OSDir


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [DISCUSS] CloudStack graceful shutdown


Rafael and Community

All is well and good and i think we are thinking along the similar lines -
the only issue that i see right now with any approach is KVM Agents (or
direct agents) and using LoadBalancer on 8250.

Here is a scenario:

You have 2 Management Server setup fronted with a VIP on 8250.
The LB Algorithm is either Round Robin or Least Connections used.
You initiate a maintenance mode operation on one of the MS servers (call it
MS1) - assume you have a long running migration job that needs 60 minutes
to complete.
We attempt to evacuate the agents by telling them to disconnect and
reconnect again
If we are using LB on 8250 with
1) Least Connection used - then all agents will continuously try to connect
to a MS1 node that is attempting to go down for maintenance. Essentially
with this  LB configuration this operation will never
2) Round Robin - this will take a while - but eventually - you will get all
nodes connected to MS2

The current limitation is usage of external LB on 8250. For this operation
to work without issue - would mean agents must connect to MS server without
an LB. This is a recent feature we've developed with ShapeBlue - where we
maintain the list of CloudStack Management Servers in the agent.properties
file.

Unless you can think of other solution - it appears we may have to forced
to bypass the 8250 VIP LB and use the new feature to maintain the list of
management servers within agent.properties.


I need to run now, let me know what your thoughts are.

Regards
ilya



On Tue, Apr 17, 2018 at 8:27 AM, Rafael Weingärtner <
rafaelweingartner@xxxxxxxxx> wrote:

