ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Valentin Kulichenko <valentin.kuliche...@gmail.com>
Subject Re: Service grid redesign
Date Thu, 19 Apr 2018 23:34:54 GMT
Denis,

I totally agree with you. I'm just not sure why do we need two methods
(init and execute) that have virtually same semantics. With the new design,
what would be the difference between these methods from user point of view,
and how one would determine what exactly should go in each of them? Unless
I'm missing something, it looks like unnecessary complication.

-Val

On Tue, Apr 17, 2018 at 1:00 AM, Denis Mekhanikov <dmekhanikov@gmail.com>
wrote:

> Val,
>
> Service initialisation is not going to happen in the discovery thread.
> It should be done asynchronously, and initialisation results should be sent
> to the coordinator over communication.
> This is described in the IEP:
> https://cwiki.apache.org/confluence/display/IGNITE/IEP-
> 17%3A+Oil+Change+in+Service+Grid#IEP-17:OilChangeinServiceGrid-
> Successfulscenario
>
> *init()* method is a validation step, making sure, that service is ready
> for work.
> And deployment shouldn't be considered successful until *init()* methods
> finish their work on the assigned nodes.
> Also *cancel() *and *init() *methods may be useful if we decide to
> implement moving existing services to new nodes
> <https://cwiki.apache.org/confluence/display/IGNITE/IEP-
> 17%3A+Oil+Change+in+Service+Grid#IEP-17:OilChangeinServiceGrid-
> Movingexistingservicestonewnodes>
>  in
> future.
> These methods can be used to save and restore service's state from cache,
> when it is rebalanced to another node.
>
> As Denis said, if we are not going to prevent nodes from starting on
> service failures, then we should at least generate corresponding events.
> Otherwise there won't be any way to react to service initialization
> failures during nodes startup.
>
> Denis
>
> вт, 17 апр. 2018 г. в 6:59, Valentin Kulichenko <
> valentin.kulichenko@gmail.com>:
>
> > I agree we shouldn't do anything synchronously within discovery threads.
> If
> > something goes wrong, we just need to properly notify the user, logging
> and
> > events seem to be right options to achieve that.
> >
> > BTW, with this design I'm not sure init() method makes sense, probably we
> > should deprecate it.
> >
> > -Val
> >
> > On Mon, Apr 16, 2018 at 12:03 PM, Denis Magda <dmagda@apache.org> wrote:
> >
> > > Denis,
> > >
> > > In general, service initialization shouldn't prevent a node from
> joining
> > > the cluster or slowing down that process. Thus, I would start the
> > > initialization routines only after the node is accepted by the cluster.
> > If
> > > the initialization fails then we need to report a respective message to
> > the
> > > logs and, probably, generate a system event the user can be subscribed
> > to.
> > >
> > > Regardless, of the service initialization time, I think we still need
> to
> > > utilize discovery SPI to avoid problems discussed later.
> > >
> > > Val, others, what do you think about that?
> > >
> > >
> > > --
> > > Denis
> > >
> > > On Mon, Apr 16, 2018 at 10:29 AM, Denis Mekhanikov <
> > dmekhanikov@gmail.com>
> > > wrote:
> > >
> > > > Basically, my question is: at which moment services should be
> deployed
> > on
> > > > connecting nodes?
> > > > Should we reject a node from being included into a topology, if
> > services,
> > > > that are assigned to it, fail to deploy?
> > > >
> > > > It would be good to be sure, that all assigned services are
> initialised
> > > and
> > > > working, when node start is finished.
> > > > Otherwise it's unclear, how to notify a user about failures in
> service
> > > > initialisation on new nodes.
> > > >
> > > > If we decide to provide such guarantee, then how are we going to do
> > that?
> > > > Is procedure, that I described, viable?
> > > > It requires hacking through the discovery protocol, which is a thing,
> > > that
> > > > should be avoided.
> > > > So, maybe there is another way to achieve the same thing?
> > > >
> > > > Denis
> > > >
> > > > сб, 14 апр. 2018 г. в 1:48, Denis Magda <dmagda@apache.org>:
> > > >
> > > > > It sounds like it's not a trivial thing to support the automatic
> > > services
> > > > > redeployment after a restart. Let's postpone it for now, guys
> > > > concentrating
> > > > > on more urgent things related to the services.
> > > > >
> > > > > Alex, Vladimir,
> > > > >
> > > > > Could you have a look at Denis question about the discovery-based
> > > > > deployment? Guess it's the only one thing that prevents us from the
> > IEP
> > > > > finalization.
