelper.java:328)
at org.apache.jackrabbit.core.fs.db.DatabaseFileSystem.getInputStream(DatabaseFileSystem.java:663)
at org.apache.jackrabbit.core.fs.BasedFileSystem.getInputStream(BasedFileSystem.java:121)
at org.apache.jackrabbit.core.fs.FileSystemResource.getInputStream(FileSystemResource.java:149)
at org.apache.jackrabbit.core.RepositoryImpl.loadRootNodeId(RepositoryImpl.java:556)
at org.apache.jackrabbit.core.RepositoryImpl.<init>(RepositoryImpl.java:325)
at org.apache.jackrabbit.core.RepositoryImpl.create(RepositoryImpl.java:673)
at org.apache.jackrabbit.core.TransientRepository$2.getRepository(TransientRepository.java:231)
at org.apache.jackrabbit.core.TransientRepository.startRepository(TransientRepository.java:279)
at org.apache.jackrabbit.core.TransientRepository.login(TransientRepository.java:375)
at com.liferay.portal.jcr.jackrabbit.JCRFactoryImpl.createSession(JCRFactoryImpl.java:67)
at com.liferay.portal.jcr.JCRFactoryUtil.createSession(JCRFactoryUtil.java:43)
at com.liferay.portal.jcr.JCRFactoryUtil.createSession(JCRFactoryUtil.java:47)
at com.liferay.documentlibrary.util.JCRHook.getFileAsStream(JCRHook.java:472)
at com.liferay.documentlibrary.util.HookProxyImpl.getFileAsStream(HookProxyImpl.java:149)
at com.liferay.documentlibrary.util.SafeFileNameHookWrapper.getFileAsStream(SafeFileNameHookWrapper.java:236)
at com.liferay.documentlibrary.service.impl.DLLocalServiceImpl.getFileAsStream(DLLocalServiceImpl.java:192)
The original size of tablespace TEMP used to be 8Gb when the error has occurred for the first time. Later on it was extended by as much as additional 7Gb to 15Gb, yet the available space was still not sufficient to fulfill subsequent requests and ORA-01652 emerged again.
-
-
- Reproduction steps ***
-
1) Create a dummy 10MB file
$ dd if=/dev/urandom of=/path/to/dummy_blob bs=8192 count=1280
1280+0 records in
1280+0 records out
10485760 bytes (10 MB) copied, 0.722818 s, 14.5 MB/s
2) Create a temp tablespace
The tablespace is created with 5Mb and automatic expansion is intentionally disabled.
SQL> CREATE TEMPORARY TABLESPACE jcr_temp
TEMPFILE '/path/to/jcr_temp_01.dbf'
SIZE 5M AUTOEXTEND OFF;
Table created.
SQL> ALTER USER jcr TEMPORARY TABLESPACE jcr_temp;
User altered.
3) Prepare the test case
For the sake of simplicity a dummy table is created (similar to Jackrabbit's FSENTRY).
SQL> create table FSENTRY(data blob);
Table created.
SQL>
CREATE OR REPLACE PROCEDURE load_blob
AS
dest_loc BLOB;
src_loc BFILE := BFILENAME('DATA_PUMP_DIR', 'dummy_blob');
BEGIN
INSERT INTO FSENTRY (data)
VALUES (empty_blob())
RETURNING data INTO dest_loc;
DBMS_LOB.OPEN(src_loc, DBMS_LOB.LOB_READONLY);
DBMS_LOB.OPEN(dest_loc, DBMS_LOB.LOB_READWRITE);
DBMS_LOB.LOADFROMFILE(
dest_lob => dest_loc
, src_lob => src_loc
, amount => DBMS_LOB.getLength(src_loc));
DBMS_LOB.CLOSE(dest_loc);
DBMS_LOB.CLOSE(src_loc);
COMMIT;
END;
/
Procedure created.
SQL> EXEC load_blob();
PL/SQL procedure successfully completed.
