Is someone already working on a port that supports Boost 1.35.0?

[David Wood]

[Additional CCs added by Aryeh noted, but left untouched]

I wanted to retitle this post, but couldn't come up with a summary of 
what I was trying to say.


In message <48181A2A.1000303 at gmail.com>, Aryeh M. Friedman 
<aryeh.friedman at gmail.com> writes
>I am top posting because my comments are general and not in 
>relationship to any given point.   I think Mezz along with Simon both 
>the right way to handle it for now.   There are several projects that I 
>am not directly involved designed to tackle this sort of issue with 
>ports 2.0 being the right long term solution but not needed to solve 
>this issue per se so I will not discuss it here.   Most of the projects 
>can be found on Ale's PortsToDo wiki (wiki.freebsd.org/PortsToDo) the 
>main one not found on there is Jim Staplton's virtual ports DB (which I 
>think is either in the committ queue or should be since both me an Ale 
>have signed off on it as being a good idea).  The issue is really not 
>the infrastruct as you state but the more patchs like this we make the 
>more likely it is the infrastruct will become problematic down the road.

Forgive me, Aryeh, but you've taken a specific point I made and 
dismissed it saying that "the infrastructure needs a complete overhaul 
and I'm the right person to head that work", though you have 
acknowledged that there is other work going on.

I know you will argue you aren't making a reply to my post, so why 
didn't you trim all my text instead of quoting it? I believe you are 
replying - and that your reply is pretty close to being non sequitur.

I am not interested in pointless points scoring; I have better things to 
do. What I want to do is understand the issue..


I am no committer - just a mere ports maintainer. Freshports doesn't dig 
up any ports you maintain, at least not with an email address beginning 
with aryeh or containing fried.

http://tinyurl.com/585dtk shows my interest. net/freeradius is a fairly 
complicated port, (optionally) depending on MySQL, PostgreSQL, OpenLDAP, 
Kerberos and python (the python dependency is controlled by the 
EXPERIMENTAL option). There's also a dependency on OpenSSL - either the 
base version, or, if it's available, ports OpenSSL.


The SQL dependencies are pretty straightforward, so long as you remember 
that you can only have one version of each of the clients on a 
particular machine.

OpenLDAP is a little tricky, in that there's not only different versions 
in ports, but also the ability to build the OpenLDAP client with and 
without SASL. However, most administrators set the appropriate knobs via 
portconf or similar; it's usually appropriate to depend on whatever is 
on the system already, or build whatever the default is after any 
administrator applied overrides and depend on that. There's LDAP 
handling in bsd.port.mk, which helps the maintainer.

Kerberos is trickier still, in that you can only have one flavour of 
Kerberos on a system, and there's no USE_KERBEROS mechanism. Instead, a 
port needs to allow the choice of one of {nothing, Heimdal, MIT 
Kerberos} - which can't be done with yes/no questions without having two 
interdependent questions. Some ports have one option to depend on MIT 
Kerberos and another to depend on Heimdal, terminating with an error if 
both are set. Other ports, like FreeRADIUS, have one option to depend on 
Kerberos, and another to choose Heimdal or MIT Kerberos. This is another 
situation where bsd.options.mk might help - it might be possible to test 
if one or other flavour of Kerberos has been installed from ports and 
set the default appropriately, though a lot of features are unavailable 
before bsd.port.pre.mk has been included.

python is the real problem amongst those dependencies - but it won't be 
for much longer. As soon as possible after 31 May, I will be sending in 
a PR to remove the current python hack in net/freeradius and use 
bsd.options.mk instead.


I do not believe that this problem is one that brings into question the 
whole infrastructure. It certainly isn't "the more patchs (sic) that we 
make the more likely it is that the infrastruct (sic) will become 
problematic down the road". Not all evolutionary software engineering is 
bad, especially if modules are rewritten when appropriate. Of course I 
accept the possibility of 'bit rot' setting in when code has been 
through many hands and has many small changes applied. In this case, I 
believe that things are getting better, not worse or less maintainable.

