Simplest approach seems to have done the trick: I've removed all output datasets from the big selectDimensions query, since they didn't constrain it anyway, and we were already not relying on that query returning dataset_ids for them.

After that, some profiling revealed that the slow overall speed of preflight (12 minutes for ci_hsc!) was not being spent in that query - it was in slow, mostly Python data ID manipulations afterwards.  All but the first commit on this branch are optimizations for that.  It's down to 5 minutes now, and while I have ideas on how to go further, I think they're out of scope for this ticket.

Andy Salnikov, could you look at the first commit?

Nate Lust, could you look at the rest?

PR is: https://github.com/lsst/daf_butler/pull/124

Glad you caught that, Andy Salnikov.  I'll look into updating GraphBuilder accordingly and make sure I test that functionality before I merge.  I like the idea of testing for dataset existence late because it should work better for use cases that involve saving the QuantumGraph and reading it back later (not that I expect the database to change frequently in between, but it would be good to at least handle any unrecoverable inconsistencies gracefully).

Andy Salnikov, I believe I've addressed your review comments and cleaned up the mess from the flurry of activity last week; please take another look.  In particular:

• The daf_butler changes are now limited to just the removal of output datasets from the big selectDimensions query, which is now optional and configurable.  Even more importantly, I've updated _convertResultRows to query for output {{dataset_id}}s at that stage when it isn't done in the big query, so now the overall behavior is the same and the configuration option to defer the lookups can be seen as just an alternate implementation.
• This also made it easy to ensure that the DatasetRefs returned by selectDimensions are always expanded to include components when they represent composites (for both implementation code paths).  I've consequently removed the workaround that previously did this in GraphBuilder in pipe_base instead.
• The data ID optimization commits in daf_butler that were previously on this branch have been moved to DM-17611, and merged from there to master.  Those also required a change in skymap, which has also been merged to master - GitHub and Jira  present a slightly confusing view of what's happened there; but you can completely ignore the apparently-merged skymap PR for this ticket.

I have approved both PRs, looks good to me, with few minor comments. Indeed new logic looks more straightforward, though performance penalty can be significant but that can be improved later.

Nate Lust, please mark as reviewed once you are satisfied.