![[Software Carpentry logo]](images/logo-software-carpentry-standard.png){kind=logo}
![[ACL Logo]](images/logo-acl-medium.png){kind=logo}
|
![[CodeSourcery Logo]](images/logo-codesourcery-medium.png){kind=logo}
|
|
|
|
|||
|
Copyright (c) 2000 Ka-Ping Yee. This material may
be distributed only subject to the terms and conditions set forth in
the Software Carpentry Open Publication License, which is available at:
|
||||
We propose an issue-tracking system called Roundup, which will manage a number of issues (with properties such as "description", "priority", and so on) and provide the ability to (a) submit new issues, (b) find and edit existing issues, and (c) discuss issues with other participants. The system will facilitate communication among the participants by managing discussions and notifying interested parties when issues are edited.
This design draws on experience from an existing implementation which we will refer to as "the Roundup prototype". The graphical interface we have in mind will resemble the main display of the prototype.
A typical software project requires the management of many tasks, usually distributed among several collaborators. In fact, any project team could use a tool for sorting out and discussing all the relevant issues. A common approach is to set up some kind of "to-do" list that people can share.
However, to address the overall problem we need much more than just a shared to-do list; we need to manage a growing body of knowledge and experience to help a team collaborate effectively on a project. The issue-tracking tool becomes a nexus for communication: the Grand Central Station of the group intelligence.
The primary focus of this design is to help developers work together well, not to provide a customer service interface to the developers. This is not to say that the design is to be made unsuitable for customers to use. Rather, it is assumed that many of the same qualities that are good for supporting development (see below) are also good for non-developers using the system. Additional niceties for providing a safe or simplified interface to clients are intentionally deferred for later consideration.
A good issue-tracking system should have at least the following properties:
|
Simplicity. It is a strong requirement that the tool be accessible and understandable. It should be fairly obvious what different parts of the interface do, and the inner mechanisms should operate in ways that most users can easily predict.
Efficiency. We aim to optimize for minimum effort to do the most common operations, and best use of resources like screen real estate to maximize the amount of information that we summarize and present.
Generality. We try to avoid making unnecessary assumptions that would restrict the applicability of the tool. For example, there is no reason why one might not also want to use this tool to manage a design process, non-software projects, or organizational decisions.
Persistence. We prefer hiding or reclassifying information to deleting it. This helps support the collection of statistics later. If records are never destroyed, there is little danger in providing access to a larger community, and logging yields accountability, which may encourage better behaviour.
|
Roundup stores a number of items, each of which can have several properties and an associated discussion. The properties can be used to classify or search for items. The discussion is a sequence of e-mail messages. Each item is identified by a unique number, and has an activity log which records the time and content of edits made on its properties. The log stays fairly small since the design intentionally provides only small data types as item properties, and encourages anything large to be attached to e-mail where it becomes part of the discussion. The next section explains how items are organized.
|
Some systems try to alleviate this problem by allowing nodes to appear at multiple locations in a tree, as with "aliases" or "symbolic links" in a filesystem, for example. This does help somewhat, but we want to be even more flexible by allowing the organization of nodes into sets that may freely intersect. Rather than putting each node at exactly one place in an overall "grand scheme", a node can belong to as many sets as are appropriate. If we choose to represent the sets themselves as nodes and set membership as a link between nodes, we're now ready to present the definition of a hyperdatabase.
|
There are several reasons for building our own kind of database for Roundup rather than using an existing one. Requiring the installation of a full-blown third-party SQL database system would probably deter many potential users from attempting to set up Roundup; yet a real relational database would be too complicated to implement on our own. On the other hand, a hyperdatabase can be implemented fairly easily using one of the Python DBM modules, so we can take the "batteries-included" approach and provide it as part of the system. It's easier to build and understand than a true relational database (in accordance with our guiding principle of simplicity), but provides most of the query functionality we want.
A hyperdatabase is well suited for finding the intersection of a number of sets in which items belong. We expect that most of the queries people want to do will be of this form, rather than complicated SQL queries. For example, a typical request might be "show me all critical items related to security". The ability to store arbitrary key-value pairs and links on nodes gives it more flexibility than an RDBMS. Users are not going to be making thousands of queries per second, so it makes sense to optimize for simplicity and flexibility rather than performance.
For our application, we store each item as a node in a hyperdatabase. The item's properties are stored as key-value pairs on its node. Four types of properties are allowed: string, date, choice, and reference.
The string type is for short, free-form strings. String properties are not intended to contain large amounts of text, and it is recommended that they be presented as one-line fields to encourage brevity.
The date type is for calendar dates and times.
The choice type denotes a single selection from a number of options. A choice property entails a link from the node possessing the property to the node representing the chosen option.
The reference type is for a list of links to any number of other nodes in the in the database. A reference property, for example, can be used to refer to related items or topic categories relevant to an item.
For Roundup, all items have five properties that are not customizable:
The description property is a short one-line description of the item. The detailed description can go in the first e-mail message of the item's discussion spool.
The superseder property is used to support the splitting, joining, or replacing of items. When several items need to be joined into a single item, all the old items link to the new item in their superseder property. When an item needs to be split apart, the item references all the new items in its superseder propety. We can easily list all active items just by checking for an empty superseder property, and trace the path of an item's origins by querying the hyperdatabase for links.
The nosy property contains a list of the people who are interested in an item. This mechanism is explained in the section on Nosy Lists.
The creation property records the item's creation time. The activity property records the last time that the item was edited or a mail message was added to its discussion spool. These two properties are managed by Roundup and are not available to be edited like other properties.
Users of the system are also represented by nodes in the hyperdatabase, containing properties like the user's e-mail address, login name, and password.
|
It is hoped that the hyperdatabase together with the specializations mentioned above for Roundup will be applicable in a variety of situations (in accordance with our guiding principle of generality).
To address the problem at hand, we need
a specific schema for items applied particularly to software development.
Again, we are trying to keep the schema simple: too many
options make it tougher for someone to make a good choice.
The schema is written here in the same form that it would
appear in a configuration file.
fixer = Reference() # people who will fix the problem
topic = Reference() # relevant topic keywords
priority = Choice("critical", # panic: work is stopped!
