User:Mstone/Commentaries/Bundles 2

From OLPC

< User:Mstone
Revision as of 01:01, 23 May 2008 by Mstone (Talk | contribs)
(diff) ← Older revision | Current revision (diff) | Newer revision → (diff)
Jump to: navigation, search
Also read the first Bundles commentary

Contents

Sugar Data Model

Broken Threads

  homunq> eben - I want to start coding new activity bundle format, do you have
          a minute to talk about your UI vision there?
    eben> homunq: Sure, let's do it.


  homunq> Is it worth including enough info to reconstruct "broken threads"?
          (recall: forks are one case of "broken threads".)
    eben> I not convinced that broken threads should be reconstructed. I
          personally feel that a broken thread should be treated as a separate
          activity. Should the "original version" be an activity on the other
          side of a break?
  homunq> The principal reason I think that broken threads are worth it is that
          they make our security model more in-line with reality and thus more
          secure - people will have less need to subvert it in order to achieve
          their goals.


  homunq> I imagine that the user has some easy way to explicitly update her
          favorite version and to explicitly update instances when a "newer"
          activity becomes available that crosses a break. This is because the
          newer version implicitly lives within an unbroken thread and you can
          explicitly say "use the original version" if it is available. I also
          imagine aggressive backups of activity versions so that if you have
          server coverage, then chances are that you can get the original
          version. Note: the original version could not be on other side of a
          break because the only way to cross a break is to open and opening
          saves a new version of the instance. Thus only the old instance would
          have an old original version.
    eben> I disagree with it on the principle of allowing "just anyone" to
          create new activites based on existing activities which appear as
          "new versions" to the rest of the kids. Let's work through an
          example:


    eben> Consider activity A.
  homunq> A exists in version Ax.1, Ax.2, and Ay.3
    eben> right, perfect.


  homunq> so you have an instance of Ax.1 and then you get Ax.2. Your instance
          will now automatically open using Ax.2.
    eben> now, Bob comes along, guts A and derives Ay.1
  homunq> No... the rules would be that version number keeps increasing. Now
          you get Ay.3. Something derived from Ax.1 is Ay.2 Or in our example
          Az.2 because y is taken.
    eben> homunq: But this is the point we disagree on.  You say that it
          increases, because the "identity" doesn't change. I say that it
          starts at one, because it's effectively a new activity.


  homunq> OK. So whose example are we following?
    eben> Well, I'm trying to contrast them.
  homunq> I say mine, because the burden of proof is on me.


  homunq> So you get Ay.3. Now if you just run your instance, it still runs as
          Ax.2 but it has an extra button in the journal called "update me".
    eben> Ay.3 did not get auto-favorited, even if Ax.2 was already favorited.
  homunq> The palette or something on that button says "Ay.3 claims to be a
          newer version, but I cannot be sure" - some message that conveys the
          alleged nature of the relationship.


    eben> OK.  Is there a need for this button?  Couldn't the kid just manually
          choose to open the instance with Ay.3 if they choose?
  homunq> Favoriting and instances are two different (but related cases), I'm
          talking instances right now - but yes, they could do it manually. I
          just think it's nice to have the button.


    eben> And why do you think it's safe to tell the kid that this is a "new
          version"?
  homunq> I think it is safe to say "claims to be a new version"
  homunq> Note:
    eben> How does it claim this?
  homunq> implementation detail


    eben> Does the activity retain the same service name?
  homunq> No.
    eben> It has a new service name?
  homunq> The service name is not user-readable. It can have same or different
          user-readable name. However, it has a list of previous service ID's
          and version-of-forks.


    eben> Who cares if it's exposed? In your model, does Ay.3 have the same
          service name as Ax.2?
  homunq> No. The identifiers you call "service names" are associated with
          "unbroken threads". They change across breaks.


Heritable Traits

    eben> How is it proper to suggest to the kid that something with a) a new
          author and b) a new service name is a "new version" of something they
          already have? For that matter, how is it fair to the author of Ax.2
          if anyone out there can come along and push (via suggestion) a "new
          version" out there which may be totally different, or buggy, or
          corrupt? I think that activity thread deserves to have a sense of
          ownership associated with it.
  homunq> I think that it is perfectly fair, if you give users explicit choice.
    eben> No, I don't think that's fair at all. You might get away with saying
          "Here's something (Ay.?) that claims to be an *alternate* version of
          the Ax.2 you have been using"...
  homunq> I think that people sometimes want to assert that and sometimes have
          good reasons for wanting it without having good reasons for getting
          the x signing key.
    eben> Saying that something is a new version implies a chain of trust to
          me, and an authorship, which doesn't apply in this case. That's why I
          personally like calling it Ay.1 (The first version of an alternate
          thread), even if it was based upon something else.
  homunq> Then what about Az.1 which inherits from Ay.1? How does it say that
          Ay.1 is a later ancestor than Ax.2? That is what the monotonic
          versions are for.


  homunq> There is another bit I have to explain. In my model, there are
          private instances. Any instance marked private cannot be opened by an
          activity which has P_NETWORK unless it has the same service ID as
          creator.
    eben> Any instance that has not been shared?


  bemasc> The key thing here is that Eben is also proposing that instances be
          explicitly mime-typed. Furthermore, the type of an instance created
          by "Ay.2" is not "Ay.2"; rather, it's "xml/tamtam-projectfile". Then,
          if you change the format in an incompatible way, then you should
          introduce a new "mime type" to describe the new format. (possibly not
          an actual "mime" format, but some kind of project file with type).


