2017-07-12 15:23 GMT+00:00 Aaron Bassett <aaron.bass...@nantomics.com>: > I have a situation where a client is GET'ing a large key (100GB) from RadosGW > and just reading the first few bytes to determine if it's a gzip file or not, > and then just moving on without closing the connection. I'm RadosGW then goes > on to read the rest of the object out of the cluster, while sending nothing > to the client as it's no longer listening. When this client does this to many > objects in quick succession, it essentially creates a DOS on my cluster as > all my rgws are reading out of the cluster as fast as they can but not > sending the data anywhere. This is on an up to date Jewel cluster, using > civetweb for the web server. > > I just wanted to reach out and see if anyone else has seen this before I dig > in more and try to find more details about where the problem may lay.
I would say your client is broken, if it is only interested in a range of the object, it should include a corresponding range header with the GET request. Though I agree that the behaviour for closed connections could probably improved, too. See http://tracker.ceph.com/issues/20166 for a similar issue, something like the opposite of your case. _______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com