Important notes on Mercurial repository access for Isabelle

===========================================================

Introduction

------------

Mercurial http://www.selenic.com/mercurial belongs to the current

generation of source code management systems that follow the so-called

paradigm of "distributed version control".  This is a terrific name

for plain revision control without the legacy of CVS or SVN.  See also

http://hginit.com/ for an introduction to the main ideas.  The

Mercurial book http://hgbook.red-bean.com/ explains many more details.

Mercurial offers great flexibility in organizing the flow of changes,

both between individual developers and designated pull/push areas that

are shared with others.  This additional power demands some additional

responsibility to maintain a certain development process that fits to

a particular project.

Regular Mercurial operations are strictly monotonic, where changeset

transactions are only added, but never deleted.  There are special

tools to manipulate repositories via non-monotonic actions (such as

"rollback" or "strip"), but recovering from gross mistakes that have

escaped into the public can be hard and embarrassing.  It is much

easier to inspect and amend changesets routinely before pushing.

The effect of the critical "pull" / "push" operations can be tested in

a dry run via "incoming" / "outgoing".  The "fetch" extension includes

useful sanity checks beyond raw "pull" / "update" / "merge".  Most

other operations are local to the user's repository clone.  This gives

some freedom for experimentation without affecting anybody else.

Mercurial provides nice web presentation of incoming changes with a

digest of log entries; this also includes RSS/Atom news feeds.  There

are add-on history browsers such as "hg view" and TortoiseHg.  Unlike

the default web view, some of these tools help to inspect the semantic

content of non-trivial merge nodes.

Initial configuration

---------------------

The official Isabelle repository can be cloned like this:

  hg clone http://isabelle.in.tum.de/repos/isabelle

This will create a local directory "isabelle", unless an alternative

name is specified.  The full repository meta-data and history of

changes is in isabelle/.hg; local configuration for this clone can be

added to isabelle/.hg/hgrc, but note that hgrc files are never copied

by another clone operation.

There is also $HOME/.hgrc for per-user Mercurial configuration.  The

initial configuration requires at least an entry to identify yourself

as follows:

  [ui]

  username = XXX

Isabelle contributors are free to choose either a short "login name"

(for accounts at TU Munich) or a "full name" -- with or without mail

address.  It is important to stick to this choice once and for all.

The machine default that Mercurial produces for unspecified

[ui]username is not appropriate.

Such configuration can be given in $HOME/.hgrc (for each home

directory on each machine) or in .hg/hgrc (for each repository clone).

Here are some further examples for hgrc.  This is how to provide

default options for common commands:

  [defaults]

  log = -l 10

This is how to configure some extension, which is called "graphlog"

and provides the "glog" command:

  [extensions]

  hgext.graphlog =

Shared pull/push access

-----------------------

The entry point http://isabelle.in.tum.de/repos/isabelle is world

readable, both via plain web browsing and the hg client as described

above.  Anybody can produce a clone, change it locally, and then use

regular mechanisms of Mercurial to report changes upstream, say via

mail to someone with write access to that file space.  It is also

quite easy to publish changed clones again on the web, using the

ad-hoc command "hg serve -v", or the hgweb.cgi or hgwebdir.cgi scripts

that are included in the Mercurial distribution, and send a "pull

request" to someone else.  There are also public hosting services for

Mercurial repositories.

The downstream/upstream mode of operation is quite common in the

distributed version control community, and works well for occasional

changes produced by anybody out there.  Of course, upstream

maintainers need to review and moderate changes being proposed, before

pushing anything onto the official Isabelle repository at TUM.  Direct

pushes are also reviewed routinely in a post-hoc fashion; everybody

hooked on the main repository is called to keep an eye on incoming

changes.  In any case, changesets need to be understandable, at the

time of writing and many years later.

Push access to the Isabelle repository requires an account at TUM,

with properly configured ssh to the local machines (e.g. macbroy20 ..

macbroy29).  You also need to be a member of the "isabelle" Unix

group.