> Ilya and others,
>
> We have been discussing this idea of graceful/nicely shutdown.  Our feeling
> is that we (in CloudStack community) might have been trying to solve this
> problem with too much scripting. What if we developed a more integrated
> (native) solution?
>
> Let me explain our idea.
>
> ACS has a table called “mshost”, which is used to store management server
> information. During balancing and when jobs are dispatched to other
> management servers this table is consulted/queried.  Therefore, we have
> been discussing the idea of creating a management API for management
> servers.  We could have an API method that changes the state of management
> servers to “prepare to maintenance” and then “maintenance” (as soon as all
> of the task/jobs it is managing finish). The idea is that during
> rebalancing we would remove the hosts of servers that are not in “Up” state
> (of course we would also ignore hosts in the aforementioned state to
> receive hosts to manage).  Moreover, when we send/dispatch jobs to other
> management servers, we could ignore the ones that are not in “Up” state
> (which is something already done).
>
> By doing this, the nicely shutdown could be executed in a few steps.
>
> 1 – issue the maintenance method for the management server you desire
> 2 – wait until the MS goes into maintenance mode, while there are still
> running jobs it (the management server) will be maintained in prepare for
> maintenance
> 3 – execute the Linux shutdown command
>
> We would need other APIs methods to manage MSs then. An (i) API method to
> list MSs, and we could even create an (ii) API to remove old/de-activated
> management servers, which we currently do not have (forcing users to apply
> changed directly in the database).
>
> Moreover, in this model, we would not kill hanging jobs; we would wait
> until they expire and ACS expunges them. Of course, it is possible to
> develop a forceful maintenance method as well. Then, when the “prepare for
> maintenance” takes longer than a parameter, we could kill hanging jobs.
>
> All of this would allow the MS to be kept up and receiving requests until
> it can be safely shutdown. What do you guys about this approach?
>
> On Tue, Apr 10, 2018 at 6:52 PM, Yiping Zhang <yzhang@xxxxxxxxxxx> wrote:
>
> > As a cloud admin, I would love to have this feature.
> >
> > It so happens that I just accidentally restarted my ACS management server
> > while two instances are migrating to another Xen cluster (via storage
> > migration, not live migration).  As results, both instances
> > ends up with corrupted data disk which can't be reattached or migrated.
> >
> > Any feature which prevents this from happening would be great.  A low
> > hanging fruit is simply checking for
> > if there are any async jobs running, especially any kind of migration
> jobs
> > or other known long running type of
> > jobs and warn the operator  so that he has a chance to abort server
> > shutdowns.
> >
> > Yiping
> >
> > On 4/5/18, 3:13 PM, "ilya musayev" <ilya.mailing.lists@xxxxxxxxx>
> wrote:
> >
> >     Andrija
> >
> >     This is a tough scenario.
> >
> >     As an admin, they way i would have handled this situation, is to
> > advertise
> >     the upcoming outage and then take away specific API commands from a
> > user a
> >     day before - so he does not cause any long running async jobs. Once
> >     maintenance completes - enable the API commands back to the user.
> > However -
> >     i dont know who your user base is and if this would be an acceptable
> >     solution.
> >
> >     Perhaps also investigate what can be done to speed up your long
> running
> >     tasks...
> >
> >     As a side node, we will be working on a feature that would allow for
> a
> >     graceful termination of the process/job, meaning if agent noticed a
> >     disconnect or termination request - it will abort the command in
> > flight. We
> >     can also consider restarting this tasks again or what not - but it
> > would
> >     not be part of this enhancement.
> >
> >     Regards
> >     ilya
> >
> >     On Thu, Apr 5, 2018 at 6:47 AM, Andrija Panic <
> andrija.panic@xxxxxxxxx
> > >
> >     wrote:
> >
> >     > Hi Ilya,
> >     >
> >     > thanks for the feedback - but in "real world", you need to
> > "understand"
> >     > that 60min is next to useless timeout for some jobs (if I
> understand
> > this
> >     > specific parameter correctly ?? - job is really canceled, not only
> > job
> >     > monitoring is canceled ???) -
> >     >
> >     > My value for the  "job.cancel.threshold.minutes" is 2880 minutes (2
> > days?)
> >     >
> >     > I can tell you when you have CEPH/NFS (CEPH even "worse" case,
> since
> > slower
> >     > read durign qemu-img convert process...) of 500GB, then imagine
> > snapshot
> >     > job will take many hours. Should I mention 1TB volumes (yes, we had
> >     > client's like that...)
> >     > Than attaching 1TB volume, that was uploaded to ACS (lives
> > originally on
> >     > Secondary Storage, and takes time to be copied over to NFS/CEPH)
> > will take
> >     > up to few hours.
> >     > Then migrating 1TB volume from NFS to CEPH, or CEPH to NFS, also
> > takes
> >     > time...etc.
> >     >
> >     > I'm just giving you feedback as "user", admin of the cloud, zero
> DEV
> > skills
> >     > here :) , just to make sure you make practical decisions (and I
> > admit I
> >     > might be wrong with my stuff, but just giving you feedback from our
> > public
> >     > cloud setup)
> >     >
> >     >
> >     > Cheers!
> >     >
> >     >
> >     >
> >     >
> >     > On 5 April 2018 at 15:16, Tutkowski, Mike <
> Mike.Tutkowski@xxxxxxxxxx
> > >
> >     > wrote:
