#48 Build ordering
Opened 2 years ago by otaylor. Modified 2 years ago

I have some mostly working code to add build ordering, but it's still rough and intertwined with the changes from PR#47, so I'll file this issue for now to get my ideas out and get some feedback.

Reasons for generating a build order

  • While module-build-service "retry until everything builds" algorithm means that an unordered module can succeed, such a build will take longer and have a more confusing build log.
  • If a module duplicates a package from a module it depends upon, or buildrequires bootstrap, then the lack of a build order can result in an incorrect build, since the library built against is not the library linked against.
  • Updating a hand-created build-order is extremely hard - it might require renumbering everything.

The Build Cycle Problem

A basic problem with trying to build order the Fedora package set is that it in general does not build order well - circular build dependencies are all over the place. However, it can be legitimate to ignore some BuildRequires when figuring out an order:

  • If the module is building against bootstrap, then BuildRequires that are only used for docs generation or similar can be ignored for the purpose of build ordering.
  • If the module is not building against bootstrap, then problematical build cycles will need to be broken by modifying the spec files. (perhaps conditionally for the modular build.) When we're figuring out the build order based on F27 package data, we can choose to ignore BuildRequires we know that we will remove or conditionalize.

Approach

A directed graph is created with the nodes being the SRPMs being rebuilt and the edges indicating that building A requires a binary package built from B to be installed. Edges are annotated with a list of BuildRequires of A that result in B being installed. We'll use an example of: fedmod resolve-deps --build-order --json pango --module=platform --module=X11-base --module=fonts. The initial set of edges is:

cairo: cairo pango harfbuzz libthai graphite2 libdatrie
graphite2: cairo pango harfbuzz libthai graphite2 libdatrie
harfbuzz: cairo graphite2
libdatrie:
libthai: libdatrie
pango: cairo harfbuzz graphite2 libthai libdatrie

The user can specify (via yaml file, likely) that certain BuildRequires should be ignored. If all the BuildRequire that contribute to an edge A=>B are ignored, then the edge is deleted.

If, after the specified ignores, cycles are found in the BuildRequires graph, then the cycles are displayed (possibly only the first N shortest cycles) and the modulemd generation errors out. In the example, if nothing is ignored, ordering errors out for cycles including:

graphite2 =[asciidoc]=> graphite2
cairo =[librsvg2-devel]=> cairo
pango =[pkgconfig(harfbuzz) >= 1.2.3]=> graphite2 =[asciidoc]=> pango
...

(By graphite2 =[asciidoc]=> graphite2, I mean graphite2, via a BuildRequires on asciidoc, causes graphite2 to be installed in the buildroot.)

These cycles (and all other cycles for this set of packages) are fixed by specifying:

    ignore-buildrequire:
        # librsvg2 is used only for a test program 
        cairo: ['librsvg2-devel']
        # graphite uses asciidoc for documentation generation.
        graphite2: ['asciidoc']

Which reduces the set of edges to:

cairo: 
graphite2:
harfbuzz: cairo, graphite2
libdatrie:
libthai: libdatrie
pango: cairo, harfbuzz, graphite2, libthai, libdatrie

If no cycles are found, the graph is used to generate a build order, by the principle that each module is built as late as possible. The last batch is everything that nothing else depends upon, the next to last batch is everything that only the last batch depends upon, and so forth.

For the example, the following buildorder is generated

Batch 0: libdatrie, graphite2, cairo
Batch 1: libthai, harfbuzz
Batch 2: pango

Downsides

  • It's fairly slow. In particular, the back-mapping from what gets installed along with the SRPM to the BuildRequires that caused that happen requires manually processing dependencies from libsolv.
  • Have to work hard to make the failure via cycles not confusing and obscure.

Improving on buildorder

The problem with buildorder: is small changes can renumber everything. Rather than specifying:

components:
    cairo: { ref: f27, buildorder: 0 } 
    graphite2: { ref: f27, buildorder: 0 }
    harfbuzz:  { ref: f27, buildorder: 1 }
    libdatrie: { ref: f27, buildorder: 0 }
    libthai:  { ref: f27, buildorder: 1 }
    pango:  { ref: f27, buildorder: 1 }

It would be better to specify something like:

components:
    cairo: { ref: f27 } 
    graphite2: { ref: f27 }
    harfbuzz:  { ref: f27, buildafter: [ cairo, graphite2 ] }
    libdatrie: { ref: f27 }
    libthai:  { ref: f27, buildafter: [ libdatrie ] }
    pango:  { ref: f27, buildafter: [ harfbuzz, libthai ] }

and let the module build service figure out the exact ordering and batches.


Tagging @ralph as I know the fm-orchestrator/MBS folks have also been thinking about this problem (e.g. in https://pagure.io/fm-orchestrator/issue/715 )

As additional potential sources of test cases, I'll point to https://github.com/sclorg/rhscl-rebuild-recipes and my own Python 3 bootstrapping recipe at https://github.com/ncoghlan/pyscl-devel/blob/master/rpmlb/python-recipe.yml

I think putting the buildorder algorithm into fedmod is the wrong approach. MBS is where all modules have to pass through, build ordering belongs there so that every module benefits from it, not only those that get created/updated with fedmod.

fedmod has a big piece of information that isn't available to MBS - the existing repodata for the global Fedora package set. The goal of rpm2module and (to be added) rpm2flatpak is to take existing packages from the global Fedora package set and modularize them.

Given that repodata, we can do an actual computation of build ordering.

MBS would have to fall back to some combination of:
Trial and error - which doesn't always work
Heuristic parsing of spec files to try and figure out dependencies
* Extra mock builds to try and figure out dependencies (see https://pagure.io/fm-orchestrator/issue/715#comment-504462)

It seems like a really hard problem.

Login to comment on this ticket.

Metadata