Re: What is a specification for? [was Re: Calculating Age Question]

Benjamin Franz writes:
>Roy - false cachable hits are MUCH worse than false uncachable hits. Your
>'doomsday scenario' of network meltdown because of the extra hits due to
>pre-maturely expired objects has NOT been demonstrated and some people
>pointed out that your preferred scheme potentially results in stale objects
>being served as fresh - which is utterly unacceptable under any
>circumstances. Caches *should* be ultra-conservative with regard to Age.

>From my perspective as one involved in Intranet and Extranet applications 
(defined as "inside the company WAN" and "inside the company WAN as tunneled 
over the Internet"), it must be remembered that HTTP is going to be used for 
some very time-critical, nay, even near real-time, applications, where 
serving false data could cost thousands or millions of (insert local 
currency unit here).  If the Age spec is actually broken (as it appears to 
be, though I haven't followed the arguments that closely), it should be 
fixed.  But it is never acceptable to serve stale data as fresh.  As a 
(possibly bad) analogy, if serving stale data as fresh were OK, relational 
database vendors would not have taken the pains they have to make sure data 
is accurate in the face of concurrent updates to that data.  Instead, 
relational DBs go through all sorts of contortions to ensure that the data 
presented is accurate as of the time of the database query.

Although I can see the possibility inherent in Roy's scenario of network 
meltdowns, any engineer should know that you don't just engineer for the 
average case -- you prepare, in one way or another, for the extreme cases. 
 (It is not acceptable, for example, for a television set to explode or 
break down if there is no signal on a channel or the TV is located in a 
house 1 door away from the television transmitter.)  If (once the Age spec 
is fixed), the New Zealand proxy "melts down" because of some (relatively 
small) fraction of HTTP responses are false uncachable hits, then it is not 
engineered well enough -- certainly not to handle the increased load that 
will be placed on it as more New Zealanders (sp?) go on-line to the Web, or 
as more New Zealand businesses use the Web more.
Mark Leighton Fisher                   Thomson Consumer Electronics                   Indianapolis, IN

Received on Monday, 2 December 1996 05:42:37 UTC