I see errors in the logs like:
LoadBalancer::commitAll 10.64.48.26 2006 MySQL server has gone away (10.64.48.26) COMMIT
Number of requested IDs (79370) is too high
The later error corresponds to a large CDN purge call (with ~80k urls). I hope that doesn't cause UDP flood once the UID limit is removed (wmf10).
Maybe the job run() method could be split up? Even internally, by batching purges with short sleep in between calling commitAll() between long running methods if needed. Alternatively, it could subdivide...but I'm not too familiar with exactly what it does to know if that would work.