Summary

Mongo::Locking is a library that enables cross-process blocking mutexes, using simple but flexible primitives to express an arbitrary graph of lock dependencies between class instances.

Background

Consider the following common scenario:

Given an object graph 1 Order -> N OrderItems -> 1 JobFlow -> N Jobs, a collection of disparate systems that operate on portions of the graph asynchronously.

If you were using a Document-oriented (e.g. Mongo) data model, you might represent the object graph as a bunch of nested objects, rooted in an Order. In an RDBMS, you usually have a collection of tables with foreign key relationships between them.

In any case, you need to enforce some notion of data integrity as portions of the graph mutate. How does one normally enforce integrity in concurrent access scenarios?

RDBMS

In the RDBMS world, you've got a couple options:

1. SELECT .. FOR UPDATE or equivalent.

   Depending on the underlying storage engine, this will write lock at minimum the given row, and in most modern RDBMS', a cluster of rows around the row you're trying to "protect". This approach tends to require breaking out of the ORM with custom SQL, and carries with it all sorts of unintended/unexpected performance/synchronization/deadlock pitfalls. It really starts to break down when there is more than one object that needs to be "locked", multiple loosely-related objects that need to be "locked", or when crossing database boundaries.

2. Rely on the ACID properties of SQL92 transactions to enforce data integrity.

   (a) Given 2 or more competing, disparate processes running asynchronously, accessing the same resources. Both enter into transactions, possibly access overlapping resources, one wins and the other (eventually) fails after attempting to commit.

   Practically, what does the erroring code do? Does it retry? Was it written in a way that even makes a retry possible? Is the context so consistently atomic and stateless that it could blindly do so? Does it just bail and fail? (Yes, most of the time.) What if it was acting on an asynchronous imperative across a message bus? Shouldn't this condition be detected, and the imperative replayed by some other code somewhere else? Wouldn't that vary depending upon the logical atomicity of the operation? Etc etc.

   (b) Given 2 competing processes, both enter into transactions, but the relationship between resources is not fully expressed in terms of RDBMS constraints. This is another very common case, as most ORMs tend to provide integrity (validation) functionality in the app layer, and only a subset actually trickle down into material RDBMS constraints. In this case, the RDBMS has no idea that logical integrity has been violated, and general misery will ensue.

   Transactions are often mistakenly perceived as a panacea for these types of of problems, and as a consequence usually compound the problems they are being used to solve with additional complexity and cost.

NoSQL

In the NoSQL world, you don't have as many options. Many folks mistakenly believe that logically embedded objects are somehow protected by that nesting. (They aren't.)

Some engines provide unique locking or pseudo-transactional primitive(s), but generally the same costs and pitfalls of transactions apply. Especially so, in distributed, partitioned environments.

A Solution

However, when certain requirements are satisfied, one mechanism can substantively bridge the gap: atomic increment/decrement. Anything that implements it can be used to build a mutual-exclusion/locking system. And that's what this library does.

Qualities:

• must be reasonably "fast" (hash-time lookup)
• must be non-blocking ("retry-able")
• must be recoverable (expiration of dead/stale locks)
• must be able to be monitored / administered

Behaviour:

• blocks for a configurable duration when acquiring a lock across execution threads
• doesn't block when (re-)acquiring a lock within the same thread of execution

.... TBC ....

Usage

Mongo::Locking makes no effort to help configure the MongoDB connection - that's what the Mongo Ruby Driver is for. However, when the collection is specified as a proc, it will be lazily resolved during the first invocation of a lock. This makes the concern of load/initialization order largely irrelevant.

Configuring Mongo::Locking with the Mongo Ruby Driver would look like this:

::Mongo::Locking.configure(:collection => proc {
    ::Mongo::Connection.new("localhost").db("somedb").collection("locks")
})

Or using Mongoid:

::Mongo::Locking.configure({
    :collection => proc { ::Mongoid.database.collection("locks") },
    :logger     => Logger.new(STDOUT),
})

While Mongo::Locking depends on Mongo, it can be applied to any arbitrary object model structure, including other ORMs. All locks have a namespace (scope) and a key (some instance-related value), classes can depend on others for their locks, and the dependency graph is resolved at invocation-time.

Consider this simplified example of using it with DataMapper:

class Order
    include ::DataMapper::Resource
    include ::Mongo::Locking

    has n, :order_items

    lockable!
end

class OrderItem
    include ::DataMapper::Resource
    include ::Mongo::Locking

    belongs_to :order
    has 1, :job_flow

    # invokes method to get "parent" lockable
    locked_by! :order
end

class JobFlow
    include ::DataMapper::Resource
    include ::Mongo::Locking

    belongs_to :order_item

    # also takes a closure, yielding some abitrary "parent" lockable
    locked_by! { |me| me.order_item }
end

Other (simplified) graph configuration imperatives:

Order.lockable! :key => :id
Order.lockable! :scope => "OtherClass"
Order.lockable! :key => proc { |me| SHA1.hexdigest(me.balls) }

OrderItem.locked_by! { |me| me.order }
OrderItem.locked_by! :parent => proc { |me| me.order }
OrderItem.locked_by! :order
OrderItem.locked_by! :parent => :order

And then, a contrived use case:

order = Order.get(1)
order.lock do
    # ...

    order.order_items.each do |item|
        item.lock do

            # this won't block even though the same lock is being acquired

        end
    end
end

Not blocking on lock re-acquisition means save hooks on models can be as defensive as controller methods operating on them: both can lock, and it will Just Work.

Pretty neat!

Testing

Testing concurrency is "difficult", especially in Ruby. So for now, here's some irb/console-level tests that you can use to test the library with:

Given:

Pn == process N
Order.id == 1
OrderItem.id == 1, OrderItem.order_id = 1

1. General race, same object

    P1: Order.first.lock { debugger }  # gets and holds lock
    P2: Order.first.lock { puts "hi" } # retries acquire, fails
   

2. General race, locked root, attempt to lock from child

    P1: Order.first.lock { debugger }      # gets and holds lock
    P2: OrderItem.first.lock { puts "hi" } # retries acquire, fails
   

3. General race, locked root from child, attempt to lock from child

    P1: OrderItem.first.lock { debugger }  # gets and holds lock
    P2: OrderItem.first.lock { puts "hi" } # retries acquire, fails
   

4. Nested lock acquisition

    P1: Order.first.lock { puts "1"; Order.first.lock { puts "2" } }
    # should see 1 and 2