4) Execute the query
SQL> select nvl(data, empty_blob()) from FSENTRY;
ERROR: ORA-01652: unable to extend temp segment by 128 in tablespace JCR_TEMP
4) Let’s see how temporary tablespace is being used
Query the size of the original LOB data.
SELECT s.segment_name,
s.segment_type,
SUM(s.blocks) AS blocks
FROM dba_segments s,
(SELECT segment_name, index_name
FROM dba_lobs l
WHERE l.table_name = 'FSENTRY'
AND l.column_name = 'DATA') ls
WHERE s.segment_name = ls.segment_name
OR s.segment_name = ls.index_name
GROUP BY s.segment_name, s.segment_type;
SEGMENT_NAME SEGMENT_TYPE BLOCKS
---------------------------------------------------------------------------
SYS_LOB0000035993C00001$$ LOBSEGMENT 1408
SYS_IL0000035993C00001$$ LOBINDEX 8
Query the size of the temporarily created LOB data.
SELECT s.sql_text, t.segtype, t.blocks
FROM v$sql s, v$tempseg_usage t
WHERE s.sql_id = t.sql_id;
SQL_TEXT SEGTYPE BLOCKS
---------------------------------------------------------------------------
select nvl(data, empty_blob()) from lobtest LOB_DATA 1408 (~11 Mb)
select nvl(data, empty_blob()) from lobtest LOB_INDEX 128 (~1 Mb)
LOB index might need some explanation: it's an internal structure that is strongly associated with LOB storage (LOB locators point to the top of the LOB index tree, where leaf blocks point to the actual LOB chunks). The bottom line is that a user may not drop/alter/rebuild LOB indexes in any way.
As a conclusion we can see here that a temporary LOB has been been created by Oracle indeed and its space requirements are quite similar to the original one.
-
-
- Analysis ***
-
The aforementioned exception is thrown in the getInputStream(...) method of class org.apache.jackrabbit.core.fs.db.DatabaseFileSystem while it’s attempting to execute that SQL statement which is denoted by selectDataSQL.
Based on this particular case, customer has the Jackrabbit repository configured to use org.apache.jackrabbit.core.fs.db.OracleFileSystem, which implies that the actual value of selectDataSQL is the following.
SELECT NVL(FSENTRY_DATA, empty_blob())
FROM J_FSASSSA_1LIFERAYFSENTRY
WHERE FSENTRY_PATH = :1
AND FSENTRY_NAME = :2
AND FSENTRY_LENGTH IS NOT NULL
The most important point here is that Oracle creates temporary LOBs, if LOB columns are used in SQL functions. From the point of view of Oracle, it’s a completely logical behaviour, since it has to evaluate the given expression and during doing so the database manager also has to store the result of that calculation.
In this case, if column FSENTRY_DATA is null an empty LOB locator is created, however it’s unclear why it is functionally required. Interestingly other database file system implementations (eg. for DB2) do not use an equivalent SQL function at the same place (eg. COALSCE in case of DB2), but they return FSENTRY_DATA directly without performing such a pre-processing.
The second part of the experienced problem is that according to the Oracle 11g Database SecureFiles and Large Objects Developer's Guide, if a temporary LOB has been returned to the application, it’s the caller (except PL/SQL program blocks) responsibility to explicitly free the received object. Having this in mind, handling temporary LOBs in a decent way can be accomplished by checking and freeing them manually.
void someMethod() {
...
ResultSet rs = ...
oracle.sql.BLOB data = (oracle.sql.BLOB) rs.getBlob(...);
if (data.isTemporary())
...
}
Apparently org.apache.jackrabbit.core.fs.db.OracleFileSystem does not take care of disposing temporary LOBs and temporary tablespace is being depleted this way in a long run.
-
-
- Conclusion ***
-
Jackrabbit definitely should at least free temporary LOBs in OracleFileSystem or avoid using the NVL function completely and moving that logic to the application instead, would make the whole phenomenon cease to exist.