ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Alexey Goncharuk <alexey.goncha...@gmail.com>
Subject Re: Cluster auto activation design proposal
Date Sat, 04 Nov 2017 18:12:09 GMT
Dmitriy,

The main point of having a BaselineNode interface is that baseline topology
will contain nodes even when they are offline. When a node is offline, most
of the methods on ClusterNode are meaningless, thus the new interface (for
example, node ID). I left on the interface only methods which will reliably
return data.

Ilya,

For now, we must keep the old AffinityFunction interface, but we can change
it in the AI 3.0.

--AG

2017-11-04 17:56 GMT+03:00 Dmitriy Setrakyan <dsetrakyan@apache.org>:

> Alexey, what is the point of BaselineNode interface? Why not just have
> ClusterNode?
>
> On Fri, Nov 3, 2017 at 11:26 PM, Alexey Goncharuk <
> alexey.goncharuk@gmail.com> wrote:
>
> > Guys,
> >
> > We are getting closer to the baseline topology finalization. As a first
> > step, I would like to request a review of the baseline topology
> management
> > API. The changes are summarized in [1]. In my opinion, changes are quite
> > simple and concise. Also, as a side note, I suggest moving cluster
> > activation methods to the IgniteCluter facade as well because the facade
> > itself looks like a good place for management API. Looks like the
> original
> > decision to place it on Ignite was wrong.
> >
> > Thanks!
> >
> > [1] https://issues.apache.org/jira/browse/IGNITE-5850
> >
> > 2017-09-04 17:46 GMT+03:00 Dmitriy Setrakyan <dsetrakyan@apache.org>:
> >
> > > On Mon, Sep 4, 2017 at 6:13 AM, Sergey Chugunov <
> > sergey.chugunov@gmail.com
> > > >
> > > wrote:
> > >
> > > > Dmitriy,
> > > >
> > > > I like the idea of ClusterActivator interface.
> > > >
> > > > From user perspective it provides the same functionality as the
> setter
> > > but
> > > > in more clear and intuitive way.
> > > >
> > >
> > > BTW, I made a naming mistake in the original email. The setter name
> > should
> > > be "setClusterActivator(...).
> > >
> > >
> > > >
> > > > Also it gives us a good place to put all the documentation about the
> > > > feature.
> > > >
> > >
> > > Agree. Another advantage is that users can now provide custom logic for
> > the
> > > initial cluster activation.
> > >
> > >
> > > >
> > > > Any other opinions?
> > > >
> > >
> > > Alexey Goncharuk, given that you provided the initial implementation of
> > the
> > > cluster activation, can you please take a look at this design and
> provide
> > > comments?
> > >
> > >
> > > >
> > > > On Fri, Sep 1, 2017 at 2:35 PM, Dmitriy Setrakyan <
> > dsetrakyan@apache.org
> > > >
> > > > wrote:
> > > >
> > > > > How about this:
> > > > >
> > > > >
> > > > > > *interface ClusterActivator {*
> > > > > > *    boolean activate(Collection<IgniteNode> nodes);**}*
> > > > >
> > > > >
> > > > > Out of the box, we can provide this implementation of the
> activation
> > > > > filter:
> > > > >
> > > > >
> > > > > >
> > > > > > *ClusterInitialActiveSet implements ClusterActivator { *
> > > > > > *    InigeInitialActiveSet(String... addresses);**}*
> > > > >
> > > > >
> > > > > Then user configuration can look as follows:
> > > > >
> > > > > *IgniteConfiguration.setActivationFilter(new
> > > > > > ClusterInitialActiveSet("1.2.3.4", "4.3.2.1", etc));*
> > > > >
> > > > >
> > > > > Thoughts?
> > > > >
> > > > > D.
> > > > >
> > > > > On Fri, Sep 1, 2017 at 1:47 PM, Sergey Chugunov <
> > > > sergey.chugunov@gmail.com
> > > > > >
> > > > > wrote:
> > > > >
> > > > > > Dmitriy,
> > > > > >
> > > > > > The idea is interesting however I cannot come up with a clear
use
> > > case
> > > > > > which can be widely adopted.
> > > > > > I would give users a simple API at first to cover 80% of their
> > needs
> > > > and
> > > > > > then collect some feedback and start thinking about adding new
> > > > > > functionality.
> > > > > >
> > > > > > Makes sense?
> > > > > >
> > > > > > Sergey.
