andy@cayuga.Stanford.EDU (Andy Freeman) (07/26/88)
In article <7397@cit-vax.Caltech.Edu> mangler@cit-vax.Caltech.Edu (Don Speck) writes: >NNTP promotes consolidation, not the decentralization that is Usenet's >hallmark. Unlike mailing lists, which can be received by any machine, >NNTP servers are one per department, with the trend being toward one >per campus. This extrapolates to one per regional network. NNTP could reduce the costs of Usenet's decentralization without adding the vulnerabilities associated with centralization. I'll come back to that after a quibble. > Do you >see much rrn'ing across BARRNET, yet? You will... and caching slave >servers will hasten the day. I use a caching version of rrn; the NNTP protocol doesn't usefully support caching slave servers because they can't be used when the master is unavailable. (NNTP doesn't let the slave server tell its clients "I don't have that article, but I can't say that it doesn't exist, so ask again later if you're still interested"; it can only say "That article/group is unavailable.") >Centralization brings us back to having critical failure points, >where loss of a single machine knocks out a department, a campus, >perhaps a regional net. I'm not talking about machine failures; >I'm talking about *administrative* shutdowns. NNTP with caching slave servers support reduces the cost of shutdowns, no matter what the cause, because it allows users/clients to change servers without "fixing" their .newsrc. (.newsrcs use article numbers whose mapping to message-ids is site dependent. NNTP with caching slave servers makes it possible for a number of computers to maintain the same mapping, that is, to behave as a single news site.) Caching slave servers also introduce substantial efficiencies, and it may be reasonable (as in, cheaper for all, and more reliable to boot) for all of the BARRNET sites, for example, to use this so that they look like a single NNTP site (as far as the rest of the world is concerned). Individual sites can always restrict their users as they see fit. Suppose an organization has a number of full-feed sites and some of them are NNTP servers for other sites. Furthermore, let's assume that some of these clients don't have the resources to take a full feed, but their users demand it, so they're clients. (This describes both the "organization" consisting BARRNET sites, as well as "real" multi-site organizations such as Stanford and Berkeley.) If NNTP supported caching servers, this could be reorganized so that large hunks of it fall apart/disappear without hurting the rest of it. This reorganization would also make it possible to better manage the loads seen by the various sites. The reorganization is as follows. One of the full-feed sites is designated the global master and all postings, either from the outside or from internal sites, come through it; the other full-feeds become local masters which copy the global master's /usr/spool/news. (This copying can be done any way that is convenient.) One local master can "feed" another - there's no need for them all of them to talk with the global master. Local masters can be added or deleted at any time. With this arrangement, any user/client can use any server, and can switch between them at will, subject to their new server's NNTP permissions file. Moreover, any client can also cache (either certain groups, or any group, but a limit on the total space used) and be a server for other machines. If the global master goes away, any one of the local masters can take over as the new global master. With a little bit of thought, there's no need for all of the local masters to keep a complete copy of the global master's state - they just have to be able to recreate it COLLECTIVELY with whatever amount of redundancy that is desired. (One could go even further and not keep the entire feed on the global master, but then the nntp servers would have to look on different machines for different groups instead of always looking one level up on the hierarchy. Of course, this complication could be restricted to the local masters, but ....) Of course, there's also a way to have newsgroups that aren't seen outside a subgraph of this whole mess. Notice that I didn't say how the outside talks to the global master; it can use NNTP, phones, or even magtapes. The global master isn't any more of a bottleneck than existing full-feed sites because it just carries all of the outside traffic and internal postings that a normal full feed site sees. The load seen by servers within the organization can also be managed as necessary. -andy UUCP: {arpa gateways, decwrl, uunet, rutgers}!polya.stanford.edu!andy ARPA: andy@polya.stanford.edu (415) 329-1718/723-3088 home/cubicle