User:Mstone/Commentaries/Releases 1
Jump to navigation
Jump to search
Chris Ball, Marco Pesenti Gritti, Michael Stone, Tomeu Visozo, and several others conducted a 2-hour planning session this morning. I've created a transcript of that discussion, reproduced below. If you're interested, please review the questions that were raised and contribute your thoughts (preferably inline in this document).
The end goal of this effort is a convincing written statement of where we want to go in the next four months, why we want to go there, and how we intend to get there.
Questions
What are our goals for the next four months
- Some advocate trying to "fix the network."
- Some advocate trying to "stabilize the UI, particularly the Journal and the Datastore."
- Some think we should deliver working power management, even if we can only do so in limited circumstances.
Should we divide our effort among several goals or not?
- We could pour labor into our most critical goal until we reach the point of diminishing returns, then let the remainder spill over to the second goal, and so on.
- Alternately, we can try to make sure that our top three goals all receive enough effort to achieve limited gains.
- Alternately, we can try to make sure that all of our subsystems receive at least minimal maintenance.
How can we get better information with which to make our decision?
???
Annotated Transcript
Initial Unhappiness
m_stone> I take it that you guys are concerned that you'll have sugar in a
releasable state say around mid-June or early July? and that you'll
be unable to make big changes for the remainder of the 1-1.5 months
until the release?
tomeu> well, I think the issue is not releasing something like last
releases; some thing that more or less works but having a better API
and based on components that are more predictable
tomeu> "something that more or less works" == "something that passes the
minimal QA work that has been done but that shows critical problems
after being deployed" - we are not happy how new features have been
rolled out
m_stone> tomeu: keep explaining, please. I really want to understand what
things look like from your vantage point. (incidentally, that sounds
more like a QA failure rather than a software team failure to me.)
tomeu> m_stone: building on an orphaned component like the DS is not very
nice, either. well, project issues that needs to be solved - we
cannot blame QA because it has been composed by interns
cjb> and is currently composed of no-one :)
tomeu> we already complained last summer about interns leaving just when
they started being useful. Also, I'm not happy with the parts of the
sugar API that I have contributed most, either
Present Unhappiness
m_stone> tomeu: help me understand how this affects the work you think we
should undertake to do for August 1 or the next release beyond.
tomeu> m_stone: we want to have a software architecture more resilient to
bugs, more maintainable, and want to give activity authors a better
API. nothing of that are new features from the POV of the project
manager, I think. just dedicate time to it ;). Then discuss, discuss,
discuss, then implement, deploy.
tomeu> so well, we have a problem deciding on what to spend time - every few
months we are promised a new DS. the activity API depends on that so
it's not evolving and really sucks because depends on important
implementation things like storing entries with delta compression. so
well, we would like to better plan things, and spend effort in
invisible features
m_stone> tomeu: cscott, cjb, and I spent several minutes yesterday debating
whether cscott should try to fix networking things or DS things. (and
consequently whether cjb should fix networking things or power
things). (and whether the kernel folks should fix power things or
???). (and whether I should concentrate on collaboration or
stability/compatibility things). we have two plans and we haven't
figured out yet which one is better.
tomeu> m_stone: right, that's also our problem. I was supposed to work in
the journal. but had to choose between sugar, read, browse, ds, etc.
at different points and performance
How should we allocate effort?
m_stone> marcopg_: plan a) is cscott, cjb, & kernel folks on networking,
mstone & collabora on collaboration, and sugar on ??? where ??? is
probably the as yet unspecified "UI features that are already in the
bag + any accessible stability". the other plan is cscott on DS, cjb
on limited cases of power, kernel folks on power, mstone on ???, and
sugar on ??? where ??? is the same as above.
cjb> I think there's a third plan of cscott, mstone and collabora on
networking, cjb and kernel on power, sugar on ???.
marcopg_> I'm not sure I understand why we are going through the two extremes
about networking
marcopg_> we certainly need someone on it, but I doubt putting the whole team
on it will help
m_stone> I'll try to answer marcopg's question and I'll try to explain the
other flaws here. first flaw: someone has to manage the release. at
present, I'm the default candidate (though feel free to suggest
others). that's not immediately a full time job. but it will turn
into one and we can't forget it. also, effort will need to shift in
the second half of the cycle toward making it happen. i.e. the
release manager will need to impose on erikos, or cjb, or... :)
m_stone> second problem: there is lots of debate about how many people can
profitably work on networking, on power, on collaboration, on UI, on
"stability", or on "compatibility". in part, debate exists because
there are different approaches to each of these problems.
