db-derby-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Mike Matrigali (JIRA)" <j...@apache.org>
Subject [jira] Issue Comment Edited: (DERBY-4577) An expanding update fails with an nospc.U error.
Date Sat, 26 Jun 2010 08:11:52 GMT

    [ https://issues.apache.org/jira/browse/DERBY-4577?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12882354#action_12882354
] 

Mike Matrigali edited comment on DERBY-4577 at 6/26/10 4:10 AM:
----------------------------------------------------------------

updated note - this fix did not pass tests.  See subsequent patch.

working copy of fix.  preliminary, not ready for commit yet.  still running tests.  After

further debugging the problem does seem to be isolated to overflow pages.  This 
patch fixes the test case which is also in the patch.  That test case was 100%
reproducible on my machine. 

The overall issue is that the current design to handle expanding updates of rows on both main
pages and overflow pages is that it is assumed that each piece can always in the worst case
of an expanding update be updated to have a piece that is just an overflow row pointer with
all the data moved to another overflow page.  The worst case size of this is 17 bytes which
is 1 byte status, 4 bytes record id, 8 bytes overflow page, 4 bytes oveflow record id.  In
some cases the space to do this is actually longer than the actual space of an existing row
(ie. something like a 1 column blob that has no data in it).  

On main pages this is taken care of by the minimum reserve space. The main intent of minimum
reserve space is to allow user to specify the amount of space to save for the "user" portion
of the row.   The system sets a minimum user reserve space of 12 for main page rows.  This
works around "reserve for overflow" problem as it means there is always space for just the
overflow pointer part to come out of the user portion does not include the record header part.
 

On overflow pages the code does not enforce minimum reserve space, because for user rows it
does not make sense.  The point of reserve space is to avoid oveflowing the row on expanding
updates to another page.   The system only maintains this reserve space on main pages.

Thus a different mechanism is needed to also insure on overflow pages that there is enough
space reserved for the "reserve for overflow" problem.  This current patch does this by just
making sure overflow rows on insert are always >= 17 bytes including reserve space for
the row.  It also changes the code that deals with reclaiming space from shrinking updates
to also maintain this 17 byte minimum.  

Going through the code the use of minimum reserve space was confusing and I believe inconsistent.
 Sometimes it was applied to user space, sometimes to the whole row.  The patch changes the
use of this in a few places to make it more consistent.  

For anyone reviewing the patch, it is probably easier to look side by side at 
compactRecord() and checkRowReservedSpace() than at the diffs.   Some reformats along with
"real" code changes went into this area as I was having a
hard time understanding the original logic.   

The goal of the change is to only affect the minimum length of rows on overflow pages, so
that short rows on main pages should remain the same length.  Also 
this fix only stops the bug from happening to future inserted rows.  Existing 
overflow rows may have the problem.  Running offline compress should fix an
existing table that runs into the problem.

      was (Author: mikem):
    working copy of fix.  preliminary, not ready for commit yet.  still running tests.  After

further debugging the problem does seem to be isolated to overflow pages.  This 
patch fixes the test case which is also in the patch.  That test case was 100%
reproducible on my machine. 

The overall issue is that the current design to handle expanding updates of rows on both main
pages and overflow pages is that it is assumed that each piece can always in the worst case
of an expanding update be updated to have a piece that is just an overflow row pointer with
all the data moved to another overflow page.  The worst case size of this is 17 bytes which
is 1 byte status, 4 bytes record id, 8 bytes overflow page, 4 bytes oveflow record id.  In
some cases the space to do this is actually longer than the actual space of an existing row
(ie. something like a 1 column blob that has no data in it).  

On main pages this is taken care of by the minimum reserve space. The main intent of minimum
reserve space is to allow user to specify the amount of space to save for the "user" portion
of the row.   The system sets a minimum user reserve space of 12 for main page rows.  This
works around "reserve for overflow" problem as it means there is always space for just the
overflow pointer part to come out of the user portion does not include the record header part.
 

On overflow pages the code does not enforce minimum reserve space, because for user rows it
does not make sense.  The point of reserve space is to avoid oveflowing the row on expanding
updates to another page.   The system only maintains this reserve space on main pages.

