- Before that change on each requests 4x we loaded plugins for authentication, this hit
many caches, db invalidation context and loaded the plugins logic each time. This was a heavy performance hit
for SVN and other backends as they needed to load that plugins many many times
- Since Authentication plugins almost never change, we'll not store the plugins listed for authentication into
the authnregistry for each process
- For AuthPlugins settings we now also flush plugins settings, and authnregistry cached plugins
- previous implementation had a bug that cached the settings in each class, caused not refreshing
the update of settings in multi-worker mode. Only restart of rhodecode loaded new settings.
- especially for mysql in case large number of data in caches there could be critical errors storing cache, and thus preventing
users from authentication. This is caused by the fact that we used single namespace for ALL users. It means it grew as number of users grew
reaching mysql single column limit.
This changes the behaviour and now we use namespace per-user it means that each user-id will have it's own cache namespace fragmenting maximum column data to
a single user cache. Which we should never reach.
- we'll skip any syncing on plugins which simply don't get any group information
- we let plugins define if they wish to sync groups
- prevent from odd cases in which someone sets user groups as syncing, and using regular
plugin. In this case memebership of that group would be wiped, and it's generaly bad behaviour.
- dependencies: bumped:
-setuptools-scm==1.15.6
-celery==4.1.0
-kombu==4.1.0
-pytz==2017.3
-Routes==2.4.1
Removed all pylons components. Refactored the code to use only pyramid
parts
- this gives a 30% speed increase in operations like svn commit
- generally uses the same mechanismy like auth cache to cache permissions for faster
access to vcs commands
- it's not recommended to log things like ldap access passwords or
other credentials.
- we expose a machanismy for each plugin to define a unsafe keys to be
removed.
This way some error messages will look nicer, since they are based on the
__str__ of the plugin instance. Using the plugin_id will give the user (and us
in the support case) the exact pointer to the affected plugins.
