Changing Isolation Level Mid-Transaction

Question

Hi,
I have a SS bean which, within a single container managed transaction, makes numerous
database accesses. Under high load, we start having serious contention issues
on our MS SQL server database. In order to reduce these issues, I would like
to reduce my isolation requirements in some of the steps of the transaction.
To my knowledge, there are two ways to achieve this: a) specify isolation at the
connection level, or b) use locking hints such as NOLOCK or ROWLOCK in the SQL
statements. My questions are:
1) If all db access is done within a single tx, can the isolation level be changed
back and forth?
2) Is it best to set the isolation level at the JDBC level or to use the MS SQL
locking hints?
Is there any other solution I'm missing?
Thanks,
Sebastien

Answer 1

Galen Boyer wrote:
On Sun, 28 Mar 2004, [email protected] wrote:
Galen Boyer wrote:
On Wed, 24 Mar 2004, [email protected] wrote:
Oracle's serializable isolation level doesn't offer what most
customers I've seen expect it to offer. They typically expect
that a serializable transaction will block any read-data from
being altered during the transaction, and oracle doesn't do
that.I haven't implemented WEB systems that employ anything but
the default concurrency control, because a web transaction is
usually very long running and therefore holding a connection
open during its life is unscalable. But, your statement did
make me curious. I tried a quick test case. IN ONE SQLPLUS
SESSION: SQL> alter session set isolation_level =
serializable; SQL> select * from t1; ID FL ---------- -- 1 AA
2 BB 3 CC NOW, IN ANOTHER SQLPLUS SESSION: SQL> update t1 set
fld = 'YY' where id = 1; 1 row updated. SQL> commit; Commit
complete. Now, back to the previous session. SQL> select *
from t1; ID FL ---------- -- 1 AA 2 BB 3 CC So, your
statement is incorrect.Hi, and thank you for the diligence to explore. No, actually
you proved my point. If you did that with SQLServer or Sybase,
your second session's update would have blocked until you
committed your first session's transaction. Yes, but this doesn't have anything to do with serializable.
This is the weak behaviour of those systems that say writers can
block readers.Weak or strong, depending on the customer point of view. It does guarantee
that the locking tx can continue, and read the real data, and eventually change
it, if necessary without fear of blockage by another tx etc.
In your example, you were able to change and commit the real
data out from under the first, serializable transaction. The
reason why your first transaction is still able to 'see the old
value' after the second tx committed, is not because it's
really the truth (else why did oracle allow you to commit the
other session?). What you're seeing in the first transaction's
repeat read is an obsolete copy of the data that the DBMS
made when you first read it. Yes, this is true.
Oracle copied that data at that time into the per-table,
statically defined space that Tom spoke about. Until you commit
that first transaction, some other session could drop the whole
table and you'd never know it.This is incorrect.Thanks. Point taken. It is true that you could have done a complete delete
of all rows in the table though..., correct?
That's the fast-and-loose way oracle implements
repeatable-read! My point is that almost everyone trying to
serialize transactions wants the real data not to
change. Okay, then you have to lock whatever you read, completely.
SELECT FOR UPDATE will do this for your customers, but
serializable won't. Is this the standard definition of
serializable of just customer expectation of it? AFAIU,
serializable protects you from overriding already committed
data.The definition of serializable is loose enough to allow
oracle's implementation, but non-changing relevant data is
a typically understood hope for serializable. Serializable
transactions typically involve reading and writing *only
already committed data*. Only DIRTY_READ allows any access to
pre-committed data. The point is that people assume that a
serializable transaction will not have any of it's data re
committed, ie: altered by some other tx, during the serializable
tx.
Oracle's rationale for allowing your example is the semantic
arguement that in spite of the fact that your first transaction
started first, and could continue indefinitely assuming it was
still reading AA, BB, CC from that table, because even though
the second transaction started later, the two transactions *so
far*, could have been serialized. I believe they rationalize it by saying that the state of the
data at the time the transaction started is the state throughout
the transaction.Yes, but the customer assumes that the data is the data. The customer
typically has no interest in a copy of the data staying the same
throughout the transaction.
Ie: If the second tx had started after your first had
committed, everything would have been the same. This is true!
However, depending on what your first tx goes on to do,
depending on what assumptions it makes about the supposedly
still current contents of that table, it may ether be wrong, or
eventually do something that makes the two transactions
inconsistent so they couldn't have been serialized. It is only
at this later point that the first long-running transaction
will be told "Oooops. This tx could not be serialized. Please
start all over again". Other DBMSes will completely prevent
that from happening. Their value is that when you say 'commit',
there is almost no possibility of the commit failing. But this isn't the argument against Oracle. The unable to
serialize doesn't happen at commit, it happens at write of
already changed data. You don't have to wait until issuing
commit, you just have to wait until you update the row already
changed. But, yes, that can be longer than you might wish it to
be. True. Unfortunately the typical application writer logic may
do stuff which never changes the read data directly, but makes
changes that are implicitly valid only when the read data is
as it was read. Sometimes the logic is conditional so it may never
write anything, but may depend on that read data staying the same.
The issue is that some logic wants truely serialized transactions,
which block each other on entry to the transaction, and with
lots of DBMSes, the serializable isolation level allows the
serialization to start with a read. Oracle provides "FOR UPDATE"
which can supply this. It is just that most people don't know
they need it.
With Oracle and serializable, 'you pay your money and take your
chances'. You don't lose your money, but you may lose a lot of
time because of the deferred checking of serializable