Thus a different mechanism is needed to also insure on overflow pages that there is enough
space reserved for the "reserve for overflow" problem.  This current patch does this by just
making sure overflow rows on insert are always >= 17 bytes including reserve space for
the row.  It also changes the code that deals with reclaiming space from shrinking updates
to also maintain this 17 byte minimum.  

Going through the code the use of minimum reserve space was confusing and I believe inconsistent.
 Sometimes it was applied to user space, sometimes to the whole row.  The patch changes the
use of this in a few places to make it more consistent.  

For anyone reviewing the patch, it is probably easier to look side by side at 
compactRecord() and checkRowReservedSpace() than at the diffs.   Some reformats along with
"real" code changes went into this area as I was having a
hard time understanding the original logic.   

The goal of the change is to only affect the minimum length of rows on overflow pages, so
that short rows on main pages should remain the same length.  Also 
this fix only stops the bug from happening to future inserted rows.  Existing 
overflow rows may have the problem.  Running offline compress should fix an
existing table that runs into the problem.
  
> An expanding update fails with an nospc.U error.
> ------------------------------------------------
>
>                 Key: DERBY-4577
>                 URL: https://issues.apache.org/jira/browse/DERBY-4577
>             Project: Derby
>          Issue Type: Bug
>    Affects Versions: 10.6.1.0
>            Reporter: Mike Matrigali
>            Assignee: Mike Matrigali
>         Attachments: derby-4577.diff, derby-4577_not_for_commit_fix.diff
>
>
> An update of a long row piece on an overflow page that has limited row used space, row
reserved space, and page free space can throw the nospc.U error on an update.  I will attach
a patch with a simple junit test that reproduces this
> issue.  
> This issue is probably the same as DERBY-2286.  Logging a new issue so that it is clear
the associated changes fix
> the attached repro.  The repro for DERBY-2286 is random and I could only get it to fail
on one machine.  If after the fix
> for this goes in and no one can repro, then we can close DERBY-2286 as a repro.
> The nospc errror is never meant to get to the user.  The code path that can raise it
is shared by insert and update.  
> Insert has code that internally catches the error and does the right thing.  The error
should never be raised for an update.  What is meant to happen for updates is that on update
on any page, including an overflow page should always
> in the worst case have enough room to transform the row piece from whatever it is to
a just a row header with an
> overflow pointer to the next piece.  
> For the attached test the row piece on the overflow page is 12 bytes reserved, and 0
bytes free on the page.  So 
> far in the code path the code has written a preliminary row header that is 4 bytes, and
then has noticed that the
> remaining 8 bytes are not enough for an overflow pointer (worst case 8 bytes for page
number + 4 bytes for record
> id).  
> I think the real problem is that not enough minimum space is being reserved on the overflow
page.  There should be
> 12 bytes reserved + the maximum size of a record header that does not include an overflow
pointer.   I think the
> code does the right thing in the case of head pages where the minimum reserved space
is 12 for the "user" portion
> of the data, but it looks like we don't apply this reserved space to just user portion
on overflow pages.
> I need to research more, but it looks like on overflow pages we apply the minimum reserved
space to the entire row
> rather than just the user space.
> The stack trace being thrown in this case is:
> 2010-03-06 00:18:40.750 GMT:
>  Booting Derby version The Apache Software Foundation - Apache Derby - 10.6.0.0 alpha
- (1): instance 3405c0cb-0127-30d6-6168-ffffe7
> 008b2c
> on database directory C:\derby\s2\systest\out\junit\system\wombat
> ^M
> Database Class Loader started - derby.database.classpath=''^M
> 2010-03-06 00:18:41.171 GMT Thread[main,5,main] (XID = 253), (SESSIONID = 1), (DATABASE