"urgent", # important, but not deadly
"bug", # lost work or incorrect results
"feature", # want missing functionality
"wish") # avoidable bugs, missing conveniences
status = Choice("unread", # submitted but no action yet
"deferred", # intentionally set aside
"chatting", # under review or seeking clarification
"need-eg", # need a reproducible example of a bug
"in-progress", # understood; development in progress
"testing", # we think it's done; others, please test
"done-cbb", # okay for now, but could be better
"resolved") # fix has been released
The fixer property assigns responsibility for an item to a person or a list of people. The topic property places the item in an arbitrary number of relevant topic sets (see the section on Browsing and Searching).
As previously mentioned, each item gets an activity log. Whenever a property on an item is changed, the log records the time of the change, the user making the change, and the old and new values of the property. This permits the later gathering of statistics (for example, the average time from submission to resolution).
We do not specify or enforce a state transition graph, since making the system rigid in that fashion is probably more trouble than it's worth. Experience has shown that there are probably two convenient automatic state transitions:
Roundup provides its services through two main interfaces: e-mail and the Web. This division is chosen to optimize the most common tasks.
E-mail is best suited for the submission of new items since most people are most comfortable with composing long messages in their own favourite e-mail client. E-mail also permits them to mention URLs or attach files relevant to their submission. Indeed, in many cases people are already used to making requests by sending e-mail to a mailing list of people; they can do exactly the same thing to use Roundup without even thinking about it. Similarly, people are already familiar with holding discussions in e-mail, and plenty of valuable usage conventions and software tools already exist for that medium.
The Web, on the other hand, is best suited for summarizing and seeking information, because it can present an interactive overview of items. Since the Web has forms, it's also the best place to edit items.
|
The system needs an address for receiving mail and an address that forwards mail to all participants. Each item has its own list of interested parties, known as its nosy list. Here's how nosy lists work:
The effect is like each item having its own little mailing list, except that no one ever has to worry about subscribing to anything. Indicating interest in an issue is sufficient, and if you want to bring someone new into the conversation, all you need to do is Cc: a message to them. It turns out that no one ever has to worry about unsubscribing, either: the nosy lists are so specific in scope that the conversation tends to die down by itself when the issue is resolved or people no longer find it sufficiently important.
Each nosy list is like an asynchronous chat room, lasting only a short time (typically five or ten messages) and involving a small group of people. However, that group is the right group of people: only those who express interest in an item in some way ever end up on the list, so no one gets spammed with mail they don't care about, and no one who wants to see mail about a particular item needs to be left out, for they can easily join in, and just as easily look at the mail spool on an item to catch up on any messages they might have missed.
We can take this a step further and permit users to monitor particular topics or classifications of items by allowing other kinds of nodes to also have their own nosy lists. For example, a manager could be on the nosy list of the priority value node for "critical", or a developer could be on the nosy list of the topic value node for "security". The recipients are then determined by the union of the nosy lists on the item and all the nodes it links to.
Using many small, specific mailing lists results in much more effective communication than one big list. Taking away the effort of subscribing and unsubscribing gives these lists the "feel" of being cheap and disposable. The transparent capture of the mail spool attached to each issue also yields a nice knowledge repository over time.
Since Roundup is intended to support arbitrary user-defined
schema for item properties, the editing interface must be
automatically generated from the schema. The configuration for
Roundup will include a template describing how to lay out the
properties to present a UI for inspecting and editing items.
For example:
<table width="100%">
<tr><td align=right>Description:</td>
<td><?property description size=70></td></tr>
<tr><td align=right>Status:</td>
<td><?property status></td></tr>
</table>
To display the editing form for an item, Roundup substitutes an HTML form widget for each <?property ...> tag, and transfers attributes (such as size=70 in the above example) from the processing tag to the form widget's tag. Each type has its own appropriate editing widget:
We foresee the use of custom date fields for things like deadlines, so input fields for date properties should support some simple way of specifying relative dates (such as "three weeks from now").
The superseder property is a special case: although it is more efficient to store a superseder property in the superseded item, it makes more sense to provide a "supersedes" edit field on the superseding item. So we need a special widget on items for this purpose (perhaps something as simple as a text field containing a comma-separated list of item numbers will do). Links in the superseder property should appear on both the superseding and superseded items to facilitate navigating an item's pedigree.
After the editing widgets, the item inspection page shows a "note" text box and then a display of the messages in the discussion spool, like the Roundup prototype. This field lets you enter a note explaining your change when you edit the item, and the note is included in the notification message that goes out to tell the interested parties on the nosy list of your edits.
The ideal we would like to achieve is to make searching as much like browsing as possible: the user simply clicks about on things that seem interesting, and the information narrows down comfortably until the goal is in sight. This is preferable to trying to digest a screen filled with widgets and buttons or entering a search expression in some arcane algebraic syntax.
|
The browsing interface presents filtering functionality for each of the properties in the schema. As with editing, the interface is generated from a template describing how to lay out the properties. Each type of property has its own appropriate filtering widget:
For a reference property like topic, one possibility is to show, as hyperlinks, the keywords whose sets have non-empty intersections with the currently displayed set of items. Sorting the keywords by popularity seems reasonable. Clicking on a keyword then narrows both the list of items and the list of keywords. This gives some of the feel of walking around a directory tree -- but without the restriction of having to select keywords in a particular hierarchical order, and without the need to travel all the way to the leaves of the tree before any items are visible.
Below the filtering form is a listing of items, with their properties displayed in a table. Rows in the table can also be generated from a template, as with the editing interface. This listing is the central overview of the system, and it should aim to maximize the density of useful information in accordance with our guiding principle of efficiency. For example, Bugzilla initially displays seven or eight items of the index, but only after the user has waded through three bewildering screens of form widgets. Jitterbug can't even fit any items at all in the first screenful, as it's taken up by artwork and adminstrative debris. In contrast, in the Roundup prototype, 25 high-priority issues are immediately visible, with most of the screen space devoted to their descriptions. Colour indicates the status of each item to help the eye sift through the index quickly.
