Re: How to make GNU Guile more successful

[Vítor De Araújo]

On 10/03/2017 17:17, Amirouche wrote:
> Héllo!
> 
> 
> Le 10/03/2017 à 03:08, Vítor De Araújo a écrit :
>> Hi,
>>
>> I'm relatively new to Guile and new to this list, but I saw this
>> thread in the archives and I'd like to make some comments.
> 
> Welcome!

Thanks :)

[...]

>> On the topic of package management, I think that, more important than
>> a central repository, is a standard format in which to distribute
>> packages for easy installation.
> I agree. The standard could be to have a tarball named using
> semantic versionning, from which one can extract the PACKAGE-NAME.
> Inside the tarball, the package manager will expect a PACKAGE-NAME.scm
> or a PACKAGE-NAME directory which can be installed into ~/.local/lib/guile
> 
> WDYT?

That's good, but it'd also have to contain a metadata file containing at
least the package's dependencies and what commands should be run to
install (for example, ./configure && make && make install, for packages
which need some compilation step before copying to ~/.local/lib/guile).

>> I've never used guildhall, and I guess
>> it defines something like this,
> Probably but it's not documented.
> 
>> but I think just having an
>> installation command where you can pass the URL of a package to
>> download and install (say, guild install http://some.url/somepackage.zip,
>> or even a git URL)
> This will still require the user to set GUILE_LOAD_PATH. Maybe we can
> add ~/.local/lib/guile/ to the default GUILE_LOAD_PATH to make things
> easier.

Yes, ~/.local/lib/guile/ (or other user-level directory) definitely
should be in the default GUILE_LOAD_PATH.

> A simple guild install https://url/to/patch can be done in mater of hours
> (if not minutes).
> 
> WDYT?

Sure. Adding dependency handling would make things somewhat more
complex, though.

I thought a little bit about the packages-as-urls idea and I had some
ideas I'll just throw out here.

First, I realized using the URL as the package _name_ is problematic,
because it hard-codes the source of the package; that complicates things
if someone wants to make a fork of the package and install it in place
of the original. However, packages can still have a normal name and a
"canonical URL" for fetching it.

The package would come with a metadata file (named something like
metadata.foopkg, where 'foopkg' is a placeholder for the name of the
package manager (and I'd rather have the name of the package manager as
an extension, because I'd rather not assume I'm the only package manager
in the world)). This metadata would contain the package name, version,
and dependencies. For each dependency, it would specify the package
name, the version, *and the canonical URL* for that package. Upoin
installation, the package manager would go through each dependency and
check if a package of that name and compatible version (assuming
semantic versioning) is already installed; if it's not, it would fetch
it from the canonical URL.

In this way, we can have decentralized package management which can
handle dependencies.

And commenting on something I said previously:

On 10/03/2017 12:08, Panicz Maciej Godek wrote:
> 2017-03-10 15:27 GMT+01:00 <address@hidden>:
>
>     Yeah, for one I don't think the URLs should go directly into the
>     code, but
>     rather in a package metadata file. The reason is I don't think we
should
>     hard-code the package manager in the code. People should be able
to use
>     the code without having the package manager, if they already have the
>     dependencies installed by some other means. But the basic idea is
>     cool. :)
>
>
> There is a reason why I like the idea of having URLs directly in the code,
> namely that it makes you independent from some big institutions/particular
> maintainers. Of course it is risky if your code refers to repositories
that
> you do not control, but you can always make your private forks.

I'm totally in favor of independence from particular institutions or
maintainers, but I think this can be achieved by putting the URLs in the
metadata file rather than directly in the code. Code would just use
(use-modules ...) as usual to load code, and the business of dowloading
and installing would be solely the task of the package manager. I think
that enables a friendlier co-existence with other package managers, and
allows the same library to be packaged for multiple package managers at
the same time (e.g., guix, compan, guildhall, etc.). It also allows
people to just unpack and install the code manually if they want to.