  homunq> Is there any support for format versioning in mime types? In
          particular, how do you define "incompatible"?
    eben> I'm not sure... do we need full ancestry across threads? If one
          thread simply uses monotonic versions and full ancestry internally,
          and perhaps also includes the thread it forked from, do we really
          have need for more?


  homunq> This is a problem for me. If you add a new bit of data that does not
          break the old activity, but is not supported by it, how is that
          indicated in mime-type-land?
  bemasc> If it breaks the spec for that filetype, it's a new filetype. That's
          the author's decision.
    eben> Incompatibility certainly comes as a responsibility, in my mind, of
          the author of the thread.
  homunq> I think that most format changes are in the grey area of
          compatibility.
  bemasc> That's why it's the author's decision.


Types vs. Inheritance

    eben> Consider taking Ax.2 (because you love the interface) and on purpose
          explicitly swapping out all of the data formats/protocols used for
          your own. This is a case where, clearly, each thread can stand alone,
          but there is no intent to make the two threads compatible with each
          other... and why should there be? They are authored by two separate
          people who may not even know each other.
  homunq> In that case, your new activity is not a child version.
    eben> I claim that this decision can only be intelligently made when you
          restrict the scope to a single thread.
  homunq> Exactly, stand alone.


  homunq> Sorry - we are talking about two things - mime type and activity
          ancestry. I do not want ancestry to be an archeological record of
          code origin.
  homunq> If you change file formats incompatibly, then you should start a "new
          activity", even if it has same user-readable name.
  bemasc> Ancestry does not imply compatibility. Witness MS Word. Conversely,
          witness AbiWord, which is by no means descended from MS Word, but
          (relatively recently) got .doc import capability. The question is:
          what can ancestry tell us? Alternately, what actions can the system
          take on the basis of ancestry? You seem to be proposing one action
          the system can take on the basis of ancestry: determining which
          Activities are capable of opening which files. Another action the
          system can take is inheriting permissions.  After some discussion at
          the "mini-conference", there was some consensus that if a future
          version of Activity is signed by the same key, then it should inherit
          the user's permission settings for that activity.
  homunq> Correct on both counts. I also think that the shell should give the
          user a prominent option to open old instances with the child
          activity.


  bemasc> I still prefer explicitly typed instance formats.
  homunq> Explicitly typed but not versioned?
  bemasc> If you've broken compatibility with the previous versions, you should
          declare a new instance format (or increment your version format
          number)
  homunq> mime types have a version number?
  bemasc> you can just add a number on the end, if you want.
  bemasc> I think formats should be versioned entirely separately from the
          activity bundle itself, and possibly with no assumption of backwards
          compatibility.


    eben> So activities that break signature, you agree, have new service
          names, do not automatically receive the same permissions, what else?
          I still fail to see an argument for treating it as a "new version" in
          the same thread.  It's effectively a new animal.
  bemasc> homunq wants to use a non-secured ancestry system to determine
          instance compatibility. This amounts to declaring that the type of an
          instance is the specific version of the activity that created it.
          Then another activity can declare "I can open that instance" by
          listing that activity version as one of its ancestors.
  homunq> I also want it because I think that if we do not provide an un-
          secured (and thus un-trusted) mechanism, then I think that developers
          will feel undue pressure to use the trusted mechanism in
          inappropriate ways.


    eben> Wait wait....I find your last comment disagreeable, bemasc. If the
          type of an instance is 1-1 with its (activity,version) pair, then we
          don't have any kind of cross activity compatibility....
  bemasc> In homunq's proposal, Activity B.8 can only read something written by
          A.3 if B.8 lists A.3 or later as an ancestor. Also, we can have that
          A.10 is a direct descendant of A.3, signed with the same key, so it
          should inherit permissions, even though format compatibility has
          since been broken.
    eben> That seems arbitrarily limiting, no?
  homunq> Note that multiple inheritance is possible. Also, Ay.10 can say "I
          save in format Ax.3".


  bemasc> Conversely, in eben's proposal, instances are handled just like
          files, in terms of typing. Thus A.10 can declare that it only reads a
          certain instance types, and the type written by A.3 might not be in
          that list so then activity A.10 might not be able to open instances
          of A.3.
    eben> What if these activities have no idea each other exists, but they
          share a common format? They should "just work", and the system should
          know that they can handle each other's formats.
  bemasc> eben: I agree, which is why I prefer your proposal. Also, homunq's
          system can be implemented easily inside eben's.


    eben> It seems absurd that a simple activity that writes a text file could
          produce something that can't be opened by any other text editor
          created in the past or future.
  homunq> Of course. That is why I keep saying "actions are not files". The
          text file can be opened by any text editor. "I emacsed that" and "I
          wrote that" are not interchangeable but the resulting text files are.
    eben> Instances with all associated metadata and goodies? I guess we
          haven't ever had the notion of an "instance type" to be honest....
  bemasc> That's because we haven't had a distinction between files and
          instances.
  homunq> In your new journal spec with actions, the difference is much more
          clear.