In both Jitterbug and Bugzilla, items are sorted by default by ID, a meaningless field. Sorting by ID puts the issues in order by ascending submission date, which banishes recent issues far away at the bottom of the list. The Roundup prototype sorts items in sections by priority, and then within sections by the date of last activity. This reveals at a glance where discussion is most active, and provides an easy way for anyone to move an issue up in the list.
The page produced by a given set of browsing options constitutes a view. The options should all be part of the query parameters in the URL so that views may be bookmarked. A view specifies:
On each sort key there is the option to use sections -- that is, instead of making the property's value a column of the table, each possible value for the property is displayed at the top of a section and all the items having that value for that property are grouped underneath. This avoids wasting screen space with redundant information.
We propose that our default view should be:
The starting URL for Roundup should immediately present the listing of items generated by this default view, with no preceding query screen.
The hyperdatabase is clearly a separable component which can be developed and tested independently to an API specification.
As soon as the API to the hyperdatabase is nailed down, the implementation of the Roundup database layer on top of the hyperdatabase can begin. (This refers to the data types and five fixed properties specific to Roundup.) This layer can also be tested separately.
When the interface to the Roundup hyperdatabase is ready, development can begin on the user interface. The mail handler and the Web interface can be developed in parallel and mostly independently of each other.
The mail handler can be set up for testing fairly easily: mail messages on its standard input can be synthesized; its output is outgoing mail, which can be captured by replacing the implementation of the "send mail" function; and its side effects appear in the hyperdatabase, which has a Python API.
The Web interface is not easily testable in its entirety, though the most important components of it can be unit tested, such as the component that translates a view specification into a list of items for display, and the component that performs replacements on templates to produce an editing or filtering interface.
The description of the hyperdatabase above avoids some issues regarding node typing that need to be better specified. It is conceivable that eventually Roundup could support multiple kinds of items with their own schemas.
To permit integration with external tools, it is probably a good idea to provide a command-line tool that exposes the hyperdatabase API. This tool will be left for a later phase of development and so isn't specified in detail here.
Generating the user interface from a template is like applying an XSL stylesheet to XML, and if there's a standard Python module for performing these transformations, we could use XML instead.
More thinking is needed to determine the best filtering interface for reference properties. The proposed interface works well for topic keywords, but it isn't clear what to do when there are too many keywords to display them all.
There has been a variety of reactions to the hyperdatabase from reviewers: some like it, some are neutral, and some would prefer a "standard" RDBMS solution. For those in the latter camp, note that it's still possible to build the Roundup database layer around an RDBMS if we really need to. The rest of the design, in particular the "nosy list" mechanism, remains intact.
The possibility of malice by registered users has been disregarded. The system is intended to be used by a co-operative group.
This design tries to address as many as possible of the suggested requirements mentioned on the contest page:
Roundup is an issue-tracking system that also functions as a communications center and a knowledge repository. It combines the strengths of e-mail and the Web to try to provide the best possible user interaction.
The use of a "hyperdatabase" as the core model for the knowledge repository gives us the flexibility to extend Roundup and apply it to a variety of domains by providing new item schemas and user-interface templates.
Roundup is self-contained and easy to set up, requiring only a webserver and a mailbox. No one needs to be root to configure the webserver or to install database software.
This design is based on an existing deployed prototype which has proven its strengths and revealed its weaknesses in heavy day-to-day use by a real development team.
My thanks are due to Christina Heyl, Jesse Vincent, Mark Miller, Christopher Simons, Jeff Dunmall, Wayne Gramlich, and Dean Tribble for reviewing this paper and contributing their suggestions.
| [Home] | [FAQ] | [License] | [Rules] | [Resources] | [Archives] |
|
|
|
This document presents a description of the components of the Roundup system and specifies their interfaces and behaviour in sufficient detail to guide an implementation. For the philosophy and rationale behind the Roundup design, see the first-round Software Carpentry submission for Roundup. This document fleshes out that design as well as specifying interfaces so that the components can be developed separately.
Lots of software design documents come with a picture of a cake. Everybody seems to like them. I also like cakes (i think they are tasty). So i, too, shall include a picture of a cake here.
|
E-mail Client |
Web Browser |
Detector Scripts |
Shell |
|
E-mail User Interface |
Web User Interface |
Detector Interface |
Command Interface |
|
Roundup Database Layer |
|||
|
Hyperdatabase Layer |
|||
|
Storage Layer |
|||
The colourful parts of the cake are part of our system; the faint grey parts of the cake are external components.
I will now proceed to forgo all table manners and eat from the bottom of the cake to the top. You may want to stand back a bit so you don't get covered in crumbs.
The lowest-level component to be implemented is the hyperdatabase. The hyperdatabase is intended to be a flexible data store that can hold configurable data in records which we call items.
The hyperdatabase is implemented on top of the storage layer, an external module for storing its data. The storage layer could be a third-party RDBMS; for a "batteries-included" distribution, implementing the hyperdatabase on the standard bsddb module is suggested.
Before we get into the hyperdatabase itself, we need a way of handling dates. The hyperdatabase module provides Timestamp objects for representing date-and-time stamps and Interval objects for representing date-and-time intervals.
As strings, date-and-time stamps are specified with the date in international standard format (yyyy-mm-dd) joined to the time (hh:mm:ss) by a period ("."). Dates in this form can be easily compared and are fairly readable when printed. An example of a valid stamp is "2000-06-24.13:03:59". We'll call this the "full date format". When Timestamp objects are printed as strings, they appear in the full date format with the time always given in GMT. The full date format is always exactly 19 characters long.
For user input, some partial forms are also permitted: the whole time or just the seconds may be omitted; and the whole date may be omitted or just the year may be omitted. If the time is given, the time is interpreted in the user's local time zone. The Date constructor takes care of these conversions. In the following examples, suppose that yyyy is the current year, mm is the current month, and dd is the current day of the month; and suppose that the user is on Eastern Standard Time.
