Search code examples
javaspringjdbcpessimistic-locking

"Select for update" and update with pessimistic locking


I'm trying to implement pessimistic locking using select for update, as I want other threads to wait until the lock on the selected row is released. The part that I have understood is after going through multiple threads Spring JDBC select for update and various similar threads is it is achievable in case select and update are happening within same method and hence they are part of same transaction.

The issue in my case is I have a JAR for DAO functionality where in a selectforUpdate method is available and a separate update method is available, both method has a finally block which contains

resultSet.close();
statement.close();
connection.close();

Now I'm struggling to find out is there a way in which I can use both the methods from outside of the JAR, maybe by annotating my method with @Transactional annotation and make it work in some way. So that lock is only released once update method has been executed.


Solution

  • You're making a mistake. Using the wrong tool for the job. Transaction levels and FOR UPDATE has the purpose of ensuring data integrity. Period. It it isn't designed for control flow and if you use it for this, it will bite you in the butt sooner rather than later.

    Let me try to explain what SELECT FOR UPDATE is for, so that, when later I tell you that it is most definitely not for what you're trying to do with it, it is easier to follow.

    Imagine a bank. Simple enough. The bank has some ATMs out front and a website where you can see your transactions and transfer money to other accounts.

    Imagine you (ABC) and I (Reinier) are trying to fleece the bank some. Here is our plan: We set it up so that you have €1000,- in your account and I have nothing.

    Then, you log into the website from your phone, and start a transfer, transferring €1000,- to my account. But, while you're doing that, right in the middle, you withdraw €10,- from the ATM.

    If the bank messed up their transactions, it's possible you end up with €990,- in your account and I have €1000,- in my account, and we fleeced the bank. This is how that could happen (and if halfway through the example you think: I already know this stuff, I know what FOR UPDATE does! - I'm not so sure you do, read it carefully)

    ATM code

    startTransaction();
    int currentBalance = sql("SELECT balance FROM account WHERE user = ?", abc);
    if (currentBalance < requestedWithdrawal) throw new InsufficientFundsEx();
    sql("UPDATE account SET balance = ? WHERE user = ?", currentBalance - requestedWithdrawal, abc);
    commit();
    moneyHopper.spitOut(requestedWithdrawal);
    

    Website code

    startTransaction();
    int balanceTo = sql("SELECT balance FROM account WHERE user = ?", reinier);
    int balanceFrom = sql("SELECT balance FROM account WHERE user = ?", abc);
    if (transfer > balanceFrom) throw new InsufficientFundsEx();
    sql("UPDATE account SET balance = ? WHERE user = ?", balanceTo + transfer, reinier);
    sql("UPDATE account SET balance = ? WHERE user = ?", balanceFrom - transfer, abc);
    commit();
    controller.notifyTransferSucceeded();
    

    How it can go wrong

    The way it goes wrong is if the balanceTo and balanceFrom are 'locked in', then the ATM withdrawal goes through, and then the update SQL statements from the website transaction go through (this wipes out the ATM withdrawal, effectively - whatever the ATM spit out is free money), or if the ATM's balance check locks in, then the transfer goes through, and then the ATM's update goes through (which gives the recipient, i.e. me their €1000,-, and ensures that the ATM code's update, setting your balance to 990, is the last thing that happens, giving us €990,- of free money.

    So what's the fix? Hint: Not FOR UPDATE

    The fix is to consider what a transaction means. The purpose of transactions is to turn operations into atomic notions. Either both your account is reduced by the transfer amount and mine is raised by the same, or nothing happens.

    It's obvious enough with statements that change things (UPDATE and INSERT). It's a bit more wonky when we talk about reading data. Should those reads be considered part of the transaction?

    One way to go is to say: No, unless you add FOR UPDATE at the end of it all, in which case, yes - i.e. lock those rows only if FOR UPDATE is applied until the transaction ends.

    But that is not the only way to ensure data integrity.

    Optimistic locking to the rescue - or rather, to your doom

    A much more common way is called MVCC (MultiVersion Concurrency Control) and is far faster. The idea behind MVCC (also called optimistic locking), is to just assume no clashes ever occur. Nothing is ever locked. Instead, [A] all changes made within a transaction are completely invisible to things running in any other transaction until you commit, and [B] when you COMMIT a transaction, the database checks if everything you have done within the span of this transaction still 'holds up' - for example, if you updated a row within this transaction that was also modified by another transaction that has committed already, you get an error when you commit, not when you ran the UPDATE statement.

