Group items matching

in title, tags, annotations or url

We need to get rid of these gratuitous dependencies.  Merely papering over them with a transaction—making them “safe”—doesn’t do anything to improve the natural parallelism that a program contains.  It just ensures it doesn’t crash.  Sure, that’s plenty important, but providing programming models and patterns to eliminate the gratuitous dependencies also achieves the goal of not crashing but with the added benefit of actually improving scalability too.  Transactions have worked so well in enabling automatic parallelism in databases because the basic model itself (without transactions) already implies natural isolation among queries.  Transactions break down and scalability suffers for programs that aren’t architected in this way.  We should learn from the experience of the database community in this regard

There will always be hidden mutation of shared state inside lower level system components.  These are often called “benevolent side-effects,” thanks to Hoare, and apply to things like lazy initialization and memorization caches.  These will be done by concurrency ninjas who understand locks.  And their effects will be isolated by convention.

Even with all of this support, we’d be left with an ecosystem of libraries like the .NET Framework itself which have been built atop a fundamentally mutable and imperative system.  The path forward here is less clear to me, although having the ability to retain a mutable model within pockets of guaranteed isolation certainly makes me think the libraries are salvageable.  Thankfully, the shift will likely be very gradual, and the pieces that pose substantial problems can be rewritten in place incrementally over time.  But we need the fundamental language and type system support first.