A response to an initial /sync can easily run to hundreds of megabytes of JSON, which takes tens of seconds of CPU time to serialise from the objects. This monopolises the Twisted reactor thread, causing all sorts of other problems (including dropped replication connections for worker-mode deployments).

To make matters worse, although we have a cache which is supposed to avoid redoing a lot of the hard work for duplicate requests to initialsync (though its effectiveness is disputed, see #3880), we cache the de-serialised version, so have to re-serialise the response for each request. (One reason we have to do this is to populate the age field).