diff options
Diffstat (limited to 'docs/requirements')
| -rw-r--r-- | docs/requirements/VNF_high_availability_across_VIM.rst | 160 | ||||
| -rw-r--r-- | docs/requirements/multisite-identity-service-management.rst | 376 | ||||
| -rw-r--r-- | docs/requirements/multisite-vnf-gr-requirement.rst | 241 |
3 files changed, 777 insertions, 0 deletions
diff --git a/docs/requirements/VNF_high_availability_across_VIM.rst b/docs/requirements/VNF_high_availability_across_VIM.rst new file mode 100644 index 0000000..1a7d41b --- /dev/null +++ b/docs/requirements/VNF_high_availability_across_VIM.rst @@ -0,0 +1,160 @@ +This work is licensed under a Creative Commons Attribution 3.0 Unported License. +http://creativecommons.org/licenses/by/3.0/legalcode + + +======================================= +VNF high availability across VIM +======================================= + +Problem description +=================== + +Abstract +------------ + +a VNF (telecom application) should, be able to realize high availability +deloyment across OpenStack instances. + +Description +------------ +VNF (Telecom application running over cloud) may (already) be designed as +Active-Standby/Active-Active/N-Way to achieve high availability, + +With a telecoms focus, this generally refers both to availability of service +(i.e. the ability to make new calls), but also maintenance of ongoing control +plane state and active media processing(i.e. “keeping up” existing calls). + +Traditionally telecoms systems are designed to maintain state and calls across +pretty much the full range of single-point failures. As listed this includes +power supply, hard drive, physical server or network switch, but also covers +software failure, and maintenance operations such as software upgrade. + +To provide this support, typically requires state replication between +application instances (directly or via replicated database services, or via +private designed message format). It may also require special case handling of +media endpoints, to allow transfer of median short time scales (<1s) without +requiring end-to-end resignalling (e.g.RTP redirection via IP / MAC address +transfers c.f VRRP). + +With a migration to NFV, a commonly expressed desire by carriers is to provide +the same resilience to any single point(s) of failure in the cloud +infrastructure. + +This could be done by making each cloud instance fully HA (a non-trivial task to +do right and to prove it has been done right) , but the preferred approach +appears to be to accept the currently limited availability of a given cloud +instance (no desire to radically rework this for telecoms), and instead to +provide solution availability by spreading function across multiple cloud +instances (i.e. the same approach used today todeal with hardware and software +failures). + +A further advantage of this approach, is it provides a good basis for seamless +upgrade of infrastructure software revision, where you can spin up an additional +up-level cloud, gradually transfer over resources / app instances from one of +your other clouds, before finally turning down the old cloud instance when no +longer required. + +If fast media / control failure over is still required (which many/most carriers +still seem to believe it is) there are some interesting/hard requirements on the +networking between cloud instances. To help with this, many people appear +willing to provide multiple “independent” cloud instances in a single geographic +site, with special networking between clouds in that physical site. +"independent" in quotes is because some coordination between cloud instances is +obviously required, but this has to be implemented in a fashion which reduces +the potential for correlated failure to very low levels (at least as low as the +required overall application availability). + +Analysis of requirements to OpenStack +=========================== +The VNF often has different networking plane for different purpose: + +external network plane: using for communication with other VNF +components inter-communication plane: one VNF often consisted of several +components, this plane is designed for components inter-communication with each +other +backup plance: this plane is used for the heart beat or state replication +between the component's active/standy or active/active or N-way cluster. +management plane: this plane is mainly for the management purpose + +Generally these planes are seperated with each other. And for legacy telecom +application, each internal plane will have its fixed or flexsible IP addressing +plane. + +to make the VNF can work with HA mode across different OpenStack instances in +one site (but not limited to), need to support at lease the backup plane across +different OpenStack instances: + +1) Overlay L2 networking or shared L2 provider networks as the backup plance for +heartbeat or state replication. Overlay L2 network is preferred, the