Project Files and Metadata

What is an instance?

    eben> Thinking about it, the whole idea was to treat the instance as having
          the type of the "primary file" object, such that one could say "open
          with X" and it would simply open that file in the new activity.
  homunq> That makes no sense to me. An action can have multiple, incompatible
          files associated.
  bemasc> You imagined a scenario in which each instance was associated with a
          "project file".
    eben> Yes. I think each instance will have a "project file", and that
          project file will have a type. In my head though, that type was TXT
          or PDF or PNG or HTML....
  bemasc> Aren't TamTamEdit's project files are going to be some complicated
          custom format?


  homunq> Also, in my head, the instances/actions had some of their format in
          metadata - or is that abusing metadata?
  bemasc> I think metadata should only be things that the user is invited to
          mess with
    eben> Sure, that's fair... metadata was to hold the state data which does
          not fit into the "well known file format" assuming there is one, or
          which the author doesn't feel is a necessary part of their new file
          format if it's not a known one already.
  homunq> What about things like screenshots? Changing it could be a security
          hole and I can't think of a positive use case.
  bemasc> Changing the screenshot won't screw up the instance. My rule of thumb
          would be, "the user should be able to write and delete metadata
          arbitrarily and at random without screwing up the instance when it
          resumes".
  homunq> bemasc: that is a fair rule.


    eben> So, in my head at the time, the point of separating the actions from
          the objects was so that the metadata *could* hold the "extra" stuff,
          and so that it *would* be possible to say "open this instance with
          some other activity" because the file associated with that instance
          would be just a well known format of some kind.
  bemasc> project files are almost always application-specific. The example in
          my head is Audacity.
    eben> I think that's about a 50/50. In the case of an Audacity instance...
  bemasc> Audacity uses an XML-based project file format to describe the
          project itself, and also stores bits of audio in another directory
          that has like 1000 little audio files in it
    eben> only audacity (right now) can read that file (even though others
          could read, for instance, the samples used or the rendered output of
          it). But, there still is a file, and it still has a format, and other
          activities could certainly adopt it in the future.
  bemasc> the project file format is specific to audacity, and is occasionally
          changed in a non-backward-compatible fashion
  homunq> good example. that stuff clearly is not metadata despite what eben
          and I vaguely thought.


Audacity example

  bemasc> So the question, for me, is UI. If someone writes an Audacity-
          compatible Activity, what should users do to continue their work in
          this new program? Is that an example of resuming your work, or an
          example of starting a new instance?


    eben> It would be prudent for the creator of that format to keep all the
          necessary data for the project inside it, and to save other info
          (perhaps UI hints about the selected tab, the current view settings,
          etc.) as state metadata.
  homunq> why not keep the tab in that project?
    eben> Because that info is specific to the UI of the activity, and the
          platform that the activity runs on.
    eben> I have audacity for OSX.  It doesn't have tabs.  It probably has a
          completely different and perhaps incompatible UI.


  homunq> OK, so actions have 4 kinds of data: files, projects, view settings,
          and genuine metadata (tags). I'd rather not assume that any of those
          are equivalent. If an activity author wants to assume that, fine.
    eben> I actually see 3 myself:  associated files (of which one is the
          project), state blob, and metadata. In practice, the state blob might
          be some piece of inaccessible metadata.
    eben> Consider a painting program. Clearly you want the "project file" to
          be simply an image, for the best portability. Storing the currently
          selected color would be metadata. (This, too depends....for instace,
          if an SVG editor decided to support .ai instead of .svg, then the
          color would be stored too, perhaps, as part of the actual project
          file....but that's all dependent upon the standards in place, or
          being set up in the case of new activities.)
  homunq> "Inaccessible" is the point. This is different in some key way from
          normal metadata. The implementation can be the same though.


  bemasc> What about undo history?
    eben> That would be part of the state too, right?
  bemasc> Well, that's the question.  An image editor clearly doesn't store the
          undo history in the image.
  bemasc> (or does it? the PSD format stores undo history, I think, and so does
          gimp's XCF)


  homunq> in a perfect world, undo history would be in olpcfs ...
  bemasc> homunq: I agree.
  homunq> I'm actually kinda serious.


UI Decisions

  bemasc> I'd like to work from the UI down. As a user, I go to an instance and
          click "resume" or whatever.  What happens?  Am I shown a list of
          activities that can resume this instance?
  homunq> There is a "just resume" button: you click it, it defaults to
          something. If you hover over it, it gives you choices.
    eben> (It resumes with the last activity you used for that instance.)
  bemasc> If there's a new version, does it use the newer version?
    eben> bemasc: I would argue yes, if it is part of the same thread.
  homunq> (If there is a new version AND the new version can open that data).
          It seems that the data has a master file which gives it its format?


    eben> Another potential option is to say that it uses the newest starred
          version, if there is a starred version in the thread, or the most
          recent otherwise. (again, assuming it can.)
  bemasc> What if there are no more activities installed in the thread, but
          there is another activity installed that can resume it?
    eben> In that case, we probably (currently) pick the next in the list.
  bemasc> what if I want to resume with a non-default activity? (or a non-
          default version)
    eben> A better solution would be to present a modal dialog presenting the
          available options, including the option to attempt to recover the
          activity which created it. If you want to do that you use the
          secondary palette of the resume button, which lets you choose
          anything that supports the type.
  bemasc> That all sounds pretty reasonable to me.
    eben> One option for the non-default version case is to say that the list
          is actually hierarchical (one level). If there are is more than one
          version of an activity available, it could yield a submenu of the
          options, and indicate which of those are starred as well.