Date intervals are specified using the suffixes "y", "m", and "d". The suffix "w" (for "week") means 7 days. Time intervals are specified in hh:mm:ss format (the seconds may be omitted, but the hours and minutes may not).
The Date class should understand simple date expressions of the form stamp + interval and stamp - interval. When adding or subtracting intervals involving months or years, the components are handled separately. For example, when evaluating "2000-06-25 + 1m 10d", we first add one month to get 2000-07-25, then add 10 days to get 2000-08-04 (rather than trying to decide whether 1m 10d means 38 or 40 or 41 days).
Here is an outline of the Date and Interval classes.
![[ACL Logo]](images//logo-acl-medium.png){kind=logo}
class Date:
def __init__(self, spec, offset):
"""Construct a date given a specification and a time zone offset.
'spec' is a full date or a partial form, with an optional
added or subtracted interval. 'offset' is the local time
zone offset from GMT in hours.
"""
def __add__(self, interval):
"""Add an interval to this date to produce another date."""
def __sub__(self, interval):
"""Subtract an interval from this date to produce another date."""
def __cmp__(self, other):
"""Compare this date to another date."""
def __str__(self):
"""Return this date as a string in the yyyy-mm-dd.hh:mm:ss format."""
def local(self, offset):
"""Return this date as yyyy-mm-dd.hh:mm:ss in a local time zone."""
class Interval:
def __init__(self, spec):
"""Construct an interval given a specification."""
def __cmp__(self, other):
"""Compare this interval to another interval."""
def __str__(self):
"""Return this interval as a string."""
Here are some examples of how these classes would behave in practice. For the following examples, assume that we are on Eastern Standard Time and the current local time is 19:34:02 on 25 June 2000.
>>> Date(".") <Date 2000-06-26.00:34:02> >>> _.local(-5) "2000-06-25.19:34:02" >>> Date(". + 2d") <Date 2000-06-28.00:34:02> >>> Date("1997-04-17", -5) <Date 1997-04-17.00:00:00> >>> Date("01-25", -5) <Date 2000-01-25.00:00:00> >>> Date("08-13.22:13", -5) <Date 2000-08-14.03:13:00> >>> Date("14:25", -5) <Date 2000-06-25.19:25:00> >>> Interval(" 3w 1 d 2:00") <Interval 22d 2:00> >>> Date(". + 2d") - Interval("3w") <Date 2000-06-07.00:34:02>
Items contain data in properties. To Python, these properties are presented as the key-value pairs of a dictionary. Each item belongs to a class which defines the names and types of its properties. The database permits the creation and modification of classes as well as items.
Each item has a numeric identifier which is unique among items in its class. The items are numbered sequentially within each class in order of creation, starting from 1. The designator for an item is a way to identify an item in the database, and consists of the name of the item's class concatenated with the item's numeric identifier.
For example, if "spam" and "eggs" are classes, the first item created in class "spam" has id 1 and designator "spam1". The first item created in class "eggs" also has id 1 but has the distinct designator "eggs1". Item designators are conventionally enclosed in square brackets when mentioned in plain text. This permits a casual mention of, say, "[patch37]" in an e-mail message to be turned into an active hyperlink.
Property names must begin with a letter.
A property may be one of five basic types:
None is also a permitted value for any of these property types. An attempt to store None into a String property stores the empty string; an attempt to store None into a Multilink property stores an empty list.
The hyperdb module provides property objects to designate the different kinds of properties. These objects are used when specifying what properties belong in classes.
class String:
def __init__(self):
"""An object designating a String property."""
class Date:
def __init__(self):
"""An object designating a Date property."""
class Link:
def __init__(self, classname):
"""An object designating a Link property that links to
items in a specified class."""
class Multilink:
def __init__(self, classname):
"""An object designating a Multilink property that links
to items in a specified class."""
Here is the interface provided by the hyperdatabase.
class Database:
"""A database for storing records containing flexible data types."""
def __init__(self, storagelocator, journaltag):
"""Open a hyperdatabase given a specifier to some storage.
The meaning of 'storagelocator' depends on the particular
implementation of the hyperdatabase. It could be a file name,
a directory path, a socket descriptor for a connection to a
database over the network, etc.
The 'journaltag' is a token that will be attached to the journal
entries for any edits done on the database. If 'journaltag' is
None, the database is opened in read-only mode: the Class.create(),
Class.set(), and Class.retire() methods are disabled.
"""
def __getattr__(self, classname):
"""A convenient way of calling self.getclass(classname)."""
def getclasses(self):
"""Return a list of the names of all existing classes."""
def getclass(self, classname):
"""Get the Class object representing a particular class.
If 'classname' is not a valid class name, a KeyError is raised.
"""
class Class:
"""The handle to a particular class of items in a hyperdatabase."""
def __init__(self, db, classname, **properties):
"""Create a new class with a given name and property specification.
'classname' must not collide with the name of an existing class,
or a ValueError is raised. The keyword arguments in 'properties'
must map names to property objects, or a TypeError is raised.
"""
# Editing items:
def create(self, **propvalues):
"""Create a new item of this class and return its id.
The keyword arguments in 'propvalues' map property names to values.
The values of arguments must be acceptable for the types of their
corresponding properties or a TypeError is raised. If this class
has a key property, it must be present and its value must not
collide with other key strings or a ValueError is raised. Any other
properties on this class that are missing from the 'propvalues'
dictionary are set to None. If an id in a link or multilink
property does not refer to a valid item, an IndexError is raised.
"""
def get(self, itemid, propname):
"""Get the value of a property on an existing item of this class.
'itemid' must be the id of an existing item of this class or an
IndexError is raised. 'propname' must be the name of a property
of this class or a KeyError is raised.
"""
def set(self, itemid, **propvalues):
"""Modify a property on an existing item of this class.
'itemid' must be the id of an existing item of this class or an
IndexError is raised. Each key in 'propvalues' must be the name
of a property of this class or a KeyError is raised. All values
in 'propvalues' must be acceptable types for their corresponding
properties or a TypeError is raised. If the value of the key
property is set, it must not collide with other key strings or a
ValueError is raised. If the value of a Link or Multilink
property contains an invalid item id, a ValueError is raised.
