Use Guice-Servlet rather than custom code and expose the Injector
This is a fairly functional variation of Gitblit with one notable
exception:
The security filters are not working properly.
This is a design flaw in Guice that I have reported upstream [1]. The
general idea is that Guice-Servlet filters are not properly wrapping the
ServletRequest. This has historically been a problem for Guice-Servlet
servlets but Google has fixed most of those issues. Unfortunately, all
the same flaws reported against the servlet delegation also exist in
Guice-Servlet filter delegation. :(
[1]: https://code.google.com/p/google-guice/issues/detail?id=807
Handle ssh keys as objects, not strings, and improve the ls and rm key commands
"gitblit keys ls" now defaults to showing an indexed list of fingerprints which almost matches the output of "sshadd -l". The indexes are useful specifying key(s) to remove using "gitblit keys rm <index>". This is an important improvement for key management.
A basic issue tracker styled as a hybrid of GitHub and BitBucket issues.
You may attach commits to an existing ticket or you can push a single
commit to create a *proposal* ticket.
Tickets keep track of patchsets (one or more commits) and allow patchset
rewriting (rebase, amend, squash) by detecing the non-fast-forward
update and assigning a new patchset number to the new commits.
Ticket tracker
--------------
The ticket tracker stores tickets as an append-only journal of changes.
The journals are deserialized and a ticket is built by applying the
journal entries. Tickets are indexed using Apache Lucene and all
queries and searches are executed against this Lucene index.
There is one trade-off to this persistence design: user attributions are
non-relational.
What does that mean? Each journal entry stores the username of the
author. If the username changes in the user service, the journal entry
will not reflect that change because the values are hard-coded.
Here are a few reasons/justifications for this design choice:
1. commit identifications (author, committer, tagger) are non-relational
2. maintains the KISS principle
3. your favorite text editor can still be your administration tool
Persistence Choices
-------------------
**FileTicketService**: stores journals on the filesystem
**BranchTicketService**: stores journals on an orphan branch
**RedisTicketService**: stores journals in a Redis key-value datastore
It should be relatively straight-forward to develop other backends
(MongoDB, etc) as long as the journal design is preserved.
Pushing Commits
---------------
Each push to a ticket is identified as a patchset revision. A patchset
revision may add commits to the patchset (fast-forward) OR a patchset
revision may rewrite history (rebase, squash, rebase+squash, or amend).
Patchset authors should not be afraid to polish, revise, and rewrite
their code before merging into the proposed branch.
Gitblit will create one ref for each patchset. These refs are updated
for fast-forward pushes or created for rewrites. They are formatted as
`refs/tickets/{shard}/{id}/{patchset}`. The *shard* is the last two
digits of the id. If the id < 10, prefix a 0. The *shard* is always
two digits long. The shard's purpose is to ensure Gitblit doesn't
exceed any filesystem directory limits for file creation.
**Creating a Proposal Ticket**
You may create a new change proposal ticket just by pushing a **single
commit** to `refs/for/{branch}` where branch is the proposed integration
branch OR `refs/for/new` or `refs/for/default` which both will use the
default repository branch.
git push origin HEAD:refs/for/new
**Updating a Patchset**
The safe way to update an existing patchset is to push to the patchset
ref.
git push origin HEAD:refs/heads/ticket/{id}
This ensures you do not accidentally create a new patchset in the event
that the patchset was updated after you last pulled.
The not-so-safe way to update an existing patchset is to push using the
magic ref.
git push origin HEAD:refs/for/{id}
This push ref will update an exisitng patchset OR create a new patchset
if the update is non-fast-forward.
**Rebasing, Squashing, Amending**
Gitblit makes rebasing, squashing, and amending patchsets easy.
Normally, pushing a non-fast-forward update would require rewind (RW+)
repository permissions. Gitblit provides a magic ref which will allow
ticket participants to rewrite a ticket patchset as long as the ticket
is open.
git push origin HEAD:refs/for/{id}
Pushing changes to this ref allows the patchset authors to rebase,
squash, or amend the patchset commits without requiring client-side use
of the *--force* flag on push AND without requiring RW+ permission to
the repository. Since each patchset is tracked with a ref it is easy to
recover from accidental non-fast-forward updates.
Features
--------
- Ticket tracker with status changes and responsible assignments
- Patchset revision scoring mechanism
- Update/Rewrite patchset handling
- Close-on-push detection
- Server-side Merge button for simple merges
- Comments with Markdown syntax support
- Rich mail notifications
- Voting
- Mentions
- Watch lists
- Querying
- Searches
- Partial miletones support
- Multiple backend options
Fixed and introduced tests for synching ldap users and groups.
Using new settings key realm.ldap.synchronize
Switched from key String to Keys class. To avoid letting tests pass with
hardcoded keys wich doesn't exist anymore. Now on Key Refactorings the
test gets compile error again.
Test the isReady behavior from LdapSyncService.
testRevlog was expecting 12 commits but getting 19. Looking at the
hello-world repository on github (on which the test executes) the 19
commit count appears to be correct.
Define manager interfaces and update all of Gitblit to use managers
These manager interfaces define how the GitBlit singleton will
eventually be split into smaller component managers. The Wicket app and
all servlets have been updated to request the needed managers. There
are _very_ few method signature changes - although there are a handful.
This is a surgical sharding of responsibility based on a proof of
concept refactor. Instead of random references to GittBlit.self()
there are now precise references to the manager interface required to
accomplish some task. Some tasks may require references to multiple
managers.
The code is now littered with calls to GitBlit.getManager(class) and
those familiar with the code-base will no doubt notice the duplication
of methods from IUserService in IUserManager and the addition of
implementation methods in the GitBlit context class. When the
GitBlit class is broken apart and the existing external authentication
user service classes are refactored to AuthenticationService classes,
this will again simplify and flatten. But in order to safely and
cleanly modularize the stable code-base we will have to live with a
little duplication for a short while.
Change-Id: I7314ec8acaab2dcc6092785ed4434cc09fdbbe16
James Moger
manisha
Kevin Walter
David Ostrovsky
Jani Averbach
Alfred Schmid
Alex Lewis