  homunq> I have a problem with that: what if you have 5 versions in the same
          thread? Do they all clog up that list?
    eben> No, the versions that are in an unbroken thread are treated as "one
          thing" until you ask for more info. This is the same as what I
          intended to represent in the activity list in home. They would be
          grouped by unbroken thread. Ax.? and Ay.? would both appear in the
          top level list, side by side (well, depending on sorting...but
          assuming they choose to keep the same readable name).


  bemasc> It sounds like the system has the following: each instance is a
          collection of files, one of which is declared the "main file".  Each
          file has an explicit type. Each instance has "application metadata"
          and "user metadata". The application can use application metadata as
          a convenience to store a small amount of simple data that is some how
          "not worth" putting in a file.
    eben> not worth or not appropriate (in the case of trying to adhere to
          already defined formats)?
  homunq> There are several issues this does not address... what if format
          tamtam-project5 can be opened imperfectly by tamtam-project3?
          Obviously tamtam3 cannot preemptively say this. What if different
          subfiles have different levels of privacy? (I am thinking about
          private signing key for an activity bundle.)
  bemasc> We could allow files to have multiple types.
    eben> In the worst case, TamTam5 could include a way to export a TamTam3
          file/instance
  bemasc> Another solution (for a paint program) would be to have the instance
          contain "image.png" in PNG format and "mysettings.paint" which is a
          JSON file with activity-specific settings. Then make image.png the
          project file.
    eben> instead of the state blob as metadata?
  bemasc> yes - the "activity metadata" could easily be stored in a file other
          than the project file, without breaking compatibility
    eben> I think the goal is to prevent those kinds of "junk" files from
          accumulating, since the kids can also browse the objects as well.
          That's not meaningful as a file itself, to the kids.
  bemasc> JSON is fairly human-readable. We could also adopt a "dotfile"
          convention like the one used in all UNIX systems.
    eben> Well, sure, but in the interest of making it easy to browse through
          the objects one has created, seeing ever other one being a settings
          file would be obnoxious.
  bemasc> Hiding things from the user is generally bad, but the "activity
          metadata" would be hidden from the user anyway.
    eben> My point is the Journal has to add knowledge for that type of thing,
          if we do that.
  bemasc> yep, but the alternative is that the datastore has to handle an
          additional metadata system
    eben> well, does that depend, too?  How is it settable?  If we don't expose
          it in the UI, does it matter if it's just normal metadata?
  bemasc> well, then the Journal has to have support for knowing what kind of
          metadata to show. there's no way around writing code for this stuff


  homunq> What if I have an tamtam5 instance and an app I like which opens
          tamtam3, but I do not have tamtam5 activity.
    eben> That is a *really* tricky question....it's part of the reason that we
          didn't use aliases at all in the past...


  bemasc> (this, by the way, sounds a lot like Core Data)
  homunq> I read the wikipedia briefly, I could not tell if it was some way to
          keep undo stack, a layer above SQL, or what...
  bemasc> it's a generic high-level data storage system
  homunq> definitely has some relation to what we are talking about, but not
          clear if it has the answers. that is not the problem we're solving
          right now.


  bemasc> eben: I think we're pretty clear on this.  Develop can implement
          homunq's semantics over a simple per-file type layer and the Journal
          just needs to know which of the files in each instance is The Main
          File
  homunq> Does the main file show in file list or just in actions list?
    eben> It's just an ordinary file - the main file might just be The Image,
          or perhaps The Text.
  bemasc> "The Main File" is the "Project File".  It's the file that defines
          the instance.
  homunq> there seems to be confusion. if you have two files, one of which is a
          project and the other is a text, then I think the project should not
          pollute the files view


Examination of Use Cases

    eben> consider Record. You take 10 photos.  You get each photo as an
          object.  You get the "roll of film" as an object.  The roll of film
          is the project file. It seems equally valid to send the whole roll to
          someone, like an album, as it would be to send one photo.
 m_stone> hang on there. the roll of film is more usefully regarded as an
          "object container", i.e. a set or list of objects. we really want
          other people to be able to process such containers.
    eben> sure, in this case it is nothing more.  In other cases, the container
          may have a bunch of other data which talks about how to put together
          the things it contains.
 m_stone> however, there's a lot less value in them building support for
          parsing records of Record-37's graphical state when it was closed.
          for sake of argument, call the latter the ui continuation.
  bemasc> right, and that's why the roll of film is the Project File for this
          instance. that's why I'm suggesting that non-critical state be put in
          a separate dotfile
    eben> right, a hidden file, or in metadata....but yes....all state not
          appropriate for the objects themselves goes elsewhere.


 m_stone> Please see the picture and descriptive paragraph at the top of
          http://wiki.laptop.org/go/User:Mstone/Bundle_commentary. Then, please
          use the already documented terminology until you demonstrate that
          this terminology is inadequate. (though please comment inline in this
          page and sign your comment if you think I left out something
          important).
    eben> where does metadata/state live in this diagram?  As "instance"?
 m_stone> UI state lives in the "instance" node.
 m_stone> give me some examples of metadata?
    eben> well, two flavors of it: creation date, participants, tags,
          description, activity specified metadata such as number of pages,
          etc. Then there is state metadata for the purpose of recovering the
          instance (the UI continuation part)
 m_stone> I tend to think about the problem relationally because I anticipate
          objects (and object containers, which might themselves be objects;
          I'm making no commitment here) being referred to from several of
          these action graphs
 m_stone> so I would say that the action has a set of time periods when you
          worked on it. likewise for the objects.


