Duncan Coutts
On Wed, 2008-06-04 at 15:14 +0100, Claus Reinke wrote:
I think that's fundamentally the wrong approach. We shouldn't have to build a "Haskell installation manager". Would you also want installation managers for Perl, Python, Ruby, C, C++, etc. each with their own different user interfaces and feature sets? I think not - you want a single package manager for the whole system from which you can install/uninstall libraries for any language.
This is something that Windows gets completely wrong. Why do I have twelve icons in my status bar all representing little programs that are running in the background checking for updates to their own bits of software? Why on earth do I have a Printer Driver Update Manager? And I'd be wondering the same thing about a "Haskell installation manager": installation and dependencies are not something specific to Haskell.
why then do we have ghc-pkg, or cabal? surely the native package managers should handle all that?
there are (at least) two dimensions of uniformity: across different kind of software on a single kind of system, and with a single kind of software across different kinds of system. platform-specific package managers hide the software-specific notions of package dependency maintainence, haskell-specific package managers hide the platform-specific notions of package dependency maintainence.
there is no need for platform- and haskell-specific tools to be entirely separate or in conflict with each other: where both exist, one can be a view on the other (if you are on linux-of-the-day, you can use its package manage, independent of whether your packages are haskell or lisp; and if you are using haskell, you should be able to use its package manager, independent of whether you are on unix-variant-of-today or on unix-variant-of-yesterday).
there seems to be a lot of confusion here, some of us not understanding the issues because we happen to be using systems where "everything just works", others among us not understanding the issues because we happen to be using systems where "such things would never work anyway", and yet others insisting on "i'll do it my way, so i know what works" (and then, of course, there are those who are actively working on improving the situation who will see my criticism as constructive, i hope!-).
1. there are no systems where "packages just work"! there are systems where a few people ensure that many people can live in such an illusion, though.
Yes indeed! :-)
2. systems with native package manager software still need help from haskell-specific toolchains (unless you want the human package managers on those systems to code all haskell-specific dependencies by hand).
Yes. As an illustration: gentoo has an "haskell-cabal.eclass" that interfaces between ebuilds and cabal as the build manager and there is a tool to generate ebuilds that use the haskell-cabal.eclass from .cabal descriptions (so we get correct deps automatically).
3. systems without native package managers (or perhaps i should say: systems on which users with unix background traditionally avoid getting acquainted with the details and usage of whatever might pass as installation management on those systems) are still in *very* wide-spread use, and if haskell users on those systems are left out in the rain, haskell developers will not be able to support those systems. this limits the user and application base of haskell on those systems, making haskell less relevant than it could be.
Eg Windows, OSX, Solaris.
4. haskell enables programming at a very high level of abstraction, with fairly good support for mostly platform independent code. but that code needs to be installed, and integrated with dependencies and dependents, and the integrated haskell installations needs to be maintained. and that should "just work", even if the developer is on (1;2) and the user is on (3), or vice versa, or if developers and users are on different flavours of (1;2) or (3).
with these clarifications out of the way, my interpretation of cabal was that it set out to provide two things
(A) a uniform platform-independent interface to such a haskell package installation manager. (B) a uniform platform-independent toolchain to support such a haskell package installation manager.
I guess so.
on systems in the (1;2) class, human package managers would use (B) to integrate haskell packages into the native package management software, so users might never even encounter cabal.
As in my example with the gentoo haskell packages above.
even so, (A) might offer a haskell-specific view on the general platform package management (when you want to see the haskell gui libs rather than all gui libs).
on systems in the (3) class, users and developers would interface with (A/B) directly, for lack of a better alternative.
and developers/users in the (4) class would simply use (A/B), without having to check whether they are "real" or just a view on the platform-specific software. it is this cross platform view of software development that i'm most interested in (i'm one of those who use bash, vim, opera, and haskell, no matter whether i'm on windows, solaris, or whatever, and the cross-platform availability of those tools has saved me many a headache;-).
returning to my earlier message, it seems that my concerns were mainly these:
- it isn't sufficient to worry about installation management, one has to worry about integration, lifetime and uninstall management as well. in short, maintain the dependency graphs over any of "install"/"upgrade"/"uninstall".
For users in (1;2) class we would expect the native package manager to do this.
- for this to work, cabal needs to maintain not only libraries as packages, but tools and compilers as well. without this, some dependencies are not recorded (this haddock depends on that ghc; to build this package i need that tool; that tool was built with this ghc version, from those sources, etc).
and if the dependencies are not even recorded, they are likely to be broken if one does install/upgrade or uninstall any haskell software, be it library, tool, or compiler.
Right. Currently cabal-install in its roll as a secondary package manager is very weak on that front. It does not record anything about what it installs. There's clearly plenty of work to be done there.
I'm wondering whether a hacked version of darcs could be used for this. Together with sandboxing and support of alien packages, it would already look like a full-fledged distribution... binary packages could be a set of zipped darcs patches. Add USE-flag derived dependencies and package and flag based stable/unstable marks and I'll switch without hesition. The question, IMHO, seems to be "How would a package manager for a posix-compilant kinetic look like?" I'd rather like a package manager to find them, rule them all and in obsoletion bind them than attempting to hack around package managers * packages incompatibilities. PS: If you don't know why sandboxing is a MUST, read http://myeve.eve-online.com/devblog.asp?a=blog&bid=526 -- (c) this sig last receiving data processing entity. Inspect headers for past copyright information. All rights reserved. Unauthorised copying, hiring, renting, public performance and/or broadcasting of this signature prohibited.