[R-pkg-devel] New test in R-devel causes existing packages to fail: "Error: connections left open"

Does R CMD check --as-cran test for newly opened connections or any
open connections?  Could the check for stray connection in
examples/vignettes be:

1. Record what connections are open
2. Attach the package
3. Record what connections are open
4. Run the example
5. Assert that no *new* connections in addition to what's recorded in
Step 3 are open
6. Unload the package
7. Assert that no *new* connections in addition to what's recorded in
Step 1 are open

Step 5 asserts that the code in the example does not leave stray
connections behind, and Step 7 that the package itself does not leave
stray connections behind.

/Henrik
On Thu, Aug 23, 2018 at 1:25 PM David B. Dahl <dahl at stat.byu.edu> wrote:

Oops, I accidentally did not "reply-all".... Here is my message:

Thanks Uwe, Duncan, and Gabor for the response, advise, and flexibility.

Regarding Uwe's suggestion:  "... there should be a function that
creates the connction and one that closes the connection," I should
clarify.  The rscala package does just that.  There is a function
(named "scala") that creates the connection (using delayedAssign) and
another the closes the function (namely an S3 close method).  The
examples for the rscala package do this full open/close semantics,
but...

The problem comes when authors of another package, let's call it the
"FooBar" package, want to implement an algorithm in Scala based on
functionality provided by the rscala package.  Let's say they write a
function called "neatAlgorithm" based on Scala.  Yes, the FooBar
package author could require that, before the user calls the
"neatAlgorithm" function, they first call a function to set up the
connection (which itself would call the "rscala::scala" function) and
then, after calling the "neatAlgorithm" function, they call a function
to close the connection.

But that is not very user friendly and exposes the user to
implementation details of the algorithm.  The user of the FooBar
package don't really care whether the "neatAlgorithm" is implemented
in pure R, C++, Scala, or whatever, much like the users of the 'lm'
function don't need to know the implementation details or do any setup
before and after calling the function.

The current approach is that the connection to Scala is transparent to
the end user of a package.  Behind the scenes, the package author
establish the connection once it is needed and the rscala package
manages the connection and explicitly closes it when 1. the package is
unloaded or 2. the R session ends.  This approach does not leave
dangling connections  --- which I believe is the point of the new test
--- yet my package is caught up in the test.

I hope that this approach is still valid.  Perhaps the test could
result in a warning (instead of an error) and CRAN could accept
packages with such a warning.

If not, a work-around is to have a \dontshow section in the examples
that will close the connection (but leave the Scala process running)
and then automatically reestablish the connection as needed.  This
would not be very efficient but, as Duncan mentioned, it mostly only
effects the package examples themselves.  Plus, it would not be too
burdensome for package developers.

Again, thanks for considering my situation.

Best regards,

-- David

On Mon, Aug 20, 2018 at 11:11 PM Uwe Ligges
<ligges at statistik.tu-dortmund.de> wrote:

My advise:

Apparently you want to have communication via sockets to scala.

So there should be a function that creates the connction and one tha
closes the connection.
Comparable to starting some parallel cluster and stopping it again.

In the meantime, you can allow for all sorts of communication.

So that's fine.

Then in your examples, simply design them to be standalone, i.e. in
*your* examples always start the connection and stop it again at the end
of one examples block, i.e. the exampels defined in one Rd file.

Best,
Uwe Ligges



On 20.08.2018 02:11, Duncan Murdoch wrote:

On 19/08/2018 12:34 PM, G?bor Cs?rdi wrote:

Sorry, missed that these were examples, so, yeah, that's harder.  G.

How about a function that checks if the connection is open before doing
anything, and then at the end you close it if it wasn't already open?
This will make all examples run slower on CRAN, but won't affect most
users who are doing their own stuff as well as running examples.

Or, how about the startup code for the package opens the connection?

Or perhaps CRAN will respond to this thread with another suggestion.

Duncan Murdoch

On Sun, Aug 19, 2018 at 6:32 PM G?bor Cs?rdi <csardi.gabor at gmail.com>
wrote:

You could just create a function to close the connection and then
people could call it at the end of their test suites. >>
Gabor
On Sun, Aug 19, 2018 at 6:22 PM David B. Dahl <dahl at stat.byu.edu> wrote:

In preparing to submit an update of my package to CRAN, I found that
R-devel has a new test regarding "connects left open" that my packages
fail.  The new test appears to have been committed by Uwe Ligges in
revisions 74959 and 74964 on 2018-07-14 and 2018-07-15, respectively.
The commit message says, "check after each example whether open
connections exist, indicating e.g. file connections were left open or
parallel clusters still running."

I am hoping for advice on how to pass "R CMD check --as-cran".  Or,
perhaps my situation will prompt a change to the test or, at least,
having it result in a warning instead of an error.

Below I describe the situation.  My rscala package allows developers
to write R packages based on Scala (much like rJava and Rcpp for Java
and C++, respectively).  Scala runs as a separate process and
interprocess communication is implemented using socket connections.

Suppose a package using rscala has functions that call Scala code.
(Such packages are 'bamboo', 'sdols', and 'shallot' on CRAN.)  The
first time a user executes an R function calling down into Scala, a
socket connect between Scala and R is established.  For the sake of
low latency, after the call to the function ends, the connection stays
open until the package is unloaded or the R session ends.  But, this
approach runs afoul of the new test mentioned above that appears to be
designed to catch connections that are *accidentally* left open.

I definitely do not want to users of my packages 'bamboo', 'sdols',
and 'shallot' to have to think about managing connection between Scala
and R.  That's an implementation detail and uing the package should be
transparent for the user (who doesn't care about the implementation
details).

On my end, I see two solutions:  1. I could try to reengineer my
approach --- establishing a new connection for every single call into
Scala --- although I am loath to do anything to increase the latency,
or 2. I could wrap all the examples in \donttest so that CRAN checks
are passed.

Or, again, perhaps my situation will prompt a reevaluation of the
test.  Perhaps it could result in a warning (instead of an error) and
the CRAN maintainers would accept packages with such a warning.

Any advise?  Thanks a lot!

-- David