Hi Reka,
I need some clarification on how the DependencyBuilder is supposed to work
(while I am trying to replace the code with the new scheme to represent the
startup order).
In the class StartupOrder.java we maintain a structure “private List<String>
startList;” which is being used later in the “DependencyBuilder.java”, see code
snipplet below. However, I can’t find anywhere in the code that the “startList”
is initialized or set (getStartList(…) being invoked). I am not really sure
what the intention is how this code is supposed to work ?
Thanks
Martin
DependencyBuilder.java
Set<StartupOrder> startupOrderSet = dependencyOrder.getStartupOrders();
ApplicationContext foundContext = null;
for (StartupOrder startupOrder : startupOrderSet) {
foundContext = null;
for (String start : startupOrder.getStartList()) {
ApplicationContext applicationContext =
ApplicationContextFactory.
getApplicationContext(start, component,
dependencyTree);
String id = applicationContext.getId(); //TODO change the id
ApplicationContext existingApplicationContext =
dependencyTree.findApplicationContextWithId(id);
if (existingApplicationContext == null) {
if (foundContext != null) {
//appending the start up order to existing
group/cluster
foundContext.addApplicationContext(applicationContext);
if (log.isDebugEnabled()) {
log.debug("Found an existing [dependency] " +
foundContext.getId() +
" and adding the [dependency] " + id +
" as the child");
}
} else {
//adding list of startup order to the dependency
tree
dependencyTree.addApplicationContext(applicationContext);
}
} else {
if (foundContext == null) {
//assigning the found context to the later use.
foundContext = existingApplicationContext;
if (log.isDebugEnabled()) {
log.debug("Found an existing [dependency] " +
id + " and setting it " +
"for the next dependency to follow");
}
} else {
//TODO Throw exception, since another same start
order already found
log.warn("Startup order is not consistent. It
contains the group/cluster " +
"which has been used more than one in
another startup order");
}
}
}
}
From: Reka Thirunavukkarasu [mailto:[email protected]]
Sent: Wednesday, October 08, 2014 9:00 AM
To: Martin Eppel (meppel)
Cc: Isuru Haththotuwa ([email protected]); [email protected]
Subject: Re: [Grouping][Part-2] Composite Application Dependency Tree building
based in Autoscaler
Hi Martin,
On Wed, Oct 8, 2014 at 3:56 AM, Martin Eppel (meppel)
<[email protected]<mailto:[email protected]>> wrote:
Hi Reka,
Just wanted to clarify, the changes suggested below should completely replace
the previous structure (StartupOrder with before, after), not only in the json
definition but also in all subsequent object models, correct ?
Yah..We need to change the json, application parser and the Topology.
Btw, what triggered this change ?
As i think, it is naming issue. If we change it to terminate, then it will be
more consistent with the naming that we currently use in cloud controller. Or
did you mean something else?
Thanks,
Reka
Thanks
Martin
From: Reka Thirunavukkarasu [mailto:[email protected]<mailto:[email protected]>]
Sent: Monday, October 06, 2014 2:22 AM
To: dev
Subject: Re: [Grouping][Part-2] Composite Application Dependency Tree building
based in Autoscaler
Hi Imesh,
On Mon, Oct 6, 2014 at 2:45 PM, Imesh Gunaratne
<[email protected]<mailto:[email protected]>> wrote:
Hi Reka,
I have a small concern on using the term "kill" in this scenario, I think it
would be much more elegant if we call it something like "terminate". WDYT?
+1. It's more appropriate to use terminate. Will change it to terminate.
Thanks,
Reka
On Mon, Oct 6, 2014 at 11:44 AM, Reka Thirunavukkarasu
<[email protected]<mailto:[email protected]>> wrote:
Hi all,
I have implemented the dependency tree as mentioned in my mail earlier. It will
return the immediate children for the start able dependencies.
FYI: a composite application has postgresGroup, php, mysqlGroup, app server
and esb as it's immediate children and their start up order is as mentioned in
the mail earlier.
"startupOrders": [
“postgresGroup, php",
"sqlGroup, tomcat",
"tomcat, apimanager",
"tomcat, esb”
]
So, if we look at the kill behaviour of this composite Application, it will be
like below:
kill-none : none of them will be returned
kill-all: all the elements in that dependency tree will be returned
For eg: if something happened to postgresGroup, then all the children of
dependency tree would be returned as php, mysqlGroup, app server and esb will
be get killed.
kill-dependent: all the children of that particular node in the dependency tree
will be returned.
For eg: If something happened to mysqlGroup, then subsequently tomcat, app
server and esb would be get killed.
Question: in case when we get more than one dependencies to be killed, can we
kill all of them in parallel or do we have to wait until it's dependent
cluster/group got killed?
Thanks,
Reka
On Thu, Oct 2, 2014 at 5:58 PM, Reka Thirunavukkarasu
<[email protected]<mailto:[email protected]>> wrote:
Hi Martin,
On Wed, Oct 1, 2014 at 11:03 PM, Martin Eppel (meppel)
<[email protected]<mailto:[email protected]>> wrote:
Hi Reka,
Are you suggesting to replace the current startupOrder definition with the one
mentioned below ?
"startupOrder" : [
{
"start":"aa",
"after":"bb"
}
]
Replaced with
"startupOrders": [
"mypostgresGroup, myphp",
"mysqlGroup, mytomcat",
"mytomcat, myapimanager",
"mytomcat, myesb"
]
I have a couple of questions,
1. If we use the cartridge alias and the group alias in the group /
application dependency definition how will it work when we auto scale groups ?