> > > > >
> > > > > --
> > > > > Denis
> > > > >
> > > > > On Fri, Apr 13, 2018 at 5:30 AM, Denis Mekhanikov <
> > > dmekhanikov@gmail.com
> > > > >
> > > > > wrote:
> > > > >
> > > > > > Vladimir,
> > > > > >
> > > > > > Currently we don't save binary metadata to disk, when persistence
> > is
> > > > > > disabled.
> > > > > > But we still persist marshaller mappings for some reason, and
I
> > > > > personally
> > > > > > believe, that we shouldn't.
> > > > > >
> > > > > > But I agree, that we should separate data and service persistence
> > > > > > configuration.
> > > > > > Right now persistence of services is configured in a pretty
> > > non-obvious
> > > > > > manner.
> > > > > > It should be a clear way to tell Ignite, whether you want
> services
> > to
> > > > be
> > > > > > persisted or not.
> > > > > >
> > > > > > I'm not sure, that we should make "statefullness" in general
> > > > > configurable.
> > > > > > Users don't care much, whether metadata is preserved on restarts,
> > or
> > > > not.
> > > > > >
> > > > > > Denis
> > > > > >
> > > > > > пт, 13 апр. 2018 г. в 14:29, Vladimir Ozerov <
> vozerov@gridgain.com
> > >:
> > > > > >
> > > > > > > Alex,
> > > > > > >
> > > > > > > I would say that we've already had this behavior for years
-
> > > > marshaller
> > > > > > > cache. I think it is time to agree that "in-memory" !=
> stateless.
> > > > > Instead
> > > > > > > "in-memory" means "data is not stored on disk".
> > > > > > > May be we can have a flag which will wipe out all metadata
on
> > node
> > > > > > restart
> > > > > > > (e.g. it could make sense for embedded clients)?
> > > > > > >
> > > > > > > On Fri, Apr 13, 2018 at 12:48 PM, Alexey Goncharuk <
> > > > > > > alexey.goncharuk@gmail.com> wrote:
> > > > > > >
> > > > > > > > Denis,
> > > > > > > >
> > > > > > > > This is a subtle question. It looks like we have now
a number
> > of
> > > > > > > use-cases
> > > > > > > > when persistent storage is required even for a pure
in-memory
> > > mode.
> > > > > One
> > > > > > > of
> > > > > > > > the use-cases is thin client authentication, the other
is
> > service
> > > > > grid
> > > > > > > > configuration persistence.
> > > > > > > >
> > > > > > > > Generally, I would agree that this is an expected
behavior.
> > > > However,
> > > > > > this
> > > > > > > > means that a user cannot simply start and stop nodes
randomly
> > > > > anymore.
> > > > > > > > Ignite start will require some sort of installation
or work
> > > folder
> > > > > > > > initialization (sort of initdb in postgres) which
is ok for
> > > > > > > > persistence-enabled modes, but I am not sure if this
is
> > expected
> > > > for
> > > > > > > > in-memory. Of course, we can run this initialization
> > > automatically,
> > > > > but
> > > > > > > it
> > > > > > > > is not always a good idea.
> > > > > > > >
> > > > > > > > If we are ok to have this restrictions for in-memory
mode,
> then
> > > > > service
> > > > > > > > persistence makes sense.
> > > > > > > >
> > > > > > > > --AG
> > > > > > > >
> > > > > > > > 2018-04-11 22:36 GMT+03:00 Denis Magda <dmagda@apache.org>:
> > > > > > > >
> > > > > > > >> Denis,
> > > > > > > >>
> > > > > > > >> I think that the service deployment state needs
be persisted
> > > > > > > cluster-wide.
> > > > > > > >> I guess that our meta-store is capable of doing
so. Alex G.,
> > > > > Vladimir,
> > > > > > > >> could you confirm?
> > > > > > > >>
> > > > > > > >> As for the split-brain scenarios, I would put
them aside for
> > now
> > > > > > > because,
> > > > > > > >> anyway, they have to be solved at lower levels
(meta store,
> > > > > discovery,
> > > > > > > >> etc.).
> > > > > > > >>
> > > > > > > >> Also, I heard that presently we store a service
> configuration
> > in
> > > > the
> > > > > > > >> system
> > > > > > > >> cache that doesn't give us a way to deploy a new
version of
> a
> > > > > service
> > > > > > > >> without undeployment of the previous one. Will
this issue be
> > > > > addressed
> > > > > > > by
> > > > > > > >> the new deployment approach?
