Prevent CDN cache inconsistency

Use the CDN to publish the page if the backend is down

This will allow the CDN to return content if backend is down

This reverts commit b3ff3ec4.

Fetch usage can be bypassed to do not use service worker when not needed.

As appcache has been dropped on Firefox, this change will improve the speed on Firefox.

No change is expected on Chrome/Safari.

Allow edition in the new UI

This make everything slow as hell and prevent to quickly save.

Remove all globals to allow multiple live tests running on the same page.
Stop reading the test output if the browser tab doesn't have the focus.
Stop hardcoding the URL in the javascript.
Stop hardcoding other fields path in the javascript (+ embed the textarea in the gadget).

Propagate all parameters via formulator.
It will allow to use this gadget from another context document.

Run live test from the RJS UI

The pagination was not working when using the limit parameter, and when
the total number of documents is equal to the pagination limit + 1.

Before:
   Float Equivalence Tester does not test the divergence even if Divergence Provider is ON
    
After:
     Float Equivalence Tester test the divergence with the epsilon-span if Divergence Provider is ON
    
Before, we needed the following condition to test the divergence at Float Equivalence Tester:
    
       "Divergence Provider is ON"  AND
       ("Absolute Tolerance has some value"  OR
        "Relative Tolerance has some value" )
    
However this default behavior was not consistent with other testers, and
difficult to notice from the user interface.
Thus test the divergence by default with the epsilon span which is defined in 
FloatEquivalenceTester class.
    
The epsilon-span is the same amount that when we put 0.0 into the ranges.
The 0.0 has been the configuration of erp5_configurator_standard_trade_template.

/reviewed-on nexedi/erp5!865

When an object is being indexed, it get wrapped so it has magic properties,
like security definitions as they will be accessed by indexation methods.
One such property is used to trigger the insertion of rows in
roles_and_users table. That property has the special feature that it is
only produced on the first object needing a new set of rows added to
roles_and_users, other objects will get a None value instead for that
property so that no further lines are inserted.
Then, SQCatalog iterates over that wrapper list to compute all indexation
method parameters. It does so via a cache (LazyIndexationParameterList) so
that unused parameters are not computed. The cache key is wrapped object's
uid.
All this together means that if an object is indexed multiple times in a
single call (which happens at site creation, for example) and it needs a
new set of rows added in roles_and_users, only one of produced wrappers
will get the appropriate value. If that object is not the first on which
that parameter is evaluated (which gets more likely with every copy of the
same object being indexed), it will be cached as None and shared between
all copies, causing no row to be added to roles_and_users.

Full investigation done by by Yusei Tahara.
wrappers will get the

As we may reach database limits, and to avoid them to be logged.

For example, if property_sheet "prop" was moved from BT 1 to BT 2 and BT 1 doesn't exist anymore.

Before this commit we have this order:
1.  install BT 2 ("prop" exist and isn't touched)
2. remove BT 1 ("prop" is removed)
=> at the end "prop" doesn't exist anymore

After this commit we have this order:
1. remove BT 1 ("prop" is removed)
2. install BT 2 ("prop" is readded)
=> at the end "prop" exists and is OK

/reviewed-on nexedi/erp5!866

old key was simply "id" so when a container is a VirtualFolder (like Web
Site), container._getOb("id") can return any object with id "id"
including an object outside the container.
    
Change the key to contain PhysicalPath of the container to get only
objects inside the container.
    
This actually fixes installation of some business templates with Web
Site inside.

/reviewed-on nexedi/erp5!867

1.auto centre for image
2.table line is centered
3. image option is left-margin 3cm under image

Not all lock uses in this code fit context managers, but use them where it
does.

Add 10 seconds timeouts on all blocking calls in this test method, which
should be a thousand times more than required on a slow machine but avoids
completely blocking tests.

Even though it should not raise, a stale lock would be a serious issue.

The intent is to be able to tell that an independently-defined group of
activity nodes may execute given activity, and no other node.
This allows more flexible parallelism control than serialization_tag.

Divide conflict hot-spot to improve write performance: now, conflict risk
will not be proportional to the number of zope processes, but only to the
number of threads within considered process (...because there is no
stable thread name, if there is one day and conflict splitting is needed,
it will be easy to implement: just concatenate that name to node name in
FragmentedLength._map).
Automatically migrate to FragmentedCount on containers larger than 1k
objects (this threshold may evolve).

Provide a way to generate ids on specific documents when they make sense.
Reduces the risk of conflicts that portal_ids's in-ZODB id generator
creates.
Reset generator state on any copy (both copy/pase and clone), but not on
move (cut/paste and setId).