= wombat), (DRDAID = null), Cleanup action s
> tarting^M
> 2010-03-06 00:18:41.171 GMT Thread[main,5,main] (XID = 253), (SESSIONID = 1), (DATABASE
= wombat), (DRDAID = null), Failed Statement
>  is: UPDATE testBadUpdate set value = ? where id = ? with 2 parameters begin parameter
#1: BLOB:Length=120000 :end parameter begin p
> arameter #2: 0 :end parameter ^M
> ERROR nospc: nospc.U^M
>     at org.apache.derby.impl.store.raw.data.StoredPage.logRow(StoredPage.java:4106)^M
>     at org.apache.derby.impl.store.raw.data.UpdateOperation.writeOptionalDataToBuffer(UpdateOperation.java:255)^M
>     at org.apache.derby.impl.store.raw.data.UpdateOperation.<init>(UpdateOperation.java:106)^M
>     at org.apache.derby.impl.store.raw.data.LoggableActions.actionUpdate(LoggableActions.java:80)^M
>     at org.apache.derby.impl.store.raw.data.StoredPage.doUpdateAtSlot(StoredPage.java:8551)^M
>     at org.apache.derby.impl.store.raw.data.BasePage.updateAtSlot(BasePage.java:1062)^M
>     at org.apache.derby.impl.store.access.conglomerate.GenericConglomerateController.replace(GenericConglomerateController.java:472)
> ^M
>     at org.apache.derby.impl.sql.execute.RowChangerImpl.updateRow(RowChangerImpl.java:523)^M
>     at org.apache.derby.impl.sql.execute.UpdateResultSet.collectAffectedRows(UpdateResultSet.java:554)^M
>     at org.apache.derby.impl.sql.execute.UpdateResultSet.open(UpdateResultSet.java:254)^M
>     at org.apache.derby.impl.sql.GenericPreparedStatement.executeStmt(GenericPreparedStatement.java:436)^M
>     at org.apache.derby.impl.sql.GenericPreparedStatement.execute(GenericPreparedStatement.java:317)^M
>     at org.apache.derby.impl.jdbc.EmbedStatement.executeStatement(EmbedStatement.java:1232)^M
>     at org.apache.derby.impl.jdbc.EmbedPreparedStatement.executeStatement(EmbedPreparedStatement.java:1673)^M
>     at org.apache.derby.impl.jdbc.EmbedPreparedStatement.executeUpdate(EmbedPreparedStatement.java:303)^M
>     at org.apache.derbyTesting.functionTests.tests.store.DerbyBugTest2.run_one(DerbyBugTest2.java:224)^M
>     at org.apache.derbyTesting.functionTests.tests.store.DerbyBugTest2.testOne(DerbyBugTest2.java:84)^M
>     at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)^M
>     at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:48)^M
>     at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:37)^M
>     at java.lang.reflect.Method.invoke(Method.java:600)^M
>     at junit.framework.TestCase.runTest(TestCase.java:154)^M
>     at junit.framework.TestCase.runBare(TestCase.java:127)^M
>     at org.apache.derbyTesting.junit.BaseTestCase.runBare(BaseTestCase.java:109)^M
>     at junit.framework.TestResult$1.protect(TestResult.java:106)^M
>     at junit.framework.TestResult.runProtected(TestResult.java:124)^M
>     at junit.framework.TestResult.run(TestResult.java:109)^M
>     at junit.framework.TestCase.run(TestCase.java:118)^M
>     at junit.framework.TestSuite.runTest(TestSuite.java:208)^M
>     at junit.framework.TestSuite.run(TestSuite.java:203)^M
>     at junit.framework.TestSuite.runTest(TestSuite.java:208)^M
>     at junit.framework.TestSuite.run(TestSuite.java:203)^M
>     at junit.extensions.TestDecorator.basicRun(TestDecorator.java:22)^M
>     at junit.extensions.TestSetup$1.protect(TestSetup.java:19)^M
>     at junit.framework.TestResult.runProtected(TestResult.java:124)^M
>     at junit.extensions.TestSetup.run(TestSetup.java:23)^M
>     at org.apache.derbyTesting.junit.BaseTestSetup.run(BaseTestSetup.java:57)^M
>     at junit.extensions.TestDecorator.basicRun(TestDecorator.java:22)^M
>     at junit.extensions.TestSetup$1.protect(TestSetup.java:19)^M
>     at junit.framework.TestResult.runProtected(TestResult.java:124)^M
>     at junit.extensions.TestSetup.run(TestSetup.java:23)^M
>     at org.apache.derbyTesting.junit.BaseTestSetup.run(BaseTestSetup.java:57)^M
>     at junit.framework.TestSuite.runTest(TestSuite.java:208)^M
>     at junit.framework.TestSuite.run(TestSuite.java:203)^M
>     at junit.textui.TestRunner.doRun(TestRunner.java:116)^M
>     at junit.textui.TestRunner.start(TestRunner.java:172)^M
>     at junit.textui.TestRunner.main(TestRunner.java:138)^M
> Cleanup action completed^M

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message