On Tue, Jul 29, 2014 at 12:55:11AM +0200, Andrew Phillips wrote: > A recent commit [1] to Glacier (thanks, Roman!) indicates that the > deleteContainer operation can take several *days*. > > How likely is this to happen? What would be the expected average and > e.g. 90th percentile time, if we have such kind of information? > > If this is not an *extremely* unlikely occurrence, does it even make > sense to support this operation? Or should we be throwing something > like an UnsupportedOpEx instead, with some description how the > desired goal could be otherwise achieved? > > Having jclouds calls that could take days to return seems pretty > scary to me... > > Thoughts?
Many Glacier calls such as list can take several hours; it is not a traditional object store and imposes some painful semantics to achieve its cost savings. Few users will be able to take an existing jclouds application written against BlobStore expectations and immediately use Glacier; rather, new applications written against Glacier semantics will be portable to traditional object stores. FWIW, Maginatics have tested deleting containers with very large number of blobs (100 million - 1 billion) and this operation also takes days to complete. -- Andrew Gaul http://gaul.org/