    In this framework, we can still talk about what SELECT even means. This, in java/JDBC, is called the Transaction Isolation Level and is configurable on a DB connection. The best level, the level the bank should be using to avoid this issue, is called the TransactionLevel.SERIALIZABLE. Serializable effectively means everything dirties everything else: If during a transaction you read some data, and when you commit, that same SELECT statement would have produced different results because some other transaction modified something, then the COMMIT just fails.

    They fail with a so-called 'RetryException'. This means literally what it says: Just start your transaction over, from the top. It makes sense if you think about that bank example: What WOULD have happened, had the bank done it right and set up serializable transaction isolation level, is that either the ATM machine's transaction or the transfer transaction would get the retryexception. Assuming the bank wrote their code right and they actually do what the exception tells you to (start over), then they would start over, and that includes re-reading the balances out. No cheating of the bank can occur now.

    Crucially, in the SERIALIZABLE model, locking NEVER occurs, and FOR UPDATE does not mean anything at all.

    Thus, usually, FOR UPDATE does literal stone cold nothing, a complete no-op, depending on how the db is setup.

    FOR UPDATE does not mean 'lock other transactions that touch this row'. No matter how much you want it to.

    Some DB implementations, or even some combination of DB engine and connection configuration may be implemented in that fashion, but that is an extremely finicky setup, and your app should include documentation that strongly recommends the operator to never change the db settings, never switch db engines, never update the db engine, never update the JDBC driver, and never mess with the connection settings.

    That's the kind of silly caveat you really, really don't want to put on your code.

    The solution is to stop buttering your toast with that chainsaw. Even if you think you can manage to get some butter on that toast with it, it's just not what it was made for, like at all, and we're all just waiting until you lose a thumb here. Just stop doing it. Get a butterknife, please.

    If you want to have one thread wait for another, don't use the database, use a lock object. If you want to have one process wait for another, don't use the database, don't use a lock object (you can't; processes don't share memory); use a file. the new java file IO has an option to make a file atomically (meaning, if the file already exists, throw an exception, otherwise make the file, and do so atomically, meaning if two processes both run this 'create atomically new file' code, you have a guarantee that one succeeds and one throws).

    If you want data integrity and that's the only reason you wanted pessimistic locking in the first place, stop thinking that way - it's the DBs job, not your job, to guarantee data integrity. MVCC/Optimistic locking DBs guarantee that the bank will never get fleeced no matter how hard you try with the shenanigans at the top of this answer and nevertheless, pessimistic locking just isn't involved.

    JDBC itself sucks (intentionally, a bit too much to get into) for 'end use' like what you are doing here. Get yourself an abstraction that makes it nice such as JDBI or JOOQ. These tools also have the only proper way to interact with databases, which is that all DB code must be in a lambda. That's because you don't want to manually handle those retry exceptions, you want your DB access framework to take care of it. This is what the bank code should really look like:

    dbAccess.run(db -> {
        int balance = db.sql("SELECT balance FROM account WHERE user =?", abc);
        if (balance < requested) throw new InsufficientBalanceEx();
        db.update("UPDATE account SET balance = ? WHERE user = ?", balance - requested, abc);
        return requested;
    };
    

    This way, the 'framework' (the code behind that run method) can catch the retryex and just rerun the lambda as often as it needs to. rerunning is tricky - if two threads on a server both cause the other to retry, which is not that hard to do, then you can get into an endless loop where they both restart and both again cause the other to retry, at infinitum. The solution is literally dicethrowing. When retrying, you should roll a random number and wait that many milliseconds, and for every further retry, the range on which you're rolling should increase. If this sounds dumb to you, know that you're currently using it: It's how Ethernet works, too (ethernet uses randomized backoff when collisions occur on the wire). Ethernet won, token ring lost. It's the exact same principle at work (token ring is pessimistic locking, ethernet is optimistic 'eh just try it and detect if it went wrong, then just redo it, with some randomized exponential backoff sprinkled in to ensure you don't get 2 systems in lock-step forever screwing up the other's attempt).