search

mancdaz / banana

test issue import
0 stars 0 forks source link

Glance image replication, or other such mechanisms #20

Open mancdaz opened 10 years ago

mancdaz commented 10 years ago

Issue by byronmccollum Thursday Aug 28, 2014 at 03:20 GMT Originally opened as https://github.com/rcbops/ansible-lxc-rpc-orig/issues/484

Glance image replicator redux...

Registering an image lands the bits on only one infra node. Spawning an instance from that image could cause nova scheduler to retry multiple times until the image fetch call goes to the correct infra node containing the image. If scheduler retries >= number of infra nodes, this should eventually success (after unnecessary delay, and artificially induced lumpy compute distribution), but with any amount of load, it's quite possible all scheduler retries go to infra nodes without the image.

So, are there plans to reintroduce glance image replicator, or other such mechanisms. Or is the preferred / recommended configuration to use Swift if you have more that one infra node?

mancdaz commented 10 years ago

Comment by claco Thursday Aug 28, 2014 at 03:29 GMT

In 9.x, the minimum number of infra nodes is 3, which as you've pointed out, re-introduces the glance scheduler/image retry scenario. Rather than reintroduce the replicator, which had it's own issues, we're opting for shared storage as the official recommended way forward. This is either Swift, CloudFiles, or an NFS share (NetApp).

mancdaz commented 10 years ago

Comment by byronmccollum Thursday Aug 28, 2014 at 03:33 GMT

Should something be added in the docs regarding this?