Frank Ludolph wrote: > Dave Miner wrote: >> Sarah Jelinek wrote: >>> Hi Sue, >>> >>> Some comments on disk selection proposals: >>>> Proposed Functional Design for Default Disk Selection >>>> ----------------------------------------------------- >>>> One disk on x86 system >>>> a) Look for solaris2 partition to use. If found, use it and create vtoc >>>> b) If partitions exist, but no solaris2 partition exists, but there >>>> is space for creating solaris2 >>>> partition, create it and use it. >>>> c1) If partitions exist, but no solaris2 partition exists, and no >>>> space exists to create one, ask user >>>> if they want to use whole disk or use a partition on disk (list the >>>> partitions). If partition is chosen, >>>> use it and make solaris 2 partition. Or user could quit and do >>>> fdisk/format. >>>> >>> >>> How exactly would we ask the user? I am a bit uncomfortable with this >>> approach since it is supposed to be a hands off install. Which means >>> we automatically succeed or we automatically fail. I would think the >>> text based installer would be used, eventually, in the event a user >>> wants an interactive install experience. >>> >>> Baking in behavior that requires user input in some scenarios seems >>> contrary to what we want to provide with AI. >>> >> >> More directly: it doesn't meet the requirements. Automated >> installations must succeed (or fail) entirely without interactive >> input. Is there any way in which this requirement is unclear? >> > The requirement is clear but the impact of the hands-off failure on the > user experience is significant. Because an install will overwrite data, > the disk/partition/slice defaults should be pretty tight which in turn > raises the likelihood of failure. > > Once an install fails it is no longer hands-off. The user must take > corrective action, in this case either reformatting a disk to match the > disk selection criteria, or more likely editing a manifest and creating > a new service. The latter is the right thing to do for a production > environment (though time might be an issue in some instances), but it > places a high hurdle for first time users evaluating UI using the > default manifest. Asking the user to interactively designate a disk and > possibly a partition/slice in the event of disk selection failure, > rather than simply failing outright, can greatly simplify the initial > out-of-box experience. >
I think this simplification only really applies in narrow cases. If we're providing for an expansive set of disk/partition/slice selection capabilities, which I think is required, then I suspect the resulting user interface would be fairly complex. Otherwise you'd seemingly have to differentiate some disk selection failures as being more important than others. Secondarily, there's the issue of auditability of the resulting installation; if it's not as expected and expressed in the AI configuration, then what do we do to close the feedback loop to make it correct for the future? > I'm not suggesting that AI should initiate user interaction in other > failure modes. Disk selection happens very early in the install process > when the user might still be observing that the installation has > properly started and is proceeding properly - there are progress > messages being sent the the target machine's console. > I'm skeptical whether this is likely to be successful; the time required to boot and initiate the installation gets into multiple minutes in most cases, and user attention spans aren't that great at that scale. Beyond that, many installations are likely to be initiated via remote agents without the console being connected at all. Which raises an earlier issue I'd brought up about reporting results back to the servers that I don't think has been discussed much. > To Sarah's question of where the interaction would be done, the answer > is the target machine's console where the progress messages are being > printed. If there are concerns about visibility of the console the > interaction could time out and fail the install after several (10?) > minutes. > Ultimately, while I'm willing to explore the possibilities here and perhaps allow for optional selection of an interactive fallback, the design needs to reflect the requirement to succeed (or fail) in any scenario predictably and automatically. Dave
