From: Charles Lindsey (chl@clw.cs.man.ac.uk)
Date: Fri Apr 28 2000 - 13:12:31 CDT
A while back we discussed possible implementations of Replaces, and
especially of the 'version number' feature. I don't think we decided
to change anything, but we did get a better understanding of how to
implement it, particularly in INN, where it turns out to be rather
easier than in CNEWS.
I have now updated the relevant text that describes implementation. Note
that 90% of this is pseudo-comment, intended for us rather than for the
final document.
-------------------------------------------------------------------------------
[Here is the implementation technique that we discussed, based on the
use of a conventional History file. This is a sanity check for our own
use, not intended to go in the final text. There are two cases to
consider:
A. Traditional implementations (e.g. CNEWS) where each History file line
includes a full message-identifier plus an item for each group in which
the article appears. Thus History file entries are of variable length,
and it is impractical to update them in situ.
B. History files made up of fixed length records (e.g. INN), which
enables entries to be overwritten in situ. The History line typically
contains a hash of the message identifier plus some pointer to an object
representing the article as stored.
We consider the traditional case first:
1A. Ensure that the implementation of DBZ is not upset if the same key
is attempted to be stored a second time, and that such a key always
retrieves the latest record indexed by that key.
2A. Additions to the History file are always made at the end. Removals
or changes to existing entries are only made by the expire program. An
entry for a Replaced (or otherwise cancelled) article will remain until,
first, the expire program removes the links to the articles that are no
longer stored, and later on removes the entire entry according to its
expiry date. For every entry containing a '$v=n' followed by random-
dollars-sequences there will be an immediately following entry identical
but for the omission of that '$v=n' and of the random-dollars-sequences.
Thus there may be several entries with identical message-ids but,
because of the change to DBZ just described, only the most recent will
ever be seen except by programs that access the History file directly,
rather than by its index.
3A. When an article is Replaced, at the same time as the successor
article is entered into the History file, with '$v=7' say, a duplicate
entry (same article list) is entered under the same key, modified by
removing any leftmost '$v=n' and the following random-dollars-sequences
from it.
For the fixed length implementations, these steps become:
1B. DBZ does not need to be changed.
2B. History file entries may be updated in situ. An entry for a Replaced
(or otherwise cancelled) article can be overwritten with that for the
new article (or with a suitable indication of cancellation). For every
entry containing a '$v=n' followed by random-dollars-sequences there
will always exist a second entry identical but for the omission of that
'$v=n' and of the random-dollars-sequences, both entries pointing to the
same article object.
3B. When an article is Replaced, at the same time as the successor
article is entered into the History file, with '$v=7' say, the existing
entry without the leftmost '$v=n' and the following random-dollars-
sequences is overwritten (with the new article and new expiry date,
after destroying the old article, of course). If no such entry exists,
one is created.
From here on, the two cases are the same:
4. Provide a call to a routine which, if asked to retrieve any message
identifier with '$v=n' and finding it missing (or rather linked to no
stored groups), immediately tries again without the '$v=n' and its
random-dollars-sequences. NOTE. We don't want this behaviour when
checking whether we already have an article offered to us by IHAVE, only
in response to an ARTICLE command. So this needs to be an extra call in
DBZ, in addition to the 'fetch' or 'dbzfetch' calls, to be used in the
proposed extension to the NNTP ARTICLE command. Observe that if the
requested '$v=n' is present and linked to stored articles (for whatever
reason) then you will be given exactly that version, even if later ones
are stored as well.
5. NOTE that I have dropped the idea of having '$v=0', because you can
never be sure that the very first issue of the FAQ used it, so you have
to provide the versionless root as well. If someone asks for '$v=0' (or
any '$v=n') t! algorithm I gave will still find it via the root. So we
don't care what people put in URLs.
6. You are supposed to cancel the replaced/superseded article. If you
REALLY want to keep the old ones around a little longer, then this
implementation will not work if you want the latest to be retrieved
automatically - you will have to invent something much more complicated.
7. Having said all that, here follows a brief account of the same thing,
but short enough to be included in our document (the convention being
that implementation issues are hinted at, rather than being described in
full detail).]
Typically, a news database will index a Replacement article both by
its "version-number" message identifier (containing a "$v=" tag
followed by a random-dollars-sequence) and by its "root" version
(without the "$v=" tag or any following random-dollars-sequence).
Thus when a request for an article comes in that is not present under
the version-number requested, any article that is present and indexed
by the corresponding root version can be retrieved instead. The
indexing mechanism needs to be such that, although the root version
may have at times referred to many different articles, it is always
the current one that is retrieved.
NOTE: The presence of a version-number in the message identifier
of an article! Replaces or Supersedes header causes no extra
action (it is just an ordinary article). Observe also that if an
article with the exact message identifier (e! it contains a
version-number) is, for whatever reason, already present on the
serving agent, that article will always be retrieved in
preference to the one indexed by any root version.
Charles H. Lindsey ---------At Home, doing my own thing------------------------
Email: chl@clw.cs.man.ac.uk Web: http://www.cs.man.ac.uk/~chl
Voice/Fax: +44 161 437 4506 Snail: 5 Clerewood Ave, CHEADLE, SK8 3JU, U.K.
PGP: 2C15F1A9 Fingerprint: 73 6D C2 51 93 A0 01 E7 65 E8 64 7E 14 A4 AB A5