| Bug #47487 | Falcon; validateUpdate asserts where an uncommitted record has no prior record. | ||
|---|---|---|---|
| Submitted: | 21 Sep 2009 15:21 | Modified: | 26 May 2010 17:46 |
| Reporter: | Kevin Lewis | Email Updates: | |
| Status: | Unsupported | Impact on me: | |
| Category: | MySQL Server: Falcon storage engine | Severity: | S3 (Non-critical) |
| Version: | OS: | Any | |
| Assigned to: | Kevin Lewis | CPU Architecture: | Any |
| Tags: | F_RECORD TREE | ||
[21 Sep 2009 15:23]
Kevin Lewis
Verified from call stack of serveral DBT2 runs.
[22 Sep 2009 15:34]
Kevin Lewis
I just caught this assert in the debugger. The record being completed by the gopher thread is the base record. It has no prior record, which means that the prior was snipped. It was most likely not pruned because the oldest active transaction was 80. I could either try to prevent pruning and snipping if the younger visible record is not yet completed (Transaction::writePending == true) or delete this assert and allow Table::validateUpdate() to assume that if there is no priorVersion, then it must have been there at one time and is now missing because it was completed ahead of this younger record version.
[22 Sep 2009 17:09]
Bugs System
A patch for this bug has been committed. After review, it may be pushed to the relevant source trees for release in the next version. You can access the patch from: http://lists.mysql.com/commits/84166 2793 Kevin Lewis 2009-09-22 Bug #47487- Falcon; validateUpdate asserts where an uncommitted record has no prior record. This situation is now possible due to pruning and most likely, snipping intermediate records. So the error condition and assert is now gone. In addition, RecordScavenge::canBeSnipped is improved to assure that any record that is snipped is not only committed, but also completed. @ storage/falcon/RecordScavenge.cpp Bug #47487 - RecordScavenge::canBeSnipped is improved to assure that any record that is snipped is not only committed, but also completed. @ storage/falcon/RecordScavenge.h Bug #47487 - RecordScavenge::canBeSnipped is improved to assure that any record that is snipped is not only committed, but also completed. @ storage/falcon/RecordVersion.cpp Bug #47487 - RecordScavenge::canBeSnipped is improved to assure that any record that is snipped is not only committed, but also completed.
[22 Sep 2009 17:27]
Bugs System
A patch for this bug has been committed. After review, it may be pushed to the relevant source trees for release in the next version. You can access the patch from: http://lists.mysql.com/commits/84169 2795 Kevin Lewis 2009-09-22 Bug#47487 - In RecordScavenge::canBeSnipped(), make sure that the transactionStates are not NULL. @ storage/falcon/RecordScavenge.cpp Bug#47487 - In RecordScavenge::canBeSnipped(), make sure that the transactionStates are not NULL.
Description: While running DBT2, The following assert happened; bool Table::validateUpdate(int32 recordNumber, TransId transactionId) { . . . Record *next = record->getPriorVersion(); if (!next) { Log::debug("Table::validateUpdate: bad record\n"); initial->print(); Record *newRecord = fetch(recordNumber); Log::debug("Table::validateUpdate: currentRecord\n"); newRecord->print(); ASSERT(false); <-- CRASHED HERE } Falcon] Error: assertion (false) failed at line 3855 in file Table.cpp Error::error(char const*, ...)+0x109) Table::validateUpdate(int,unsigned int)+0x12c) Section::updateRecord(int,Stream*, unsigned int, bool)+0x4c) Dbb::updateRecord(int,int, Stream*, unsigned int, bool)+0x45) SRLUpdateRecords::commit()+0x1aa) SerialLogTransaction::commit()+0x131) Gopher::gopherThread()+0x165) How to repeat: Run DBT2 Suggested fix: This new assertion may be related to the new feature of the scavenger within Falcon that snips invisible intermediate records. It was recently added and this assert was not seen before that.