> > > > > >
> > > > > > On Thu, Aug 31, 2017 at 3:55 AM, Dmitriy Setrakyan <
> > > > > dsetrakyan@apache.org>
> > > > > > wrote:
> > > > > >
> > > > > > > Hm... Can we also ask user to optionally provide a predicate
> > which
> > > > will
> > > > > > > receive a collection of nodes started so far and return
true if
> > the
> > > > > > > activation should happen? Will it be useful?
> > > > > > >
> > > > > > > On Wed, Aug 30, 2017 at 6:28 PM, Sergey Chugunov <
> > > > > > > sergey.chugunov@gmail.com>
> > > > > > > wrote:
> > > > > > >
> > > > > > > > Nick,
> > > > > > > >
> > > > > > > > As I summed up in this thread above, calling setter
for
> initial
> > > > > > > activation
> > > > > > > > nodes is not the only option:
> > > > > > > >
> > > > > > > >    1. user starts up new cluster of desired number
of nodes
> and
> > > > > > activates
> > > > > > > >    it using existing API.
> > > > > > > >    BLT is created with all nodes presented in the
cluster at
> > the
> > > > > moment
> > > > > > > of
> > > > > > > >    activation, no API is needed;
> > > > > > > >
> > > > > > > >    2. user prepares BLT using web-console or visor
CMD tools
> > and
> > > > sets
> > > > > > it
> > > > > > > to
> > > > > > > >    the cluster. New API setter is needed:
> > > > > > > >    Ignite.activation().setInitialActivationNodes(Collection<
> > > > > > ClusterNode>
> > > > > > > >    nodes);
> > > > > > > >
> > > > > > > >    3. user provides via static configuration a list
of nodes
> > that
> > > > are
> > > > > > > >    expected to be in the cluster.
> > > > > > > >    User starts nodes one by one; when all preconfigured
nodes
> > are
> > > > > > started
> > > > > > > >    cluster is activated and BLT is created.
> > > > > > > >    As list of nodes may be huge it is provided via
separate
> > file
> > > to
> > > > > > avoid
> > > > > > > >    flooding main configuration.
> > > > > > > >
> > > > > > > > So the option you proposed is already in the list.
> > > > > > > >
> > > > > > > > As for idea of activating cluster based only on number
of
> nodes
> > > may
> > > > > be
> > > > > > > > risky.
> > > > > > > > E.g. if user starts up with data stored on disk and
> unexpected
> > > node
> > > > > > joins
> > > > > > > > the topology.
> > > > > > > > Cluster will get activated with N-1 nodes where all
the data
> is
> > > > > > presented
> > > > > > > > and one node completely empty. Data loss may happen
in such
> > > > scenario.
> > > > > > > >
> > > > > > > > Thanks,
> > > > > > > > Sergey.
> > > > > > > >
> > > > > > > >
> > > > > > > > On Wed, Aug 30, 2017 at 4:23 PM, Nick Pordash <
> > > > nickpordash@gmail.com
> > > > > >
> > > > > > > > wrote:
> > > > > > > >
> > > > > > > > > How is a user expected to produce a collection
of
> ClusterNode
> > > > prior
> > > > > > to
> > > > > > > > all
> > > > > > > > > of the expected nodes joining? Users don't create
instances
> > of
> > > > > this,
> > > > > > so
> > > > > > > > as
> > > > > > > > > far as I can tell it would have to be retrieved
from
> > > > IgniteCluster.
> > > > > > > > > However, would doing that and calling the proposed
method
> be
> > > > really
> > > > > > any
> > > > > > > > > different than calling Ignite.activate and using
the
> current
> > > set
> > > > of
> > > > > > > > server
> > > > > > > > > nodes as that collection?
> > > > > > > > >
> > > > > > > > > From a user's perspective is it really necessary
that
> > specific
> > > > > nodes
> > > > > > > need
> > > > > > > > > to be identified vs saying that they expect N
server nodes
> to
> > > be
> > > > in
> > > > > > the
> > > > > > > > > cluster for auto activation?
> > > > > > > > >
> > > > > > > > > -Nick
> > > > > > > > >
> > > > > > > > > On Wed, Aug 30, 2017, 1:23 AM Sergey Chugunov
<
> > > > > > > sergey.chugunov@gmail.com
> > > > > > > > >
> > > > > > > > > wrote:
> > > > > > > > >
> > > > > > > > > > Dmitriy,
> > > > > > > > > >
> > > > > > > > > > Now I see your point and I think you're
right.
