ORACLE ON LINUX HAS SERIOUS BUG! IS NOT FAILSAFE COMMIT!!!

Question

I made one experiment to make sure in failsafety on Oracle/Linux.
I tried to turn off the computer after commit, when "Commit
complete" message was appeared.
See a link on my last report:
http://f2.dejanews.com/oracletechnet/nonmembers/index.cgi?do=show
msg&conf=Early.Linux&msg=3953&show=unread&depth=all
This experiment finished with bad result:
Oracle for Linux can not be used for any more or less
responsible applications!
I tried to use filesystem with sync option and I tried to use
waitfor_sync parameter in "init.ora" file.
All of above without result. Oracle does not remember the data
last commited transactions. Oracle does not write commited data
on disk during about 1-3 sec period after the message "Commit
complete" appeared.
What do the Oracle Team say about this bug?
null

Answer 1

Has anyone reported this to Oracle? I haven't heard about this
bug, but it may be in the pipeline for the next patch.
Mark Malakanov (guest) wrote:
: StE,
: you are right.
: Oracle MUST write into current redo log file after every
: commit.
: A confirmation "Commited" MUST be appeared after a write
: operation finished and sync'ed.
: Unfortunately this rule does not supported in Oracle8.0.5.1 for
: Linux.
: I made strace against LGWR process. It really open log file
: without O_SYNC key.
: But I tried to set a Sync attribute with chattr command against
: logfiles. Also I tried to do mount with sync option against
: whole filesystem. This ways brings system to open files as
: O_SYNC key used.
: All of above doesnt help. Oracle looses last commited
: transactions. Only very bad performance.
: Also I make the simple test in C. Open file without O_SYNC,
: Write 5Mb, fsync and message. The message appears only after
: fsync operation is finished. I understood - Linux is not the
: cause of error.
: I did find no fsync calls in strace's log file.
: I think it is the cause of transactions are loosed.
: Mark
: StE (guest) wrote:
: : Jay Walters (guest) wrote:
: : : It is expected behavior to not write to the datafiles until
: a
: : : checkpoint is performed. Did you see this problem for
: : : logfiles? Or just on data files? I am looking into moving
: to
: : : ORACLE on Linux but need data integrity...
: : : Can anybody confirm if this is really a problem as it seems
: : : like a show stopper.
: : Using strace I've observed the RDBMS opening the online redo
: : logs without O_SYNC and, on a commit, checkpoint, or log
: change,
: : I cannot see fsync being called on the file handle.
: : Maybe this is a bug in strace, or a blip of stupidity on my
: : part, but someone testing has managed to lose a committed
: : transaction after deliberately powering off their machine.
: That
: : shouldn't happen.
: : -michael
null