  homunq> OK, sorry, I need to go back for a second. I agree with the decision
          of this conversation that there is no UI-necessary information in
          version history, that is not in "what actions can I open". However, I
          still like version history because it is compatible with the model of
          forking versions of an arbitrary object, not just of an activity. My
          feeling is that the journal UI should support this forking-version
          model in some way.
  homunq> OK, it was good for me to say that - it made me see that they are
          separate issues.


  bemasc> really?  I thought we were going for copy-on-write semantics.
    eben> that's true in general but in practice, it's complicated. Consider a
          Journal entry which is just an action "you sent file X to your friend
          Y". That's an action which references X, but does not copy it. In
          fact, it's not a write. But that doesn't mean that we won't have
          multiple actions referring to the same object (actually, version
          thereof)
  bemasc> Then we shouldn't say actions=instances anymore. I would argue that
          that action does not have any instance associated with it.
 m_stone> depends whether the action is resumable. actions which can be resumed
          have instances associated because instances are the actual material
          which permits resuming
    eben> which is why the graph m_stone pointed two has those three nodes in a
          little triangle and why the bundle (needed for resuming) is a child
          of the instance.
 m_stone> however, even though I can't resume the (completed) action of sending
          (A,B,C) to Z, I still want to be able to snag A, (A, B, C), or Z...
  bemasc> I think that makes sense, so the action should retain pointers to the
          appropriate objects


Grammar and Criticism

  bemasc> what we have further been proposing is that each instance refer also
          to some specific object as the Main Object
 m_stone> and, at face value, that sounds nuts. (though perhaps I haven't
          understood your use case or your design)
  bemasc> the use case is determining which activities are capable of resuming
          an action
 m_stone> resuming an action? perhaps you meant 'processing an object
          container'?
  bemasc> I'm speaking from the user's perspective
 m_stone> don't. one makes a new action, the other doesn't... apologies, the
          "don't" was a visceral reaction - a segment of "your statement does
          not typecheck"


    eben> we meant 'resuming the instance'. The question is: "which bundles,
          apart from those linked to the instance within a given action, could
          be used instead if desired?"
  bemasc> Suppose I've just used Record to take 10 pictures. I now want to view
          them in a gallery that understands Record's photoset file format. How
          should we think about this?
 m_stone> (perhaps because it's a standard List format used throughout
          Sugar...)
  bemasc> one option is to offer a number of activities that can resume this
          instance
 m_stone> I say: "You View the photoset."
 m_stone> That requires finding a Viewer (an activity), instantiating it and
          telling the instance to process the photoset, and recording the
          action of Viewing. (Recall that Activities are usefully regarded as
          prototypes of instances)
    eben> right, so we're talking about how to "find the viewer"


  bemasc> the broader question is: when do we resume instances, and when do we
          open files and thereby start new instances? We are trying to come up
          with a way to make this determination. One suggestion is that each
          instance should identify one of its files as the "project file", and
          any activity that can open this file can be shown in the UI under
          "Resume with..."
 m_stone> I see two formulations. I don't yet have a strong preference between
          them. a) Activities can be used to synthesize instances ready to
          process some objects. It's just part of what Activities are. so we
          might use this ability to get an instance attached to our roll of
          photos which we can "sume" for the first time (since we're not really
          RE-suming it...). ;).  The other option is to say that Activities can
          be performed on objects and that this creates an Action which, just
          as a quirk of activity implementation, probably has an instance
          associated with it that can be resumed.
  bemasc> I don't understand, as usual.  Do you mean Activities can be
          performed on object containers?
 m_stone> I mean that the shell should understand object containers for me so
          that I can Paint on individual photos from a roll of photos of
          butterflies that I Recorded yesterday.


 m_stone> eben: have I lost you as well?
    eben> Only inasmuch as I don't know what point you're driving at with the
          example.
 m_stone> I'm saying that "resume with" is either nonsensical (because
          instances resume and do nothing else) or that what "resume with"
          should mean is "use <___> as a prototype to make a new instance
          which, when resumed, will be attached to objects <---->" since
          activities are logically quite similar to prototype instances in that
          they offer some fixed UI continuations which can be combined with
          zero or more objects. (i.e. they tend to have a default "initial UI"
          that they use when instantiated.)


  bemasc> nobody else here as written a program in a prototype-based language,
          so please use a different metaphor
 m_stone> never used javascript?
  bemasc> oh, didn't know that
 m_stone> here's another small example:
          http://www.iolanguage.com/scm/git/checkout/Io/docs/IoGuide.html
          #Objects-Overview (for future reference.) :)
     mtd> I think I see the continuation analogy (or philosophy), but pretty
          much in practice you're not describing something much different than
          what bemasc is. given the stack = the metadata and the runtime = a
          new instance of an activity.


  bemasc> "resume this instance with Activity A" literally means "start
          Activity A and tell it to open this object container"
 m_stone> resume this "instance" with activity A? According to my grammar, you
          should mean "resume this instance|".
  bemasc> I do not support your grammar.
 m_stone> The "with acitivity A" seems nonsensical. Instances are not
          parametric over activities.
  bemasc> they are if we decide they are.
 m_stone> why is the resulting UI a good one?
  bemasc> mostly because we have no distinction between object containers and
          instances in the UI
    eben> agreed with m_stone regarding "resume with" being the wrong speech
          for the intended meaning of the operation. We considered "as activity
          A" instead...