> >     >
> >     > > Wow, there’s been a lot of good details noted from several people
> > on how
> >     > > this process works today and how we’d like it to work in the near
> > future.
> >     > >
> >     > > 1) Any chance this is already documented on the Wiki?
> >     > >
> >     > > 2) If not, any chance someone would be willing to do so (a flow
> > diagram
> >     > > would be particularly useful).
> >     > >
> >     > > > On Apr 5, 2018, at 3:37 AM, Marc-Aurèle Brothier <
> > marco@xxxxxxxxxxx>
> >     > > wrote:
> >     > > >
> >     > > > Hi all,
> >     > > >
> >     > > > Good point ilya but as stated by Sergey there's more thing to
> > consider
> >     > > > before being able to do a proper shutdown. I augmented my
> script
> > I gave
> >     > > you
> >     > > > originally and changed code in CS. What we're doing for our
> > environment
> >     > > is
> >     > > > as follow:
> >     > > >
> >     > > > 1. the MGMT looks for a change in the file /etc/lb-agent which
> > contains
> >     > > > keywords for HAproxy[2] (ready, maint) so that HA-proxy can
> > disable the
> >     > > > mgmt on the keyword "maint" and the mgmt server stops a couple
> of
> >     > > > threads[1] to stop processing async jobs in the queue
> >     > > > 2. Looks for the async jobs and wait until there is none to
> > ensure you
> >     > > can
> >     > > > send the reconnect commands (if jobs are running, a reconnect
> > will
> >     > result
> >     > > > in a failed job since the result will never reach the
> management
> >     > server -
> >     > > > the agent waits for the current job to be done before
> > reconnecting, and
> >     > > > discard the result... rooms for improvement here!)
> >     > > > 3. Issue a reconnectHost command to all the hosts connected to
> > the mgmt
> >     > > > server so that they reconnect to another one, otherwise the
> mgmt
> > must
> >     > be
> >     > > up
> >     > > > since it is used to forward commands to agents.
> >     > > > 4. when all agents are reconnected, we can shutdown the
> > management
> >     > server
> >     > > > and perform the maintenance.
> >     > > >
> >     > > > One issue remains for me, during the reconnect, the commands
> > that are
> >     > > > processed at the same time should be kept in a queue until the
> > agents
> >     > > have
> >     > > > finished any current jobs and have reconnected. Today the
> little
> > time
> >     > > > window during which the reconnect happens can lead to failed
> > jobs due
> >     > to
> >     > > > the agent not being connected at the right moment.
> >     > > >
> >     > > > I could push a PR for the change to stop some processing
> threads
> > based
> >     > on
> >     > > > the content of a file. It's possible also to cancel the drain
> of
> > the
> >     > > > management by simply changing the content of the file back to
> > "ready"
> >     > > > again, instead of "maint" [2].
> >     > > >
> >     > > > [1] AsyncJobMgr-Heartbeat, CapacityChecker, StatsCollector
> >     > > > [2] HA proxy documentation on agent checker:
> > https://cbonte.github.io/
> >     > > > haproxy-dconv/1.6/configuration.html#5.2-agent-check
> >     > > >
> >     > > > Regarding your issue on the port blocking, I think it's fair to
> >     > consider
> >     > > > that if you want to shutdown your server at some point, you
> have
> > to
> >     > stop
> >     > > > serving (some) requests. Here the only way it's to stop serving
> >     > > everything.
> >     > > > If the API had a REST design, we could reject any
> POST/PUT/DELETE
> >     > > > operations and allow GET ones. I don't know how hard it would
> be
> > today
> >     > to
> >     > > > only allow listBaseCmd operations to be more friendly with the
> > users.
> >     > > >
> >     > > > Marco
> >     > > >
> >     > > >
> >     > > > On Thu, Apr 5, 2018 at 2:22 AM, Sergey Levitskiy <
> > serg38l@xxxxxxxxxxx>
> >     > > > wrote:
> >     > > >
> >     > > >> Now without spellchecking :)
> >     > > >>
> >     > > >> This is not simple e.g. for VMware. Each management server
> also
> > acts
> >     > as
> >     > > an
> >     > > >> agent proxy so tasks against a particular ESX host will be
> > always
> >     > > >> forwarded. That right answer will be to support a native
> > “maintenance
> >     > > mode”
> >     > > >> for management server. When entered to such mode the
> management
> > server
> >     > > >> should release all agents including SSVM, block/redirect API
> > calls and
> >     > > >> login request and finish all async job it originated.
> >     > > >>
> >     > > >>
> >     > > >>
> >     > > >> On Apr 4, 2018, at 5:15 PM, Sergey Levitskiy <
> > serg38l@xxxxxxxxxxx
> >     > > <mailto:
> >     > > >> serg38l@xxxxxxxxxxx>> wrote:
> >     > > >>
> >     > > >> This is not simple e.g. for VMware. Each management server
> also
> > acts
> >     > as
> >     > > an
> >     > > >> agent proxy so tasks against a particular ESX host will be
> > always
> >     > > >> forwarded. That right answer will be to a native support for
> >     > > “maintenance
> >     > > >> mode” for management server. When entered to such mode the
> > management
> >     > > >> server should release all agents including save,
> block/redirect
> > API
> >     > > calls
> >     > > >> and login request and finish all a sync job it originated.
> >     > > >>
> >     > > >> Sent from my iPhone
> >     > > >>
> >     > > >> On Apr 4, 2018, at 3:31 PM, Rafael Weingärtner <
> >     > > >> rafaelweingartner@xxxxxxxxx<mailto:rafaelweingartner@
> gmail.com
> > >>
> >     > wrote:
> >     > > >>
> >     > > >> Ilya, still regarding the management server that is being shut
> > down
> >     > > issue;