Power Management
m_stone> i.e. one way to do power is to make it work in limited circumstances.
I think of this as the "low-hanging fruit" approach. It mainly
requires cjb + some kernel, we think.
marcopg_> some kernel == not full time kernel person?
cjb> Yeah. I'd be surprised if we can't come up with an acceptable
limited/incremental use of power that takes less than a day to
implement. I'll try to write something about this. For example, we
don't do power management if you're currently connected to a jabber
server, else we do.
m_stone> the other way to deal with power (over a longer time frame) is to
comprehend the networking stack (so that we know what timeouts &
wakeups are needed), to fix the kernel to make those timeouts &
wakeups occur, then to do more userland power management. similarly
for networking.
cjb> (There are still the timeouts. It's hard to know how to tackle them,
when no particular misbehavior has ever been reported due to them.)
Networking
m_stone> one way to approach networking is to write down a believable
networking architecture, then to follow through. (where a believable
networking architecture would contain a small number of protocols
that "stretch" to cover all of our present use cases). the other way
is to try to follow the critical path toward one use case (e.g.
reliable read sharing on a school server wifi with K/N laptops) and
to fix every damn bug you find along the way until it works or until
you recognize defeat. the latter might get you a single working use
case (or maybe 3), but it's not going to get you reliability,
stability, or leverage. (they also take different amounts of time)
Focus or Divide and Conquer?
morgs> Are the approaches mutually exclusive from a resource point of view?
cjb> morgs: my third approach does a fairly even split of three groups
into three things. we can argue that this *sort* of approach has been
the main problem with OLPC software development and has led to power
management that doesn't really work, a mesh that doesn't really work,
collaboration that doesn't really work, and a datastore that doesn't
really work.
m_stone> and, as cjb alluded, we presently feel like we have N demos rather
than 1 solid product. some people believe this is because we sent N
people in N+2 directions rather than either sending 10N people in N+2
directions or sending N people in, say, 3 directions
marcopg_> I don't think that's matter of people; it's matter of too many
ambitious goals
tomeu> I don't think that having twice the number of people would have been
a waste of resources - in many projects it is, but in our case...
m_stone> marcopg_: the complete absence of a QA team definitely made itself
felt here... (and the absence of software attacking the QA problem,
e.g. test suites)
tomeu> the absence of a DS maintainer, too
tomeu> the problem is that it depends on what we aim to do and, regarding
sugar, we just don't know right now... perhaps later today
marcopg_> let me rephrase: I don't think moving everyone to work on the network
will help the network a lot, at least on 4 months schedule. gradually
adding people to the current team would be much more sensible, ihmo
m_stone> marcopg_: a) I think scott proposes to make sure that every aspect of
networking has at least one person working on it; not to make
everyone work on it regardless of whether they are doing something
useful
m_stone> marcopg_: b) why, precisely, do you anticipate that (a) would fail?
(or would be too expensive to pursue?)
marcopg_> because it takes time for people to figure out how to work together
in a productive way. if we have uncovered areas then we should cover
them. which aspects of networking we are not covering right now?
tomeu> m_stone: do you think that many of our orphaned or under-resourced
areas can only be tackled efficiently for people already in the team?
(in this case, without wasting most of the new resources' potential)
m_stone> depends what you mean by "efficiently"; specifically: I think that we
have individuals who have the experience required to 'begin labor' on
many areas but I don't feel that they'll have the ability to close
noticeable numbers of bugs
tomeu> if we talk about how to distribute resources, we need to characterize
every area as in how high is the entry point. some areas will require
a longer effort to start being productive, some less
OLE Nepal: Journal & Datastore
BryanWB> hey guys, if you dont' mind i have a few ideas of where OLPC should
put its resources. I think OLPC should focus on the datastore and
Sugar over the next serveral months, whether that means putting all
or most of your human resources towards that. W/ the
datastore/Journal being most important. the networking and power work
well enough but it is fairly difficult to use the Journal to find
stuff. Also, in my week's worth of QA, Sugar just does unexpected
stuff. Over the last 2 weeks of working w/ os699, then 702, now 703,
I have had the most problems w/ the Journal/datastore then Sugar
weirdness.
marcopg_> I tend to agree with BryanWB, simply because the Journal is more a
base/fundamental use case then network sharing; i.e. losing your work
is more critical then not being able to share an activity
cjb> I think my counter-argument would be that if you live somewhere with
ample power, have a wireless AP and only one XO, of course you're
going to say that networking and power work fine. :)
BryanWB> cjb: we are using an AP not active antenna
cjb> BryanWB: Yes. That's why I say it's natural that you would think it
works fine. What *doesn't* work is the mesh (non-AP) layer.
m_stone> cjb: rather, our choice of protocols and implementations is
incompatible with the mesh. (I speak relationally because I suspect
that we are more likely to change the protocols & implementations
than we are to change the firmware)
BryanWB> cjb: I know, but that doesn't work anyway when you've got 70 XO's at
one school, AFAIK. mesh is less important than basic usability of
datastore
tomeu> BryanWB: that's another problem we have, knowing the very different
problems of every deployment location. I think peru's next
deployments will use mainly mesh, but I'm not sure.