guarantees.
Other than that, the clunky way that oracle saves temporary
transaction-bookkeeping data in statically- defined per-table
space causes odd problems we have to explain, such as when a
complicated query requires more of this memory than has been
alloted to the table(s) the DBMS will throw an exception
saying it can't serialize the transaction. This can occur even
if there is only one user logged into the DBMS.This one I thought was probably solved by database settings,
so I did a quick search, and Tom Kyte was the first link I
clicked and he seems to have dealt with this issue before.
http://tinyurl.com/3xcb7 HE WRITES: serializable will give you
repeatable read. Make sure you test lots with this, playing
with the initrans on the objects to avoid the "cannot
serialize access" errors you will get otherwise (in other
databases, you will get "deadlocks", in Oracle "cannot
serialize access") I would bet working with some DBAs, you
could have gotten past the issues your client was having as
you described above.Oh, yes, the workaround every time this occurs with another
customer is to have them bump up the amount of that
statically-defined memory. Yes, this is what I'm saying.
This could be avoided if oracle implemented a dynamically
self-adjusting DBMS-wide pool of short-term memory, or used
more complex actual transaction logging. ? I think you are discounting just how complex their logging
is. Well, it's not the logging that is too complicated, but rather
too simple. The logging is just an alternative source of memory
to use for intra-transaction bookkeeping. I'm just criticising
the too-simpleminded fixed-per-table scratch memory for stale-
read-data-fake-repeatable-read stuff. Clearly they could grow and
release memory as needed for this.
This issue is more just a weakness in oracle, rather than a
deception, except that the error message becomes
laughable/puzzling that the DBMS "cannot serialize a
transaction" when there are no other transactions going on.Okay, the error message isn't all that great for this situation.
I'm sure there are all sorts of cases where other DBMS's have
laughable error messages. Have you submitted a TAR?Yes. Long ago! No one was interested in splitting the current
message into two alternative messages:
"This transaction has just become unserializable because
of data changes we allowed some other transaction to do"
or
"We ran out of a fixed amount of scratch memory we associated
with table XYZ during your transaction. There were no other
related transactions (or maybe even users of the DBMS) at this
time, so all you need to do to succeed in future is to have
your DBA reconfigure this scratch memory to accomodate as much
as we may need for this or any future transaction."
I am definitely not an Oracle expert. If you can describe for
me any application design that would benefit from Oracle's
implementation of serializable isolation level, I'd be
grateful. There may well be such.As I've said, I've been doing web apps for awhile now, and
I'm not sure these lend themselves to that isolation level.
Most web "transactions" involve client think-time which would
mean holding a database connection, which would be the death
of a web app.Oh absolutely. No transaction, even at default isolation,
should involve human time if you want a generically scaleable
system. But even with a to-think-time transaction, there is
definitely cases where read-data are required to stay as-is for
the duration. Typically DBMSes ensure this during
repeatable-read and serializable isolation levels. For those
demanding in-the-know customers, oracle provided the select
"FOR UPDATE" workaround.Yep. I concur here. I just think you are singing the praises of
other DBMS's, because of the way they implement serializable,
when their implementations are really based on something that the
Oracle corp believes is a fundamental weakness in their
architecture, "Writers block readers". In Oracle, this never
happens, and is probably one of the biggest reasons it is as
world-class as it is, but then its behaviour on serializable
makes you resort to SELECT FOR UPDATE. For me, the trade-off is
easily accepted.Well, yes and no. Other DBMSes certainly have their share of faults.
I am not critical only of oracle. If one starts with Oracle, and
works from the start with their performance arcthitecture, you can
certainly do well. I am only commenting on the common assumptions
of migrators to oracle from many other DBMSes, who typically share
assumptions of transactional integrity of read-data, and are surprised.
If you know Oracle, you can (mostly) do everything, and well. It is
not fundamentally worse, just different than most others. I have had
major beefs about the oracle approach. For years, there was TAR about
oracle's serializable isolation level *silently allowing partial
transactions to commit*. This had to do with tx's that inserted a row,
then updated it, all in the one tx. If you were just lucky enough
to have the insert cause a page split in the index, the DBMS would
use the old pre-split page to find the newly-inserted row for the
update, and needless to say, wouldn't find it, so the update merrily
updated zero rows! The support guy I talked to once said the developers
wouldn't fix it "because it'd be hard". The bug request was marked
internally as "must fix next release" and oracle updated this record
for 4 successive releases to set the "next release" field to the next
release! They then 'fixed' it to throw the 'cannot serialize' exception.
They have finally really fixed it.( bug #440317 ) in case you can
access the history. Back in 2000, Tom Kyte reproduced it in 7.3.4,
8.0.3, 8.0.6 and 8.1.5.
Now my beef is with their implementation of XA and what data they
lock for in-doubt transactions (those that have done the prepare, but
have not yet gotten a commit). Oracle's over-simple logging/locking is
currently locking pages instead of rows! This is almost like Sybase's
fatal failure of page-level locking. There can be logically unrelated data
on those pages, that is blocked indefinitely from other equally
unrelated transactions until the in-doubt tx is resolved. Our TAR has
gotten a "We would have to completely rewrite our locking/logging to
fix this, so it's your fault" response. They insist that the customer
should know to configure their tables so there is only one datarow per
page.
So for historical and current reasons, I believe Oracle is absolutely
the dominant DBMS, and a winner in the market, but got there by being first,
sold well, and by being good enough. I wish there were more real market
competition, and user pressure. Then oracle and other DBMS vendors would
be quicker to make the product better.
Joe