Bug-Tracking In Fossil


A bug-report in fossil is called a "ticket". Tickets are tracked separately from code check-ins.

Some other distributed bug-tracking systems store tickets as files within the source tree and thereby leverage the syncing and merging capabilities of the versioning system to sync and merge tickets. This approach is rejected in fossil for three reasons:

  1. Check-ins in fossil are immutable. So if tickets were part of the check-in, then there would be no way to add new tickets to a check-in as new bugs are discovered.
  1. Any project of reasonable size and complexity will generate thousands and thousands of tickets, and we do not want all those ticket files cluttering the source tree.
  1. We want tickets to be managed from the web interface and to have a permission system that is distinct from check-in permissions. In other words, we do not want to restrict the creation and editing of tickets to developers with check-in privileges and an installed copy of the fossil executable. Casual passers-by on the internet should be permitted to create tickets.

Recall that a fossil repository consists of an unordered collection of artifacts. (See the file format document for details.) Some artifacts have a special format, and among those are Ticket Change Artifacts. One or more ticket change artifacts are associated with each ticket. A ticket is created by a ticket change artifact. Each subsequent modification of the ticket is a separate artifact.

The "push", "pull", and "sync" algorithms share ticket change artifacts between repositories in the same way as every other artifact. In fact, the sync algorithm has no knowledge of the meaning of the artifacts it is syncing. As far as the sync algorithm is concerned, all artifacts are alike. After the sync has occurs, the individual repositories must make sense of the meaning of the various artifacts for themselves.

Interpretation Of Ticket Change Artifacts

Note: The following is implementation detail which can be safely ignored by casual users of fossil.

Every ticket change artifact contains (among other things)

The current state of a ticket is found by replaying all ticket change artifacts with the same ticket ID in timestamp order. For a given ticket, all values are initially NULL. As each ticket change artifact is encountered, values are either replaced or appended, according to a flag on the name/value pair. The current values for the fields of a ticket are the values that remain at the end of the replay process.

To create a new ticket, one inserts a ticket change artifact with a new ID. The ticket ID is a random 40-character lower-case hexadecimal number. The "tktnew" page in the fossil web interface creates new ticket IDs using a good source of randomness to insure uniqueness. The name/value pairs on the initial ticket change artifact are the initial values for the fields in the ticket.

Amending a ticket means simply creating a new artifact with the same ticket ID and with name/value pairs for those fields which are changing. Fields of the ticket which are not being modified should not appear as name/value pairs in the new artifact.

This approach to storing ticket state means that independently entered changes are automatically merged together when artifacts are shared between repositories. Tickets do not branch. This approach also makes it trivial to track the historic progression of changes to a ticket.

In order for this scheme to work, the system clocks on machines that add new ticket changes artifacts have to be set close to reality. It is OK for a ticket change artifact timestamp to be off by a few minutes or even a few hours. But if a timestamp on a ticket change artifact is off by months or years, it can seriously confuse the replay algorithm for determining the current ticket state. There are techniques available to cause rogue artifacts to be ignored by fossil. So if a ticket change artifact with a bad timestamp does get into a repository, it can be removed by an administrator. But the best approach is to take steps to insure that timestamps are approximately correct in the first place.

Local Configuration

The ticket change artifacts are part of the global state for a project. The global state is that which is shared between repositories during a sync. Each repository also has local state which is not (normally) shared. The name/value pairs for a ticket are part of global state, but the interpretation and display of that information is local state. Hence, each repository is free to set up its own ticket display and input formats and reporting rules according to its own particular needs.

Each repository defines its own TICKET table in its database. There is one row in the TICKET table for each unique ticket ID. The names of columns in the TICKET table correspond to the names in the name/value pairs of ticket change artifacts. When running the replay algorithm, if a name/value pair is seen which has no corresponding column in the TICKET table, then that name/value pair is ignored. Columns can be added or removed from the TICKET table at any time. Whenever the TICKET table is modified, the replay algorithm automatically reruns to repopulate the table using the new column names. Note that the TICKET table schema and content is part of the local state of a repository and is not shared with other repositories during a sync, push, or pull.

Each repository also defines scripts used to generate web pages for creating new tickets, viewing existing tickets, and modifying an existing ticket. These scripts consist of HTML with an embedded scripts written a Tcl-like language called "TH1". Every new fossil repository is created with default scripts. Administrators wishing to customize their ticket entry, viewing, and editing screens should modify the default scripts to suit their needs. These screen generator scripts are part of the local state of a repository and are not shared with other repositories during a sync, push, or pull.

To be continued...