BryanWB> tomeu: they can buy a small AP and power backup for Peruvian schools.
Power backup is pretty ubiquitous anywhere they have electricity and
load-shedding. ** at least in south and east asia ** don't know
south america
cjb> BryanWB: what about when the kids go home from school, though?
erikos> i think both are important use cases mesh and datastore and should
both be addressed - i hope that for august we can do both - at least
parts. we once started to provide mesh usability - and should
continue to push for it; i think the statement that the datastore
should work is out of question
cjb> the problem is that no-one is confident we can get there by pushing,
rather than by throwing out what we have and redesigning. I guess
then we come to a question of whether the datastore can also be made
to work properly incrementally.
BryanWB> cjb: it needs to work at school first, that's where the other kids
are and they spend 8+ hours per day. Home should be a lower priority
cjb> since cscott's time is one of the main variables here, and he'd be
doing olpcfs.
m_stone> cjb: well, we might be able to have two people work on it. for
example, one person could start implementing the testing!
erikos> cjb: sure - i mean if we feel we have to throw something out - we
might have to keep a beta solution in the hands as well. in the case
of the datastore - if we decide we rewrite (i think that is what
people feel) we should start now
cjb> I think the short answer is that we shouldn't have to be making these
decisions ourselves. But if we write them up clearly enough, we can
pass the decision up the chain.
tomeu> having to choose between mesh and datastore is like choosing between
drinking or eating - both are needed
Feedback from Peru?
m_stone> so two questions: 1) what have the peruvian teachers discovered about
703? carla was clearly upset about a mesh bug.
marcopg_> m_stone: that's something that the deployment team should figure out
- we just don't have enough information at the moment to say it
m_stone> marcopg_: in which case the priority should be to get that
information, yes? or to change the decision so that it's no longer
necessary.
marcopg_> m_stone: I think that should be the priority, yeah
What is "compatibility"?
m_stone> marcopg_: you say the "compatibility" requirement is "unknown"?
marcopg_> m_stone: I don't know why it has been suddenly added to the list of
top reqs. (I'm not against, I'd just like to understand better)
m_stone> marcopg_: first, let's make sure we mean the same thing.
compatibility means two things: 1) our software running on other
people's platforms and 2) other people's software running on our
platform.
m_stone> we mean both of these, I think.
marcopg_> but there are a lot of possible approaches with pretty different end
results and that's the main reason I want to know why it was pushed -
to figure out solutions that meets the exact requirements
bemasc> cscott specifically referred to #2, not #1
Why do we want it?
m_stone> agreed that the reasoning behind making it a priority has not been
explained - I'll try to articulate the justification. 1) we wish more
people were supporting our platform. We thing that there are two ways
to do this - first, to bring our software to them, and second, to
give them an immediate userbase running on our platform and 2) we're
concerned, from time to time, that the days of funded support for our
platform are limited; therefore, we wish it to be less expensive to
maintain.
marcopg_> how much limited? 1-2 years or 1-2 months? ;)
m_stone> marcopg_: we've been greatly reassured on this point in recent weeks.
(specifically by the fact that kim's budget was accepted). but I
think that anxiety is a driving factor in the desire to increase
compatibility. as for your question: "both, with different
liklihoods"
tomeu> we are having to guess too much, considering we have kids using the
laptops in the thousands and adding anxiety to the guessing...
m_stone> tomeu: I say the same thing all the time but I don't know what to do
about it. the anxiety has greatly abated. we are now confident enough
in survival to begin hiring.
