Daniel Shahaf wrote on Wed, Feb 23, 2011 at 16:02:49 +0200: > julianf...@apache.org wrote on Thu, Feb 17, 2011 at 21:20:33 -0000: > > Author: julianfoad > > Date: Thu Feb 17 21:20:33 2011 > > New Revision: 1071795 > > > > URL: http://svn.apache.org/viewvc?rev=1071795&view=rev > > Log: > > * notes/wc-ng/pristine-store > > Update with initial feedback from danielsh. > > > > Modified: > > subversion/trunk/notes/wc-ng/pristine-store > > > > Modified: subversion/trunk/notes/wc-ng/pristine-store > > URL: > > http://svn.apache.org/viewvc/subversion/trunk/notes/wc-ng/pristine-store?rev=1071795&r1=1071794&r2=1071795&view=diff > > ============================================================================== > > --- subversion/trunk/notes/wc-ng/pristine-store (original) > > +++ subversion/trunk/notes/wc-ng/pristine-store Thu Feb 17 21:20:33 2011 > > @@ -1,22 +1,36 @@ > > A. THE PRISTINE STORE > > ===================== > > > > === A-1. Introduction === > > > > +1 on your changes here, and (having read through) virtually everywhere > else in this commit. > > I made a minor clarification in r1073749, and I have one question: > > > (d) To read a pristine text, the reader must: > > - 1. Ensure no pristine-remove txn is in progress while querying and > > - opening it. > > + 1. Query the SDB and open the file within the same SDB txn (to ensure > > + that no pristine-remove txn (A-3(b)) is in progress at the same > > + time). > > 2. Ensure the pristine text remains in the store continuously from > > opening it for the duration of the read. (Perhaps by ensuring > > refcount remains >= 1 and/or by cooperating with the clean-up > > code. Under spec B, the clean-up code is controlled by rule > > B-3(c), so holding any WC lock would prevent a pristine from being > > - deleted.) > > + deleted. An alternative is to use the operating system's ability > > + to keep the file available for reading as long as the file handle > > + is open, even if the file's directory entry is removed.) > > + > > +(e) To clean up "orphan" pristine files: > > + 1. > > + > > +###? > > I guess it would be: > > (e) To clean up an "orphan" pristine file: > 0. Acquire a 'RESERVED' lock. > 1. Add a table row with refcount=0. > 2. Follow the procedure A-3(b) for removing a pristine. >
You may have noticed that I glossed over the "How to determine which pristines are orphaned" detail. Coming back to that detail, it seems that iterating all on-disk pristines is going to be expensive --- that's O(# files in wc) pristines, and I haven't mentioned the DB queries. Would it make sense to route the unlink() calls via the work queue, in order to avoid having to do that expensive scan for orphans? > And then we can also add an invariant A-2(c); > > A-2(c) Pristines will only be added to or removed from the store by an > entity that holds an SDB lock. > > Thoughts? > > (I think it's useful to document the high level "What locks are needed > for what operations" --- I was looking for such an invariant but didn't > find any in (A).)