> > > > > > > > > >
> > > > > > > > > > We can give end-user a simple setter like
> > > > > > > > > > Ignite::activation::setInitialActivationNodes(
> > > > > > > Collection<ClusterNode>
> > > > > > > > > > nodes) to provide collection of nodes that
grid must
> reach
> > to
> > > > > > > activate
> > > > > > > > > > automatically.
> > > > > > > > > >
> > > > > > > > > > And then using the collection we'll create
> BaselineTopology
> > > > > > > abstraction
> > > > > > > > > > internally.
> > > > > > > > > >
> > > > > > > > > > As a result user won't be exposed to our
internal
> > > abstractions
> > > > > and
> > > > > > > will
> > > > > > > > > > work with intuitive concept of collection
of nodes.
> > > > > > > > > >
> > > > > > > > > > Thanks,
> > > > > > > > > > Sergey.
> > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > On Mon, Aug 28, 2017 at 4:39 PM, Dmitriy
Setrakyan <
> > > > > > > > > dsetrakyan@apache.org>
> > > > > > > > > > wrote:
> > > > > > > > > >
> > > > > > > > > > > Sergey, the interface you are suggesting
is internal,
> not
> > > > > > external.
> > > > > > > > Why
> > > > > > > > > > > should user ever see it or care about
it?
> > > > > > > > > > >
> > > > > > > > > > > D.
> > > > > > > > > > >
> > > > > > > > > > > On Mon, Aug 28, 2017 at 3:18 PM, Sergey
Chugunov <
> > > > > > > > > > > sergey.chugunov@gmail.com>
> > > > > > > > > > > wrote:
> > > > > > > > > > >
> > > > > > > > > > > > Dmitriy,
> > > > > > > > > > > >
> > > > > > > > > > > > It was my misunderstanding, I
believe that setter is
> > not
> > > > > enough
> > > > > > > and
> > > > > > > > > we
> > > > > > > > > > > need
> > > > > > > > > > > > a full-fledged entity.
> > > > > > > > > > > >
> > > > > > > > > > > > We should also be able to check
if BLTs are
> compatible.
> > > > > > Interface
> > > > > > > > > looks
> > > > > > > > > > > > like this and use case for this
functionality is
> > > described
> > > > > > below.
> > > > > > > > > > > >
> > > > > > > > > > > > interface BaselineTopology {
> > > > > > > > > > > >    Collection<ClusterNode>
nodes();
> > > > > > > > > > > >    boolean isCompatibleWith(BaselineTopology
blt);
> > > > > > > > > > > > }
> > > > > > > > > > > >
> > > > > > > > > > > > Let's consider the following scenario:
> > > > > > > > > > > >
> > > > > > > > > > > >    1. We have a grid with N nodes:
it is up, active
> and
> > > has
> > > > > > data
> > > > > > > in
> > > > > > > > > it.
> > > > > > > > > > > ->
> > > > > > > > > > > >    BLT #1 created.
> > > > > > > > > > > >    2. We shutdown the grid. Then
divide it into two
> > > parts:
> > > > > > > > Part1_grid
> > > > > > > > > > and
> > > > > > > > > > > >    Part2_grid.
> > > > > > > > > > > >    3. We start and activate Part1_grid
. Topology has
> > > > changed
> > > > > > ->
> > > > > > > > > BLT#2
> > > > > > > > > > > >    created.
> > > > > > > > > > > >    After that we shutdown that
Part1_grid.
> > > > > > > > > > > >    4. We start and activate Part2_grid.
Topology also
> > has
> > > > > > changed
> > > > > > > > ->
> > > > > > > > > > > BLT#3
> > > > > > > > > > > >    created.
> > > > > > > > > > > >    5. Then we start Part1_grid
and it's nodes try to
> > join
> > > > > > > > Part2_grid.
> > > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > > > If join is successful we have
an undefined state of
> the
> > > > > > resulting
> > > > > > > > > grid:
> > > > > > > > > > > > values for the same key may (and
will) differ between
> > > grid
> > > > > > parts.
> > > > > > > > > > > >
> > > > > > > > > > > > So to prevent this we should keep
nodes with BLT#2
> from
> > > > > joining
> > > > > > > the
> > > > > > > > > > grid
> > > > > > > > > > > > with BLT#3. And we should fail
nodes with an error
> > > message.
> > > > > > > > > > > >
> > > > > > > > > > > > Thanks,
> > > > > > > > > > > > Sergey.
> > > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > > > On Wed, Aug 23, 2017 at 5:47 AM,
Dmitriy Setrakyan <
> > > > > > > > > > > dsetrakyan@apache.org>
> > > > > > > > > > > > wrote:
> > > > > > > > > > > >
> > > > > > > > > > > > > Sergey, I am still confused.
