8 Apr
2017
8 Apr
'17
4:36 a.m.
On Fri, Apr 07, 2017 at 03:08:48PM -0700, Steve Johnson wrote:
One lesson learned that I've never forgotten is
how smooth it is to
evolve a language using the following process:
* Announce that the change is coming and explain why
* Change the compiler to accept both the old and new syntax
* Produce a simple warning message when the old syntax was used, but
make it still work
* Produce a more complicated, verbose message, but still make it work
* Produce a message that says "After date xxxx, the old stuff won't
work any more"
* On the date, change the warning to fatal, but keep recognizing the
old syntax and emit "Error: You used the old xxx, change to the new
one"
* Eventually, stop recognizing the old syntax and remove the message.
Dennis was a master at this strategy, so things like the otherwise
painful evolution of changing =+ to += went well.
That's interesting that that sort of thing dates back (at least) to the
Labs.
We did a distributed source management system which has
- a file format (I can think of 3 different major versions)
- a network protocol (also had major revisions)
- various per repository features
All that started back in 1998 and we were extremely good about backwards
compatibility.
That said, we eventually had too many things to be compat with and we
took the approach of supporting
- our original ascii file format which was SCCS compat plus extensions
- our original (once stable) network protocol
and
- our latest and greatest stuff
Every version of the software supports a
bk clone --downgrade repo repo.old
which gets you back to the old ascii format, and then there is a
bk clone --upgrade repo.old repo
which gets you to the latest and greatest.
What did this buy us? We only had two targets, the original (not moving)
formats and the latest and greatest. If you had some halfway repo you
could use the old release to clone it down to the old format and the new
release to clone it forward to the latest and greatest.
It worked really, really well.
Sort of off topic but thought I'd share, I wish someone had told me this
picture when we started, especially the feature bit idea (which was that
if a repo uses some optional feature you stuck that on disk. If all
versions of the software know to read in the feature bits, see if they
know all of them, if not, spit out "This repo uses feature XYZ which is
not understood by this version of BitKeeper". We called them bits, they
were actually strings, but internally they were bits).