Move data mask cleaning from rlang to dplyr

[lionel-]

As we're no longer cleaning the mask by default.

Also investigate whether we can avoid active binding crashes without cleaning the data mask.

[krlmlr]

We might be able to avoid cleaning the data mask with weak references (WEAKREFSXP), need to look into it. But:

At the time of writing no CRAN or Bioconductor package uses weak references.

That's not very promising...

[krlmlr]

On the other hand, we might also use a C++ implementation of weak references.

To recap: We need to clean the data mask because otherwise a pointer to an object with limited lifetime might leak. We are not in control of the lifetime of the objects for which we share an XPtr. We can overcome this by sharing only proxy objects that contain an internal weak reference to the real C++ object; when the real C++ object goes out of scope, the proxy returns some default value when accessed.

[lionel-]

Maybe an alternative is to add a stale flag in the data mask. When that flag is set a default value would be returned?

Or: adding a finalizer to the environment containing the active bindings and clean up the relevant dplyr objects from that finalizer?

[krlmlr]

One way or another, the proxy object needs to know that the data source isn't available anymore. I prefer controlling lifetime of internal dplyr classes from dplyr code.

[krlmlr]

This needs an enhancement to bindrcpp, which I meant to release anyway before updating dplyr: krlmlr/bindrcpp#7. The current interface only supports naked pointers, I need to give up ownership at least for a proxy object.

[lionel-]

Is it useful to keep bindrcpp as a separate package?

[krlmlr]

It has two downstream dependencies ;-)

[krlmlr]

And, yes, the ping-pong between R and C++ is pretty convoluted, it helps to isolate it from the rest of the codebase.

[krlmlr]

I have an implementation that uses boost::weak_ptr, will share soon.

[krlmlr]

@lionel-: Is there functionality in rlang that can replace bindr and bindrcpp? I kept these packages separate to avoid the Rcpp dependency in bindr.

[lionel-]

I don't understand completely what those packages do but I don't think so.

[krlmlr]

bindr creates an environment full of almost identical active bindings. bindrcpp is the Rcpp interface to bindr.

[lionel-]

I think this functionality is too specific for rlang.

[lock]

This old issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with reprex) and link to this issue. https://reprex.tidyverse.org/