(A problem that still remains if packages have names separate from the
URLs is avoiding name clashes. The wiki would go some way towards that
(people could check more easily if the name they want is already in
use), but maybe some conventions for globally unique names (e.g.,
prepending a domain prefix like Java people do) could ameliorate this. I
don't think this is really important at this point.)

[...]
> FWIW, I've started a new effort to build a package manager for GNU
> Guile. You can find a demo at http://guildhall.hypermove.net/
> 
> It started as a guildhall web frontend but I noticed that guildhall:
> 
> 1) doesn't work with guile 2.2 and I can not make it work
> 2) The solver is too complicated, even if it's based on aptitude algorithm
>   it's still complicated.
> 3) use a lot of other scheme libraries [0] (like foof-loop) which
>   doesn't help noobs like me to dive into the code. Maybe those
>   libraries are *very* neat but why not include them in Guile proper
>   then and make them default.
> 
> [0] https://github.com/ijp/guildhall/tree/master/guildhall/ext
> 
> For all this reasons I forked the effort.
> 
> I address each previous point as follow:
> 
> 1) I only use guile 2.2, right now guile 2.1 is not supported. I think it's
> bad this should be addressed at some point. AFAIR the issue is in
> guile-wiredtiger (or it's an issue with setlocale (anyway this must be
> addressed)))
> 
> 2) I use a logic programming library called minikanren to solve the
> dependency problem. This is the first real problem I can use minikanren
> to solve a problem that is craft. This is the logic programming
> library in scheme. While it's not very advanced compared to core.logic,
> the version I use called microkaren has a straight forward
> implementation that can
> hold all at once in a single head. Logic programming, like probabilistic
> programming (cf. ) are two research areas in programming languages
> that could improve the way we craft algorithms today. The logic approach
> being more useful to me (but someone doing machine learning stuff,
> might find the probabilistic scheme very useful. Ask me for a link!).
> OpenCog
> does have a logic language embedded in it's language in their hyper
> graph database
> (AFAIU it works using backward / forward inference but implements on top
> of it higher logic constructs, like abduction, anaphora resolution...).
> It's
> something they call PLN as Probabilistic Logic Network. AFAIK it's some
> kind
> of probabilistic Datalog (which is one of the API Datomic support).
> There is
> also probabilistic minikanren (albeit not useable). So, to sum up (because
> I can (because I am not a robot)):
> 
> programming language research + guile + logic = minikanren.
> 
> Mind the fact that there is also a guile-log. But I still don't get what
> guile-log does (hints: http://c-lambda.se/) whereas I understand how
> microkanren does its stuff (I still need to benchmark it).
> 
> 3) I don't use foof-loop but I bring my own dependencies. I use my
> own database library that is based on wiredtiger. I've been working
> on this database library for 3 years, it's well documented and various
> example codes. That said, AFAIK, I am the only one to use it. I've built
> a clone of Datomic (with the persistence part, patch welcome) which
> use a query engine similar to Datomic I guess based on minikanren.
> 
> It's performance on a middle end laptop run Guile 2.2 are the followings:
> 
> - 1500 document reads per seconds
> - 1000 to 500 document write per second
> 
> Documents are scheme assoc, and those are inserted 1000 per 1000 and
> read 1000 per 1000 until it reach 50Go of data. At that point writes
> take 1/500
> seconds or 0.002 seconds.
> 
> The biggest dataset I *loaded* into that database is wikidata which is
> 50Go, I
> don't remember how much time it takes to load it.
> 
> I had issues with wiredtiger using the Python bindings but only during
> reads
> over a gigantic dataset of 60Giga. I say gigantic for a blog  not for
> Google
> scale. Also upstream can solve issues if we can have a way to reproduce the
> issues (which I plan to do once guile 2.2 is out (which means I will
> redo the
> benchmarks against wikidata and read/write)). I understand the problem
