fix migration from pre-3.13.1 versions (closes #2846978) Need to check the existence of the asource column before the first call to eid_type_source.

[constraint] more robust unicity constraint failures reporting for end-users Postgres or Sqlserver have limits on the index names (around resp. 64 and 128 characters). Because `logilab.database` encodes the `unique together` constraint rtypes in the index names, we sometimes get truncated index names, from which it is impossible to retrieve all rtypes. In the long run, the way such index are named should be changed. In the short term, we try to reduce the end-user confusion resulting from this design flaw: * in source/native, the regex filtering ``IntegrityError`` message does not impose an `_idx` suffix, which indeed may be absent (the result being an UI message that resembles a catastrophic failure), * also we avoid including a trailing " (double quote) from the error message * in entities/adapters, the well-named ``IUserFriendly`` adapter is made a bit smarter about how to handle missing rtypes. * the adapter also always produces a global message explaining the issue (and the fact that sometimes, the user is not shown all the relevant info) * i18n is updated Closes #2793789

[hooks/security] Streamline attributes default permission check. The current default permission on attributes delegates the check to the entity permission update policy. Since this is already checked it can be skipped. The equality comparison will work, even with a deserialized schema, because the default update perm is:: ('managers', ERQLExpression(Any X WHERE U has_update_permission X, X eid %(x)s, U eid %(u)s)) which will always be deserialized in this order (groups first). However this is a slight semantic change: entity type level 'update' permissions can now be effectively used to encode update-time rules if the default attribute permissions are used (before this change, the 'update' rules at entity type level were fired at creation time). Closes #2930861.

[test/schemaserial] swap got/expected to get nicer unittest2 diagnostics Prepares #2965518.

[repository] drop safe attribute on ``internal_cnx`` People that need to disable hook can do it explicitly anyway.

[dbapi] deprecated the dbapi

[repoapi] deprecate dbapi compat method

[testlib] deprecated the older api to access the repo.

[connection] deprecated free_cnset and set_cnxset

[session] deprecate all Connection related method on session

[session] drop dead _current_cnx_id Not used anymore for a few commit.

[session] privatise get_cnx and close_cnx The only user, repoapi now use ``new_cnx``

[doc] add documentation for the new API in test

[testlib] add an default testcase.adminaccess (and use it for default session) This adminaccess is the new offical way to get connection, and request on a repo.

[testlib] introduce a RepoAccess class to easily create connection and request Each RepoAccess hold a session for a user and three helper function to help create Connection, ClientConnection of WebRequest related to this session. related to #2920299

[repoapi] make ClientConnection.__enter__ return self This allow the standard idiom:: with repoapi.connect(repo, login='babar', passwork='elephant') as cnx: cnx.execute(…)

[documentation] describe repoapi and web side change. Short version explaining what object replace what and that BC existes for a few version.

[testlib] use internal_cnx instead of internal_session internal_session is deprecated.

[repository] add an ``internal_cnx`` method to replace ``internal_session`` Accessing the repo through a Session is deprecated. We need an easy replacement for ``internal_session``. This API change was a good occasion to stop disabling integrity hook by default. This is huge source of bug in user-code. related to #2503918

[connection] enforce that a connection must be open to be used The same than for ClientConnection, we ensure the connection is used inside a its context. .. note:: We may rely on that for ClientConnection and remove de dedicated code in Client Connection but I prefer the current explicite and duplicated version for now.

use standalone Connection to Client Connection

[session] add a new_cnx factory Having user-code importing cubicweb.server.session.Connection is inconvenient. We add a simple factory fonction on the session.

[connection] allow simple instantiation of standalone Connection Such connection will automatically pick a connection id. Note, They are not automatically closed on session close. But they will fails to grab new cnxset once the session is closed.

[connection] handle and explicitly life cycle on Connection Like ClientConnection, Connection object need to be explicitly started and stop. They aims to be used as context manager.

[sesion] distinction between Connection handled by the session and other. Not mixing the new and backward compat approach seems a good idea. Let's enforce it.

[session] replace _clear_thread_storage with close_cnx There is not good reason to keep two distinct method.

[session] explicitly take Connection object in close_cnx Now that ClientConnection explicitly reference and use the Connection object we do not need to use connectionid here. We can safely change this signature, ClientConnection is the only use of close_cnx for now.

[session] drop the Session._clear_cnx_storage method It is just cnx.clear() now.

[session] wrap too long line Too long line is too long. (Confucius 503 BC)

[session] gather close_cnx with get_cnx and set_cnx They do the same kind of operation and deserve to be grouped together.

[client-connection] remove the _srv_cnx usage It does not do anything special now that we use explicite Connection object with automatic cnx_set handling.

[client-connection] explicitly check that the client-connection is open The check is also perform by the _srv_cnx property. But we do not need those property anymore.

[client-connection] handle the lack of connection id while not open the connection id is computed at opening and forgotten when closing. We can't rely on it in various messages … like the "connection closed" exception.

[connection] transparent cnx_set handling Connection object while take cares of there cnxset themself (as dbapi connection does). The ``set_cnxset`` and ``free_cnxset`` operation are still available for backward compatibility purpose. The ``_auto_free_cnx_set`` is introduced to handle mixed usage. A new context manager ``connection.ensure_cnx_set`` is added for code that access ``cnx.cnxset`` directly and are not wrapped in any specific ``Connection`` method. A ``_with_cnx_set`` decorator is used on all Connection method that need a cnxset.

[client-connection] explicit the client part in __repr__ Now that we have real server side connection we need to remove ambiguity.

[ClientConnection] directly use the Connection object to access the database Now that Connection are a full featured standalone object we can directly reference and use it in the ClientConnection instead of using the session. The session object is kept around for a while to perform various utility role.

[connection] invert __init__ parameter Takes session first. At some point, the connection_id will become optional for Connection created explicitly.

[service] enforce that Service argument and return are json-serialisable The call_service API need to be able to run through RPC. So we ensure front start that it is possible serialise both input and output.

[connection] move call_service on Connection This is the last step toward standalone transaction.

[connection] move the commit method on Connection object One step closer of standalone Connection!

[connection] move the rollback method on Connection object One step closer of standalon Connection.

[connection] pass a Connection object to RQLRewriter RQLRewriter can now directly use a Connection object. No need for specific handling session side.

[connection] Connection.execute touch Session timestamp This allow to keep Session alive while using Connection object only for database access.

[connection] Connection now call _set_user to CWUser object linked to itseld Session.user is linked to the session.

[server/session] better faking of user for InternalManager

[connection] move execute All necessary method are now available on the Connection object. So the Connection can handle RQL execution itself.

[connection] mark Connection as "not a request" The ORM use this information. Having a wrong value will lead to crash.

[server/session] ensure appobject obtained from the session are linked to the session

[connection] have rql_rewriter accession on Connection Code expecting a session search the RQLRewriter on this attribute.

[connection] move local_perm_cache management on Connection It belongs here anyway.

[session] use a rich object for timestamp This allows the Transaction to keep a reference to it

[migraction] do not access session.data directly use set_shared_data and get_shared_data instead

[connection] move relation management method

[connection] give access to shared data method A reference to session data are kept

[connection] move relation cache method to Connection needed for execute (and need execute themself...)

[connection] give access to session execute from connection This allows multiple function that both requires an execute method and are required by the execute method.