We were indiscriminately retrieving the node's provenance data in the
`node_get()` function which was accounting for approximately a third of
the total runtime of `node_get()` which dominates pre-scheduler time in
`bst build`.  This change ameliorates that situation by only retrieving
the provenance data when it's actually needed.

Signed-off-by: Daniel Silverstone <daniel.silverstone@codethink.co.uk>

We know that nodes are typically one of:
str, list, dict, bool, tuple, NoneType or our ChainMap
Of these, dict and ChainMap are Mapping, only list is list
and the rest are returned unchanged.  We can reduce/defer our use
of isinstance here, dramatically, improving performance.

Signed-off-by: Daniel Silverstone <daniel.silverstone@codethink.co.uk>

A non-trivial proportion of time pre-scheduler in `bst build` is spent
copying (or chain-copying) nodes.  Approximately a quarter of the time
spent in that effort is in `isinstance()`.  This removes that CPU load.

Signed-off-by: Daniel Silverstone <daniel.silverstone@codethink.co.uk>

By re-using the isinstance replacements from earlier commits and
using a tuple of the string constants for checking for composition
markers, we reduce the cost of node_final_assertions by two thirds
in basic testing.

Signed-off-by: Daniel Silverstone <daniel.silverstone@codethink.co.uk>

While get_strict() doesn't look expensive per-se, it is called so many
times that it is valuable to cache the result once computed.  Since I
don't think it can change once it is computable, cache it immediately
that becomes possible and we save 20s in my test case.

Signed-off-by: Daniel Silverstone <daniel.silverstone@codethink.co.uk>

Further optimisations

See merge request !1131

Since junction check-outs are now persistent, there is no need to clean
up temporary directories once finished.
Now, junction checkouts are stored within the top-level project dir and
will be cleaned-up when that directory is removed.

This involves changes in:
* loader.py: Remove cleanup logic and passing-around of tmpdir.
* _project.py: Remove passing around tmpdir and calling loader's
               cleanup.

Stage junctions into .bst instead of a tmpdir

Closes #895

See merge request !1134

Valentin David authored Feb 11, 2019 and Valentin David committed Feb 12, 2019

This was broken by 5ef19a0b.

buildstream/_cas/cascache.py: Set 0644 rights to pulled files

See merge request !1144

This commit removes the method ArtifactCache.get_artifact_fullname()
and replaces it with Element.get_artifact_name()

Given a key, we are now able to construct the full name of any of an
element's artifacts.

This object should be used when we want to handle artifact refs directly
from the command line.
An ArtifactElementError has also been added to _exceptions.py

* There is no need for this method to use a cas object.
* Search for artifact globs in the project's element path
* An artifact key is always 64 chars long

A CasBasedDirectory object of an artifacts logs can be obtained
with ArtifactCache.get_artifacts_log(). This ultimately calls
CASCache.get_top_level_dir() to obtain a CasBasedDirectory
of an artifact's subdirectory (or subdirectories).

The loading of elements and the handling of artifacts does not belong
in this module. Such logic should be invoked using the Stream API

This method has been moved to Stream, where is it used there exclusively.

This commit ensures that CASCache.list_refs(), and
ArtifactCache.list_artifacts(), can both handle glob expressions.

_classify_artifacts() no longer filters glob expressions by first
obtaining a list of ALL refs locally cached. We now only obtain refs
specified by the glob expression.

Furthermore, the Project.element_path is used to start searching for
globbed elements as opposed to the Project.directory.

Refactor artifact log command

See merge request !1101

Most split rules already included the relevant directories themselves in
addition to the directory contents.  Add the missing bin, sbin, and
libexec directories.

This is required to fix tests with the following commit that changes
list_relative_paths() to return all directories.

This is required to fix tests with the following commit that changes
list_relative_paths() to return all directories.

Returning only empty directories leads to inconsistencies when computing
a manifest by combining results from multiple list_relative_paths()
calls as done by the compose plugin.

I.e., the same directory may be empty in one dependency and non-empty in
another dependency.  The merged file list will still contain that
directory even though it's no longer empty.

This inconsistency causes problems when calculating differences between
manifests. Returning all directories fixes these inconsistencies.

This is a change in API behavior.

This matches the change in utils.list_relative_paths().