My current understanding is that to get group scaling to work we would need 2
parameters – group name (==group.name<http://group.name>) and group instance id
(== group.alias), one static and one dynamic. So I would think we’ll have to
define the application dependencies and group dependencies based on the name
and not the alias, but, during run time we have to calculate the dependencies
based on the alias.
I think is important to make the distinction between group type (or name,) and
group instance Id, without it we won’t be able to implement group scaling, wdyt
?
Thanks for pointing this out..Yah..As you have mentioned, if we are to scale
the groups by creating new groups, then we will be unable to use the groups
alias in place of startuporders. But stratos is tightly coupled with
subscription to cluster as one to one mapping and also, load balancer uses one
to one mapping between cluster and hostname. So, if we are to bring up new
clusters/groups, then things might get complicated in stratos. As i explained
in the [part-1] discussion, we thought of achieving scale by group member and
scale by group using constructing the deployment policy in a more advanced
manner. I will start a separate thread on that. According to all of our
opinion, we can decide on how to follow that up.
IMHO the startupOrders in composite application and group definitions (json )
should look like
"startupOrders": [
“postgresGroup, php",
"sqlGroup, tomcat",
"tomcat, apimanager",
"tomcat, esb”
]
while the runtime representation of the logical relationship model for each
group or cartridge should use the corresponding aliases (or instance Id) so
the monitor will reference the aliases (or instance Ids) while the json
application / group definition will reference the group name (or type) and
cartridge type to define the dependencies, WDYT ?
2. If for example a cartridge has multiple dependencies we would just add
another line to the the startupOrders :
e.g. postgresGroup depends on php and abc would be represented
by:
"startupOrders": [
"postgresGroup, php",
"postgresGroup, “abc”
….
]
Otherwise I think the proposal looks good, +1
Thanks,
Reka
Thanks
Martin
From: Reka Thirunavukkarasu [mailto:[email protected]<mailto:[email protected]>]
Sent: Wednesday, October 01, 2014 5:58 AM
To: dev
Cc: Lakmal Warusawithana; Isuru Haththotuwa; Martin Eppel (meppel); Udara
Liyanage
Subject: [Grouping][Part-2] Composite Application Dependency Tree building
based in Autoscaler
Hi
As you aware, in the composite application we can define the depencies between
groups/cartridges. Autoscaler's responsible is to parse this dependencies and
build up a logical relationship model in order to handle the dependency
information among the child nodes. As we have the hierarchical monitors in
autoscaler, i propose to have dependencies information in each monitor that
they aware of (the immediate child only). In that monitor, we need to identify
the group/cartridge which can be started in parallel. So that a monitor can
look at it's dependency and control it's immediate children based on that. Once
all the children are active, it can pass the control to it's parent. For Eg:
If we take the top level in Composite application which has mysqlGroup,
postgresGroup, php, tomcat, apimanager and esb. If they have an alias saying my
+ cartridge/groupName then we can define the dependency information as follows:
- myPhp depends on myPostgresGroup (means postgresGroup should be
started before php)
- myTomcat depends on myMysqlGroup
- myApiManager depends on myTomcat
- myEsb depends on myTomcat
Like wise, groups will define their own dependency as well.
In that way, we need to represent these dependency information as part of
Composite Application definition/GroupDefinition. In order to represent this
dependency information given above for Composite Application, i would suggest
to have the following in Composite Application definition.
"startupOrders": [
"mypostgresGroup, myphp",
"mysqlGroup, mytomcat",
"mytomcat, myapimanager",
"mytomcat, myesb"
]
You can use the same format in GroupDefinition to define dependencies in a
group.
As per the example, autoscaler will build a dependency tree for
ApplicationMonitor as below in order to identify the parallel and dependent
ones. So that Autoscaler will start up same level children monitors as in
parallel.
[cid:[email protected]]
As above, ApplicationMonitor will start GroupMonitors of myPostgresGroup and
myMysqlGroup in parallel. Once postgres becomes active, ApplicationMonitor will
start ClusterMonitor for myPhp. Once myPostgresGroup becomes active,
ApplicationMonitor will start the immediate child myTomcat. Once myTomcat
becomes active, ApplicationMonitor will start the myAppServer and myEsb in
parallel. This will be applicable for GroupMonitors as well. They can look at
their own dependency tree and will start their children.
Please share your suggestions on the above model to handle the Dependency
information of Composite Application in autoscaler.
Thanks,
Reka
--
Reka Thirunavukkarasu
Senior Software Engineer,
WSO2, Inc.:http://wso2.com,
Mobile: +94776442007<tel:%2B94776442007>
--
Reka Thirunavukkarasu
Senior Software Engineer,
WSO2, Inc.:http://wso2.com,
Mobile: +94776442007<tel:%2B94776442007>
--
Reka Thirunavukkarasu
Senior Software Engineer,
WSO2, Inc.:http://wso2.com,
Mobile: +94776442007<tel:%2B94776442007>
--
Imesh Gunaratne
Technical Lead, WSO2
Committer & PMC Member, Apache Stratos
--
Reka Thirunavukkarasu
Senior Software Engineer,
WSO2, Inc.:http://wso2.com,
Mobile: +94776442007<tel:%2B94776442007>
--
Reka Thirunavukkarasu
Senior Software Engineer,
WSO2, Inc.:http://wso2.com,
Mobile: +94776442007
