Mailing-List: contact derby-dev-help@db.apache.org; run by ezmlm
Precedence: bulk
Reply-To: "Derby Development" <derby-dev@db.apache.org>
Received-SPF: neutral (hermes.apache.org: local policy)
Message-ID: <41B8EC11.4060506@sbcglobal.net>
Date: Thu, 09 Dec 2004 16:21:37 -0800
From: Mike Matrigali <mikem_app@sbcglobal.net>
User-Agent: Mozilla Thunderbird 0.7.2 (Windows/20040707)
MIME-Version: 1.0
To: Derby Development <derby-dev@db.apache.org>
Subject: Re: [jira] Commented: (DERBY-96) partial log record writes that occur
 because of out-of order writes need to be handled by recovery.
References: <62271451.1102635496226.JavaMail.apache@nagoya>
In-Reply-To: <62271451.1102635496226.JavaMail.apache@nagoya>
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit

I think that some fix to this issue should be implemented for the next
release.  The order of my preference is #2, #1, #3.

I think that the option #2 can be implemented in the logging system and
require very little if no changes to the rest of the system processing
of log records.  Log record offsets remain efficient, ie. they can use
LSN's directly.  Only the boot time recovery code need look for the
new log record and do the work to verify checksums, online abort is
unaffected.

I would like to see some performance numbers on the checksum overhead
and if it is measurable then maybe some discussion on checksum choice.
An obvious first choice would seem to be the standard java provided one
used on the data pages.  If I had it to do over, I would probably have
used a different approach on the data pages.  The point of the checksum
on the data page is not to catch data sector write errors, the system
expects the device to catch those, the only point is to catch
inconsistent sector writes (ie. 1st and 2nd 512 byte sector but not
3rd and 4th), for this the current checksum is overkill.  For this one
need not checksum every byte on the page,
one can guarantee a consistent write with 1 bit per sector in the page.

In the future we may want to revisit #3 if it looks like the stream log
is an I/O bottleneck which can't be addressed by striping or some other
hardware help like smart caching controllers.  I see it as a performance
project rather than a correctness project.  It also is a lot more work
and risk.  Note that this could be a good project for someone wanting to
do some research in this area as it is implemented as a derby module
where an alternate implementation could be dropped in if available.

While I believe that we should address this issue, I should also note
that in all my time working on cloudscape/derby I have never received a
problem database (in that time any log related error would have come
through me), that resulted from this out of order/imcomplete log
write issue - this of course does not mean it has not happened just that
it was not reported to us and/or did not affect the database in a
noticable way.  We have actually never seen an out of order write from
the data pages also - we have seen a few checksum errors but all of
those were caused by a bad disk.

On the upgrade issue, it may be time to start an upgrade thread.  Here
are just some thoughts.  If doing option #2, it would be nice if the
new code could still read the old log files and then optionally
write the new log record or not.  Then if users wanted to run a
release in a "soft" upgrade mode where they needed to be able to
go back to the old software they could - they just would not get
this fix.  On a "hard" upgrade the software should continue to read
the old log files as they are currently formatted, and for any new
log files it should begin writing the new log record.  Once the new
log record make's it way into the log file accessing the db with the
old software is unsupported (it will throw an error as it won't know
what to do with the new log record).

Suresh Thalamati (JIRA) wrote:

>      [ http://nagoya.apache.org/jira/browse/DERBY-96?page=comments#action_56482 ]
>      
> Suresh Thalamati commented on DERBY-96:
> ---------------------------------------
> 
> Some thoughts on how this problem could be solved:
> 
> To identify the partial writes, some form of checksum has to be added to the log data written to the file. On recovery using the checksum information partial written log records could be identified and thrown away.  Checksum information has to be included 
> with the log data before it is written to the disk. Now the issue is when do we calculate the checksum and write to the disk. 
> 
> Following are some logical points when the checksum can be calculated and written along with log informaton:
> 
> 1)Calculate the checksum for each log records and store the information as part of log record data structure.  Disadvantage of this approach, storing checksum with each log records could be expensive with respect to the amount of space and time spent to calculate. 
> 
> 2)Calculate checksum for group of log records in the log buffers before writing the buffer to the disk and also write an addition log records that will have the checksum information and the length of the data. This log records (LogCheckSum) will be prefixes to the log buffer. The reason  checksum log records are to be written in the beginning  is it is easier to find to how much data has to be read during recovery to verify the checksum. 
> 
>     Log data is written only when log buffers is full or make sure WAL protocol is not violated.  Size of the data that is part of the checksum can potentially be 32K or whatever log buffer size is. Overhead with this approach is less compared to the first approach. 
> 
> 
> 3)Block-based log i/0: Idea is to group log record data into 4k/8K pages with a checksum on each page. During recovery checksum will be recalculated for  each 
> and match it one on the disk, if checksum does not match it is possibly as partial write. 
>  
> This approach is liked to have more overhead compared to the second one.  But this approach also has the benefit of making log writes aligned.  Not sure yet whether there is any performance by doing so. (Please see aligned Vs Non-Aligned e-mail thread on derby list). 
> 
> I should also bring to the attention this approach will likely require more changes  than 1 & 2 , reasons for that are :
> 
> a)Current system assumed LSN to file offset. If  the data is written in page format , that will no longer be true. 
> b)To strict to WAL protocol , it may be required that an unfilled page needs to be written.  If this unfilled page happened to have a COMMITTED log records it can not be simply rewritten; If the rewrite is incomplete log records with committed information will be thrown away.  To avoid this issue,  log pages can not be written , which could lead to of unused space in the log file or implement safe-write mechanism(ping-pong algorithm). 
> 
> 
> Upgrade: 
> Irrespective of what approach is used to solve this problem, I believe new type of information (checksum) has to be written to the disk, which will not be understood by Old versions.  
> 
> 
> Any comments/suggestion ?
> 
> 
> -suresh
> 
> 
> 
>>partial log record writes that occur because of out-of order writes need to be handled by recovery.
>>---------------------------------------------------------------------------------------------------
>>
>>         Key: DERBY-96
>>         URL: http://nagoya.apache.org/jira/browse/DERBY-96
>>     Project: Derby
>>        Type: New Feature
>>  Components: Store
>>    Versions: 10.0.2.1
>>    Reporter: Suresh Thalamati
> 
> 
>>Incomplete log record write that occurs because of
>>an out of order partial writes gets recognized as complete during
>>recovery if the first sector and last sector happens to get written.
>> Current system recognizes incompletely written log records by checking
>>the length of the record that is stored in the beginning and end.
>> Format the log records are written to disk is:
>>  +----------+-------------+------------------+
>>  | length     |  LOG RECORD |    length   |
>>  +----------+-------------+------------------+
>>This mechanism works fine if sectors are written in sequential manner or
>>log record size is less than 2 sectors. I  believe on SCSI types disks
>>order is not necessarily sequential, SCSI disk drives may sometimes do a
>>reordering of the sectors to optimize the performance.  If a log record
>>that spans multiple disk sectors is being written to SCISI type of
>>devices,  it is possible that first and last sector written before the
>>crash; If this occurs recovery system will incorrectly  interpret the
>>log records was completely written and replay the record. This could
>>lead to recovery errors or data corruption.
>>-
>>This problem also will not occur if a disk drive has write cache with a
>>battery backup which will make sure I/O request will complete.
> 
>