Sharing a locally modified clone then works as follows, using your

user name instead of "wenzelm":

  hg out ssh://wenzelm@macbroy20//home/isabelle-repository/repos/isabelle

In fact, the "out" or "outgoing" command performs only a dry run: it

displays the changesets that would get published.  An actual "push",

with a lasting effect on the Isabelle repository, works like this:

  hg push ssh://wenzelm@macbroy20//home/isabelle-repository/repos/isabelle

Default paths for push and pull can be configured in

isabelle/.hg/hgrc, for example:

  [paths]

  default = ssh://wenzelm@macbroy20//home/isabelle-repository/repos/isabelle

Now "hg pull" or "hg push" will use that shared file space, unless a

different URL is specified explicitly.

When cloning a repository, the default path is set to the initial

source URL.  So we could have cloned via that ssh URL in the first

place, to get exactly to the same point:

  hg clone ssh://wenzelm@macbroy20//home/isabelle-repository/repos/isabelle

Simple merges

-------------

The main idea of Mercurial is to let individual users produce

independent branches of development first, but merge with others

frequently.  The basic hg merge operation is more general than

required for the mode of operation with a shared pull/push area.  The

"fetch" extension accommodates this case nicely, automating trivial

merges and requiring manual intervention for actual conflicts only.

The fetch extension can be configured via $HOME/.hgrc like this:

  [extensions]

  hgext.fetch =

  [defaults]

  fetch = -m "merged"

To keep merges semantically trivial and prevent genuine merge

conflicts or lost updates, it is essential to observe to following

general discipline wrt. the public Isabelle push area:

  * Before editing, pull or fetch the latest version.

  * Accumulate private commits for a maximum of 1-3 days.

  * Start publishing again by pull or fetch, which normally produces

    local merges.

  * Test the merged result as usual and push back in real time.

Piling private changes and public merges longer than 0.5-2 hours is

apt to produce some mess when pushing eventually!

Content discipline

------------------

The following principles should be kept in mind when producing

changesets that are meant to be published at some point.

  * The author of changes needs to be properly identified, using

    [ui]username configuration as described above.

    While Mercurial also provides means for signed changesets, we want

    to keep things simple and trust that users specify their identity

    correctly (and uniquely).

  * The history of sources is an integral part of the sources

    themselves.  This means that private experiments and branches

    should not be published by accident.  Named branches are not

    allowed on the public version.  Note that old SVN-style branching

    is replaced by regular repository clones in Mercurial.

    Exchanging local experiments with some other users can be done

    directly on peer-to-peer basis, without affecting the central

    pull/push area.

  * Log messages are an integral part of the history of sources.

    Other people will have to inspect them routinely, to understand

    why things have been done in a certain way at some point.

    Authors of log entries should be aware that published changes are

    actually read.  The main point is not to announce novelties, but

    to describe faithfully what has been done, and provide some clues

    about the motivation behind it.  The main recipient is someone who

    needs to understand the change in the distant future to isolate

    problems.  Sometimes it is helpful to reference past changes via

    changeset ids in the log entry.

  * The standard changelog entry format of the Isabelle repository

    admits several (vaguely related) items to be rolled into one

    changeset.  Each item is then described on a separate line that

    may become longer as screen line and is terminated by punctuation.

    These lines are roughly ordered by importance.

    This format conforms to established Isabelle tradition.  In

    contrast, the default of Mercurial prefers a single headline

    followed by a long body text.  The reason for that is a somewhat

    different development model of typical "distributed" projects,

    where external changes pass through a hierarchy of reviewers and

    maintainers, until they end up in some authoritative repository.

    Isabelle changesets can be more spontaneous, growing from the

    bottom-up.

    The web style of http://isabelle.in.tum.de/repos/isabelle/

    accommodates the Isabelle changelog format.  Note that multiple

    lines will sometimes display as a single paragraph in HTML, so

    some terminating punctuation is required.  Do not squeeze multiple

    items on the same line in the original text!

Building a repository version of Isabelle

-----------------------------------------

Compared to a proper distribution or development snapshot, it is

relatively hard to build from the raw repository version.  Essential

contributing components are missing and need to be reconstructed by

running the Admin/build script by hand.  Afterwards the main Isabelle

system and logic images can be compiled via the toplevel ./build

script.  Note that the repository lacks some textual version

identifiers in the sources and scripts; this implies some changed

behavior when processing settings etc.

There is no guarantee that the NEWS file is up to date at an arbitrary

point in history.