Noticed that publishing a dataset with a handle takes some minutes in production. (Note that this is after the patch for skipping file registration has been applied; so there is just one dataset-level handle, that needs to be looked up and possibly re-registered during the publication process).
It appears that there are timeouts when Dataverse tries to communicate to the central handlenet service:

[edu.harvard.iq.dataverse.HandlenetServiceBean] 
  Caught exception trying to process lookup request
HandleException (CANNOT_CONNECT_TO_SERVER) 132.151.1.179: java.io.InterruptedIOException: Connection timed out.
        at net.handle.hdllib.HandleResolver.sendHttpRequest(HandleResolver.java:1959)
        ...
        at edu.harvard.iq.dataverse.HandlenetServiceBean.isHandleRegistered(HandlenetServiceBean.java:197)
        ... etc.

(the dataset eventually gets published).

My best guess is that this is because Dataverse is built with an ancient version of the handlenet library (the ip addresses of the central servers are somewhere in that jar).

        <dependency>
            <groupId>net.handle</groupId>
            <artifactId>handle</artifactId>
            <version>2006-06-16-generated</version>
        </dependency>

(that's 2006 in the date, yes)

So we may need to upgrade it.

(error handling inside the handlenet service bean may be problematic too; for example, it looks like it can't differentiate between the lookup connection timing out, and the handle not being registered...)