> that wiredtiger being part of MongoDB is problematic as MongoDB might not
> care much for the same problem as ours. They will always be interested by
> bigger free software database, tho. As it make free publicity of how
> successful
> is their software.
> 
> Also there is much documentation about this library. I created several
> Guile projects
> using it (albeit not big) they document several layers of the library
> and one
> Guile user reported using it for doing human/social science research.
> 
> Search for guile wiredtiger in you favorite duckduckgo search engine.
> 
> Some people claim that PostgreSQL has all the required feature that
> someone wants
> to store data and that this mongodb/wiredtiger is a bad. I recognize
> that wiredtiger
> can be poor man's database right now. I don't have the required
> expertise to verify
> whether it's good enough for your usecase. PostgreSQL is used in
> "production" in
> all the world using multiple workload and stuff. PostgreSQL is good.
> 
> 1) There is no dynamic ffi bindings for PostgreSQL yet, otherwise said you
>     can't use pgsql from Guile RIGHT NOW.
> 2) wiredtiger is GOOD enough, benchmark it before saying it's not good
> ENOUGH.
> 3) it's not SQL, it's Schemey! It's inspired from datomic which the goto
> database
>     for clojure with multiple implementation in the browser. This is
> killer feature
>     of the clojure ecosystem that is closed source.
> 4) I find it more funny to fiddle with logic programming that set pseudo
> theory of SQL.
> 5) The first thing I do when pgsql bindings will be out, is to port
> feature space to PostgreSQL
> and compare both pgsql and wiredtiger.
> 6) wiredtiger will always be simpler to use that Postgresql it's like
> comparing sqlite
> and Datomic.
> 
> What happens next? You choose to hack on this project and find something
> interesting in this
> TODO list:
> 
> - fix guile-wiredtiger to work on guile 2.0
> 
> - make it possible in guile-wiredtiger to stream the results of a look
> up using traversi using multiple cursors
> 
> - implement disequality in microkanren (for implementing the package
> dependency resolver
>   and for fun because logic is awesome).
> 
> - implement a pastebin service using GNU Guile, yes because I think this
> a good bad reason to start
>   another simple project that is not a static blog generator.
> 
> - Port the wiki of http://culturia.one to use feature space library.
> 
> - implement or find a scheme library that does the diff two lists (to
> finish wiki implementation
>   and for the package dependency resolver).
> 
> - Help Matt Wette to complete his ffi helper
> http://lists.gnu.org/archive/html/guile-user/2017-03/msg00085.html
>   (because we will need it for guix).
> 
> - add guile-fibers as a submodule and use it
> 
> - there is also a simple boolean keyword search engine that must be
> ported to feature space
> 
> - find the code that implements trigram transderivational search and put
> it in the repository. Bridging
>   the gap with tsearch2.
> 
> - index the wiki
> 
> - index the packages
> 
> - index the pastebin
> 
> - index the web
> 
> And add gnunet to the mix.
> 
> There is not a lot of get-together pure guile project out there. There
> is guix. I hope this new project
> can be the occasion for new guilers to submit patches or ideas.
> 
> The code is currently hosted at
> https://framagit.org/a-guile-mind/culturia.one/tree/guildhall/
> 
> The codename is "primary".
> 
> Also I'd like to point out, that I don't need primary. I do it for the
> community only. The
> road is fascinating, tho. I could make mistakes so please input your
> feedback if you want
> to use a tool like that.
> 
> You can:
> 
> guix package -i wiredtiger
> git clone https://framagit.org/a-guile-mind/culturia.one.git
> git checkout guildhall
> emacs culturia.one/src/webui.scm
> 
> or
> 
> cd culturia.one/src/ && guile -L . webui.scm

There are certainly lots of fun subprojects to tackle here, but I'm
afraid it would take really long to get ready. I'd rather have a really
simple package manager, preferably with zero dependencies other than
Guile itself. But then again I think multiple package managers should be
able to co-exist, so there's no problem for different people to tackle
multiple approaches to the problem.

-- 
Vítor De Araújo