Hi Pinaki,

I think the issue is whether a LockModeType.READ holds for the entire transaction (the subject tx), from the moment that the lock is obtained until the moment when the transaction has successfully committed. By "hold", I mean that either another tx cannot successfully commit a change to an object that the subject tx has locked until the subject's tx ends, or that the subject tx will fail if another transaction has successfully committed a change prior to the subject tx's end.

In the case of the OpenJPA implementation and the time sequence under discussion, the lock would hold if the implementation obtained a database row level lock (SELECT FOR UPDATE) when it checked the locked object's version.

A peripheral question is whether the spec requires that a read lock hold for the entire tx (as defined above.) If it does, it certainly doesn't test for that compliance, and OpenJPA is not in compliance.

A clear downside to locking the row when checking the version for a read lock is that two or more transactions with no incompatible changes but a variety of read locks for unchanged objects could end up in deadlock.

My take is that the tradeoff is worthwhile, especially since LockModeType.WRITE will give the consistency desired. SFAIK, there are not a lot of implementation options to make a read lock hold as defined. However, the expert group, currently discussing lock mode types, should make clear exactly what can be expected for all lock modes, and have TCK tests to ensure compliance. Intentional ambiguity in a spec is like infidelity in a marriage: it's a knife in the heart of reasonable expectations.

Cheers,

David



Pinaki Poddar wrote:
The expressed view relates to Philip's use case by his own observation:
It "works" if I run with non-enhanced classes, since then there is no
change detection and all rows get > written and version checked.

The point is the way OpenJPA decides what is flushed in a commit is not the
entire set {A,B,C} but only {B,C} that are dirty. So 'transaction
consistency' is ensured but not 'database consistency' because another
transaction may have committed {A,B}. And that breaks the parity invariance
of the entire set {A,B,C}.


dezzio wrote:
Hi Pinaki,

Actually, much as I like your concepts, I don't yet see how they illuminate the issue.

Cheers,

David


Pinaki Poddar wrote:
Hi David & Philip,
   I have not had the time to pay attention to the use case it deserves
--
but reading the case brings up certain aspects that I will like to share
with your experience.
This interesting use case can fail and is failing. But the issue it reveals goes beyond locking semantics.
Behavior of lock is described at datum level -- the levels of warranty
for shared access/mutation to a *single datum* in a consistent way. Transaction goes to the next stage and describes the level of warranty of
a
set of datum as an atomic 'unit of work'. But this test case demands a even higher level of warranty -- consistency
or
invariance of a set-based property (in this case the odd-even parity of 3
instances), which is neither the property of an individual datum nor the
property of a unit of work.

Of course, optimism of optimistic transaction model results in weaker
warranty of set-based invariance. To ensure set-based property
invariance, a
transaction must commit all 3 instances (with consistent odd-even parity)
as
a unit of work, but what it does is that it reads {A,B,C} and writes only
{B, C}.
I will refrain from describing what flags of which OpenJPA configuration
property can be tweaked to get there because let me first hear your
comments
on the expressed views in this posts.
Philip Aston wrote:
Hi David,

Thanks for confirming this. So to summarise where we are, we have:

 1. A reasonable use case that can fail with some unlucky timing.

 2. A technical test case demonstrating the problem that does not rely
on unlucky timing.

 3. A disagreement in our readings of whether 1 and 2 are spec.
compliant. Personally, I don't share your reading of the spec. In my
reading, read locks are safe and provide a concrete guarantee that if
locked entity is changed by another transaction, the locking transaction
will not complete.

(This is a different QoS compared to a write lock - if a write lock is
obtained and the pc flushed, the transaction knows that it will not fail
due to another transaction updating the locked entity. Read locks are
"more optimistic" and can support higher concurrency if there is minimal
contention - many transactions can hold read locks, only one can hold
right locks.).


How can I convince you to change your interpretation of the spec? Anyone
else have an opinion?

FWIW, EclipseLink passes the test case.

- Phil

dezzio (via Nabble) wrote:
Hi Philip,

Let's take a closer look.

We have two bank accounts, Account[1] and Account[2], shared
jointly by customers Innocent[1] and Innocent[2]. The bank's
business rule is that no withdrawal can be made the draws
the combined total of the accounts below zero. This rule is
enforced in the server side Java application that customer's
use.

At the start of the banking day, the accounts stand at:

     Account[1] balance 100.
     Account[2] balance 50.

Innocent[1] wants to draw out all the money, and asks the
application to take 150 from Account[1]. Innocent[2] also
wants to draw out all the money, and asks the application to
take 150 from Account[2]. By itself, either transaction
would conform to the bank's business rule.

The application implements the withdrawal logic by doing the
following for each transaction.

For Innocent[1], read Account[1] and Account[2]. Obtain a
read lock on Account[2]. Refresh Account[2]. Deduct 150 from
Account[1]. Verify business rule, result, sum of balances =
0. Call JPA commit.

For Innocent[2], read Account[1] and Account[2]. Obtain a
read lock on Account[1]. Refresh Account[1]. Deduct 150 from
Account[2]. Verify business rule, result, sum of balances =
0. Call JPA commit.

Within JPA commit, as seen over the JDBC connections, the
following time sequence occurs. (Other time sequences can
yield the same result.)

Innocent[1]: Check version of Account[2]: passes.

Innocent[2]: Check version of Account[1]: passes.
Innocent[2]: Update balance of Account[2], withdraw 150,
                 setting balance to -100: does not block.
Innocent[2]: commit: successful
Innocent[2]: Receives 150.

Innocent[1]: Update balance of Account[1], withdraw 150,
                 setting balance to -50: does not block.
Innocent[1]: commit: successful.
Innocent[1]: Receives 150.

After the two transactions:

Account[1]: balance -50
Account[2]: balance -100

Clearly the bank would not be happy. What's a developer to
do?

I think the developer needs an education about what is meant
by the JPA spec. What JPA is guaranteeing is that when JPA
commit is called, the objects with read locks will have
their versions checked. The objects with write locks will
have their versions checked and changed. The objects that
have been modified will have their versions checked, their
information updated, and their versions changed. Clearly all
of these rules were enforced in the above example.

If the developer had used write locks, both transactions
would not have succeeded. In fact, for the above example and
a similar time sequence, if write locks had been used in
place of read locks, there would have been deadlock.

Now, if in fact, I'm wrong about my interpretation of the
JPA spec (and it wouldn't be the first time) then you have a
case. I'd be curious to know whether other JPA
implementations pass your elegant test case, and what they
are doing differently that makes it so.

Also, if I am wrong about my interpretation, then the JPA
TCK needs a test case that will snag this failure, because
OpenJPA passes the current JPA TCK.

Cheers,

David

Philip Aston wrote:

Oh yeah - my bad. Try this one instead:

Suppose there are a set of Accounts, and a business rule that says
that the net balance must be positive.

Innocent wants to draw down on Account 1 as far as possible. It read
locks the of Accounts, sums up the value, and and subtracts the
positive total from Account 1. Innocent begins its commit, and its
read locks are validated.

Meanwhile InnocentToo does the same for Account 2, and commits.

Innocent updates Account 1 and finishes its commit.

The total in account summary is now negative, violating the business
rule. If read locks worked as I think they should, Innocent would have
received an OptimisticLockException.



dezzio wrote:
Hi Philip,

When two transactions read the same version of AccountSummary, both
cannot successfully update its sum.  Only one will successfully
commit.

David



Reply via email to