"""
def retire(self, itemid):
"""Retire an item.
The properties on the item remain available from the get() method,
and the item's id is never reused. Retired items are not returned
by the find(), list(), or lookup() methods, and other items may
reuse the values of their key properties.
"""
def history(self, itemid):
"""Retrieve the journal of edits on a particular item.
'itemid' must be the id of an existing item of this class or an
IndexError is raised.
The returned list contains tuples of the form
(date, tag, action, params)
'date' is a Timestamp object specifying the time of the change and
'tag' is the journaltag specified when the database was opened.
'action' may be:
'create' or 'set' -- 'params' is a dictionary of property values
'link' or 'unlink' -- 'params' is (classname, itemid, propname)
'retire' -- 'params' is None
"""
# Locating items:
def setkey(self, propname):
"""Select a String property of this class to be the key property.
'propname' must be the name of a String property of this class or
None, or a TypeError is raised. The values of the key property on
all existing items must be unique or a ValueError is raised.
"""
def getkey(self):
"""Return the name of the key property for this class or None."""
def lookup(self, keyvalue):
"""Locate a particular item by its key property and return its id.
If this class has no key property, a TypeError is raised. If the
'keyvalue' matches one of the values for the key property among
the items in this class, the matching item's id is returned;
otherwise a KeyError is raised.
"""
def find(self, propname, itemid):
"""Get the ids of items in this class which link to a given item.
'propname' must be the name of a property in this class, or a
KeyError is raised. That property must be a Link or Multilink
property, or a TypeError is raised. 'itemid' must be the id of
an existing item in the class linked to by the given property,
or an IndexError is raised.
"""
def list(self):
"""Return a list of the ids of the active items in this class."""
def count(self):
"""Get the number of items in this class.
If the returned integer is 'numitems', the ids of all the items
in this class run from 1 to numitems, and numitems+1 will be the
id of the next item to be created in this class.
"""
# Manipulating properties:
def getprops(self):
"""Return a dictionary mapping property names to property objects."""
def addprop(self, **properties):
"""Add properties to this class.
The keyword arguments in 'properties' must map names to property
objects, or a TypeError is raised. None of the keys in 'properties'
may collide with the names of existing properties, or a ValueError
is raised before any properties have been added.
"""Here is an example of how the hyperdatabase module would work in practice.
>>> import hyperdb >>> db = hyperdb.Database("foo.db", "ping") >>> db <hyperdb.Database "foo.db" opened by "ping"> >>> hyperdb.Class(db, "status", name=hyperdb.String()) <hyperdb.Class "status"> >>> _.setkey("name") >>> db.status.create(name="unread") 1 >>> db.status.create(name="in-progress") 2 >>> db.status.create(name="testing") 3 >>> db.status.create(name="resolved") 4 >>> db.status.count() 4 >>> db.status.list() [1, 2, 3, 4] >>> db.status.lookup("in-progress") 2 >>> db.status.retire(3) >>> db.status.list() [1, 2, 4] >>> hyperdb.Class(db, "issue", title=hyperdb.String(), status=hyperdb.Link("status")) <hyperdb.Class "issue"> >>> db.issue.create(title="spam", status=1) 1 >>> db.issue.create(title="eggs", status=2) 2 >>> db.issue.create(title="ham", status=4) 3 >>> db.issue.create(title="arguments", status=2) 4 >>> db.issue.create(title="abuse", status=1) 5 >>> hyperdb.Class(db, "user", username=hyperdb.Key(), password=hyperdb.String()) <hyperdb.Class "user"> >>> db.issue.addprop(fixer=hyperdb.Link("user")) >>> db.issue.getprops() {"title": <hyperdb.String>, "status": <hyperdb.Link to "status">, "user": <hyperdb.Link to "user">} >>> db.issue.set(5, status=2) >>> db.issue.get(5, "status") 2 >>> db.status.get(2, "name") "in-progress" >>> db.issue.get(5, "title") "abuse" >>> db.issue.find("status", db.status.lookup("in-progress")) [2, 4, 5] >>> db.issue.history(5) [(<Date 2000-06-28.19:09:43>, "ping", "create", {"title": "abuse", "status": 1}), (<Date 2000-06-28.19:11:04>, "ping", "set", {"status": 2})] >>> db.status.history(1) [(<Date 2000-06-28.19:09:43>, "ping", "link", ("issue", 5, "status")), (<Date 2000-06-28.19:11:04>, "ping", "unlink", ("issue", 5, "status"))] >>> db.status.history(2) [(<Date 2000-06-28.19:11:04>, "ping", "link", ("issue", 5, "status"))]
For the purposes of journalling, when a Multilink property is set to a new list of items, the hyperdatabase compares the old list to the new list. The journal records "unlink" events for all the items that appear in the old list but not the new list, and "link" events for all the items that appear in the new list but not in the old list.
The Roundup database layer is implemented on top of the hyperdatabase and mediates calls to the database. Some of the classes in the Roundup database are considered item classes. The Roundup database layer adds detectors and user items, and on items it provides mail spools, nosy lists, and superseders.
Internal to this layer we reserve three special classes of items that are not items.
Users are stored in the hyperdatabase as items of class "user". The "user" class has the definition:
hyperdb.Class(db, "user", username=hyperdb.String(),
password=hyperdb.String(),
address=hyperdb.String())
db.user.setkey("username")E-mail messages are represented by hyperdatabase items of class "msg". The actual text content of the messages is stored in separate files. (There's no advantage to be gained by stuffing them into the hyperdatabase, and if messages are stored in ordinary text files, they can be grepped from the command line.) The text of a message is saved in a file named after the message item designator (e.g. "msg23") for the sake of the command interface (see below). Attachments are stored separately and associated with "file" items. The "msg" class has the definition:
hyperdb.Class(db, "msg", author=hyperdb.Link("user"),
recipients=hyperdb.Multilink("user"),
date=hyperdb.Date(),
summary=hyperdb.String(),
files=hyperdb.Multilink("file"))The "author" property indicates the author of the message (a "user" item must exist in the hyperdatabase for any messages that are stored in the system). The "summary" property contains a summary of the message for display in a message index.