Return all directories in list_relative_paths()

See merge request !1139

_context.py: Add cache_buildtrees global user context, the default
of which is set to by default to 'always' via the addition of
cache-buildtrees to userconfig.yaml cache group. 'failure' & 'never'
can be given as valid options.

app.py & cli.py: Add --cache-buildtrees as a bst main option, which
when passed with a valid option can override the default or user
defined context for cache_buildtrees.

tests/completions/completions.py: Update for the added flag.

not _cached_sucess() could resolve to true if the element wasn't
cached at all. switch to _cached_failure() to ensure condition
reflects expected artifact state

Artifacts can be cached explicitly with an empty `build tree` when
built via the cli main options or user config for all or only
successful build artifacts. Default behaviour is to still create
and cache all expected buildtrees.

element.py: _cache_artifact() Check if context for cache_buildtrees
has been set to always or failure with a corresponding build
result, if not skip attempting to export the build-root. Element
types without a build-root are cached with an empty buildtree
regardless. Update _stage_sources_at() to warn the user that the
buildtree import is empty.

tests/integration: Add test to artifact.py for the optional caching
of buildtree content from bst build. Rename build-tree.py to
shellbuildtrees.py to reflect included test cases, add test for
empty buildtree warning and failure option.

NEWS: Add entry for new option.

Optional creation of buildtrees

Closes #896

See merge request !1135

This fix a problem with the garbage collector not being able to
clean the MetaElements that are loaded.

On small projects this is not a problem, but in bigger projects, this
can save a few hundred of MBs at runtime

The reason behind this is, whenever we have a "stack" element, which
has no stack.yaml configuration, since it doesn't need it, we would
get an exception thrown when initiating the first one, as loading the
yaml file would fail.

This would capture the frame in which this command was executed, which
references meta_elements. Therefore, as long as another exception is not
thrown, the garbage collector would not be able to clean all the
MetaElements.

Cleanup MetaElement local state

See merge request !1147

This affects the cache key version (updated to 7) and introduces
a dependency on `ujson` which is BSD licenced as of the version
locked in `requirements.txt`

Signed-off-by: Daniel Silverstone <daniel.silverstone@codethink.co.uk>

Update cache keys to use JSON

See merge request !1151

We can easily rebiuld MetaSource from any source so there is no
reason to keep them around.

This will slightly improve memory usage.

Don't keep MetaSource around in Source

See merge request !1150

Resolving symlinks during staging causes various issues:

* Split rules may not work properly as the resolved paths will differ
  depending on whether another artifact with a directory symlink has
  been staged in the same root directory or not, e.g., as part of
  compose.

* The order of symlinks in file lists is difficult to get right to
  guarantee consistent and predictable behavior as paths in a file list
  might rely on symlinks in the same file list. See #647 and #817.

* Staging order differences can lead to surprising results. See #390.

* Difficult to properly support absolute symlinks. Absolute symlinks are
  currently converted to relative symlinks, however, this doesn't always
  work. See #606 and #830.

This will require changes in projects that rely on the current behavior.
However, the changes are expected to be small and are often a sign of
buggy element files. E.g., elements that don't fully obey `bindir` or
`sbindir` variables.

This matches the change in utils._process_list().

This also removes the _Resolver class as it is now unused. We may want
to support controlled symlink resolution in the future, in which case
the _Resolver class can be resurrected from this commit.

Copy symlinks as they are, absolute or relative. We no longer resolve
symlinks when copying files, which makes this safe.

Do not resolve or mangle symlinks during staging

See merge request !1140

Chandan Singh authored Feb 13, 2019 and Chandan Singh committed Feb 15, 2019

This script leverages the recently added format strings
(`%{build-deps}`, `%{runtime-deps}`) to `bst show` to print a graph in
DOT format. This requires users to have the `graphviz` python package
installed.

Additionally, users can also render the graph using the `--format`
option if they have the `graphviz` command line tool installed.

contrib/bst-graph: Add script to print graph in DOT format

Closes #705

See merge request !1148

Rather than constantly using regular expressions and retrieval from YAML
nodes, pre-parse expansion strings into a list representation cached
for reuse, and then expand them as simple string concatenation.

Signed-off-by: Daniel Silverstone <daniel.silverstone@codethink.co.uk>

Variables: Rework how expansion strings work

See merge request !1152

Angelos Evripiotis authored Feb 14, 2019 and Angelos Evripiotis committed Feb 15, 2019

Change the default value of mtime when doing export_to_tar() from `0` to
`_utils._magic_timestamp`. This avoids problems in other software, which
assume that an mtime of `0` means the file does not exist.

There is more than one example where files with an mtime of zero are
treated as non-existant by other software. e.g. [ninja][1] and [Template
Toolkit][2].

The OSTree project also [express an intention][3] to move from an mtime
of 0 to an mtime of 1:

> For this reason, OSTree acts as though all timestamps are set to
> time_t 0, so that comparisons will be considered up-to-date. Note that
> for a few releases, OSTree used 1 to fix warnings such as GNU Tar
> emitting "implausibly old time stamp" with 0; however, until we have a
> mechanism to transition cleanly to 1, for compatibilty OSTree is
> reverted to use zero again.

From the comments on export_to_tar(), the motivation for having an mtime
of 0 was to have reproducible results, rather than it specifically being
the value 0.

