Questions/suggestions about new staged installation

Thu, Apr 25, 2019 5:41 AM

On 4/25/19 2:20 PM, Pages, Herve wrote:

On 4/25/19 04:57, Tomas Kalibera wrote:

On 4/25/19 3:11 AM, Pages, Herve wrote:

Hi,

I was playing around with inotifywait (great tool!) to see the new
staged installation of source packages in action. In one terminal I'm
monitoring the create/delete/move events of the installation library
with:

 ? ? inotifywait -m --timefmt '%F %T' --format '%T -- %w %e %f' -e create
-e delete -e move path/to/R/library/

While in another terminal I install CRAN package abc with:

 ? ? R CMD INSTALL abc_2.1.tar.gz

All deps are already installed.

## With R 3.5

When installing abc the first time with R 3.5 inotifywait reports the
following events:

2019-04-24 16:57:42 -- /home/hpages/R/R-3.5.r75051/library/ CREATE,ISDIR
00LOCK-abc
2019-04-24 16:57:42 -- /home/hpages/R/R-3.5.r75051/library/
CREATE,ISDIR abc
2019-04-24 16:57:45 -- /home/hpages/R/R-3.5.r75051/library/ DELETE,ISDIR
00LOCK-

Then on subsequent times:

2019-04-24 16:58:14 -- /home/hpages/R/R-3.5.r75051/library/ CREATE,ISDIR
00LOCK-abc
2019-04-24 16:58:14 -- /home/hpages/R/R-3.5.r75051/library/
MOVED_FROM,ISDIR abc
2019-04-24 16:58:14 -- /home/hpages/R/R-3.5.r75051/library/
CREATE,ISDIR abc
2019-04-24 16:58:17 -- /home/hpages/R/R-3.5.r75051/library/ DELETE,ISDIR
00LOCK-abc

IIUC the additional 'MOVED_FROM,ISDIR abc'? event on subsequent installs
is triggered by the backing up of the earlier installation in case the
new installation fails.

## With R 3.6

First installation:

2019-04-24 17:09:04 -- /home/hpages/R/R-3.6.r76385/library/ CREATE,ISDIR
00LOCK-abc
2019-04-24 17:09:04 -- /home/hpages/R/R-3.6.r76385/library/
CREATE,ISDIR abc
2019-04-24 17:09:08 -- /home/hpages/R/R-3.6.r76385/library/
MOVED_TO,ISDIR abc
2019-04-24 17:09:09 -- /home/hpages/R/R-3.6.r76385/library/ DELETE,ISDIR
00LOCK-abc

Subsequent installations:

2019-04-24 17:09:59 -- /home/hpages/R/R-3.6.r76385/library/ CREATE,ISDIR
00LOCK-abc
2019-04-24 17:09:59 -- /home/hpages/R/R-3.6.r76385/library/
MOVED_FROM,ISDIR abc
2019-04-24 17:09:59 -- /home/hpages/R/R-3.6.r76385/library/
CREATE,ISDIR abc
2019-04-24 17:10:02 -- /home/hpages/R/R-3.6.r76385/library/
MOVED_TO,ISDIR abc
2019-04-24 17:10:03 -- /home/hpages/R/R-3.6.r76385/library/ DELETE,ISDIR
00LOCK-abc

IIUC the new 'MOVED_TO,ISDIR abc' event is triggered by the package
installation folder being moved from the temporary location to the final
location.

However I wonder about the CREATE,ISDIR abc event. It looks like even
for a staged installation tools:::.install_packages() still creates the
empty abc dir. This should no longer be necessary because this empty
folder is replaced later by the 'MOVED_TO,ISDIR abc' event. Am I missing
somethin

Yes, the final directory location gets created before the
installation. I did this to minimize the added cognitive complexity to
the code. I could explore if this could be changed, if the
installation could work without that directory (which would mean
implementing and testing on all CRAN/BIOC, installing in different
modes, so certainly some effort).? Is the current implementation
causing any problem?

Another thing: the times indicate that the 'MOVED_TO,ISDIR abc' event
(move from temp to final location) happens about 3-4 seconds after the
'MOVED_FROM,ISDIR abc' event (backing up). In this interval of time, the
abc package is missing. Wouldn't it be safer to leave the earlier
installation where it is and to replace it when the new installation
folder is moved from the temporary location to the final location? Said
otherwise, the backing up step no longer seems necessary with staged
installs. So inotifywait would report something like this:

2019-04-24 17:09:59 -- /home/hpages/R/R-3.6.r76385/library/ CREATE,ISDIR
00LOCK-abc
2019-04-24 17:10:02 -- /home/hpages/R/R-3.6.r76385/library/
MOVED_TO,ISDIR abc
2019-04-24 17:10:03 -- /home/hpages/R/R-3.6.r76385/library/ DELETE,ISDIR
00LOCK-abc

That is, only 3 events. On first and subsequent installation.

I don't think that would be possible. The package is test-loaded also
from the final installation directory, and if that fails, restored
from a back-up. This test-loading is useful as it can detect some
problems with hard-coded directories possibly not detected before on
test-loading from the temporary installation directory.

Makes sense.

So backing up is still needed.

But couldn't the gap of 3-4 seconds where the abc package is missing be
reduced by backing up right before moving the newly installed package
from the temporary location to the final location?

In other words, instead of:

  ??? back up, then install to temp loc, then move to final loc

do:

  ??? install to temp loc, then back up, then move to final loc

This would reduce the gap to almost 0 seconds on a Unix file system
where moving a folder is very cheap (even if the folder is big), at
least when moving doesn't cross partition borders.

Not a big deal for a light package like abc but some Bioconductor
packages take about 20 min. to compile!

It should be possible, probably. But why is the gap a problem? And why 
is a shorter gap better than long gap? (the time depends on how long it 
takes to load the package, which can be >10s for some Bioconductor 
packages).

It is not safe to run concurrently with the installation any code on the 
same library that would be accessing the packages being installed. Such 
code could access the old version of the package mixed with the new 
version. A longer gap could actually (perhaps accidentally but still) 
make such problems more visible, and easier to detect, than a short gap.

Tomas

Questions/suggestions about new staged installation

Thread (4 messages)