Manipulating Object Containers

  bemasc> the question is really "what should resume mean from the user's
          perspective?", and I think the answer is "I want to keep working on
          this abstract project" where the abstract project is concretely
          represented by an object container, not by any particular activity.
 m_stone> totally agreed. however, "project" should concretely mean "crazy UI
          data that I never expect anyone else to deal with" rather than
          "useful group of objects which lots of other things can deal with".
    eben> Actually, project in his meaning was "the file which describes what
          to do with the useful group of objects"
 m_stone> a simple collection of objects does not innately specify an activity
          I can use for processing those files.
    eben> That's the point of the "project file" in the first place.  There
          *is* something that takes on that role. (And which advertises it's
          mime-type so other activities can be instantiated from the objects.)
 m_stone> note: an instance is navigable to the activity that is its prototype.
          that way, if I want to have another instance of recording on "fish"
          instead of "butterflies", I can get there from the instance
          associated with the action of recording butterflies. However, I'm not
          connected solely to Record from the collection of objects. In fact,
          if Record has no photoediting capabilities, I'm might not ever find
          it adjacent to the photos in the UI.
    eben> each instance has "a collection of objects" associated with it. one
          of those is flagged as "the primary object", such that other
          activities can be instantiated with it as necessary.  The primary
          object knows what to do with "the collection of files"


Website Example

  bemasc> Suppose I write a webpage using the Dreamweaver activity.  It's a
          collection of files, with one HTML file but also CSS and jpeg's. Then
          I decide that I want to view it in the Browser. then I decide that I
          want to edit the CSS by hand in a text editor and then I decide I
          want to open it with dreamweaver again. first: how do I open the
          webpage with Browse?
 m_stone> you've got tree-structured data. All of our "object containers" to
          date can be represented as plain old directories.
  bemasc> right, but how does the Journal know that Browse can open this object
          container?
    eben> I don't think that's true. the photos example is a case where the
          container is empty apart from its contents.... however, an HTML
          instance might have several images, a few sounds, and a video in its
          object collection, with a HTML file treated as the "object
          container", including the necessary code to put it all together. You
          can open the HTML file in the Dreamweaver activity again if you want.
          You can open it with Browse which will know how to display it.  You
          could open it with SimpleWrite, which would ignore the other objects
          and just show you the source code.
  bemasc> but the HTML file alone will not open correctly in Browse, and this
          is my key point


 m_stone> I suggest you draw some labelled graphs of the progression of data
          through your use case. it will be much easier to discuss specific
          diagrams.
  bemasc> I am much more interested in working from the user's perspective
 m_stone> then draw your diagrams from the user's perspective first and
          translate beneath them into the dataflow perspective. begin the
          action of Browsing the Web-page.
  bemasc> So which button would that be? what should a user have to do, after
          having put together a webpage in a creation activity, to view it in
          the Browse activity?
 m_stone> I'm sorry, but text in IRC is not cutting it for me. I'm just not
          understanding what you're getting at. Please provide a diagram, or
          adapt my diagram and show where the fault occurs.
  bemasc> I cannot diagram this
 m_stone> can you write a paragraph instead?


    eben> why wouldn't it open in Browse?
  bemasc> because the HTML file is missing the CSS and pictures. so it might
          open, but it would be missing lots of pieces.
    eben> no, all that stuff is contained in the "useful objects collection" of
          the instance, right?
  bemasc> Browse needs to be able to open the whole "object container"
    eben> Why shouldn't browse be able to reference these?
  bemasc> because HTML refers to images by paths, but those pictures won't even
          be visible in the filesystem to Browse
    eben> Why not?  If not, how are they even accessible to Dreamweaver when
          you resume it!?
    eben> If I have an HTML file and an image on my Desktop, the former
          referencing the latter, and I open that HTML file, it will work fine.
          What's the difference here?
  bemasc> the difference is that in this case, that image won't even be in the
          filesystem as seen from the perspective of Browse. activities can
          only see the specific files that they have been told to open.
    eben> That's the whole point, here, right?  Store the objects out
          individually so other things can use them, but keep the references so
          that things work with respect to instances.
  bemasc> you can't modify the HTML file to include references, or it won't be
          HTML. so instead, we create a "container"
  bemasc> I imagine it as a .tar file containing the HTML file, the pictures,
          and the CSS
    eben> :-/  Then what good is this?
  bemasc> if you open the .tar file with Browse, it works
  bemasc> because it has all the components
  bemasc> if you open just the HTML file, you get just the HTML file
    eben> But the tar file doesn't have type HTML
  bemasc> but it does, because we designate the HTML file as the "project file"
          for this container
    eben> No, if that were the project file, then that's what would open with
          Browse.
  bemasc> Browse advertises the ability to read HTML, and this .tar file is
          listed in the datastore as containing a project file of type HTML so
          the Journal offers the user the ability to open this .tar file with
          Browse
    eben> So you're proposing that every "collection of objects" is a tar file
          instead. when the user accepts, the datastore unpacks the tarfile
          into Browse's directory and then points it at the project file, which
          is HTML
  bemasc> not specifically; there are many other implementations
    eben> You're talking the technical side.  The user doesn't know this was
          ever tarred. They can still see each, and interact with each object.
 m_stone> he's just saying it to give your mind the flavor of the packaging-of-
          related-stuff-together. The Browse-displaying-website example is
          really the "interact with this html file in the context of this
          larger collection of stuff"


