* Deal with clock skew, both forwards and backwards, when logging
information to the git-annex branch.
* GIT_ANNEX_VECTOR_CLOCK can now be set to a fixed value (eg 1)
rather than needing to be advanced each time a new change is made.
* Misuse of GIT_ANNEX_VECTOR_CLOCK will no longer confuse git-annex.
When changing a file in the git-annex branch, the vector clock to use is now
determined by first looking at the current time (or GIT_ANNEX_VECTOR_CLOCK
when set), and comparing it to the newest vector clock already in use in
that file. If a newer time stamp was already in use, advance it forward by
a second instead.
When the clock is set to a time in the past, this avoids logging with
an old timestamp, which would risk that log line later being ignored in favor
of "newer" line that is really not newer.
When a log entry has been made with a clock that was set far ahead in the
future, this avoids newer information being logged with an older timestamp
and so being ignored in favor of that future-timestamped information.
Once all clocks get fixed, this will result in the vector clocks being
incremented, until finally enough time has passed that time gets back ahead
of the vector clock value, and then it will return to usual operation.
(This latter situation is not ideal, but it seems the best that can be done.
The issue with it is, since all writers will be incrementing the last
vector clock they saw, there's no way to tell when one writer made a write
significantly later in time than another, so the earlier write might
arbitrarily be picked when merging. This problem is why git-annex uses
timestamps in the first place, rather than pure vector clocks.)
Advancing forward by 1 second is somewhat arbitrary. setDead
advances a timestamp by just 1 picosecond, and the vector clock could
too. But then it would interfere with setDead, which wants to be
overrulled by any change. So it could use 2 picoseconds or something,
but that seems weird. It could just as well advance it forward by a
minute or whatever, but then it would be harder for real time to catch
up with the vector clock when forward clock slew had happened.
A complication is that many log files contain several different peices of
information, and it may be best to only use vector clocks for the same peice
of information. For example, a key's location log file contains
InfoPresent/InfoMissing for each UUID, and it only looks at the vector
clocks for the UUID that is being changed, and not other UUIDs.
Although exactly where the dividing line is can be hard to determine.
Consider metadata logs, where a field "tag" can have multiple values set
at different times. Should it advance forward past the last tag?
Probably. What about when a different field is set, should it look at
the clocks of other fields? Perhaps not, but currently it does, and
this does not seems like it will cause any problems.
Another one I'm not entirely sure about is the export log, which is
keyed by (fromuuid, touuid). So if multiple repos are exporting to the
same remote, different vector clocks can be used for that remote.
It looks like that's probably ok, because it does not try to determine
what order things occurred when there was an export conflict.
Sponsored-by: Jochen Bartl on Patreon
git-annex get when run as the first git-annex command in a new repo did not
populate unlocked files. (Reversion in version 8.20210621)
I am not entirely happy with this, because I don't understand how
428c91606b caused the problem in the first
place, and I don't fully understand how skipping calling scanAnnexedFiles
during autoinit avoids the problem.
Kept the explicit call to scanAnnexedFiles during git-annex init,
so that when reconcileStaged is expensive, it can be made to run then,
rather than at some later point when the information is needed.
Sponsored-by: Brock Spratlen on Patreon
I'm now reasonably sure I've identified both cases where this can
happen. v8 upgrades and certian filesystems eg NFS. Both are handled as
well as can be, though it may involve some extra checksumming work.
Eg, showImprecise 1 1.99 returned "1.1" rather than "2". The 9 rounded
upward to 10, and that was wrongly used as the decimal, rather than
carrying the 1.
Sponsored-by: Jack Hill on Patreon
An easy way to see this in action is to have an unlocked file, and touch the
object file.
While all code that compares inode caches for object files needs to be
prepared for this kind of problem and fall back to verification, having
fsck notice it and correct it is cheap (as long as fsck is being run
anyway) and ensures that if it happens for some unusual reason, there's a
way for the user to notice that it's happening.
Not that, when annex.thin is in use, the earlier call to isUnmodified
(and also potentially earlier calls to inAnnex in eg, verifyLocationLog)
will fix up the same problem silently. That might prevent the warning
being displayed, although probably it still will be, because the
Database.Keys write of the InodeCache will be queued but will not have
happened yet. I can't see a way to improve this, but it's not great.
Sponsored-by: Dartmouth College's Datalad project
Fix bug that caused some transfers to incorrectly fail with "content
changed while it was being sent", when the content was not changed.
While I don't know how to reproduce the problem that several people
reported, it is presumably due to the inode cache somehow being stale.
So check isUnmodified', and if it's not modified, include the file's
current inode cache in the set to accept, when checking for modification
after the transfer.
That seems like the right thing to do for another reason: The failure
says the file changed while it was being sent, but if the object file was
changed before the transfer started, that's wrong. So it needs to check
before allowing the transfer at all if the file is modified.
(Other calls to sameInodeCache or elemInodeCaches, when operating on inode
caches from the database, could also be problimatic if the inode cache is
somehow getting stale. This does not address such problems.)
Sponsored-by: Dartmouth College's Datalad project
