Jeromy I already experienced a similar problem. A host had a problem of connectivity with a storage, ACS did not know (I think it does check this right now) and it was kind of odd. Because some VMs could not be started, because of their last host id was the host with the connectivity problem.
I believe it would be a great thing to improve these checks and to treat this type of problem in ACS. My two cents: Change the name of what you are proposing. It is not high availability per se; it is more like a storage health check/heartbeat thing. It may give people very high expectations (I am not trying to undermine the work that is required). For instance, the first time I read your email, I thought about real HA, with some sort of block replication and redundancy in the storage system. That means, I was expecting if a whole storage goes down, ACS would have to have a way of solving this problem without losing data and without interrupting VMs. What you want to do is a little different, if a host has some connectivity problem, we simply add it to the avoid list, and then we migrate/move the VMs that were running there to a healthy host. About the idea of being for KVM only; even though we might not be able to run scripts on some hypervisors (Dom0 not being exposed), we could use their API and check if the storage is responding. For instance, in XenServr you could call “xe sr-scan uuid=<storage uuid>” (or maybe other command that uses storage), if there is any communication problem, it will report an error. The same type of feature will probably be available on all other hypervisors. On Tue, Mar 14, 2017 at 1:46 PM, Jeromy Grimmett <jer...@cloudbrix.com> wrote: > If all networking is lost, then obviously there is a bigger problem > there. The monitor is designed to do a true read/write test to the storage > and report back a pass/fail. Through this discussion, there was a ping > suggestion, which I think we will include. > > The way this came about was that one of our hosts had a problem with a > single primary storage, but all other hosts were 100% good across that > storage and all others. That troubled host was having problems with just > the single storage device, but according to CloudStack, all was well and > everything was good. The way I am looking at this, is that what we are > attempting to do is a much better and far more accurate test of storage > availability compared to how Cloudstack currently does it based our > experience. > > Make more sense? > > j > > Jeromy Grimmett > P: 603.766.3625 > jer...@cloudbrix.com > www.cloudbrix.com > > > -----Original Message----- > From: Simon Weller [mailto:swel...@ena.com] > Sent: Tuesday, March 14, 2017 10:01 AM > To: dev@cloudstack.apache.org > Cc: Alex Bonilla <a...@cloudbrix.com> > Subject: Re: [Proposal] - StorageHA > > So a few questions come to mind here. > > > So if all networking is lost, how are you going to being able reliably > fence the VM on the hosts? > > Are you assuming you still have out of band IPMI connectivity? > > If you're running bonded interfaces to different switches, what scenario > would occur where the host loses network connectivity? > > > - Si > > ________________________________ > From: Tutkowski, Mike <mike.tutkow...@netapp.com> > Sent: Tuesday, March 14, 2017 8:25 AM > To: dev@cloudstack.apache.org > Cc: Alex Bonilla > Subject: Re: [Proposal] - StorageHA > > Thanks for your clarification. I see now. You were referring to a > networking problem where one host could not see the storage (but the > storage was still up and running). > > On 3/13/17, 10:31 PM, "Jeromy Grimmett" <jer...@cloudbrix.com> wrote: > > I apologize for the delay on the response, let me clarify the points > requested: > > Mike asked: > > "What I was curious about is if you plan to exclusively build your > feature as a set of scripts and/or if you plan to update the CloudStack > code base, as well." > > JG: My idea was to do this separately as a plugin, then add it to the > code base down the road. > > "Also, if a primary storage actually goes offline, I'm not clear on > how starting an impacted VM on a different compute host would help. Could > you clarify this for me?" > > JG: The VM would be started on another host that still has access to > the storage. Individually a host can have problems and lose its > connectivity to a primary storage device. The solution we are working on > would help to get the VM back and up running much faster than waiting for > Cloudstack to make a decision to restart the VM on a different host. > > Paul asked: > > " 1. We can't/don't run scripts on vSphere hosts (not sure about > Hyper-V)" > > JG: I should have been more clear, this is for KVM hosts. > > "2. I know of one failure scenario (which happened) where MTU issues > in intermediate switches meant that small amounts of data could pass, but > anything that was passed as jumbo frames then failed. So it would be > important to exercise that." > > JG: I have faced this Jumbo Frame issue as well, perhaps we need to > have an option that would indicate Jumbo Frames are being used to access > that storage and the test result would reflect a failure to access using > Jumbo Frames. > > "3. You need to be very sure of failures before shutting hosts down. > Also a host is likely to be connected to multiple storage pools, so you > wouldn't want to shut