Before NGS and for about a week after NGS, replication packets were very reliably produced within the first couple of minutes past the hour. Occasionally they would be delayed (I suspect by the full dumps), but replicating at :02 worked most of the time and replicating any time after :05 worked nearly all of the time.
About a week after NGS, in late May 2011, things started to get later and less consistent... The averages remained under :05 until about January 2012 and under :10 until about November 2012 although less reliably because there was a lot more variation. In the last couple of months, the average time has more than doubled, with none of the days in the last week having an average of less than :20.
When packets will be produced is now so unreliable that I've switched to running the replication script every five minutes, so has Ian. AFAIK Lukáš has been doing that for a while already. Since the live data feed is one of the things people pay us for, it seems like this is problematic and it should be behaving more consistently than this.
(attaching two scattergraphs of data extracted from the timestamps of files in /pub/musicbrainz/data.classic/replication/ and /pub/musicbrainz/data/replication/, one is of all replication times, the other is of the daily average of times)