[Rd] Objects not gc'ed due to caching (?) in R's S3 dispatch mechanism

Tomas Kalibera tomas.kalibera at gmail.com
Tue Mar 27 12:16:28 CEST 2018


On 03/27/2018 11:53 AM, Iñaki Úcar wrote:
> 2018-03-27 11:11 GMT+02:00 Tomas Kalibera <tomas.kalibera at gmail.com>:
>> On 03/27/2018 09:51 AM, Iñaki Úcar wrote:
>>> 2018-03-27 6:02 GMT+02:00  <luke-tierney at uiowa.edu>:
>>>> This has nothing to do with printing or dispatch per se. It is the
>>>> result of an internal register (R_ReturnedValue) being protected. It
>>>> gets rewritten whenever there is a jump, e.g. by an explicit return
>>>> call. So a simplified example is
>>>>
>>>> new_foo <- function() {
>>>>     e <- new.env()
>>>>       reg.finalizer(e, function(e) message("Finalizer called"))
>>>>         e
>>>>         }
>>>>
>>>> bar <- function(x) return(x)
>>>>
>>>> bar(new_foo())
>>>> gc() # still in .Last.value
>>>> gc() # nothing
>>>>
>>>> UseMethod essentially does a return call so you see the effect there.
>>> Understood. Thanks for the explanation, Luke.
>>>
>>>> The R_ReturnedValue register could probably be safely cleared in more
>>>> places but it isn't clear exactly where. As things stand it will be
>>>> cleared on the next use of a non-local transfer of control, and those
>>>> happen frequently enough that I'm not convinced this is worth
>>>> addressing, at least not at this point in the release cycle.
>>> I barely know the R internals, and I'm sure there's a good reason
>>> behind this change (R 3.2.3 does not show this behaviour), but IMHO
>>> it's, at the very least, confusing. When .Last.value is cleared, that
>>> object loses the last reference, and I'd expect it to be eligible for
>>> gc.
>>>
>>> In my case, I was using an object that internally generates a bunch of
>>> data. I discovered this because I was benchmarking the execution, and
>>> I was running out of memory because the memory wasn't been freed as it
>>> was supposed to. So I spent half of the day on this because I thought
>>> I had a memory leak. :-\ (Not blaming anyone here, of course; just
>>> making a case to show that this may be worth addressing at some
>>> point). :-)
>>  From the perspective of the R user/programmer/package developer, please do
>> not make any assumptions on when finalizers will be run, only that they
>> indeed won't be run when the object is still alive. Similarly, it is not
>> good to make any assumptions that "gc()" will actually run a collection (and
>> a particular type of collection, that it will be immediately, etc). Such
>> guarantees would too much restrict the design space and potential
>> optimizations on the R internals side - and for this reason are typically
>> not given in other managed languages, either. I've seen R examples where
>> most time had been wasted tracing live objects because explicit "gc()" had
>> been run in a tight loop. Note in Java for instance, an explicit call to
>> gc() had been eventually turned into a hint only.
>>
>> Once you start debugging when objects are collected, you are debugging R
>> internals - and surprises/changes between svn versions/etc should be
>> expected as well as changes in behavior caused very indirectly by code
>> changes somewhere else. I work on R internals and spend most of my time
>> debugging - that is unfortunately normal when you work on a language
>> runtime. Indeed, the runtime should try not to keep references to objects
>> for too long, but it remains to be seen whether and for what cost this could
>> be fixed with R_ReturnedValue.
> To be precise, I was not debugging *when* objects were collected, I
> was debugging *whether* objects were collected. And for that, I
> necessarily need some hint about the *when*.
They would be collected eventually if you were running a non-trivial 
program (because there would be a jump inside).
> But I think that's another discussion. My point is that, as an R user
> and package developer, I expect consistency, and currently
>
> new_foo <- function() {
>    e <- new.env()
>    reg.finalizer(e, function(e) message("Finalizer called"))
>    e
> }
>
> bar <- function(x) return(x)
>
> bar(new_foo())
> gc() # still in .Last.value
> gc() # nothing
>
> behaves differently than
>
> new_foo <- function() {
>    e <- new.env()
>    reg.finalizer(e, function(e) message("Finalizer called"))
>    e
> }
>
> bar <- function(x) x
>
> bar(new_foo())
> gc() # still in .Last.value
> gc() # Finalizer called!
>
> And such a difference is not explained (AFAIK) in the documentation.
> At least the help page for 'return' does not make me think that I
> should not expect exactly the same behaviour if I write (or not) an
> explicit 'return'.
As R user and package developer, you should have consistency in 
_documented_ behavior. If not, it is a bug and has to be fixed either in 
the documentation, or in the code. You should never depend on 
undocumented behavior, because that can change at any time. You cannot 
expect that different versions of R would behave exactly the same, not 
even the svn versions, that is not possible and would not be possible 
even if we did not change any code in R implementation, because even the 
OS, C compiler, hardware, and third party libraries have their specified 
and unspecified behavior.

Best
Tomas
>
> Regards,
> Iñaki
>
>> Best
>> Tomas
>>



More information about the R-devel mailing list