researchv10no/cmd/dist/talk/outline - view

File: [Research Unix] / researchv10no / cmd / dist / talk / outline
Revision 1.1.1.1 (vendor branch): download - view: text, annotated - select for diffs
Tue Apr 24 17:21:35 2018 UTC (8 years, 1 month ago) by root
Branches: belllabs, MAIN
CVS tags: researchv10, HEAD

researchv10 Norman

"Automatic Software Distribution and the Tower of Bable" (really the Tower of Babble, or more accurately, the Network of Babble) ### slide: chopped down version of this outline * motivation and history * consistency across a network of machines * first publicly described for unix by Koenig (ship) * rdist (Berkeley) * coda (Rich $alz, BBN) * previous work * samples of ship, rdist, and coda ### slide: samples of ship, rdist, and coda * simple ship command, command options * rdist Distfile * coda Codafile * differences between ship rdist and Coda * ship is command-oriented * rdist and coda work from a configuration file * ship and rdist work from master machines to slave machines * coda provides a service--client machines in control * problems with previous designs * ship and rdist have too much control * ship usage is strange * rdist and coda config files are hard to use for simple things * config files do not interact well with the rest of the system * rdist and coda are monolithic * ship, rdist, and coda are network dependent * the design of dist * problems to solve * distribution on a heterogeneous network of heterogenous machines * cascaded redistribution ### slide: picture of cascaded redistribution * completing the square ### slide: square containing ship, rdist, dist, and coda * dist is command-line driven * dist provides a service machine for client machines * example of dist usage * dist operates on multiple networks * implementation ### slide: picture of distribution crossing a network * cleanly dividing along functional lines * packaging * queueing and transmission * unpackaging * packaging is decoupled from transmission * written mostly as a collection of shell programs * problems and their solutions * problem: whining about networks ### slide: sample v10 & Berkeley network code, sample machine names * network programming interfaces * incompatible naming conventions * solutions: isolate network dependencies * we want reliable byte streams but don't care how * isolate network dialers from the rest of the program * provide easily-replaced heuristics to make sense of names * problem: atomic installation * hypothetical case study of partial installation * want dist to succeed completely or fail completely * solution: install package script in the unix style ### slide: excerpt from insdist shell script * build a backup copy of stuff to be replaced * install new stuff * remove any remaining old stuff not replaced by new stuff * problem: writing bulletproof programs * bulletproof is necessary for security * funny names confuse people and are insecure * funny names confuse programs and are insecure * solution: new shell utilities * qargs and uargs * serendepitous interaction with xargs * conclusions and lessons learned * problem decomposition wins again * dist is smaller than ship * network pain is isolated and easily replaced with future upgrades * security issues are isolated and hence more easily controlled * careful use of existing tools allowed atomic installation to be coded and tested within an hour * shell programs are a tremendous benefit * they aid and in fact encourage clean problem decomposition * contrary to popular belief, can easily be made bulletproof * shell programs are very easy to make portable * future work * demonstrate cascaded redistribution (use dist to maintain itself) * focus on the notifier (isolate security/policy dependencies here)

This archive runs on limited infrastructure. Preserving old code on modern bandwidth. Automated agents are requested to crawl responsibly.