I’ve been using MySQL (and recently MariaDB) for many years – it must be something like 14 by now – but every now and then I learn something new about it. Recently, I’ve learned about named locks and how you can use to use your already-there MySQL server as a mean to create distributed locks which are not related to a specific DB transaction.
Here is an example of a Python function who’s internal code will never execute concurrently, even in a multi-process, multi-machine distributed environment, as long as all processes talk to the same MySQL database:
NOTE: this code uses SQLAlchemy-like session semantics, but can easily be applied to any Python MySQL client.
That’s very nice! This code will try to obtain a lock for 5 seconds before resuming execution. If the lock is obtained (meaning no other MySQL client has requested to lock this specific lock), execution will resume and when finished the lock will be released. If the lock cannot be obtained within the given timeout, meaning some other client is currently running this code, an exception will be thrown. In any case the code will never run more than once at any given time. Oh, and MySQL named locks are connection-bound, meaning they are released if the connection dies or is explicitly closed – but again this should not happen while the code is executing, but will keep us safe if the entire program crashes, for example.
Before I knew about this feature, we used to do some custom logical locking in our code (which never feels like a solid solution) or use transaction-level row / table locking which coupled our application’s logic with DB operations too much; MySQL named locks are decoupled of any actual data in your tables – its just a mean to get centralized app-level locking. And while there might be other, more lean mechanisms to achieve that, if you already use MySQL I believe this is a very good solution. To clarify, the Python code between
RELEASE_LOCK can be anything and does not need to tie in to the database.
However, the code example above is not very clean and has a few disadvantages:
- It does not handle exceptions properly. If an exception is thrown after the lock was acquired and before it was released, we are most likely going to end up with the lock not being released until the MySQL connection is closed, and we don’t know when that’s going to happen. Not good.
- No clear separation of concerns – we have a single function that handles both application logic (the part between the locks) and provides the locking implementation. This can be solved in several ways, but I believe the way I’ll demonstrate below to be most elegant.
- No code reusability, which is somewhat tied to the previous point. We cannot reuse the locking mechanism in other code paths very easily, and need to retype it. We also cannot reuse the application logic between the locks in a non-locked context – or even in unit tests for that matter.
We can solve all these issues very elegantly using the with statement and context managers. These features are one of my favorites idioms more or less unique to the Python programming language, and it’s these sort of features that I believe really help make Python code very clean and elegant without being too verbose.
We’ll start by creating a context manager for MySQL named lock:
And then proceed to use it in our function:
So what does this do? The
@contextmanager annotation help us easily create a context-managed resource using generator-like semantics; The wrapper ensures that no matter what happens, the lock is released as we leave the managed context whether it is because the code executed successfully or because an exception was thrown.
The semantics of using the
locking_context.named_lock context manager are extremely simple and readable, and reusing the locking context manager is a matter of an import statement and a single line of code. By injecting a mock or monkey-patched object as the first argument of
named_lock(), we can also easily test the context manager itself and any code using it. In addition, if we ever need to switch from MySQL-based locking to some other implementation, it can be done more easily.
While the same flow can be achieved in many other languages supporting, for example,