> > > > > > > >>
> > > > > > > >> --
> > > > > > > >> Denis
> > > > > > > >>
> > > > > > > >> On Wed, Apr 11, 2018 at 1:28 AM, Denis Mekhanikov
<
> > > > > > > dmekhanikov@gmail.com>
> > > > > > > >> wrote:
> > > > > > > >>
> > > > > > > >> > Denis,
> > > > > > > >> >
> > > > > > > >> > Sounds reasonable. It's not clear, though,
what should
> > happen,
> > > > if
> > > > > a
> > > > > > > >> joining
> > > > > > > >> > node has some services persisted, that are
missing on
> other
> > > > nodes.
> > > > > > > >> > Should we deploy them?
> > > > > > > >> > If we do so, it could lead to surprising
behaviour. For
> > > example
> > > > > you
> > > > > > > >> could
> > > > > > > >> > kill a node, undeploy a service, then bring
back an old
> > node,
> > > > and
> > > > > it
> > > > > > > >> would
> > > > > > > >> > make the service resurrect.
> > > > > > > >> > We could store some deployment counter along
with the
> > service
> > > > > > > >> > configurations on all nodes, that would show
how many
> times
> > > the
> > > > > > > service
> > > > > > > >> > state has changed, i.e. it has been undeployed/redeployed.
> > It
> > > > > should
> > > > > > > be
> > > > > > > >> > kept for undeployed services as well to avoid
situations
> > like
> > > I
> > > > > > > >> described.
> > > > > > > >> >
> > > > > > > >> > But it still leaves a possibility of incorrect
behaviour,
> if
> > > > there
> > > > > > > was a
> > > > > > > >> > split-brain situation at some point. I don't
think we
> should
> > > > > precess
> > > > > > > it
> > > > > > > >> > somehow, though. If we choose to tackle it,
it will
> > > > overcomplicate
> > > > > > > >> things
> > > > > > > >> > for a sake of a minor improvement.
> > > > > > > >> >
> > > > > > > >> > Denis
> > > > > > > >> >
> > > > > > > >> > вт, 10 апр. 2018 г. в 0:55, Valentin
Kulichenko <
> > > > > > > >> > valentin.kulichenko@gmail.com>:
> > > > > > > >> >
> > > > > > > >> > > I was responding to another Denis :)
Agree with you on
> > your
> > > > > point
> > > > > > > >> though.
> > > > > > > >> > >
> > > > > > > >> > > -Val
> > > > > > > >> > >
> > > > > > > >> > > On Mon, Apr 9, 2018 at 2:48 PM, Denis
Magda <
> > > > dmagda@apache.org>
> > > > > > > >> wrote:
> > > > > > > >> > >
> > > > > > > >> > > > Val,
> > > > > > > >> > > >
> > > > > > > >> > > > Guess we're talking about other
situations. I'm
> bringing
> > > up
> > > > > the
> > > > > > > case
> > > > > > > >> > > when a
> > > > > > > >> > > > service was deployed dynamically
and has to be brought
> > up
> > > > > after
> > > > > > a
> > > > > > > >> full
> > > > > > > >> > > > cluster restart w/o user intervention.
To achieve this
> > we
> > > > need
> > > > > > to
> > > > > > > >> > persist
> > > > > > > >> > > > the service's configuration somewhere.
> > > > > > > >> > > >
> > > > > > > >> > > > --
> > > > > > > >> > > > Denis
> > > > > > > >> > > >
> > > > > > > >> > > > On Mon, Apr 9, 2018 at 1:42 PM,
Valentin Kulichenko <
> > > > > > > >> > > > valentin.kulichenko@gmail.com>
wrote:
> > > > > > > >> > > >
> > > > > > > >> > > > > Denis,
> > > > > > > >> > > > >
> > > > > > > >> > > > > EVT_CLASS_DEPLOYED should
be fired every time a
> class
> > is
> > > > > > > deployed
> > > > > > > >> or
> > > > > > > >> > > > > redeployed. If this doesn't
happen in some cases, I
> > > > believe
> > > > > > this
> > > > > > > >> > would
> > > > > > > >> > > > be a
> > > > > > > >> > > > > bug. I don't think we need
to add any new events.
> > > > > > > >> > > > >
> > > > > > > >> > > > > -Val
> > > > > > > >> > > > >
> > > > > > > >> > > > > On Mon, Apr 9, 2018 at 10:50
AM, Denis Magda <
> > > > > > dmagda@apache.org
> > > > > > > >
> > > > > > > >> > > wrote:
> > > > > > > >> > > > >
> > > > > > > >> > > > > > Denis,
> > > > > > > >> > > > > >
> > > > > > > >> > > > > > I would encourage us
to persist a service
> > > configuration
> > > > in
> > > > > > the
> > > > > > > >> meta
> > > > > > > >> > > > store
> > > > > > > >> > > > > > and have this capability
enabled by default.