Who wants it and how?
marcopg_> m_stone: do we expect countries to actually deploy machines with
standard applications installed? and are we going to encourage or
discourage it?
marcopg_> the degree of compatibility we need is also an important factor for
the datastore work even though I think we will be forced to do
*something* about the datastore situation by august
tomeu> m_stone: which legacy software needs to be run in the laptops and
which problems it has? if it's flash apps, then the compatibility
stuff gains a totally different meaning if it's java, we may have
different problems than APIs
m_stone> marcopg_, tomeu: excellent questions which I do not have
authoritative answers for.
marcopg_> m_stone: ok, that's a question we need to answer I think
How does it affect the DS (and other APIs)?
tomeu> m_stone: I wonder why the compatibility thing is so tied to a new DS
m_stone> tomeu: because of the desire to use the journal as the only
navigation tool pointing to local persistent data
tomeu> from my pov, the same mechanisms that olpcfs has for that, could be
added on top of the current implementation. we could add a fuse
thingy on top of the current DS
m_stone> tomeu: I'm skeptical that the current DS would withstand that much
stress...
marcopg_> yeah what m_stone said
tomeu> m_stone: ok, I'm only concerned of mixing so many different issues
m_stone> tomeu: it's a good concern. I have no good answers for you :(
tomeu> we talk about compatibility, but do something because of performance
marcopg_> there is value in the idea of reusing the API though and to start by
fixing the implementation
m_stone> marcopg_: except for the fact that it's an API that we know is
terrible.
cjb> I don't think we should be too afraid about breaking APIs, if we have
a principled reason for doing it. It happens.
bemasc> I don't think there is very much legacy software worth using marcopg_> m_stone: it is, but we have tons of activities using it out there m_stone> marcopg_: not very many, actually. marcopg_> so in some way we will probably need to keep supporting it
tomeu> I'm perhaps the person who has suffered more because of the DS. I
have been promised lots of DS replacements - let me count them... 4!
Four times, I have been promised a DS, and have received only half!
Can we drop the current DS?
marcopg_> m_stone: do you think we can drop it and ask authors to port? (real
question)
m_stone> marcopg_: yes.
cjb> And we can offer to update all activity code in our git repo to the
new one.
cjb> marcopg_: I think so too.
marcopg_> ok, I tend to agree. (I'm sure there will be people that don't
though). but that's fine
m_stone> marcopg_: I'm not sure that we should commit to doing much else. but
I think we could be confident of achieving that thing.
tomeu> I think there are cheap alternatives to breaking existing activities,
worth discussing, at least
m_stone> tomeu: I don't think there are cheap alternatives to breaking
activities that will noticeably improve stability. I don't really
believe it. But I'm happy to hear your thoughts.
marcopg_> yeah that's a good point
tomeu> m_stone: using python namespaces, having daemons exporting two
different dbus interfaces, etc
m_stone> tomeu: what does that have to do with the causes of failure in the
present system? (which I believe are primarily bitrot and invalid
assumptions)
tomeu> m_stone: was referring to maintain API compatibility, while revamping
all of it
marcopg_> personally I think that if we are confident cscott can come up with
something solid by august, then it's totally worth to have him focus
on the ds
m_stone> marcopg_: I'm very confident in him (modulo the risk that he gets
sick, gets run over by a bus, gets "promoted", etc.) :) - recall
olpc-update... once he understand a problem, he's quite difficult to
stop.
marcopg_> then let's do it ;) and let's figure out how to properly resource
network without Scott, it doesn't seem impossible
erikos> m_stone: i think as well that it is worth to have at least one person
focusing on the ds reimplementation
bemasc> the high-level API will (read(), save(), etc.) will presumably remain
unchanged, so most activities won't require any modification
tomeu> bemasc: we have some problems with that API, I'm not sure yet how to
solve them... well, we overcome those by redesigning, and use those
techniques for old activities not breaking afterwards
m_stone> marcopg_: let's see if we can get some better information from the
deployment & sales folks first. even if it takes us a week to make a
decision, it will be time well spent if it winds up being a better
decision. many times over.
marcopg_> m_stone: I agree
In more detail, what do we want?
bemasc> m_stone: it would be very nice to have a list of desired legacy-Linux
programs. the only ones I know of are gnumeric and inkscape
m_stone> bemasc: agreed. (though tomeu & marcopg rightly suggest that flash &
java might also be good compatibility targets)
bemasc> m_stone: well, flash doesn't write to disk, so it's immaterial. It's
essentially out of our hands
m_stone> bemasc: totally false. we can decide to ask developers to work on
gnash or not. rsavoye has repeatedly said that gnash is shorthanded.
we've just never tried to actively assist him.
tomeu> about gnumeric, embedding it like we do with abiword shouldn't be
very complicated
marcopg_> tomeu: it would require to refactor the code heavily
tomeu> marcopg_: yeah, but how many months would take it to you?
marcopg_> tomeu: that doesn't scale - we can't embed every possible application
m_stone> tomeu: this would not improve compatibility though.