"Act on ___ with ___"

  bemasc> right, but interacting with the HTML file itself is very different
          from interacting with the whole collection of files. Context is key.
    eben> indeed! but, this poses some problems for the UI.... consider a
          Journal entry for this instance...this HTML instance. It might read:
          "Today you created This_HTML_File, which contains Image_1 and
          Image_2". Now, the way to resume this directly is to click on
          This_HTML_File. One would expect clicking that to resume the
          instance....not open it independent of the images. What, then, if I
          hover and wait for the palette to appear, to do "resume as" (or
          whatever we call it)?
 m_stone> eben: "act on ___ with ___"
 m_stone> I agree that Browse can use some hints. But I think that Browse _has_
          to be able to deal with non-hinted material because it's certainly
          going to encounter it both because of design changes and because of
          dumb media like USB keys. I tend to think of the hint as an xattr,
          keyed by mime type, attached to the top-level dir representing the
          container of stuff - i.e. it says "dear people looking for HTML -
          start looking here; love, the Dreamweaver activity"
  bemasc> if you choose to resume the action with Browse, you get all 4 files
          and the webpage consists of a combination of all 4 files
 m_stone> bemasc: ARGH. be grammatical. "if you choose to act on all four files
          with Browse, you want to get a webpage full of pictures and CSS."
  bemasc> alternatively, if you choose to act on just the HTML file with
          Browse, you'll get just the text, no formatting or pictures. if you
          choose to act on just the picture with Browse, you'll get just the
          picture. and you can't open just-the-css with Browse, because it
          doesn't support that type.


    eben> back up a step.  "choose to resume with Browse"... how do you do
          that?
  bemasc> I'm actually not sure, because I don't see any way to resume an
          action in Designs/Journal
    eben> My point here is, looking at the mockup, "Rain Forest" is the "roll
          of film", and it is likewise the Project File
  bemasc> in other words, the project file is hidden
    eben> Well, right.  The main point to recognize there is that the instances
          are represented as the activity icons.  The objects are thumbnails
          (or a similar treatment, as seen in object view, which contains the
          activity icon on a "page")
 m_stone> my suggestion is that you pull the representation of the object
          container (holding the four objects) [sic: dir holding four files]
          onto the Browse activity. Or you select Browse from the list of
          activities to act on this data with, (where we suggest it either
          because of an explicit hint or because we suggest all activities
          which can process any file contained in the group)
  bemasc> or one of the pictures is designated as the project file
    eben> The intent was to make resuming an instance as simple as clicking on
          the primary file which represents it. Of course, each object
          (including the primary file) has a little "details" button next to
          it, which takes one to the detail for the object (independent of any
          instance at all) for acting directly on the object itself.


Hints

 m_stone> that's a really bad design. you're going to want more than one of
          them.
  bemasc> I don't know what you're referring to. I'm just attempting to figure
          out what's going on in slide 2
 m_stone> I'm saying that the idea of "a single designated Primary Object" is
          folly.
  bemasc> it's necessary
 m_stone> for two reasons. 1) because you're going to want more than one of
          them. 2) because you can't guarantee that it will exist in the
          presence of dumb media. in my opinion, you have no choice but to
          design a UI which does not depend on their existence in order to
          function. (you should, however, design a UI which takes hints that it
          can find)
    eben> in dumb media, every "instance" is associated with one and only one
          object. What does it mean to want more than one?  Doesn't that just
          imply you have two instances?
 m_stone> take the web-page example. it has some html, some css, and some
          photos. maybe you're lucky enough to have AwesomeEditor [etoys?] that
          handles all three. however, in practice, you're likely to want to mix
          and match. right?
    eben> m_stone: mix and match in what sense?  One could "open" any of the
          objects independently.
 m_stone> but some of them must be opened in context in order to make sense.
          (an essay containing labelled diagrams is an easy example.) but let's
          go with the web page for the moment.
  bemasc> the abstract "Webpage" consists of all 4 files.  If you want to start
          an activity instance to view the webpage, that single instances needs
          to have access to all 4 objects. so one object = one instance won't
          work
    eben> in what context?  Either I care to look at the collection as a whole,
          for instance, in Dreamweaver, or I care to look at a single object
          from the collection.
  bemasc> the problem is, the UI has no way to open a collection (that I know
          of)
    eben> But you just got through proposing the tar approach for getting
          around that internally
  bemasc> sure, but I'm talking about in the UI. the only way one ever opens a
          collection, in the mockups, is to "resume" an activity


