Figure out how portal_transforms handles equally short transform paths and policies

[lukasgraf]

The portal_transforms tool is largely undocumented.

Currently I'm not 100% clear about how decides which transform to pick if more than one transform with a direct path from INPUT to OUTPUT exists, and how transform policies factor into pathfinding.

Figure out how all this works, and modfiy our installation and uninstallation functions in setuphandlers.py accordingly.

[lukasgraf]

• Only one transform policy for a given ouput type can be installed at a time. See Products.PortalTransforms.TransformEngine:532

When trying to install a policy for an output type that already has a policy, a TransformException is raised.

[lukasgraf]

• One-hop transforms (aka filters) from the same INPUT to OUTPUT type are effectively disabled:

Products.PortalTransforms.TransformEngine.TransformTool.convertTo:139

        ## fastpath
        # If orig_mt and target_mt are the same, we only allow
        # a one-hop transform, a.k.a. filter.
        # XXX disabled filtering for now
        if orig_mt == str(target_mt):
            data.setData(orig)
            md = data.getMetadata()
            md['mimetype'] = str(orig_mt)
            if object is not None:
                cache.setCache(str(target_mimetype), data)
            return data

[lukasgraf]

• If a transform policy for OUTPUT_TYPE exists, the transform engine first searches only for paths that contain all of the transforms listes in that policy. Only of no paths, even long-winded ones, containing all those transforms are found, the engine falls back to searching for the shortest path ignoring the policy.

This is done by first looking for a subpath leading from INPUT to INPUT_OF_FIRST_TRANSFORM_IN_CHAIN. If such a path is found, TransformTool._findPath() is called recursively with the remaining transforms from the chain as requirements, working its way trough the chain.

So to summarize: If a policy exists, and one or more paths exists to satisfy the policy, the shortest path satisfying that policy is chosen. Choosing the shortest path from the candidates is done using the same logic that's used when no policies exist.

[lukasgraf]

• The default logic for finding the shortest path uses a Breadth-first search. Ordering of the tree depends on the order of the internal structure TransformTool._mtmap. Which is a PersistentMapping (dictionary), so its order is undefined.

=> Without a policy, which of two equally short paths will be selected is undetermined.

[lukasgraf]

• In a fresh Plone 4.3 install, there's no policy for text/plain installed.

[lukasgraf]

Conclusion:

• We need a policy to ensure ftw.tika's transform is chosen with the highest priority
• We don't have to expect a policy for text/plain to be installed on a fresh Plone
• We therefore don't need to remove any existing policies, but should instead fail (with a decent exception message) if there already is one present anyway, and let the integrator handle it (uninstall the other policy first).
• We should uninstall earlier versions (from previous installs) of our own policy to guarantee repeated installs of the default profile don't fail.