playing nice with parallel evaluation?

[bbolker]

I tried to evaluate multiple odin runs across cores using the future package and got

Error: future::evalFuture() failed on haroun (pid 306702) at 2026-05-23T16:39:36. Using package 'future' v1.70.0. Possible other reasons: Failed to attach one or more packages: there is no package called ‘odin.twostrain14ff673c2’ [future ; on 02945ea5dcb41927b3269b6ae7efe879@haroun<306702>]

• is there a recommended way to handle this?
• should I migrate to odin2 ?
• is this of sufficient interest that I should provide a reprex (I would guess that the same happens with plain-vanilla parallel::parapply etc., but haven't checked)?

[richfitz]

Hi Ben - sorry, I've moved on from Imperial and we were never that good at tracking issues anyway.

Short version: use odin2 if you possibly can, it's better in every way we think, especially running in parallel in a reasonable way. There's a migration doc and old code can usually be migrated automatically.

The parallelism in the new version is much better behaved with things like random number generation (if you're using stochastic models). The section in the odin-monty book focuses mostly on fitting to data (parallel chains and parallelism within the particle filter) but the basic dust2 docs have probably what you are looking for.

If you need to wrap at a higher level, then you can probably make use of future without modification because the dynamically generated pseudo-packages that we create have enough weird hacks in them to bootstrap themselves when loaded by R's multi-process based parallel frameworks.

You might get away with keeping with odin v1 if you package your model. This was never very well described but there's some docs here. If you just have an old project to get over the line try that but for new work, odin2 is the way forward.

I hope that helps! @edknock or @weshinsley are continuing the development of these packages and may have more to add.