What is the
> > > BaselineTopology
> > > > > > > > interface
> > > > > > > > > > in
> > > > > > > > > > > > your
> > > > > > > > > > > > > example? I thought that you
agreed with me that we
> > > simply
> > > > > > need
> > > > > > > a
> > > > > > > > > > setter
> > > > > > > > > > > > for
> > > > > > > > > > > > > activation nodes, no?
> > > > > > > > > > > > >
> > > > > > > > > > > > > D.
> > > > > > > > > > > > >
> > > > > > > > > > > > > On Tue, Aug 22, 2017 at 4:47
AM, Sergey Chugunov <
> > > > > > > > > > > > > sergey.chugunov@gmail.com>
> > > > > > > > > > > > > wrote:
> > > > > > > > > > > > >
> > > > > > > > > > > > > > Dmitriy,
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > As I understand you
use the term
> > > > "minimalActivationNodes"
> > > > > > as
> > > > > > > a
> > > > > > > > > > > synonym
> > > > > > > > > > > > > for
> > > > > > > > > > > > > > BaselineTopology concept.
> > > > > > > > > > > > > > In that case I agree
with you that we can replace
> > > both
> > > > > > > > > "establish*"
> > > > > > > > > > > > > methods
> > > > > > > > > > > > > > with a simple setter
method (see below in
> summary).
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > Summing up the whole
discussion I see the
> > > functionality
> > > > > as
> > > > > > > > > > following:
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > New concept BaselineTopology
is introduced. The
> > main
> > > > > > features
> > > > > > > > it
> > > > > > > > > > > > enables
> > > > > > > > > > > > > > are:
> > > > > > > > > > > > > >
> > > > > > > > > > > > > >    1. automatic activation
of cluster;
> > > > > > > > > > > > > >
> > > > > > > > > > > > > >    2. easy management
of cluster topology changes
> > > > > (planned
> > > > > > > > nodes
> > > > > > > > > > > > > >    maintenance, adding
new nodes etc);
> > > > > > > > > > > > > >
> > > > > > > > > > > > > >    3. eliminating of
rebalancing traffic on
> > > short-term
> > > > > node
> > > > > > > > > > failures.
> > > > > > > > > > > > > >
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > Use cases to create
BLT:
> > > > > > > > > > > > > >
> > > > > > > > > > > > > >    1. user starts up
new cluster of desired
> number
> > of
> > > > > nodes
> > > > > > > and
> > > > > > > > > > > > activates
> > > > > > > > > > > > > >    it using existing
API. BLT is created with all
> > > nodes
> > > > > > > > presented
> > > > > > > > > > in
> > > > > > > > > > > > the
> > > > > > > > > > > > > >    cluster at the moment
of activation, no API is
> > > > needed;
> > > > > > > > > > > > > >
> > > > > > > > > > > > > >    2. user prepares
BLT using web-console or
> visor
> > > CMD
> > > > > > tools
> > > > > > > > and
> > > > > > > > > > sets
> > > > > > > > > > > > it
> > > > > > > > > > > > > to
> > > > > > > > > > > > > >    the cluster. New
API setter is needed:
> > > > > > > > > > > > > >    Ignite.activation().setBaselin
> > > > > > eTopology(BaselineTopology
> > > > > > > > > blt);
> > > > > > > > > > > > > >
> > > > > > > > > > > > > >    3. user provides
via static configuration a
> list
> > > of
> > > > > > nodes
> > > > > > > > that
> > > > > > > > > > are
> > > > > > > > > > > > > >    expected to be in
the cluster.
> > > > > > > > > > > > > >    User starts nodes
one by one; when all
> > > preconfigured
> > > > > > nodes
> > > > > > > > are
> > > > > > > > > > > > started
> > > > > > > > > > > > > >    cluster is activated
and BLT is created.
> > > > > > > > > > > > > >    As list of nodes
may be huge it is provided
> via
> > > > > separate
> > > > > > > > file
> > > > > > > > > to
> > > > > > > > > > > > avoid
> > > > > > > > > > > > > >    flooding main configuration.
> > > > > > > > > > > > > >
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > Igniters, does this
description match with your
> > > > > > understanding
> > > > > > > > of
> > > > > > > > > > > > > > functionality? If it
does I'll create a set of
> > > tickets
> > > > > and
> > > > > > > > start
> > > > > > > > > > > > working
> > > > > > > > > > > > > on
> > > > > > > > > > > > > > implementation.