This problem was known about some while back and there's no need to do 
any fresh engineering. The delay in having a solution available was the 
need for changes in the base system to support bsd.options.mk. Base 
system changes mean waiting until all versions of FreeBSD that lack the 
necessary support are End of Life. After 31 May, ports will be able to 
process their options and set a dependency on python before including 
bsd.pre.port.mk.

A simple and elegant solution will soon be available to solve what is 
essentially a circular dependency at the moment.



I see two issues here, both stemming from the complexities of dependency 
between ports. Firstly, there's a need to look for the best possible 
solutions with the tools we have now as well as to look for continuing 
improvement. Secondly, there's the need to educate both maintainers and 
administrators on how best to handle the complexities that can occur.

I think everyone would agree that the current documentation isn't the 
best. In stating this, I believe I'm restating Mark Linimon's well known 
criticism of the current Porters' Handbook. It is not my intention to 
hurt anyone by what I say. The way out of this is, of course, for people 
to pay attention to the documentation - but those best placed to do that 
are some of the busiest people already within the FreeBSD project. 
Having worked alongside technical authors, I realise what a specialist 
(and impressive) skill set they have.

At the moment, it is hard for a maintainer to discover current best 
practice. I suspect this is partly why the same issue is handled in 
different ways by different ports (as I mentioned above with Kerberos). 
Meanwhile, it is hard for system administrators to discover how best to 
use the richness of ports.


System administrators need the freedom to make whatever decisions they 
feel are appropriate (as I said, I wouldn't want ports unnecessarily 
depending on OpenSSL or the Cyrus SASL library). However, features that 
help administrators manage this richness, such as portconf, are not that 
well known about, especially by the less technical users.



Aryeh - you seem to have something against slave ports. At times, they 
are very useful. They make the creation and maintenance of client/server 
ports easy - for example, databases/mysql50-client is a slave port of 
databases/mysql50-server.

net/freeradius-mysql was created as a slave port of net/freeradius for 
work being done with pfSense. The need was for a FreeRADIUS package with 
MySQL support built in. The easiest way to ensure the necessary package 
is available from the FreeBSD servers is with a slave port. There are 
times when a slave port that enables one option makes sense.

However, slave ports that enable a single option in their master port 
can be troublesome. The example that comes to mind here is 
devel/subversion and the slave ports devel/subversion-perl, 
devel/subversion-python and devel/subversion-ruby. All these slave ports 
do is enable the appropriate language binding in Subversion. The options 
they enable aren't mutually exclusive, but these ports all conflict with 
each other, which can lead to problems. The language bindings aren't the 
defaults because of the dependency on a sizeable language port that 
isn't installed by default (there's also a fourth optional binding, 
which is Java).


The Subversion slave ports illustrate the problems caused by not 
necessarily being able to determine the options used for a particular 
port or package. edwin@ has already covered this at
http://wiki.freebsd.org/RewriteOfOptions by noting the wish for ports to 
be able to depend on (or at least test) OPTIONS of other ports. Another 
wish expressed there is a way of representing

That, by itself, doesn't need a new ports infrastructure - though it 
does need carefully thinking through in both the port and package cases. 
Options 2.0 is certainly being thought about - and whilst it touches on 
dependency graphing and other things you've mentioned in connection with 
your proposed "Ports 2.0", Aryeh, it doesn't necessarily need to change 
the way dependencies are currently handled - just enhance them.



The reason for both my posts is a feeling that you haven't understood 
*why* the current situation is as it is, Aryeh. A reply along the lines 
of "we're working on a complete replacement of the whole infrastructure, 
and that will solve it" doesn't help with understanding the specific 
issue that has arisen. You can't offer any guarantee that your new 
system will solve all the problems - especially if you do not take the 
time to understand the weaknesses in the current system as well as its 
strengths.

If you bring forward coherent proposals and a proof of concept for Ports 
2.0, I will certainly give what input I can at that stage. For now, we 
live with and continue to improve what we have, rather than looking only 
to what might replace it if it's found to be better after careful 
evaluation.


Maintainers have to work with OPTIONS as they are. They are a valuable 
feature, even if they don't offer all the functionality wished for.



Best wishes,




David
-- 
David Wood
david at wood2.org.uk