Enable fit$save_object also to serialise directly to R

[MansMeg]

I would like to save a cmdstanr object and use the serialisation done in save_object(), but I would like to wrap it myself. Hence, something like this:

obj <- fit$save_object()
saveRDS(list(obj, my_metadata=list(...)))

I think this would be done, e.g. by supplying format=NULL or similar in the current function?

cmdstanr/R/fit.R

Line 140 in cb6d218

save_object <- function(file, format = c("rds", "qs2"), ...) {

I'm happy to open a PR for this if you think this is a good idea. Just tell me what you think would be a good API.

[jgabry]

I think format=NULL is OK and a PR would be great. But how would this work internally? Right now save_object does two things:

1. Forces some things into memory that are important for saving the object (draws, sampler diagnostics, etc)
2. Calls saveRDS or qs2::qs_save depending format

What would format = NULL do? Just step 1?

[VisruthSK]

I think a NULL format just doing step 1 makes sense and fits neatly with the existing API. Alternatively we could split step 1 into a small function and make save_object() call it.

Edit: not sure how idiomatic it is, but maybe we could write a method for force() that loads the whole thing properly into memory?

[MansMeg]

Yes. It would just do step 1 and return the object. The function already returns the serialised object invisbly so this is a very small change.

Edit:
I think force is a good idea to read everything into memory to store it later on.

So this would mean:
fit$force() would return a fit object I can then wrap and store using eg saveRDS()?

[jgabry]

I think force is typically used to force the evaluation of an argument, right? So maybe it would be confusing to repurpose it for this. But I like the idea. Any other ideas for names we could use? (I guess force could be ok, but if we can come up with something that doesn't already have another meaning in R that would be ideal I think)

[VisruthSK]

@MansMeg yes that is the API I was thinking, the same as your example except with a different func call and no format option at all.

Since fit objects can be seen as kinda sorta loosely lazy, I thought force() might be okay but you're right. How about these: load(), store(), prep()?

[jgabry]

I think of those I prefer load. You guys?

[VisruthSK]

Of those I also prefer load but I don't know if is descriptive enough.

[jgabry]

Yeah it's not very descriptive. fit$load_everything_into_memory() is a big much though, and I'm not actually sure it loads everything (we'd need to check). Part of me thinks we should just just add format = NULL instead of trying to figure out how to name this function ;) But I'm also ok with adding it.

[VisruthSK]

The documentation is about saving to disk, but I guess we could tweak that. I'm now vaguely leaning towards a new function being cleaner and reserving save_object() for writing to a file. It does seem easier to just offer format = NULL though :)

[MansMeg]

Maybe fit$collect_content(), fit$collect_output() or $materialize()?

[VisruthSK]

I like materialize the most!

[jgabry]

I like it too

[MansMeg]

Now a PR that passes the unit tests.