> That's
> > > > > > essential
> > > > > > > >> for
> > > > > > > >> > > > > services
> > > > > > > >> > > > > > started dynamically.
Moreover, we support similar
> > > > behavior
> > > > > > for
> > > > > > > >> > > caches,
> > > > > > > >> > > > > > indexes, and other DDL
changes happened at
> runtime.
> > > > > > > >> > > > > >
> > > > > > > >> > > > > > --
> > > > > > > >> > > > > > Denis
> > > > > > > >> > > > > >
> > > > > > > >> > > > > > On Mon, Apr 9, 2018 at
9:34 AM, Denis Mekhanikov <
> > > > > > > >> > > > dmekhanikov@gmail.com>
> > > > > > > >> > > > > > wrote:
> > > > > > > >> > > > > >
> > > > > > > >> > > > > > > Another question,
that I would like to discuss
> is
> > > > > whether
> > > > > > > >> > services
> > > > > > > >> > > > > should
> > > > > > > >> > > > > > > be preserved on
cluster restarts.
> > > > > > > >> > > > > > >
> > > > > > > >> > > > > > > Currently it depends
on persistence
> configuration.
> > > If
> > > > > > > >> persistence
> > > > > > > >> > > for
> > > > > > > >> > > > > any
> > > > > > > >> > > > > > > data region is enabled,
then services will be
> > > > persisted
> > > > > as
> > > > > > > >> well.
> > > > > > > >> > > This
> > > > > > > >> > > > > is
> > > > > > > >> > > > > > a
> > > > > > > >> > > > > > > pretty strange way
of configuring this
> behaviour.
> > > > > > > >> > > > > > > I'm not sure, if
anybody relies on this
> > > functionality
> > > > > > right
> > > > > > > >> now.
> > > > > > > >> > > > Should
> > > > > > > >> > > > > > we
> > > > > > > >> > > > > > > support it at all?
If yes, should we make it
> > > > > configurable?
> > > > > > > >> > > > > > >
> > > > > > > >> > > > > > > Denis
> > > > > > > >> > > > > > >
> > > > > > > >> > > > > > > пн, 9 апр.
2018 г. в 19:27, Denis Mekhanikov <
> > > > > > > >> > > dmekhanikov@gmail.com
> > > > > > > >> > > > >:
> > > > > > > >> > > > > > >
> > > > > > > >> > > > > > > > Val,
> > > > > > > >> > > > > > > >
> > > > > > > >> > > > > > > > Sounds reasonable.
I just think, that user
> > should
> > > > have
> > > > > > > some
> > > > > > > >> way
> > > > > > > >> > > to
> > > > > > > >> > > > > > know,
> > > > > > > >> > > > > > > > that new version
of a service class was
> > deployed.
> > > > > > > >> > > > > > > > One way to
do it is to listen to
> > > > *EVT_CLASS_DEPLOYED.
> > > > > > *I'm
> > > > > > > >> not
> > > > > > > >> > > > sure,
> > > > > > > >> > > > > > > > whether it
is triggered on class redeployment,
> > > > though.
> > > > > > If
> > > > > > > >> not,
> > > > > > > >> > > then
> > > > > > > >> > > > > > > another
> > > > > > > >> > > > > > > > event type
should be added.
> > > > > > > >> > > > > > > >
> > > > > > > >> > > > > > > > I don't think,
that a lot of people will
> > implement
> > > > > their
> > > > > > > own
> > > > > > > >> > > > > > > > *DeploymentSpi*-s,
so we should make work with
> > > > > > > >> > *UriDeploymentSpi*
> > > > > > > >> > > > as
> > > > > > > >> > > > > > > > comfortable
as possible.
> > > > > > > >> > > > > > > >
> > > > > > > >> > > > > > > > Denis
> > > > > > > >> > > > > > > >
> > > > > > > >> > > > > > > > пт, 6 апр.
2018 г. в 23:40, Valentin
> Kulichenko
> > <
> > > > > > > >> > > > > > > > valentin.kulichenko@gmail.com>:
> > > > > > > >> > > > > > > >
> > > > > > > >> > > > > > > >> Yes, the
class deployment itself has to be
> > > > explicit.
