I will be interested to see if you can
successfully map the OSGi uses concept into the SAT solver p2 uses. I
briefly looked at that a long time ago when we were refactoring the Equinox
framework (Luna) and were replacing the old Equinox resolver. It
was far from obvious how you would achieve this. At that time I opt'ed
to collaborate with a common resolver in Felix for the Equinox framework.
But this is no magic implementation. There are still cases
where the OSGi resolver algorithm will blow up. In Equinox we try
to minimize that possibility by avoiding the resolution of all (10000)
bundles at once. But as Pascal states, this does leave out some possible
valid solutions that you will then not discover while resolving.
If you do focus on how to map uses into
the SAT solver in p2 I would be interested in collaborating to see if such
a resolver would outperform the Felix resolver we use at runtime. My
hope at the time I was looking into this was to map an OSGi Resolver service
on top of the SAT solver implementation. But I cannot remember how
the SAT solver is driven. I suspect the split between the OSGI Resovler
and the OSGi ResolveContext will not fit well into the SAT implementation
Todor Boev <rinsvind@xxxxxxxxx> To:
mailing list <equinox-dev@xxxxxxxxxxx> Date:
11/17/2016 02:22 AM Subject:
Convergence between p2 and the OSGi resolver+repository Sent by:
- Regarding batch resolution: Ultimately I think the batch processing is about performance.
At provisioning time where finding the best solution trumps speed the resolver
can be executed against the entire set. But I have to try this. After than
the equinox runtime should be able to re-create a correct (maybe not identical)
resolution from the much smaller set of resources. I have tried the resolver
against about 700 bundles and it did okay, but this is well short of 10,000.
More research required....some day.
- Regarding the additional p2 concepts: Can you point me to the documentation of how the resolution
problem is converted to a SAT formula?
On Thu, Nov 17, 2016 at 6:20 AM, Pascal Rapicault <pascal@xxxxxxxxxxxxx>
wrote: On 11/16/2016 10:49 AM, Todor Boev wrote: - Regarding resolver behavior: The goal is actually to replace the behavior of
the objective function with the behavior of the resolver. This is the best
way to guarantee that both p2 and the OSGi runtime agree on what is a consistent
set of bundles. For example p2 does not take into account package uses
constraints which leads to p2 selecting bundles that later fail to resolve
at runtime. It does not matter which way to resolve is better, so long
as they agree. Since the OSGi resolver is very unlikely to change it falls
on p2 to match it's behavior. My current company (software ag) has had
quite a number of issues where essentially p2 sets up the resolver to fail.
- Regarding resolver scalability: The resolution is split between the resolver which
processes the current set of resources and the resolver context which selects
candidates when asked. If the goal is to support a very high number of
candidates - a resolver context impl optimized for searches in a large
candidate space can be provided. If the goal is to produce a solution that
includes a very high number of resources - more research is required. Even
if the initial set is 10,000 the resolver can be asked to process them
not all at once, but incrementally in batches or even one by one. Which
is in fact what equinox does today. The thing is that if you look at a
subset of the available bundles, you may find a solution that is not the
optimal one. p2 will consider all the possible candidates in one resolution
I am trying to determine if it makes sense to invest effort in prototyping
this given that subtle changes in behavior are in fact a goal, rather than
an issue. Even though on the surface p2 resolver
looks similar to what the OSGi resolver does, p2 has at least 2 additional
concepts: 1) the _expression_ of strict negation 2) the concept of patch
I'm tempted to think that it is probably simpler to add support for the
uses-clause in p2 (this has been a known issue for years, but I can't seem
to find the bug tonight) than it is to replace the resolver completely
and get all the tests to pass. The encoding of dependencies to a SAT formula
is well documented and so are the optimization functions.
On Wed, Nov 16, 2016 at 4:44 AM, Pascal Rapicault <pascal@xxxxxxxxxxxxx>
wrote: On 11/15/2016 12:52 PM, Todor Boev wrote: Hello,
Are there any plans to bring together p2 and OSGi resolver+repository
standards? There is no immediate plan for this.
It should be beneficial to have similar (maybe identical?) dependency resolution
at provisioning time and later at runtime. The install time and runtime resolvers
resolve a slightly different problem because the install time resolver
has to look for candidates in a large space, whereas the runtime one has
to connect as many components together. I have not tried replacing the p2 resolver with the
new OSGi resolver so I can't tell how it would perform.
Specifically: - Is it possible to publish the bundle generic capabilities/requirements
to the p2 repository? Yes this should be possible. The underlying
p2 capability / requirement model is really extensible and the current
limitation is only the serialized format.
- Is it possible to use the equinox Resolver inside the
p2 Planner? It is possible to get something going
but I'm not sure if this will scale (p2 resolver is able to perform seamlessly
on 10's of thousands of IUs), nor if you will be able to replicate the
subtleties that result from having an objective function.
- Even if the equinox Resolver can not be used is
it possible for p2 to handle generic requirements/capabilities? Yes. This should not be too much work.