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > Thanks,
> > > > > > > > > > > > > > Sergey.
> > > > > > > > > > > > > >
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > On Sat, Aug 19, 2017
at 5:41 PM, Dmitriy
> Setrakyan
> > <
> > > > > > > > > > > > > dsetrakyan@apache.org>
> > > > > > > > > > > > > > wrote:
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > > I still do not
see why anyone would explicitly
> > call
> > > > > > these 2
> > > > > > > > > > > methods:
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > *Ignite::activation::
> > establishBaselineTopology();*
> > > > > > > > > > > > > > > *Ignite::activation::
> establishBaselineTopology(
> > > > > > > > > BaselineTopology
> > > > > > > > > > > > > > bltTop);*
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > For example, if
a web console, or some other
> > admin
> > > > > > process,
> > > > > > > > > want
> > > > > > > > > > to
> > > > > > > > > > > > > > > automatically set
currently started nodes as
> the
> > > > > baseline
> > > > > > > > > > topology,
> > > > > > > > > > > > > > > shouldn't they
just call a setter for
> > > > > > > minimalActivationNodes?
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > D.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > On Fri, Aug 18,
2017 at 10:18 AM, Alexey
> > Dmitriev <
> > > > > > > > > > > > > > admitriev@gridgain.com>
> > > > > > > > > > > > > > > wrote:
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > API is proposed
in the head of the thread by
> > > > Sergey,
> > > > > > as I
> > > > > > > > > > > > understood:
> > > > > > > > > > > > > > > > ______________________________
> > > > > ________________________
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > API for BaselineTopology
manipulation may
> look
> > > like
> > > > > > this:
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > *Ignite::activation::
> > > establishBaselineTopology();*
> > > > > > > > > > > > > > > > *Ignite::activation::
> > establishBaselineTopology(
> > > > > > > > > BaselineTopology
> > > > > > > > > > > > > > > bltTop);*
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > Both methods
will establish BT and activate
> > > cluster
> > > > > > once
> > > > > > > it
> > > > > > > > > is
> > > > > > > > > > > > > > > established.
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > The first
one allows user to establish BT
> using
> > > > > current
> > > > > > > > > > topology.
> > > > > > > > > > > > If
> > > > > > > > > > > > > > any
> > > > > > > > > > > > > > > > changes happen
to the topology during
> > > establishing
> > > > > > > process,
> > > > > > > > > > user
> > > > > > > > > > > > will
> > > > > > > > > > > > > > be
> > > > > > > > > > > > > > > > notified and
allowed to proceed or abort the
> > > > > procedure.
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > Second method
allows to use some
> > > > > > monitoring'n'management
> > > > > > > > > tools
> > > > > > > > > > > like
> > > > > > > > > > > > > > > > WebConsole
where user can prepare a list of
> > > nodes,
> > > > > > using
> > > > > > > > them
> > > > > > > > > > > > create
> > > > > > > > > > > > > a
> > > > > > > > > > > > > > BT
> > > > > > > > > > > > > > > > and send to
the cluster a command to finally
> > > > > establish
> > > > > > > it.
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > From high
level BaselineTopology entity
> > contains
> > > > only
> > > > > > > > > > collection
> > > > > > > > > > > of
> > > > > > > > > > > > > > > nodes:
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > *BaselineTopology
{*
> > > > > > > > > > > > > > > > *  Collection<TopologyNode>
nodes;*
> > > > > > > > > > > > > > > > *}*
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > *TopologyNode*
here contains information
> about
> > > > node -
> > > > > > its
> > > > > > > > > > > > consistent
> > > > > > > > > > > > > id
> > > > > > > > > > > > > > > and
> > > > > > > > > > > > > > > > set of user
attributes used to calculate
> > affinity
> > > > > > > function.
> > > > > > > > > > > > > > > > ____________________________________________
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > --
> > > > > > > > > > > > > > > > View this
message in context:
> > > > http://apache-ignite-
> > > > > > > > > > > > > > > > developers.2346864.n4.nabble.
> > > > > > > com/Cluster-auto-activation-
> > > > > > > > > > > > > > > > design-proposal-tp20295p21066.html
> > > > > > > > > > > > > > > > Sent from
the Apache Ignite Developers
> mailing
> > > list
> > > > > > > archive
> > > > > > > > > at
> > > > > > > > > > > > > > > Nabble.com.
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > >
> > > > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > >
> > > > >
> > > >
> > >
> >
>

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