> > > > > > > I.e.,
> > > > > > > >> > there
> > > > > > > >> > > > has
> > > > > > > >> > > > > > to
> > > > > > > >> > > > > > > be
> > > > > > > >> > > > > > > >> a manual
step where user updates the class,
> and
> > > the
> > > > > > exact
> > > > > > > >> step
> > > > > > > >> > > > > > required
> > > > > > > >> > > > > > > >> would depend
on DeploymentSpi implementation.
> > But
> > > > > then
> > > > > > > >> Ignite
> > > > > > > >> > > > takes
> > > > > > > >> > > > > > care
> > > > > > > >> > > > > > > >> of
> > > > > > > >> > > > > > > >> everything
else - service redeployment and
> > > restart
> > > > is
> > > > > > > >> > automatic.
> > > > > > > >> > > > > > > >>
> > > > > > > >> > > > > > > >> Dmitriy
Pavlov, all this is going to be
> > disabled
> > > if
> > > > > > > >> > > DeploymentSpi
> > > > > > > >> > > > is
> > > > > > > >> > > > > > not
> > > > > > > >> > > > > > > >> configured.
In this case service class
> > > definitions
> > > > > have
> > > > > > > to
> > > > > > > >> be
> > > > > > > >> > > > > deployed
> > > > > > > >> > > > > > > on
> > > > > > > >> > > > > > > >> local classpath
and can't be updated in
> > runtime.
> > > > Just
> > > > > > > like
> > > > > > > >> it
> > > > > > > >> > > > works
> > > > > > > >> > > > > > > right
> > > > > > > >> > > > > > > >> now.
> > > > > > > >> > > > > > > >>
> > > > > > > >> > > > > > > >> -Val
> > > > > > > >> > > > > > > >>
> > > > > > > >> > > > > > > >> On Fri,
Apr 6, 2018 at 10:20 AM, Dmitriy
> > > Setrakyan
> > > > <
> > > > > > > >> > > > > > > dsetrakyan@apache.org
> > > > > > > >> > > > > > > >> >
> > > > > > > >> > > > > > > >> wrote:
> > > > > > > >> > > > > > > >>
> > > > > > > >> > > > > > > >> > On
Fri, Apr 6, 2018 at 9:13 AM, Dmitry
> > Pavlov <
> > > > > > > >> > > > > > dpavlov.spb@gmail.com>
> > > > > > > >> > > > > > > >> > wrote:
> > > > > > > >> > > > > > > >> >
> > > > > > > >> > > > > > > >> > >
Hi Igniters,
> > > > > > > >> > > > > > > >> > >
> > > > > > > >> > > > > > > >> > >
I like automatic redeploy which can be
> > > disabled
> > > > > by
> > > > > > > >> config
> > > > > > > >> > if
> > > > > > > >> > > > > user
> > > > > > > >> > > > > > > >> wants
> > > > > > > >> > > > > > > >> > to
> > > > > > > >> > > > > > > >> > >
control this process. What do you think?
> > > > > > > >> > > > > > > >> > >
> > > > > > > >> > > > > > > >> >
> > > > > > > >> > > > > > > >> > I
do not think we should have anything
> > > automatic
> > > > > when
> > > > > > > it
> > > > > > > >> > comes
> > > > > > > >> > > > to
> > > > > > > >> > > > > > > >> > deployment,
everything should be explicit.
> > > > However,
> > > > > > if
> > > > > > > we
> > > > > > > >> > use
> > > > > > > >> > > > the
> > > > > > > >> > > > > > > >> > deployment
SPI, then a user should be able
> to
> > > do
> > > > > > "hot"
> > > > > > > >> > > redeploy,
> > > > > > > >> > > > > > > where a
> > > > > > > >> > > > > > > >> > new
service will be deployed if the user
> > drops
> > > an
> > > > > > > updated
> > > > > > > >> > jar.
> > > > > > > >> > > > > > > >> >
> > > > > > > >> > > > > > > >> > We
should not create anything new here.
> > Ignite
> > > > > > already
> > > > > > > >> has a
> > > > > > > >> > > > > > > deployment
> > > > > > > >> > > > > > > >> SPI
> > > > > > > >> > > > > > > >> > and
it already works in a certain way.
> Let's
> > > not
> > > > > > change
> > > > > > > >> it.
> > > > > > > >> > > > > > > >> >
> > > > > > > >> > > > > > > >> > D.
> > > > > > > >> > > > > > > >> >
> > > > > > > >> > > > > > > >>
> > > > > > > >> > > > > > > >
> > > > > > > >> > > > > > >
> > > > > > > >> > > > > >
> > > > > > > >> > > > >
> > > > > > > >> > > >
> > > > > > > >> > >
> > > > > > > >> >
> > > > > > > >>
> > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message