Submitted files are represented by hyperdatabase items of class "file". Like e-mail messages, the file content is stored in files outside the database, named after the file item designator (e.g. "file17"). The "file" class has the definition:
hyperdb.Class(db, "file", user=hyperdb.Link("user"),
name=hyperdb.String(),
type=hyperdb.String())The "user" property indicates the user who submitted the file, the "name" property holds the original name of the file, and the "type" property holds the MIME type of the file as received.
All items have the following standard properties:
title=hyperdb.String()
messages=hyperdb.Multilink("msg")
files=hyperdb.Multilink("file")
nosy=hyperdb.Multilink("user")
superseder=hyperdb.Multilink("item")Also, two Date properties named "creation" and "activity" are fabricated by the Roundup database layer. By "fabricated" we mean that no such properties are actually stored in the hyperdatabase, but when properties on items are requested, the "creation" and "activity" properties are made available. The value of the "creation" property is the date when an item was created, and the value of the "activity" property is the date when any property on the item was last edited (equivalently, these are the dates on the first and last records in the item's journal).
The interface to a Roundup database delegates most method calls to the hyperdatabase, except for the following changes and additional methods.
class Database:
# Overridden methods:
def __init__(self, storagelocator, journaltag):
"""When the Roundup database is opened by a particular user,
the 'journaltag' is the id of the user's "user" item."""
def getclass(self, classname):
"""This method now returns an instance of either Class or
ItemClass depending on whether an item class is specified."""
# New methods:
def getuid(self):
"""Return the id of the "user" item associated with the user
that owns this connection to the hyperdatabase."""
class Class:
# Overridden methods:
def create(self, **propvalues):
def set(self, **propvalues):
def retire(self, itemid):
"""These operations trigger detectors and can be vetoed. Attempts
to modify the "creation" or "activity" properties cause a KeyError.
"""
# New methods:
def audit(self, event, detector):
def react(self, event, detector):
"""Register a detector (see below for more details)."""
class ItemClass(Class):
# Overridden methods:
def __init__(self, db, classname, **properties):
"""The newly-created class automatically includes the "messages",
"files", "nosy", and "superseder" properties. If the 'properties'
dictionary attempts to specify any of these properties or a
"creation" or "activity" property, a ValueError is raised."""
def get(self, itemid, propname):
def getprops(self):
"""In addition to the actual properties on the item, these
methods provide the "creation" and "activity" properties."""
# New methods:
def addmessage(self, itemid, summary, text):
"""Add a message to an item's mail spool.
A new "msg" item is constructed using the current date, the
user that owns the database connection as the author, and
the specified summary text. The "files" and "recipients"
fields are left empty. The given text is saved as the body
of the message and the item is appended to the "messages"
field of the specified item.
"""
def nosymessage(self, itemid, msgid):
"""Send a message to the members of an item's nosy list.
The message is sent only to users on the nosy list who are not
already on the "recipients" list for the message. These users
are then added to the message's "recipients" list.
"""
The default schema included with Roundup turns it into a typical software bug tracker. The database is set up like this:
pri = Class(db, "priority", name=hyperdb.String(), order=hyperdb.String())
pri.setkey("name")
pri.create(name="critical", order="1")
pri.create(name="urgent", order="2")
pri.create(name="bug", order="3")
pri.create(name="feature", order="4")
pri.create(name="wish", order="5")
stat = Class(db, "status", name=hyperdb.String(), order=hyperdb.String())
stat.setkey("name")
stat.create(name="unread", order="1")
stat.create(name="deferred", order="2")
stat.create(name="chatting", order="3")
stat.create(name="need-eg", order="4")
stat.create(name="in-progress", order="5")
stat.create(name="testing", order="6")
stat.create(name="done-cbb", order="7")
stat.create(name="resolved", order="8")
Class(db, "keyword", name=hyperdb.String())
Class(db, "issue", fixer=hyperdb.Multilink("user"),
topic=hyperdb.Multilink("keyword"),
priority=hyperdb.Link("priority"),
status=hyperdb.Link("status"))
(The "order" property hasn't been explained yet. It gets used by the Web user interface for sorting.)
The above isn't as pretty-looking as the schema specification in the first-stage submission, but it could be made just as easy with the addition of a convenience function like Choice for setting up the "priority" and "status" classes:
def Choice(name, *options):
cl = Class(db, name, name=hyperdb.String(), order=hyperdb.String())
for i in range(len(options)):
cl.create(name=option[i], order=i)
return hyperdb.Link(name)
Detectors are Python functions that are triggered on certain kinds of events. The definitions of the functions live in Python modules placed in a directory set aside for this purpose. Importing the Roundup database module also imports all the modules in this directory, and the init() function of each module is called when a database is opened to provide it a chance to register its detectors.
There are two kinds of detectors:
When the Roundup database is about to perform a create(), set(), or retire() operation, it first calls any auditors that have been registered for that operation on that class. Any auditor may raise a Reject exception to abort the operation.
If none of the auditors raises an exception, the database proceeds to carry out the operation. After it's done, it then calls all of the reactors that have been registered for the operation.
The audit() and react() methods register detectors on a given class of items.
class Class:
def audit(self, event, detector):
"""Register an auditor on this class.
'event' should be one of "create", "set", or "retire".
'detector' should be a function accepting four arguments.
"""
def react(self, event, detector):
"""Register a reactor on this class.
'event' should be one of "create", "set", or "retire".
'detector' should be a function accepting four arguments.
"""
Auditors are called with the arguments:
where db is the database, cl is an instance of Class or ItemClass within the database, and newdata is a dictionary mapping property names to values. For a create() operation, the itemid argument is None and newdata contains all of the initial property values with which the item is about to be created. For a set() operation, newdata contains only the names and values of properties that are about to be changed. For a retire() operation, newdata is None.audit(db, cl, itemid, newdata)
Reactors are called with the arguments:
where db is the database, cl is an instance of Class or ItemClass within the database, and olddata is a dictionary mapping property names to values. For a create() operation, the itemid argument is the id of the newly-created item and olddata is None. For a set() operation, olddata contains the names and previous values of properties that were changed. For a retire() operation, itemid is the id of the retired item and olddata is None.react(db, cl, itemid, olddata)
Here is an example of detectors written for a hypothetical project-management application, where users can signal approval of a project by adding themselves to an "approvals" list, and a project proceeds when it has three approvals.