down a host due to one pool becoming unavailable." > > JG: The script wouldn't shut down any hosts at all. Just force stop > the affected VMs on that specific host and then start them on a host that > is not having the issue with storage. > > "4. Environments can have hundreds of storage pools, so watch out for > spamming the logs with updates." > > JG: The polling/testing time increments are configurable, so I am > hoping that can help with that. The results are pretty small and should be > relatively negligible. > > "5. The primary storage pools have a 'state' which should get updated > and used by the deployment planners" > > JG: I have copied Alex on this email to make sure he sees this > suggestion. We will figure out how to incorporate that 'state' field. > > "6. Secondary storage pools don't have a 'state' - but it would be > great if that were added in the DB and reflected in the UI." > > JG: For now, I think this might be a feature request that maybe we > should submit through the normal Cloudstack request process. Otherwise, we > can definitely include that into our work when we start to add it into the > code base. > > To take this a step further, we are also working on a KVM host load > balancer that will be used as a factor when moving the VMs. We have a > number of little projects we are working on. > > Thank you all for reviewing the information. All suggestions are > welcome. > > Jeromy Grimmett > P: 603.766.3625 > jer...@cloudbrix.com > www.cloudbrix.com<http://www.cloudbrix.com> > > > -----Original Message----- > From: Paul Angus [mailto:paul.an...@shapeblue.com] > Sent: Saturday, March 11, 2017 2:43 AM > To: dev@cloudstack.apache.org > Subject: RE: [Proposal] - StorageHA > > Hi Jeromy, > > I love the idea, I'm not really a developer, so those guys will look > at things a different way, but... > > These would be by my initial comments: > > > 1. We can't/don't run scripts on vSphere hosts (not sure about > Hyper-V) > 2. I know of one failure scenario (which happened) where MTU issues > in intermediate switches meant that small amounts of data could pass, but > anything that was passed as jumbo frames then failed. So it would be > important to exercise that. > 3. You need to be very sure of failures before shutting hosts > down. Also a host is likely to be connected to multiple storage pools, so > you wouldn't want to shut down a host due to one pool becoming unavailable. > 4. Environments can have hundreds of storage pools, so watch out > for spamming the logs with updates. > 5. The primary storage pools have a 'state' which should get > updated and used by the deployment planners > 6. Secondary storage pools don't have a 'state' - but it would be > great if that were added in the DB and reflected in the UI. > > > > Kind regards, > > Paul Angus > > > paul.an...@shapeblue.com > www.shapeblue.com<http://www.shapeblue.com> > 53 Chandos Place, Covent Garden, London WC2N 4HSUK @shapeblue > > > > From: Jeromy Grimmett [mailto:jer...@cloudbrix.com] > Sent: 10 March 2017 15:28 > To: dev@cloudstack.apache.org > Subject: [Proposal] - StorageHA > > Hello, > > I am new to the mailing list, and we are glad to be a part of the > CloudStack community. We are looking to develop plugins and modules that > will help grow and expand the adoption and use of CloudStack. So as part > of my introductory email, I'd like to introduce a little project we have > been working on; a StorageHA Monitor. The Monitor would allow CloudStack > and the hosts to test, communicate and resolve VM availability issues when > storage (primary and/or secondary) availability becomes apparent. This is > a small write up about how it would work: > > Consists of two scripts/programs: > > The host script runs on the host servers and checks to see if the > primary and secondary storage is available by doing a read/write test then > reports to the master script that runs on the Cloudstack server. The host > script will test a read and a write to the storage every 5 seconds > (configurable), and if it fails 3 times (configurable) then it will be > recorded by the master script. > > The master script will monitor the results of the host script. If the > test is good, nothing happens and the results are logged and so that we can > track the history of the test results. If the test reports back as failed, > then it will perform the following actions: > > > * Secondary Storage - It will simply generate and send an alert > that the failure has occurred. > > > * Primary Storage - The script will perform the following tasks: > * Generate and send an alert that the failure has occurred. > * Force the VMs on that host to shutdown. > * Determine which host to move the VMs to. > * Start the VMs on the healthy host. > > We have already started working on some code, and the solution seems > to be testing well. Any thoughts/ideas/input are(is) welcome. Should > there are a solution out there already, then please forgive our ignorance, > and point us in the right direction. We look forward to further > collaboration with you all. > > Regards, > j > > Jeromy Grimmett > [cb-sig-logo2] > 155 Fleet Street > Portsmouth, NH 03801 > Direct: 603.766.3625 > Office: 603.766.4908 > Fax: 603.766.4729 > jer...@cloudbrix.com<mailto:jer...@cloudbrix.com> > www.cloudbrix.com<http://www.cloudbrix.com/> > > > > -- Rafael Weingärtner
