*** x00350071 has quit IRC | 06:08 | |
*** x00350071 has joined #openstack-smaug | 06:09 | |
*** yuval has joined #openstack-smaug | 06:51 | |
*** gampel has joined #openstack-smaug | 07:30 | |
*** c00281451_ has joined #openstack-smaug | 07:38 | |
*** c00281451 has quit IRC | 07:38 | |
openstackgerrit | Yuval Brik proposed openstack/smaug: BankSection: implement get sub-section https://review.openstack.org/311614 | 07:38 |
---|---|---|
openstackgerrit | Yuval Brik proposed openstack/smaug: BankSection: implement get sub-section https://review.openstack.org/311614 | 10:29 |
*** saggi has joined #openstack-smaug | 10:50 | |
saggi | yuval: Please review ( https://review.openstack.org/#/c/229722/17/specs/newton/replication-group.rst ) | 10:53 |
yuval | saggi: thanks | 10:54 |
*** gampel has quit IRC | 11:08 | |
*** gampel has joined #openstack-smaug | 11:08 | |
*** yuval_ has joined #openstack-smaug | 11:08 | |
*** yuval has quit IRC | 11:08 | |
*** saggi has quit IRC | 11:09 | |
*** saggi has joined #openstack-smaug | 11:10 | |
saggi | yuval: Cinder's weekly meeting is Wed at 19:00 IST | 11:14 |
*** yuval_ is now known as yuval | 11:33 | |
saggi | yuval: wrt your question about why delete marks for deletion and than deletion happens in the background instead of actually deleting. | 11:48 |
saggi | any operation can fail, especially a complex operation like deletion | 11:49 |
saggi | But the user would like it to always succeed. | 11:49 |
saggi | Also, this is to prevent collision between multiple sites. | 11:50 |
saggi | Since this doesn't require a lease. | 11:50 |
yuval | Writing to 'indice' bank section does require a lease. | 11:51 |
saggi | not for deletion. | 11:52 |
saggi | well not exactly | 11:53 |
yuval | and when the GC actually removes the checkpoint, does it remove it from the 'delete checkpoints' index? | 11:53 |
saggi | If a lease exists I fail | 11:53 |
saggi | But I don't need to acquire one | 11:53 |
yuval | (that means that delete can fail, if a lease exists) | 11:53 |
saggi | So this means that if there is no lease on a checkpoint. Anyone can just mark it for deletion. This means that after a certain amount of time I can be assured that no one is restoring from it. | 11:54 |
saggi | Yes, but immediately. Not because of some arcane storage problem. And it will never be in a half state. | 11:54 |
saggi | If it fails during actual deletion (or purge) than we the admin will have to fix it. But the checkpoint has been invalid for restore for a long time. | 11:55 |
saggi | The main issue is if we are restoring on one site and deleting on another. | 11:55 |
yuval | I see | 11:55 |
saggi | So even if you marked for deletion. The restore can succeed since no data is being deleted, | 11:56 |
saggi | You just can't start new restores. | 11:56 |
saggi | Once we know all sites know about the deletion we can start deleting | 11:56 |
saggi | There might be a collision during purge | 11:56 |
saggi | but this is OK | 11:56 |
saggi | since we can just try again later. | 11:56 |
saggi | hopefully it will converge. | 11:57 |
saggi | But it's not something that the tenant cares about. | 11:57 |
saggi | (except for his qoutas still being used). | 11:57 |
yuval | If the GC runs in a fixed interval, it might 'hit' the exact moment after someone marked a checkpoint for deletion, and start GC the data | 11:58 |
yuval | we also might have an issue with multiple protection services whose GCs might collide | 12:00 |
saggi | As I said, we don't care if multiple GCs collide. The cleanup should handle that and just be OK if someone deleted the resource for you. | 12:04 |
saggi | The GC will need to only delete entities that have been marked a certain amount of time ago. | 12:04 |
saggi | It's all a bit flaky but we are trying to avoid having cross site distributed locks so we configure the maximum time for global replication and use that. | 12:04 |
saggi | DLM over WAN is not a thing you can scale. | 12:05 |
saggi | It's also a very rare edge case so even if there are some places where it might break it | 12:07 |
saggi | it's not that bad since we are talking about something that is being deleted. | 12:07 |
saggi | You really should only delete things you know you don't need. | 12:07 |
saggi | We also need the GC since cleanup might involve multiple sites (though we are not supporting this in V1). | 12:08 |
saggi | So all sites will clean up and when someone arrive to the conclusion that everything is free actually remove the index. | 12:08 |
saggi | yuval: ^^^^ | 12:12 |
yuval | saggi: I see, thanks | 12:12 |
*** yuval has quit IRC | 12:29 | |
*** yuval_ has joined #openstack-smaug | 12:29 | |
*** gampel1 has joined #openstack-smaug | 12:29 | |
*** saggi has quit IRC | 12:29 | |
*** saggi has joined #openstack-smaug | 12:29 | |
*** gampel has quit IRC | 12:30 | |
*** yuval_ is now known as yuval | 12:34 | |
openstackgerrit | Yuval Brik proposed openstack/smaug: BankSection: implement get sub-section https://review.openstack.org/311614 | 14:26 |
openstackgerrit | Yuval Brik proposed openstack/smaug: BankSection: implement get sub-section https://review.openstack.org/311614 | 14:34 |
*** yuval has quit IRC | 14:55 | |
*** chenying has quit IRC | 15:29 | |
*** chenying has joined #openstack-smaug | 15:30 | |
*** gampel has joined #openstack-smaug | 15:51 | |
*** zhonghua-lee has quit IRC | 16:06 | |
*** zhonghua-lee has joined #openstack-smaug | 16:07 | |
*** saggi has quit IRC | 16:09 | |
*** gampel has quit IRC | 17:54 | |
*** saggi has joined #openstack-smaug | 19:25 | |
*** saggi has quit IRC | 21:11 | |
*** gampel1 has quit IRC | 23:25 | |
*** gampel has joined #openstack-smaug | 23:39 |
Generated by irclog2html.py 2.14.0 by Marius Gedminas - find it at mg.pov.lt!