Remixing Objects

    eben> The only way to *use* the collection of objects is to do so from its
          corresponding action entry, naturally. In these designs, the implicit
          assumption was that the "primary object" served as a standin for the
          instance itself, which when clicked on resumes the instance. In the
          same manner, that object would be the thing which had a "resume this
          instance as activity X" option in its palette. In other words, the
          representation of the primary object in the design is *as* the
          instance...
 m_stone> eben: how do you mean? consider turning a roll of photos into a
          slideshow.
  bemasc> you mean a stand-in _in the UI_. I hadn't thought of that. but every
          object can be opened by other activities. so one object can be opened
          or resumed, and these mean different things?
    eben> so there's a hole in the system, right? You can act on that thing
          either as the instance, or as the object.  So maybe that's busted.
          Maybe we need to in fact represent it "twice" in the UI. Maybe, for
          instance, that slide should also have a "Rain Forest" *object* down
          below, with a preview...(a contact sheet, let's say, with the 4
          images on it)
  bemasc> eben: what I want is a way for the user to open a collection.
          Currently, collections are only represented, as actions, so I've been
          using the word "resume" to refer to opening a collection
    eben> currently, meaning in the designs we're looking at, right?
  bemasc> eben: yes


 m_stone> well, how does the photoroll -> slideshow example work? furthermore,
          suppose I resume the action of photographing butterflies. Presumably,
          the next time I edit my slideshow, I'd like to be reminded that there
          are new butterfly photos. (and, in future work, pay attention to
          changes made to that collection [or to new versions of it that
          appear, etc, etc])
    eben> Right, as an aside, another missing component of the current design
          (a big one) is that there is no way to say "show my all versions of
          this instance". The current design only really has versioning for
          objects, not instances.
 m_stone> don't you mean 'for objects, not _collections_'? at the big Journal
          meeting w/ marco, tomeu, bertf, &co. we definitely discussed being
          able to use collections as part of several actions.
    eben> Ummm, I don't know? I mean, if I resume an instance 5 times, make
          changes each time, and then want to see the chain of actions I took
          to get to where I am.


Conclusions

  bemasc> it depends on what the metaphor is. let's not forget the problem of
          figuring out which version of an activity can resume an instance,
          which is how this conversation started.
 m_stone> (a problem with at least one easy solution, namely, to regard
          instances as executables derived from some prototype...)
    eben> I feel like we've gotten about nowhere in an hour.
 m_stone> we had a nice synchronization of terminology. we know of a use case -
          record butterflies, make slide show, record more butterflies, update
          slide show - that it's not quite clear how the present design would
          represent.
    eben> Perhaps we need to have an in Cambridge meeting to resolve all of
          this, going through the posted Journal designs and hammering it until
          we have a solution for all these issues.
  bemasc> (I thought you guys already _had_ this meeting, 7 months ago)
    eben> We did.  It resulted in the mockups you see.  Which still aren't
          perfect!
 m_stone> they don't need to be perfect. they've already allowed significant
          progress. they're mockups for a reason! :)


 m_stone> we also have the "make website - view website - edit images - edit
          html - view website" sequence. and the thoughts derived from
          consideration of that use-case, namely, that our UI will work much
          better if we attach some hints to collections but that we cannot rely
          on the existence of these hints.
    eben> right, both of those use cases are not covered, at present, by the
          design. Mostly because we don't have aliases/references in that sense
          at all.
 m_stone> so we should explain that clearly on the design and send some email
          so the rest of the world can think about it.
    eben> certainly, that's my point, and why I think it's worth another
          meeting to make them better. As an interesting option, by the way, we
          could consider making the way you act on an object contextual to
          where you initiate the action from.  For instance, if I "open" an
          image in an HTML instance from that instance, maybe it actually saves
          "in place" by just creating a new version of that instance, with the
          appropriate file changed? Whereas "opening" the object from within
          object view would treat it wholly independently, and use the copy-on-
          write approach to create a new instance for the changes.
  bemasc> eben: I think I get it.  Interesting... it sounds like it might be
          easy to break things though.  For example, a Desktop Publishing
          application might record the dimensions of all its images, and fail
          to load them properly if the dimensions change unexpectedly
 m_stone> eben: I didn't quite follow that comment and I need to relocate to
          1cc anyway. sounds like it's worthing reflecting on though.
HoboPrimate> can't the use case of creating a new instance of an object be taken
             care by copy+paste and/or "Keep as New" within the activity?


  homunq> OK I'm back. Just so you know, I think it was a very productive
          conversation. I definitely think that forking versions in the journal
          is something out-of-scope but worthwhile. but in-scope, I now see
          what is needed for activity bundles. The html problem (html stores
          references as paths) is I think soluble - every "open this" is either
          alone or in a symlinked directory of the bundle. internally, I think
          the files would be separate - not a tar -but there would be a way for
          sugar to create a tar to share with others. in fact, I think it would
          create a tgz. Honestly, this whole idea is just a few steps short of
          turning activity bundles into actions rather than objects. The
          MANIFEST could be the "master file". one missing part would be, where
          do you keep the private signing key. I think that if you edit the
          html, the UI continuation will be out-of-date.
    eben> This issue of "editing subparts with other activities" was never
          formally discussed, and currently has no solution.  I'm not sure what
          to do with that.
 m_stone> first, write that it exists. Then, spend a few days thinking about
          it. It may turn out to be nothing more than "future work".
    eben> I've thought about it at length, without coming to any meaningful
          conclusions.
 m_stone> don't give up. you'll work it out. (with help from many others, I'm
          sure)
    eben> It's a complement to the problem of "embedding an image in a Write
          document, editing that image in Paint, and later updating the version
          in Write automatically". And it also boils down to the copy-on-write
          vs. reference debate.
Personal tools
  • Log in
  • Login with OpenID
About OLPC
About the laptop
About the tablet
Projects
OLPC wiki
Toolbox