Skip to content

I'm taking a break from Promise for a post or two to jot down some stuff that I've been thinking about while discussing future enhancements to MindTouch Dream with @bjorg. In Dream all service to service communication is done via HTTP (although the traffic may never hit the wire). This is very powerful and flexible, but also has performance drawbacks, which have led to many data sharing discussions.

Whether you are using data as a message payload or even just putting data in a cache, you want sender and receiver to be unable to see each others interaction with that data, which would happen if the data was a shared, mutable instance. If you were to allow shared modification on purpose or on accident can have very problematic consequences:

  1. Data corruption: Unless you wrap the data with a lock, two threads could try to modify the data at the same time
  2. Difference in distributed behavior: As soon as the payload crosses a process boundary, it ceases to be shared so changing topology, changes data behavior

There are a number of different approaches for dealing with this, each a trade-off in performance and/or usability. I'll use caching as the use case, since it's a bit more universal than message passing, but the same patterns applies.

Cloning

A naive implementation of a cache might just be a dictionary. Sure, you've wrapped the dictionary access with a mutex, so that you don't get corruption accessing the data. But multiple threads would still have access to the same instance. If you aren't aware of this sharing, expect to spend lots of time trying to debug this behavior. If you are unlucky it's not causing crashes but causes strange data corruption that you won't even know about until your data is in shambles. If you are lucky the program crashes because of an access violation of some sort.

Easy, we'll just clone the data going into the cache. Hrm, but now two threads getting the value are still messing with each other. Ok, fine, we'll clone it coming out of the cache. Ah, but if the orignal thread is still manipulating its copy data while others are getting the data, the cache keeps changing. That kind of invalidates the purpose of caching data.

So, with cloning we have to copy the data going in and coming back out. That's quite a bit of copying and in the case that the data goes into the cache and expires before someone uses it, it's a wasted copy to boot.

Immutability

If you've paid any attention to concurrency discussions you've heard the refrain from the functional camp that data should be immutable. Every modfication of the data should be a new copy with the orginal unchanged. This is certainly ideal for sharing data without sharing state. It's also a degenerative version of the cloning approach above, in that we are constantly cloning, whether we need to or not.

Unless your language supports immutable objects at a fundamental level, you are likely to be building this by hand. There's certainly ways of mitigating its cost, using lazy cloning, journaling, etc. i.e. figuring out when to copy what in order to stay immutable. But likely you are going to be building a lot of plumbing.

But if the facilities exist and if the performance characteristics are acceptable, Immutability is the safest solution.

Serialization

So far I've ignored the distributed case, i.e. sending a message across process boundaries or sharing a cache between processes. Both Cloning and Immutability rely on manipulating process memory. The moment the data needs to cross process boundaries, you need to convert it into a format that can be re-assembled into the same graph, i.e. you need to serialize and deserialize the data.

Serialization is another form of Immutability, since you've captured the data state and can re-assemble it into the original state with no ties to the original instance. So Serialization/Deserialization is a form of Cloning and can be used as an engine for immutability as well. And it goes across the wire? Sign me up, it's all i need!

Just like Immutability, if the performance characteristics are acceptable, it's a great solution. And of course, all serializers are not equal. .NET's default serializer, i believe, exists as a schoolbook example of how not to do it. It's by far the slowest, biggest and least flexible ones. On other end of scale, google's protobuf is the fastest and most compact I've worked with, but there are some flexibility concessions to be made. BSON is a decent compromise when more flexibility is needed. A simple, fast and small enough serializer for .NET that i like is @karlseguin's Metsys.Little. Regardless of serializer, even the best serializer is still a lot slower than copying in-process memory, never mind not even having to copy that memory.

Freeze

It would be nice to avoid the implicit copies and only copy or serialize/deserialize when we need to. What we need is for a way for the originator to be able to declare that no more changes will be made to the data and for the receivers of the data to declare whether they intend to modify the retrieved data, providing the folowing usage scenarios:

  • Originator and receiver won't change the data: same instance can be used
  • Originator will change data, receiver won't: need to copy in, but not coming out
  • Originator won't change the data, receiver will: can put instance in, but need to copy on the way out

In Ruby, freeze is a core language concept (a_lthough I profess my ignorance of not knowing how to get a mutable instance back again or whether this works on object graphs as well._) To let the originator and receiver declare their intended use of data in .NET, we could require data payloads to implement an interface, such as this:

public interface IFreezable<T> {
  bool IsFrozen { get; }

  void Freeze(); // freeze instance (no-op on frozen instance)
  T FreezeDry(); // return a frozen clone or if frozen, the current instance
  T Thaw();      // return an unfrozen clone (regardless whether instance is frozen)
}

On submitting the data, the container (cache or message pipeline) will always call FreezeDry() and store the returned instance. If the originator does not intend to modify the instance submitted further, it can Freeze() it first, turning the FreezeDry() that the container does into a no-op.

On receipt of the data, the instance is always frozen, which is fine for any reference use. But should the receiver need to change it for local state tracking, or submitting the changed version, it can always call Thaw() to get a mutable instance.

While IFreezable certainly offers some benefits, it'd be a pain to add to every data payload we want to send. This kind of plumbing is a perfect scenario for AOP, since its a concern of the data consumer not of the data. In my next post, I'll talk about some approaches to avoid the plumbing. In the meantime, the WIP code for that post can be found on github.