Comment by vlovich123
15 hours ago
I don’t think we read the same thing.
> Models can be pulled along other axes, however, such as whether memory locations must be tagged in order to be used in a transaction or not, etc. Haskell requires this tagging (via TVars) so that side-effects are evident in the type system as with any other kind of monad. We quickly settled on unbounded transactions.
Snip
> In hindsight, this was a critical decision that had far-reaching implications. And to be honest, I now frequently doubt that it was the right call. We had our hearts in the right places, and the entire industry was trekking down the same path at the same time (with the notable exception of Haskell)
So basically not that TM isn’t workable, but unbounded TM is likely a fool’s errand but Haskell’s is bounded TM that requires explicit annotation of memory that will participate in atomicity.
> unbounded TM is likely a fool’s errand
It's the whole language, not just the TM code. Other languages have no way of opting out of the TM code, whereas Haskell does.
Well, in a sense other languages opt-in to TM code by interfacing with a relational database. Similar to how in Haskell you put TVar in front of the relevant bits to opt-in.
In Haskell it's just more economic and better integrated with the rest of the language and its typesystem.
Having worked a bit on a hobby STM in C++ (spun out of a DB startup) I would have to agree. Fully transparent STM that depends on a "sufficiently smart compiler" for an imperative language with unrestricted side effects is hopeless. But I do think that a much humbler version of STM is feasible for C++ or Rust, requiring much more explicit cooperation from the programmer. I haven't worked on this for 3 years but hope to revisit it someday.
Haskell still needs TVar and it’s not an imperative language with unrestricted side effects. I think it’s bounded vs unbounded. Side effects make it more complicated perhaps but it sounds like even in a JIT language you could have done it.
It's possible (I've done it) in C++ but there are huge footguns that I'm still ambivalent about. I agree that the bounded/unbounded distinction is the key: data must explicitly be tagged as transactional. One of the benefits of bootstrapping an STM from an existing DB as I did is that this explicitness is already present.