When the git log contained merges, the changelog generator didn't
necessarily work correctly. First, it would attempt to parse the message
in merge commits which was always wrong and always printed an error
message. Now, the changelog generator is told to just skip those commits
and move on since merge commits won't contain the message we're looking
for anyway.

The second happened with patch releases and feature branches that were
under work for more than a release. If a release was performed on the
maint branch and another commit was done on master before merging
maint back to master, the tree could be constructed in a way where the
previous version commit could not be clearly found such as in this
example from the `v0.31.0` release.

    * 3204dd2 ci: add codecov.io support to flux (#1273)
    *   f0b6850 Merge branch 'maint'
    |\
    | * ec37117 (tag: v0.30.0) feat(universe): support for dynamic queries (#1284)
    * | 2cce6fe feat(universe): support for dynamic queries (#1284)
    |/
    * 32fb2fb (tag: v0.29.0) fix(execute): properly use RefCount to reference count tables (#1216)

If the changelog generator encountered this pattern, it was undefined
exactly how it would behave. In this specific scenario, it parsed every
commit on the left branch going back to the root commit and then found
the previous commit on the second side of the branch. The opposite could
happen too and then it would have missed adding the commit from the
other branch.

The changelog generator has now been modified to find the merge base
before it attempts to read the log to prevent searching back through the
entire history. In the above situations, `v0.29.0` would be found and
act as a final constraint for parent commits.

The changelog parser has also been modified to use a breadth-first
search instead of depth-first search and it does not use the `Log()`
method from the go-git library. This is because the go-git library does
not have a way to stop at a specific commit and stop looking at parent
commits. We now limit the breadth-first search of commits to either the
merge base or the previous version commit which should be the same
commit if the release process was followed correctly. A breadth-first
search is used to ensure we get all commit messages from any merges and
don't skip any commits because we reached the terminating commit first.

When the git log contained merges, the changelog generator didn't
necessarily work correctly. First, it would attempt to parse the message
in merge commits which was always wrong and always printed an error
message. Now, the changelog generator is told to just skip those commits
and move on since merge commits won't contain the message we're looking
for anyway.

The second happened with patch releases and feature branches that were
under work for more than a release. If a release was performed on the
maint branch and another commit was done on master before merging
maint back to master, the tree could be constructed in a way where the
previous version commit could not be clearly found such as in this
example from the `v0.31.0` release.

    * 3204dd2 ci: add codecov.io support to flux (#1273)
    *   f0b6850 Merge branch 'maint'
    |\
    | * ec37117 (tag: v0.30.0) feat(universe): support for dynamic queries (#1284)
    * | 2cce6fe feat(universe): support for dynamic queries (#1284)
    |/
    * 32fb2fb (tag: v0.29.0) fix(execute): properly use RefCount to reference count tables (#1216)

If the changelog generator encountered this pattern, it was undefined
exactly how it would behave. In this specific scenario, it parsed every
commit on the left branch going back to the root commit and then found
the previous commit on the second side of the branch. The opposite could
happen too and then it would have missed adding the commit from the
other branch.

The changelog generator has now been modified to find the merge base
before it attempts to read the log to prevent searching back through the
entire history. In the above situations, `v0.29.0` would be found and
act as a final constraint for parent commits.

The changelog parser has also been modified to use a breadth-first
search instead of depth-first search and it does not use the `Log()`
method from the go-git library. This is because the go-git library does
not have a way to stop at a specific commit and stop looking at parent
commits. We now limit the breadth-first search of commits to either the
merge base or the previous version commit which should be the same
commit if the release process was followed correctly. A breadth-first
search is used to ensure we get all commit messages from any merges and
don't skip any commits because we reached the terminating commit first.

When the git log contained merges, the changelog generator didn't
necessarily work correctly. First, it would attempt to parse the message
in merge commits which was always wrong and always printed an error
message. Now, the changelog generator is told to just skip those commits
and move on since merge commits won't contain the message we're looking
for anyway.

The second happened with patch releases and feature branches that were
under work for more than a release. If a release was performed on the
maint branch and another commit was done on master before merging
maint back to master, the tree could be constructed in a way where the
previous version commit could not be clearly found such as in this
example from the `v0.31.0` release.

    * 3204dd2 ci: add codecov.io support to flux (#1273)
    *   f0b6850 Merge branch 'maint'
    |\
    | * ec37117 (tag: v0.30.0) feat(universe): support for dynamic queries (#1284)
    * | 2cce6fe feat(universe): support for dynamic queries (#1284)
    |/
    * 32fb2fb (tag: v0.29.0) fix(execute): properly use RefCount to reference count tables (#1216)

If the changelog generator encountered this pattern, it was undefined
exactly how it would behave. In this specific scenario, it parsed every
commit on the left branch going back to the root commit and then found
the previous commit on the second side of the branch. The opposite could
happen too and then it would have missed adding the commit from the
other branch.