> >     > > >> if other MSs/or maybe system VMs (I am not sure to know if
> they
> > are
> >     > > able to
> >     > > >> do such tasks) can direct/redirect/send new jobs to this
> > management
> >     > > server
> >     > > >> (the one being shut down), the process might never end because
> > new
> >     > tasks
> >     > > >> are always being created for the management server that we
> want
> > to
> >     > shut
> >     > > >> down. Is this scenario possible?
> >     > > >>
> >     > > >> That is why I mentioned blocking the port 8250 for the
> >     > > “graceful-shutdown”.
> >     > > >>
> >     > > >> If this scenario is not possible, then everything s fine.
> >     > > >>
> >     > > >>
> >     > > >> On Wed, Apr 4, 2018 at 7:14 PM, ilya musayev <
> >     > > ilya.mailing.lists@xxxxxxxxx
> >     > > >> <mailto:ilya.mailing.lists@xxxxxxxxx>>
> >     > > >> wrote:
> >     > > >>
> >     > > >> I'm thinking of using a configuration from
> >     > > "job.cancel.threshold.minutes" -
> >     > > >> it will be the longest
> >     > > >>
> >     > > >>    "category": "Advanced",
> >     > > >>
> >     > > >>    "description": "Time (in minutes) for async-jobs to be
> > forcely
> >     > > >> cancelled if it has been in process for long",
> >     > > >>
> >     > > >>    "name": "job.cancel.threshold.minutes",
> >     > > >>
> >     > > >>    "value": "60"
> >     > > >>
> >     > > >>
> >     > > >>
> >     > > >>
> >     > > >> On Wed, Apr 4, 2018 at 1:36 PM, Rafael Weingärtner <
> >     > > >> rafaelweingartner@xxxxxxxxx<mailto:rafaelweingartner@
> gmail.com
> > >>
> >     > wrote:
> >     > > >>
> >     > > >> Big +1 for this feature; I only have a few doubts.
> >     > > >>
> >     > > >> * Regarding the tasks/jobs that management servers (MSs)
> > execute; are
> >     > > >> these
> >     > > >> tasks originate from requests that come to the MS, or is it
> > possible
> >     > > that
> >     > > >> requests received by one management server to be executed by
> > other? I
> >     > > >> mean,
> >     > > >> if I execute a request against MS1, will this request always
> be
> >     > > >> executed/threated by MS1, or is it possible that this request
> is
> >     > > executed
> >     > > >> by another MS (e.g. MS2)?
> >     > > >>
> >     > > >> * I would suggest that after we block traffic coming from
> >     > > >> 8080/8443/8250(we
> >     > > >> will need to block this as well right?), we can log the
> > execution of
> >     > > >> tasks.
> >     > > >> I mean, something saying, there are XXX tasks (enumerate
> tasks)
> > still
> >     > > >> being
> >     > > >> executed, we will wait for them to finish before shutting
> down.
> >     > > >>
> >     > > >> * The timeout (60 minutes suggested) could be global settings
> > that we
> >     > > can
> >     > > >> load before executing the graceful-shutdown.
> >     > > >>
> >     > > >> On Wed, Apr 4, 2018 at 5:15 PM, ilya musayev <
> >     > > >> ilya.mailing.lists@xxxxxxxxx<mailto:ilya.mailing.lists@
> > gmail.com>
> >     > > >>
> >     > > >> wrote:
> >     > > >>
> >     > > >> Use case:
> >     > > >> In any environment - time to time - administrator needs to
> > perform a
> >     > > >> maintenance. Current stop sequence of cloudstack management
> > server
> >     > will
> >     > > >> ignore the fact that there may be long running async jobs -
> and
> >     > > >> terminate
> >     > > >> the process. This in turn can create a poor user experience
> and
> >     > > >> occasional
> >     > > >> inconsistency  in cloudstack db.
> >     > > >>
> >     > > >> This is especially painful in large environments where the
> user
> > has
> >     > > >> thousands of nodes and there is a continuous patching that
> > happens
> >     > > >> around
> >     > > >> the clock - that requires migration of workload from one node
> to
> >     > > >> another.
> >     > > >>
> >     > > >> With that said - i've created a script that monitors the async
> > job
> >     > > >> queue
> >     > > >> for given MS and waits for it complete all jobs. More details
> > are
> >     > > >> posted
> >     > > >> below.
> >     > > >>
> >     > > >> I'd like to introduce "graceful-shutdown" into the
> > systemctl/service
> >     > of
> >     > > >> cloudstack-management service.
> >     > > >>
> >     > > >> The details of how it will work is below:
> >     > > >>
> >     > > >> Workflow for graceful shutdown:
> >     > > >> Using iptables/firewalld - block any connection attempts on
> > 8080/8443
> >     > > >> (we
> >     > > >> can identify the ports dynamically)
> >     > > >> Identify the MSID for the node, using the proper msid - query
> >     > > >> async_job
> >     > > >> table for
> >     > > >> 1) any jobs that are still running (or job_status=“0”)
> >     > > >> 2) job_dispatcher not like “pseudoJobDispatcher"
> >     > > >> 3) job_init_msid=$my_ms_id
> >     > > >>
> >     > > >> Monitor this async_job table for 60 minutes - until all async
> > jobs for
> >     > > >> MSID
> >     > > >> are done, then proceed with shutdown
> >     > > >>  If failed for any reason or terminated, catch the exit via
> trap
> >     > > >> command
> >     > > >> and unblock the 8080/8443
> >     > > >>
> >     > > >> Comments are welcome
> >     > > >>
> >     > > >> Regards,
> >     > > >> ilya
> >     > > >>
> >     > > >>
> >     > > >>
> >     > > >>
> >     > > >> --
> >     > > >> Rafael Weingärtner
> >     > > >>
> >     > > >>
> >     > > >>
> >     > > >>
> >     > > >>
> >     > > >> --
> >     > > >> Rafael Weingärtner
> >     > > >>
> >     > >
> >     >
> >     >
> >     >
> >     > --
> >     >
> >     > Andrija Panić
> >     >
> >
> >
> >
>
>
> --
> Rafael Weingärtner
>