GitLab

If the context manager is exited via an exception, the original
permissions should be restored.

Closes #7170830

SQLServer refuses to index an unlimited text column, so we should drop
unique_together constraints (which imply an index) before we drop size
constraints.  Closes #5560601.

Until now we would:

- remove the old size constraint from the in-memory schema
- call update_rdef_column which removes the size restriction from the
  column's type
- add the new constraint object
- call update_rdef_column which adds the size restriction back

This breaks on SQL Server when the column is involved in an index (e.g.
as part of a multi-column unique constraint), because in the
intermediate stage the column's type is "nvarchar(max)", which is not
indexable.

Of course we must still detect the case where a size constraint is
really dropped and update the db schema accordingly.

Closes #5557633.

On entity creation, if there are some local permissions on the relation,
we have no way of checking them since neither the subject nor the object
of the relation exists yet. In such a case, we should add the link by
default, for consistency (see other places where we use
`may_have_permission`).

Closes #6711900

Prevents the "modify" action from being shown when there is nothing to
edit.

When selecting an inlined creation form, we should catch the
NoSelectable exception that will be raised if the user cannot add
entities of the target type (this is not and cannot be verified earlier)
or if some other custom selector prevents the form from being selected.

Closes #6510921

Otherwise a simple sync_schema(syncprops=False) turns the subsequent
sync_schema() calls into a no-op.

Closes #6510918.

Without the fix in migractions.py, the introduced test crashes with an
operational error because of an attempt to write to a non-existing
relation table (computed relations are not materialized).

Closes #6304946

as this method may be used from outsite the class (eg in UpdateFTIHook), this is
the proper way to consider it anyway.

Closes #6206636

devtools stores info about existing dbs in the db handler, but in the
case of postgresql that doesn't take into account the path to the
cluster's datadir.  Which means if we run two test modules (in the same
test run), we'll create a "__default_empty_db__" for the first one,
cache its existence, and then when moving on to the other module,
believe the template already exists (but since the datadir depends on
the test module's path, it does not).  This patch is a bit of a kludge,
and it would be better to make the cache key include enough data to not
need this, but I'm not sure how to do that.

There is a special handling for datetime with tzinfo, where value was stored in
the query cache. The implementation of merge_args was simply overwriting
parameters of the query with those in the query cache, expecting no collision.

To fix this:
* handle replacement of tzinfo in merge_args, not at sql generation time
* add an assertion to ensure we've actually no collision

Closes #6978316

Closes #7173702

SQL Server's UNIQUE is not standards compliant, it doesn't accept two
NULL values.  Related to #7173702.

"msg % row" dies with "not enough arguments for format string" if row is
a pyodbc.Row.

SQL Server doesn't allow indexing (and thus unique constraints) on
nvarchar(max).  Related to #7173702.

Seems confusing to have __eq__ but not __ne__.

eg. with pyramid.

It doesn't seem to have ever been used.

Simply skip the action (rename_relation_type) with an early return.

- stop using *.install.in files, just use wildcards in *.install
- switch to X-Python-Version (preferred for dh_python2)
- also call dh_python2 for /usr/share/cubicweb, not just public modules
- drop long obsolete ginco and erudi migration code from cubicweb-ctl.postinst
- fix cubicweb-ctl cleanup: purge code belongs in postrm, not prerm
- don't compress changelog.html