9.6k link to arpa link, only 2 ports for everybody to share, $5k for 1 tty port
Observed people just wanted email. Solution forward messages between networks
design principles:Only 1 programmer, don’t redesign user agents, don’t change local mail store, make delivermail adapt to the world
Problems: compiled in config, no address translate,
Transittion to sendmail – UCB awarded contract , adding SMTP support, needed to be build in, forced introduction of queue, queues harder than look
released 1982 with 4.1a BSD – SMTP, Queueing, runtime config
sendmail picked up my most unix vendors. Unix wars raged on – vendors added value and incompatible added up.
Returned to Berkley, sendmail rewrite -> sendmail8
Added in extensions developed elsewhere. Revision of SMTP, DSNsm other systems, 8bit mail, new configuration package
sendmail bat book dramatic increased uptake of sendmail8
Commercial sendmail – Sendmail Inc in 1998. 1st commercial open source/commercial hybrid companies
Added by commercial company – encryption & auth, milter, virtual hosting, LDAP, lots more checking
Part 2 Lessons Learned
Changes in requirements over time:
Reliability – deliver or bounce
Functionality and performance
Then became about protection
Next legal/regulatory issues
Increasingly controlling costs
Design decisions and evolution
Overly general: can do anything, reality world was in flux, easier to build tool than solution, still true today to extent, retrospect would do again
rewrite rules: good idea, using tabs as active character was wrong. concept right but syntax and control flow could have been better
Message munging – essential for interoperability at the time, still used – to you pass through transh, reject or fix
Syntax of config – flat (no nesting) , to much used of single characters, ugly but not fundimentally flawed, would use something like apache todday
smtp and queuing embedded in sendmail – good idea
structure of queue – two files per message, requires lots of scans, ascii simplified debugging, correct for time, wouldn’t do now
Using m4 for config – painful syntax, damn “dnl” ‘s aren’t even necessary. Retrospec some tool was needed unclear if m4 was the right one.
Extending vs changing features. Example masquerading. Wrong first time, reimplemented but added new feature instead of fixing. Didn’t want to break existing sites. retrospec wrong and big part of why sendmail hard to configure, should have provided upgrade tools
Accepting and fixing bogus input. permits broken software to exist. right at the time unclear if it’s the right thing today.
What do diff today –
Fix problems asap, v7 mailbox format, tabs in config files, bogus features
use modern tools
more privilege separation
create string abstraction
create string separation
separate mailbox names from unix user ids
cleaner config file
What would I do the same –
use C as the implementation language
Bit things off in smallish chunks
syslog
rewriting rules (without tab char)
Net rely too heavily on outside tools
Takeaways –
KISS actually works ( sendmail was a “2nd System” )
If you don’t know what you are doing advance designs don’t help much
The world is a messy place
Flexibility trumps performance
Fix things early – if you succeed installed base only gets bigger
ASCII is great for internal files and protocols
documentation is key to broad acceptance
the design space is always changing
Q: if you were not Eric and you were starting today which mail server would you use? A: “postfix”