The changelog generator has now been modified to find the merge base
before it attempts to read the log to prevent searching back through the
entire history. In the above situations, `v0.29.0` would be found and
act as a final constraint for parent commits.

The changelog parser has also been modified to use a breadth-first
search instead of depth-first search and it does not use the `Log()`
method from the go-git library. This is because the go-git library does
not have a way to stop at a specific commit and stop looking at parent
commits. We now limit the breadth-first search of commits to either the
merge base or the previous version commit which should be the same
commit if the release process was followed correctly. A breadth-first
search is used to ensure we get all commit messages from any merges and
don't skip any commits because we reached the terminating commit first.

…1319)

When the git log contained merges, the changelog generator didn't
necessarily work correctly. First, it would attempt to parse the message
in merge commits which was always wrong and always printed an error
message. Now, the changelog generator is told to just skip those commits
and move on since merge commits won't contain the message we're looking
for anyway.

The second happened with patch releases and feature branches that were
under work for more than a release. If a release was performed on the
maint branch and another commit was done on master before merging
maint back to master, the tree could be constructed in a way where the
previous version commit could not be clearly found such as in this
example from the `v0.31.0` release.

    * 3204dd2 ci: add codecov.io support to flux (#1273)
    *   f0b6850 Merge branch 'maint'
    |\
    | * ec37117 (tag: v0.30.0) feat(universe): support for dynamic queries (#1284)
    * | 2cce6fe feat(universe): support for dynamic queries (#1284)
    |/
    * 32fb2fb (tag: v0.29.0) fix(execute): properly use RefCount to reference count tables (#1216)

If the changelog generator encountered this pattern, it was undefined
exactly how it would behave. In this specific scenario, it parsed every
commit on the left branch going back to the root commit and then found
the previous commit on the second side of the branch. The opposite could
happen too and then it would have missed adding the commit from the
other branch.

The changelog generator has now been modified to find the merge base
before it attempts to read the log to prevent searching back through the
entire history. In the above situations, `v0.29.0` would be found and
act as a final constraint for parent commits.

The changelog parser has also been modified to use a breadth-first
search instead of depth-first search and it does not use the `Log()`
method from the go-git library. This is because the go-git library does
not have a way to stop at a specific commit and stop looking at parent
commits. We now limit the breadth-first search of commits to either the
merge base or the previous version commit which should be the same
commit if the release process was followed correctly. A breadth-first
search is used to ensure we get all commit messages from any merges and
don't skip any commits because we reached the terminating commit first.

…nfluxdata#1319)

When the git log contained merges, the changelog generator didn't
necessarily work correctly. First, it would attempt to parse the message
in merge commits which was always wrong and always printed an error
message. Now, the changelog generator is told to just skip those commits
and move on since merge commits won't contain the message we're looking
for anyway.

The second happened with patch releases and feature branches that were
under work for more than a release. If a release was performed on the
maint branch and another commit was done on master before merging
maint back to master, the tree could be constructed in a way where the
previous version commit could not be clearly found such as in this
example from the `v0.31.0` release.

    * 3204dd2 ci: add codecov.io support to flux (influxdata#1273)
    *   f0b6850 Merge branch 'maint'
    |\
    | * ec37117 (tag: v0.30.0) feat(universe): support for dynamic queries (influxdata#1284)
    * | 2cce6fe feat(universe): support for dynamic queries (influxdata#1284)
    |/
    * 32fb2fb (tag: v0.29.0) fix(execute): properly use RefCount to reference count tables (influxdata#1216)

If the changelog generator encountered this pattern, it was undefined
exactly how it would behave. In this specific scenario, it parsed every
commit on the left branch going back to the root commit and then found
the previous commit on the second side of the branch. The opposite could
happen too and then it would have missed adding the commit from the
other branch.

The changelog generator has now been modified to find the merge base
before it attempts to read the log to prevent searching back through the
entire history. In the above situations, `v0.29.0` would be found and
act as a final constraint for parent commits.

The changelog parser has also been modified to use a breadth-first
search instead of depth-first search and it does not use the `Log()`
method from the go-git library. This is because the go-git library does
not have a way to stop at a specific commit and stop looking at parent
commits. We now limit the breadth-first search of commits to either the
merge base or the previous version commit which should be the same
commit if the release process was followed correctly. A breadth-first
search is used to ensure we get all commit messages from any merges and
don't skip any commits because we reached the terminating commit first.