Avoid materializing ALTREP vectors by subsetting them first #499
Conversation
… some unit tests for variables display values.
| // For ALTREP objects we first subset the first `MAX_DISPLAY_VALUE_LENGTH` values | ||
| // and only then try to format. This is due to a possible bug in the implementation of | ||
| // deferred string, for which once you try to acess a single element, it tries to allocate | ||
| // the full vector. |
There was a problem hiding this comment.
I'm pretty sure its not a bug
When they hit the ALTREP Elt method the first time, they allocate a STRSXP of the final size but with all CHARSXP elements set to NULL (the C pointer value). Then they just coerce and update the 1 Elt value they need to extract.
The reason they do it this way is because someone has to PROTECT() the CHARSXP element that gets generated. It ends up being the STRSXP they allocate the first time Elt is called.
It's different for Extract_subset because they decide that when they extract a subset, they actually extract the subset from the underlying object they are delaying the conversion to string on, and then they wrap that result in another deferred string. It's up to the caller to protect that new deferred string
There was a problem hiding this comment.
Indeed, the Error: vector memory limit of 32.0 Gb reached, see mem.maxVSize() made me think it was materializing the vector, but they are actually just allocating an empty STRSXP, which shouldn't be that problematic.
Addresses posit-dev/positron#4537
When calling
STRING_ELTin an ALTREP vector, it might try to allocate the entire vector. With this PR, we first subset the vector, which will callExtract_subsetALTREP method before trying to allocate the entire vector.Also added some unit tests for variables display values.