Bugzilla Bug 275217

Another tale in the quest for a easier to follow portage code.

Created an attachment (id=195594) [details]
resolve the circular dep between Scheduler and depgraph; mark all functions as
local that aren't used outside of depgraph

Created an attachment (id=195595) [details]
move changelog stuff stuff out of depgraph (needs changelog.py)

Created an attachment (id=195597) [details]
changelog.py for part2

(In reply to comment #1)
> Created an attachment (id=195594) [edit] [details]
> resolve the circular dep between Scheduler and depgraph; mark all functions as
> local that aren't used outside of depgraph

Thanks, this is in svn r13676.

(In reply to comment #2)
> Created an attachment (id=195595) [edit] [details]
> move changelog stuff stuff out of depgraph (needs changelog.py)

Thanks, this is in svn r13677.

Created an attachment (id=195653) [details]
splits depgraph.select_files into select_files and _resolve

In the long term all decision making shall be made in _resolve. select_files
converts its input to args and _resolve uses the args to pull in the root nodes
now.

(In reply to comment #5)
> Created an attachment (id=195653) [edit] [details]
> splits depgraph.select_files into select_files and _resolve

Thanks, this is in svn r13685.

Created an attachment (id=196212) [details]
Move all memeber variables of _emerge.depgraph into frozen_config or
dynamic_config.

This patch creates two classes in _emerge/depgraph.py: _frozen_depgraph_config
and _dynamic_depgraph_config. They are supposed to hold all member variables of
_emerge.depgraph. Those that may change after initialization belong to
_dynamic_depgraph_config, others to _frozen_depgraph_config. depgraph itself
holds only instances of these two classes.

(In reply to comment #7)
> Created an attachment (id=196212) [edit] [details]
> Move all memeber variables of _emerge.depgraph into frozen_config or
> dynamic_config.

Thanks, this is in svn r13747.

Just for reference:

What is left? 
1) Write wrapper functions for _dynamic_config to track changes to it.
2) Write a function that searches the point where offending packages were
selected and reverts the dynamic_config to this point.
3) Write a function that figures out what to mask if something goes wrong.
4) modify _resolve to be somehow state less, so we can simply restart from any
state.

As discussed on irc, the simplest way implement backtracking (at least
initially) is to try saving choices from depgraph, and then creating a whole
new depgraph with slightly different choices. Right now I'm working on making
the depgraph constructor take a _frozen_depgraph_config instance as a
parameter, so that we can use a single instance for each depgraph we create.

(In reply to comment #10)
> Right now I'm working on making
> the depgraph constructor take a _frozen_depgraph_config instance as a
> parameter, so that we can use a single instance for each depgraph we create.

This is done in svn r13760, and it's used to optimize resume_depgraph() in
r13761.

I looked at your backtracking patch with the _pref_list_lookup() function and I
think that it would be better to use an approach similar to the one used to
solve bug #253904, in which _dep_check_composite_db._visible() is used to
influence the behavior of dep_zapdeps(). When a package is masked by
_dep_check_composite_db._visible(), it appears to dep_zapdeps() that the
package is unavailable, and so any choice which pulls in that package will be
avoided.

Created an attachment (id=196543) [details]
Make emerge mask the first package pulled into the tree in case of a slot
conflict and restart the dependency calculation

(In reply to comment #13)
> Created an attachment (id=196543) [edit] [details]
> Make emerge mask the first package pulled into the tree in case of a slot
> conflict and restart the dependency calculation

Thanks, that's in svn r13769. I added some additional features:

 * display message about missed updates
 * cache frozen_config instance, to optimize performance
 * disable backtracking if it fails, fall back to a normal

This is fixed in 2.2_rc34.

Reopening until this is an unmasked release.

This is fixed in 2.1.7.