Additionally, the reproducible builds project has a [page on
archives][4]; it mentions the benefits of setting all the files to have
the same mtime, or clamping the mtime. It makes no mention of a
motivation for the mtime to be specifically 0.

Fixes #914

[1]: https://github.com/ninja-build/ninja/issues/1120
[2]: https://github.com/abw/Template2/blob/8d7d37200af436f1ad43628278d3caad257c8e27/lib/Template/Provider.pm#L635
[3]: https://ostree.readthedocs.io/en/latest/manual/repo/
[4]: https://reproducible-builds.org/docs/archives/

storage.Directory.export_to_tar: default mtime=utils._magic_timestamp

Closes #914

See merge request !1149

Use sets when checking for existence of an element

See merge request !1154

Angelos Evripiotis authored Feb 05, 2019 and Angelos Evripiotis committed Feb 15, 2019

Remove the need for the 'really-workspace-close-project-inaccessible'
config option, as well as the option itself.

As agreed on the mailing list [1], all the 'are you sure?' prompts on
workspace reset and close were removed. While that discussion was going
on, this new prompt and option was added. At the 2019 BuildStream
Gathering, it was verbally agreed between myself and Tristan VB that we
would also remove this instance.

It was also agreed that we should have a notice to let the user know
what they'd done, this was already in place if interactive. Moved it to
be unconditional so that there's no difference in non-interactive
behaviour. Made it output to stderr, as it's diagnostic meant for the
user. Made it the last thing echo'd so it's next to the prompt - it's
very relevant to what they type next. Added a test to make sure the text
makes it to stderr in the appropriate case, and not in an inappropriate
one.

This is the last instance of any prompt configuration, so BuildStream
can also forget all of that machinery.

[1] https://mail.gnome.org/archives/buildstream-list/2018-December/msg00111.html

userconfig: rm really-workspace-close-project-inaccessible

Closes #726 and #744

See merge request !1130

See buildstream-docker-images#26
for detailed discussion around this.

`buildstream/buildstream-fedora` is now considered deprecated. Switch to
`buildstream/buildstream` image. This image also offers more tags that
will provide users more flexibility.

Now that the `buildstream/buildstream` image has 9 variants, let's make
it easier to choose the desired tag, using a command-line option.

This is otherwise possible by specifying the full image name `image:tag`
using the `-i` option. But, this will make it easier to specify just the
tag using `-j`.

The following two invocations of `bst-here` are now equivalent:

    bst-here -i buildstream/buildstream:dev
    bst-here -j dev

contrib/bst-here: Allow users to specify image variant

See merge request !1153

[ci skip]

README.rst: Add table with distros with packaged buildstream

See merge request !1143

So that other people can also make releases.

CONTRIBUTING.rst: Documenting the release process

See merge request !1155

The directory needs to be serialized after the Digest for the
subdirectory `caller` has been updated.

ArtifactCache.commit() is no longer the only expensive part and even
more so with the upcoming change to use CASBasedDirectory.

Cache artifacts with virtual directories instead of filesystem.

Closes #787

See merge request !991

Having `pkg-resources` in requirements files is never the right thing as
it is not really a package that one can install. This is basically an
artifact of Debian-based systems providing incorrect metadata to pip.
See [this pip issue](https://github.com/pypa/pip/issues/4022) and
[this Ubuntu issue](https://bugs.launchpad.net/ubuntu/+source/python-pip/+bug/1635463)
for more background.

requirements/Makefile: exclude pkg-resources

Closes #918

See merge request !1156

Tristan Van Berkom authored Feb 18, 2019 and Tristan Van Berkom committed Feb 18, 2019

The test jobs which create coverage reports need to be listed
in the dependencies of the coverage job, as highlighted in
issue #356.

.gitlab-ci.yml: Some new test environments were added, but not collecting coverage.

See merge request !1157

This will allow cross-junction dependencies to be listed as strings on a
single line.

As part of this, some logic around initializing `Dependency()` objects
have been moved out of `_extract_depends_from_node()` method into the
constructor of `Dependency()` class, to keep all related code in one
place.

* _loader/types.py: While initializing `Dependency` objects, attempt to
                    split filenames, only if no `junction` was specified
                    explicitly. If a `junction` was specified, then
                    filenames with `:` in their names will result in an
                    error.

* _loader/loadelement.py: Refactor logic to initialize `Dependency()`
                          objects to move it to the `Dependency()`
                          constructor.

* tests/frontend/buildcheckout.py: Add tests to ensure the above.

* _versions.py: Bump BST_FORMAT_VERSION.

Fixes #809.

In the previous commit, we added support to express cross-junction
dependencies inline as simple strings. Document it along with the
version in which the feature was added.