# Permit users only to add themselves to the "approvals" list.
def check_approvals(db, cl, id, newdata):
if newdata.has_key("approvals"):
if cl.get(id, "status") == db.status.lookup("approved"):
raise Reject, "You can't modify the approvals list " \
"for a project that has already been approved."
old = cl.get(id, "approvals")
new = newdata["approvals"]
for uid in old:
if uid not in new and uid != db.getuid():
raise Reject, "You can't remove other users from the "
"approvals list; you can only remove yourself."
for uid in new:
if uid not in old and uid != db.getuid():
raise Reject, "You can't add other users to the approvals "
"list; you can only add yourself."
# When three people have approved a project, change its
# status from "pending" to "approved".
def approve_project(db, cl, id, olddata):
if olddata.has_key("approvals") and len(cl.get(id, "approvals")) == 3:
if cl.get(id, "status") == db.status.lookup("pending"):
cl.set(id, status=db.status.lookup("approved"))
def init(db):
db.project.audit("set", check_approval)
db.project.react("set", approve_project)Here is another example of a detector that can allow or prevent the creation of new items. In this scenario, patches for a software project are submitted by sending in e-mail with an attached file, and we want to ensure that there are text/plain attachments on the message. The maintainer of the package can then apply the patch by setting its status to "applied".
# Only accept attempts to create new patches that come with patch files.
def check_new_patch(db, cl, id, newdata):
if not newdata["files"]:
raise Reject, "You can't submit a new patch without " \
"attaching a patch file."
for fileid in newdata["files"]:
if db.file.get(fileid, "type") != "text/plain":
raise Reject, "Submitted patch files must be text/plain."
# When the status is changed from "approved" to "applied", apply the patch.
def apply_patch(db, cl, id, olddata):
if cl.get(id, "status") == db.status.lookup("applied") and \
olddata["status"] == db.status.lookup("approved"):
# ...apply the patch...
def init(db):
db.patch.audit("create", check_new_patch)
db.patch.react("set", apply_patch)The command interface is a very simple and minimal interface, intended only for quick searches and checks from the shell prompt. (Anything more interesting can simply be written in Python using the Roundup database module.)
A single command, roundup, provides basic access to the hyperdatabase from the command line.
Property values are represented as strings in command arguments and in the printed results:
When multiple items are specified to the roundup get or roundup set commands, the specified properties are retrieved or set on all the listed items.
When multiple results are returned by the roundup get or roundup find commands, they are printed one per line (default) or joined by commas (with the -list) option.
To find all messages regarding in-progress issues that contain the word "spam", for example, you could execute the following command from the directory where the database dumps its files:
shell% for issue in `roundup find issue status=in-progress`; do > grep -l spam `roundup get $issue messages` > done msg23 msg49 msg50 msg61 shell%
Or, using the -list option, this can be written as a single command:
shell% grep -l spam `roundup get \ \`roundup find -list issue status=in-progress\` messages` msg23 msg49 msg50 msg61 shell%
The Roundup system must be assigned an e-mail address at which to receive mail. Messages should be piped to the Roundup mail-handling script by the mail delivery system (e.g. using an alias beginning with "|" for sendmail).
Incoming messages are examined for multiple parts. In a multipart/mixed message or part, each subpart is extracted and examined. In a multipart/alternative message or part, we look for a text/plain subpart and ignore the other parts. The text/plain subparts are assembled to form the textual body of the message, to be stored in the file associated with a "msg" class item. Any parts of other types are each stored in separate files and given "file" class items that are linked to the "msg" item.
The "summary" property on message items is taken from the first non-quoting section in the message body. The message body is divided into sections by blank lines. Sections where the second and all subsequent lines begin with a ">" or "|" character are considered "quoting sections". The first line of the first non-quoting section becomes the summary of the message.
All of the addresses in the To: and Cc: headers of the incoming message are looked up among the user items, and the corresponding users are placed in the "recipients" property on the new "msg" item. The address in the From: header similarly determines the "author" property of the new "msg" item. The default handling for addresses that don't have corresponding users is to create new users with no passwords and a username equal to the address. (The web interface does not permit logins for users with no passwords.) If we prefer to reject mail from outside sources, we can simply register an auditor on the "user" class that prevents the creation of user items with no passwords.
The subject line of the incoming message is examined to determine whether the message is an attempt to create a new item or to discuss an existing item. A designator enclosed in square brackets is sought as the first thing on the subject line (after skipping any "Fwd:" or "Re:" prefixes).
If an item designator (class name and id number) is found there, the newly created "msg" item is added to the "messages" property for that item, and any new "file" items are added to the "files" property for the item.
If just an item class name is found there, we attempt to create a new item of that class with its "messages" property initialized to contain the new "msg" item and its "files" property initialized to contain any new "file" items.
Both cases may trigger detectors (in the first case we are calling the set() method to add the message to the item's spool; in the second case we are calling the create() method to create a new item). If an auditor raises an exception, the original message is bounced back to the sender with the explanatory message given in the exception.
A standard detector is provided that watches for additions to the "messages" property. When a new message is added, the detector sends it to all the users on the "nosy" list for the item that are not already on the "recipients" list of the message. Those users are then appended to the "recipients" property on the message, so multiple copies of a message are never sent to the same user. The journal recorded by the hyperdatabase on the "recipients" property then provides a log of when the message was sent to whom.
The e-mail interface also provides a simple way to set properties on items. At the end of the subject line, propname=value pairs can be specified in square brackets, using the same conventions as for the roundup set shell command.
The web interface is provided by a CGI script that can be run under any web server. A simple web server can easily be built on the standard CGIHTTPServer module, and should also be included in the distribution for quick out-of-the-box deployment.