tomeu> inkscape, had some memory usage problems that I don't know if have
been solved yet
m_stone> tomeu: scott apparently looked at inkscape when he was writing his
icon editor - he thinks it would be quite hard to sugarize. (I don't
fully understand why)
bemasc> but for both gnumeric and inkscape, the datastore is still not the
problem because both of them use the standard OS-provided open/save
dialogs, which sugar can substitute for DS-based dialogs
tomeu> marcopg_: I have heard gnumeric and inkscape, only
bemasc> marcopg_: there really aren't very many applications
marcopg_> it's not much a problem of existing applications
tomeu> marcopg_: we are getting back to the same point: we need to have
someone from the ministries of education to tell us which legacy
applications they feel the need for. we are struggling to provide
features that maybe nobody will appreciate, while failing to provide
the important ones
marcopg_> if I want to write a cool application for sugar, I know that it will
be always confined to that platform and honestly I can see how it's
not an appealing perspective. we have the possibility to change this
situation without a lot of effort and so I think we should do it.
bemasc> marcopg_: yes. It's just a question of priority.
homunq> marcopg_++ on should, no opinion on can
tomeu> marcopg_: you, as a good software developer, will abstract the
platform dependent code.
marcopg_> tomeu: don't design your platform for very good software developers
tomeu> marcopg_: my point is not if we should do something or not, what I
mean is that I'm not convinced of the urgency of this code-on-xo run-
everywhere requirement. the current DS API is insufficient no matter
what we decide. and we have already devised a scheme for legacy
applications to store its files in the journal. where are we
differing?
What compatibility might we try to provide?
marcopg_> I think it would make for a much better experience for developers
which start working on sugar. it would put us in a better position in
the case of a end-of-funding and it will make more likely that our
applications are reused outside sugar. I think those are pretty
strong reasons and I argue the effort is minimal.
bemasc> tomeu: (I think we need to distinguish between legacy FD.o compliant
applications and legacy non-FD.o-compliant)
tomeu> marcopg_: we can make files dropped in ~/Documents to appear in the
journal as well as possible. but those apps won't be able to give a
so good experience as an activity. no matter what we do, "You created
the document My Dog with the application Gimp" is not as good as "You
drawed My Dog with Tim and Silvia".
marcopg_> tomeu: apps not specifically designed for sugar will never give you
an as good experience as an activity, but that's fine
tomeu> as I said, we agree on what needs to be done
marcopg_> then let's just do it :P
How should we try to get there?
tomeu> and I think we agree in that the current DS API will need to be
scraped, right? well, the problem is that nobody is discussing what I
think needs to be discussed
marcopg_> tomeu: what do you think needs to be discussed... :)
tomeu> marcopg_: what will be the new API(s)
marcopg_> my suggestion about that would be to look into scott prototype and
build a prototype UI on the top of it. that would give us a much
better base to discuss on. Scott is in charge of designing the API -
the more feedback he gets, better the API will be. I'm not really a
fan of a long abstract discussion of an API, without experimenting
with it.
tomeu> marcopg_: well, activity authors may be able to provide something
more close to their real needs - I don't think we are at the point
where activity authors feedback is useful, yet
marcopg_> when olpcfs is more mature, that will certainly be useful
Why is the current DS API unhappy?
Blaketh> tomeu: Can you point me towards some failings of the current DS API?
bemasc> activity authors are pretty much happy with the current high-level
API, which is what almost all of them use. the only significant issue
is an inability to save multiple files as a single entry
marcopg_> bemasc: well that's your opinion. I heard a bunch of complaints about
the current API, and I've seen a bunch of confused developers
tomeu> Blaketh: see http://wiki.laptop.org/go/DatastoreOpenIssues. being
able to update a file without having to copy it around is one missing
point.
tomeu> bemasc: the issue here is API. right now, activity authors need to
copy the whole file to a temp dir, modify it, and submit again to the
DS. it's the same at the end, but would be good if activity authors
weren't concerned about how the DS stores data. the DS passes a path
where the activity can read or write to so the DS should make the
copy. after the activity is happy with it, commits to the DS, no
matter if the activity is Read, VideoEditor, etc
bemasc> anyone thinking about legacy compatibility should know about
http://www.pygtk.org/docs/pygtk/class-gtkfilechooser.html - Sugar can
provide its own filechooser, thus integrating any GTK application
with the datastore without fancy filesystem footwork, and without
ever showing paths to the user
tomeu> bemasc: well, using the fs as a means for getting things into the
journal is better than that, right?