* Fix bug that could make git-annex importfeed not see recently recorded
state when configured with annex.alwayscommit=false.
* importfeed: Made "checking known urls" phase run 12 times faster.
The massive speedup is because it no longer queries for metadata
accompanying each url. Instead it processes the whole git-annex branch and
checks all metadata files for feed item ids, and uses any it finds.
This could result in a behavior change, in an unlikely situation: If a feed
id is recorded in a key's metadata, but the url gets removed, the old code
would not see that item id and would re-download it if it finds an url for
it in a feed, while the new code will see the item id. I don't think
the old behavior was intentional, and it may be that the new behavior is
better. Not gonna worry about this.
This only partly fixes importfeed to see journalled files, since it
separately cats metadata directly from the branch. Held off on a
changelog for a bug fix until that's dealt with.
Fix bug caused by recent optimisations that could make git-annex not see
recently recorded status information when configured with
annex.alwayscommit=false.
This does mean that --all can end up processing the same key more than once,
but before the optimisations that introduced this bug, it used to also behave
that way. So I didn't try to fix that; it's an edge case and anyway git-annex
behaves well when run on the same key repeatedly.
I am not too happy with the use of a MVar to buffer the list of files in the
journal. I guess it doesn't defeat lazy streaming of the list, if that
list is actually generated lazily, and anyway the size of the journal is
normally capped and small, so if configs are changed to make it huge and
this code path fire, git-annex using enough memory to buffer it all is not a
large problem.
When downloading content from a remote, if the content is able to be
verified during the transfer, skip checksumming it a second time.
Note that in this case, the fsck output does not include "(checksum)"
which it does when the checksumming is done separately from the download.
This commit was sponsored by Brock Spratlen on Patreon.