[PATCH 1 of 4] revlog: use absolute_import

[Matt Mackall]

On Sat, 2015-07-18 at 13:25 -0700, Gregory Szorc wrote:
> On Sat, Jul 18, 2015 at 1:13 PM, Gregory Szorc <gregory.szorc at gmail.com>
> wrote:
> 
> > On Sat, Jul 18, 2015 at 1:07 PM, Gregory Szorc <gregory.szorc at gmail.com>
> > wrote:
> >
> >> On Thu, Jul 16, 2015 at 10:28 AM, Matt Mackall <mpm at selenic.com> wrote:
> >>
> >>> On Tue, 2015-07-14 at 13:51 -0700, Gregory Szorc wrote:
> >>> > # HG changeset patch
> >>> > # User Gregory Szorc <gregory.szorc at gmail.com>
> >>>
> >>> > # Date 1436894831 25200
> >>> > #      Tue Jul 14 10:27:11 2015 -0700
> >>> > # Node ID 5154b8421ebd670d6155659c5811f193f5fe1555
> >>> > # Parent  35fa7c77c754aa4d156c42abfdb61ca178468872
> >>> > revlog: use absolute_import
> >>> >
> >>> > This is needed for Python 3 compatibility. We have static analysis
> >>> > checks in place to ensure files using absolute_import conform to the
> >>> > style adopted by this patch.
> >>>
> >>> More weird test failures on 2.6:
> >>>
> >> ........
> >>> --- /home/mpm/hg-test-/tests/test-help.t
> >>> +++ /home/mpm/hg-test-/tests/test-help.t.err
> >>> @@ -1244,6 +1244,15 @@
> >>>
> >>>    $ cd "$TESTDIR"/../doc
> >>>    $ python check-seclevel.py
> >>> +  Traceback (most recent call last):
> >>> +    File "check-seclevel.py", line 168, in <module>
> >>> +      if checkhghelps():
> >>> +    File "check-seclevel.py", line 99, in checkhghelps
> >>> +      mod = extensions.load(None, name, None)
> >>> +    File "../mercurial/extensions.py", line 86, in load
> >>> +      ui.debug('could not import hgext.%s (%s): trying %s\n'
> >>> +  AttributeError: 'NoneType' object has no attribute 'debug'
> >>> +  [1]
> >>>    $ cd $TESTTMP
> >>>
> >>>  #if serve
> >>>
> >>> ERROR: test-help.t output changed
> >>> !......
> >>> --- /home/mpm/hg-test-/tests/test-extension.t
> >>> +++ /home/mpm/hg-test-/tests/test-extension.t.err
> >>> @@ -949,12 +949,15 @@
> >>>    $ if [ -z "$hgver" ]; then
> >>>    >   echo "unable to fetch a mercurial version. Make sure __version__
> >>> is correct";
> >>>    > fi
> >>> +  unable to fetch a mercurial version. Make sure __version__ is correct
> >>>    $ rm -f throw.pyc throw.pyo
> >>>    $ hg --config extensions.throw=throw.py throw 2>&1 | egrep '^\*\*'
> >>> -  ** unknown exception encountered, please report by visiting
> >>> -  ** http://mercurial.selenic.com/wiki/BugTracker
> >>> -  ** Python * (glob)
> >>> -  ** Mercurial Distributed SCM (*) (glob)
> >>> +  ** Unknown exception encountered with possibly-broken third-party
> >>> extension throw
> >>> +  ** which supports versions unknown of Mercurial.
> >>> +  ** Please disable throw and try your action again.
> >>> +  ** If that fixes the bug please report it to http://example.com/bts
> >>> +  ** Python 2.6.6 (r266:84292, Jan 22 2014, 09:42:36) [GCC 4.4.7
> >>> 20120313 (Red Hat 4.4.7-4)]
> >>> +  ** Mercurial Distributed SCM (version +0-)
> >>>    ** Extensions loaded: throw
> >>>
> >>>  Patch version is ignored during compatibility check
> >>>
> >>>
> >> I'm able to reproduce these on Linux with Python 2.6.9. However, the
> >> problem goes away if `make local` is executed (both run-tests.py and
> >> run-tests.py -l work with a local build in place). However, the failure
> >> persists if the local install was built against Python 2.7. It certainly
> >> appears like the local install is leaking into the test environment
> >> somehow, at least for these tests.
> >>
> >
> > doc/check-seclevels.py is doing weirdness with sys.path foo. The error
> > message about ui not having a "debug" attribute is a red herring because
> > demandimport is mucking about with things. This is the actual error:
> >
> > $ HGDEMANDIMPORT=disable ~/.pyenv/versions/2.6.9/bin/python
> > check-seclevel.py
> > Traceback (most recent call last):
> >   File "check-seclevel.py", line 13, in <module>
> >     from mercurial.commands import table
> >   File "../mercurial/commands.py", line 13, in <module>
> >     import hg, scmutil, util, revlog, copies, error, bookmarks
> >   File "../mercurial/hg.py", line 13, in <module>
> >     import localrepo, bundlerepo, unionrepo, httppeer, sshpeer,
> > statichttprepo
> >   File "../mercurial/localrepo.py", line 10, in <module>
> >     import peer, changegroup, subrepo, pushkey, obsolete, repoview
> >   File "../mercurial/subrepo.py", line 13, in <module>
> >     import config, util, node, error, cmdutil, scmutil, match as matchmod
> >   File "../mercurial/cmdutil.py", line 11, in <module>
> >     import util, scmutil, templater, patch, error, templatekw, revlog,
> > copies
> >   File "../mercurial/revlog.py", line 29, in <module>
> >     from . import (
> > ImportError: cannot import name parsers
> >
> > This is probably a result of the 'sys.path.insert(0, "..")' line in
> > check-seclevel.py. That will almost certainly throw off relative imports.
> >
> 
> The rabbit hole says this is a problem with mercurial/pure being in
> sys.path. Before "import parsers" would pick up the C extension from
> mercurial/ (if available) and fall back to mercurial/pure/parsers.py if
> there was no C extension. With absolute imports, we need to explicitly
> qualify imports.
> 
> The obvious solution is to catch ImportError when native extensions can't
> be imported:
> 
> try:
>     from . import parsers
> except ImportError:
>     from .pure import parsers

Historically (circa 0.6 or so), the reason we stopped falling back to
pure Python was that people would fail to run any of the build scripts..
and then publish benchmarks of the unoptimized code vs Git that were off
by an order of magnitude. This was, to put it mildly, aggravating.

We care a bit less about that now, since we have some packaging in place
these days, but I'd still like people to not be able to accidentally
waste their time and everyone else's running pure/ by accident. We can
probably come up with something less hacky than the current system, but
it should still require an explicit opt-in of some sort.

-- 
Mathematics is the supreme nostalgia of our time.