Branko Čibej wrote on Mon, Feb 13, 2012 at 17:52:01 +0100: > On 13.02.2012 17:09, Daniel Shahaf wrote: > > Thanks for your thoughts. One comment: > > > > Branko Čibej wrote on Sun, Feb 12, 2012 at 21:52:16 +0100: > >> The idea is that we'd always maintain the complete index, i.e., in order > >> to determine if path@15 exists, one only needs to search the index for > >> (path, rev <= 15, !deleted) -- which is trivial in a properly ordered > >> index. (Yes, this assumes that we record deletions in the index as well > >> as insertions.) > >> > >> All of this can be done with an append-only index representation. What > >> you can't do with append-only is represent forward history links, but -- > >> we're not representing them very well right now, are we. :) > > Within the append-only constraint one can implement forward links if one > > has, say, a separate file per revision. (So asking "where was foo@15 > > copied to" involves a linear scan of $REPOS/db/forward-links/15.) > > > > stsp and I even started on such a design on some branch somewhere --- > > but that branch has been abandoned as the motivation for it was rename > > tracking, which stsp figured he could solve better without the new FS > > feature. > > A separate file per revision has other problems, too. You'd quickly run > out of inodes on ext-derived file systems, for example. >
Yeah, I was in theory mode in the previous mail. On the branch we didn't ignore those "implementation considerations" and went for 3 files per shard of 4k revisions. https://svn.apache.org/repos/asf/subversion/branches/fs-successor-ids/BRANCH-README > -- Brane