The user interface is constructed from a number of template files containing mostly HTML. Among the HTML tags in templates are interspersed some nonstandard tags, which we use as placeholders to be replaced by properties and their values.
There are two main kinds of views: index views and item views. An index view displays a list of items of a particular class, optionally sorted and filtered as requested. An item view presents the properties of a particular item for editing and displays the message spool for the item.
A view specifier is a string that specifies all the options needed to construct a particular view. It goes after the URL to the Roundup CGI script or the web server to form the complete URL to a view. When the result of selecting a link or submitting a form takes the user to a new view, the Web browser should be redirected to a canonical location containing a complete view specifier so that the view can be bookmarked.
Properties appear in the user interface in three contexts: in indices, in editors, and as filters. For each type of property, there are several display possibilities. For example, in an index view, a string property may just be printed as a plain string, but in an editor view, that property should be displayed in an editable field.
The display of a property is handled by functions in a displayers module. Each function accepts at least three standard arguments -- the database, class name, and item id -- and returns a chunk of HTML.
Displayer functions are triggered by <display> tags in templates. The call attribute of the tag provides a Python expression for calling the displayer function. The three standard arguments are inserted in front of the arguments given. For example, the occurrence of
<display call="plain('status', max=30)">
when displaying item 13 in the "issue" class. The displayer functions can accept extra arguments to further specify details about the widgets that should be generated. By defining new displayer functions, the user interface can be highly customized.plain(db, "issue", 13, "status", max=30)
Some of the standard displayer functions include:
An index view contains two sections: a filter section and an index section. The filter section provides some widgets for selecting which items appear in the index. The index section is a table of items.
An index view specifier looks like this (whitespace has been added for clarity):
/issue?status=unread,in-progress,resolved&
topic=security,ui&
:group=+priority&
:sort=-activity&
:filters=status,topic&
:columns=title,status,fixer
The index view is determined by two parts of the specifier: the layout part and the filter part. The layout part consists of the query parameters that begin with colons, and it determines the way that the properties of selected items are displayed. The filter part consists of all the other query parameters, and it determines the criteria by which items are selected for display.
The filter part is interactively manipulated with the form widgets displayed in the filter section. The layout part is interactively manipulated by clicking on the column headings in the table.
The filter part selects the union of the sets of items with values matching any specified Link properties and the intersection of the sets of items with values matching any specified Multilink properties.
The example specifies an index of "issue" items. Only items with a "status" of either "unread" or "in-progres" or "resolved" are displayed, and only items with "topic" values including both "security" and "ui" are displayed. The items are grouped by priority, arranged in ascending order; and within groups, sorted by activity, arranged in descending order. The filter section shows filters for the "status" and "topic" properties, and the table includes columns for the "title", "status", and "fixer" properties.
Associated with each item class is a default layout specifier. The layout specifier in the above example is the default layout to be provided with the default bug-tracker schema described above in section 4.4.
The template for a filter section provides the filtering widgets at the top of the index view. Fragments enclosed in <property>...</property> tags are included or omitted depending on whether the view specifier requests a filter for a particular property.
Here's a simple example of a filter template.
<property name=status>
<display call="checklist('status')">
</property>
<br>
<property name=priority>
<display call="checklist('priority')">
</property>
<br>
<property name=fixer>
<display call="menu('fixer')">
</property>The template for an index section describes one row of the index table. Fragments enclosed in <property>...</property> tags are included or omitted depending on whether the view specifier requests a column for a particular property. The table cells should contain <display> tags to display the values of the item's properties.
Here's a simple example of an index template.
<tr>
<property name=title>
<td><display call="plain('title', max=50)"></td>
</property>
<property name=status>
<td><display call="plain('status')"></td>
</property>
<property name=fixer>
<td><display call="plain('fixer')"></td>
</property>
</tr>String and Date values are sorted in the natural way. Link properties are sorted according to the value of the "order" property on the linked items if it is present; or otherwise on the key string of the linked items; or finally on the item ids. Multilink properties are sorted according to how many links are present.
An item view contains an editor section and a spool section. At the top of an item view, links to superseding and superseded items are always displayed.
An item view specifier is simply the item's designator:
/patch23
The editor section is generated from a template containing <display> tags to insert the appropriate widgets for editing properties.
Here's an example of a basic editor template.
<table>
<tr>
<td colspan=2>
<display call="field('title', size=60)">
</td>
</tr>
<tr>
<td>
<display call="field('fixer', size=30)">
</td>
<td>
<display call="menu('status')>
</td>
</tr>
<tr>
<td>
<display call="field('nosy', size=30)">
</td>
<td>
<display call="menu('priority')>
</td>
</tr>
<tr>
<td colspan=2>
<display call="note()">
</td>
</tr>
</table>
As shown in the example, the editor template can also request the display of a "note" field, which is a text area for entering a note to go along with a change.
When a change is submitted, the system automatically generates a message describing the changed properties. The message displays all of the property values on the item and indicates which ones have changed. An example of such a message might be this:
title: Polly Parrot is dead priority: critical status: unread -> in-progress fixer: (none) keywords: parrot,plumage,perch,nailed,dead
If a note is given in the "note" field, the note is appended to the description. The message is then added to the item's message spool (thus triggering the standard detector to react by sending out this message to the nosy list).
The spool section lists messages in the item's "messages" property. The index of messages displays the "date", "author", and "summary" properties on the message items, and selecting a message takes you to its content.
The design described above should be general enough to permit the use of Roundup for bug tracking, managing projects, managing patches, or holding discussions. By using items of multiple types, one could deploy a system that maintains requirement specifications, catalogs bugs, and manages submitted patches, where patches could be linked to the bugs and requirements they address.
My thanks are due to Christy Heyl for reviewing and contributing suggestions to this paper and motivating me to get it done, and to Jesse Vincent, Mark Miller, Christopher Simons, Jeff Dunmall, Wayne Gramlich, and Dean Tribble for their assistance with the first-round submission.
| [Home] | [FAQ] | [License] | [Rules] | [Configure] | [Build] | [Test] | [Track] | [Resources] | [Archives] |