reason is: +a. Support legacy compatibility: Some telecom app with built-in internal L2 +network, for easy to move these app to VNF, it would be better to provide L2 +network b. Support IP overlapping: multiple VNFs may have overlaping IP address +for cross OpenStack instance networking +Therefore, over L2 networking across Neutron feature is required in OpenStack. + +2) L3 networking cross OpenStack instance for heartbeat or state replication. +For L3 networking, we can leverage the floating IP provided in current Neutron, +so no new feature requirement to OpenStack. + +3) The IP address used for VNF to connect with other VNFs should be able to be +floating cross OpenStack instance. For example, if the master failed, the IP +address should be used in the standby which is running in another OpenStack +instance. There are some method like VRRP/GARP etc can help the movement of the +external IP, so no new feature will be added to OpenStack. + + +Prototype +----------- + None. + +Proposed solution +----------- + + requirements perspective It's up to application descision to use L2 or L3 +networking across Neutron. + + For Neutron, a L2 network is consisted of lots of ports. To make the cross +Neutron L2 networking is workable, we need some fake remote ports in local +Neutron to represent VMs in remote site ( remote OpenStack ). + + the fake remote port will reside on some VTEP ( for VxLAN ), the tunneling +IP address of the VTEP should be the attribute of the fake remote port, so that +the local port can forward packet to correct tunneling endpoint. + + the idea is to add one more ML2 mechnism driver to capture the fake remote +port CRUD( creation, retievement, update, delete) + + when a fake remote port is added/update/deleted, then the ML2 mechanism +driver for these fake ports will activate L2 population, so that the VTEP +tunneling endpoint information could be understood by other local ports. + + it's also required to be able to query the port's VTEP tunneling endpoint +information through Neutron API, in order to use these information to create +fake remote port in another Neutron. + + In the past, the port's VTEP ip address is the host IP where the VM resides. +But the this BP https://review.openstack.org/#/c/215409/ will make the port free +of binding to host IP as the tunneling endpoint, you can even specify L2GW ip +address as the tunneling endpoint. + + Therefore a new BP will be registered to processing the fake remote port, in +order make cross Neutron L2 networking is feasible. RFE is registered first: +https://bugs.launchpad.net/neutron/+bug/1484005 + + +Gaps +==== + 1) fake remote port for cross Neutron L2 networking + + +**NAME-THE-MODULE issues:** + +* Neutron + +Affected By +----------- + OPNFV multisite cloud. + +References +========== + diff --git a/docs/requirements/multisite-identity-service-management.rst b/docs/requirements/multisite-identity-service-management.rst new file mode 100644 index 0000000..b411c28 --- /dev/null +++ b/docs/requirements/multisite-identity-service-management.rst @@ -0,0 +1,376 @@ +This work is licensed under a Creative Commons Attribution 3.0 Unported +License. +http://creativecommons.org/licenses/by/3.0/legalcode + + +======================================= + Multisite identity service management +======================================= + +Glossary +======== + +There are 3 types of token supported by OpenStack KeyStone + **UUID** + + **PKI/PKIZ** + + **FERNET** + +Please refer to reference section for these token formats, benchmark and +comparation. + + +Problem description +=================== + +Abstract +------------ + +a user should, using a single authentication point be able to manage virtual +resources spread over multiple OpenStack regions. + +Description +------------ + +- User/Group Management: e.g. use of LDAP, should OPNFV be agnostic to this? + Reusing the LDAP infrastructure that is mature and has features lacking in +Keystone (e.g.password aging and policies). KeyStone can use external system to +do the user authentication, and user/group management could be the job of +external system, so that KeyStone can reuse/co-work with enterprise identity +management. KeyStone's main role in OpenStack is to provide +service(Nova,Cinder...) aware token, and do the authorization. You can refer to +this post https://blog-nkinder.rhcloud.com/?p=130.Therefore, LDAP itself should +be a topic out of our scope. + +- Role assignment: In case of federation(and perhaps other solutions) it is not + feasible/scalable to do role assignment to users. Role assignment to groups + is better. Role assignment will be done usually based on group. KeyStone + supports this. + +- Amount of inter region traffic: should be kept as little as possible, + consider CERNs Ceilometer issue as described in +http://openstack-in-production.blogspot.se/2014/03/cern-cloud-architecture-update-for.html + +Requirement analysis +=========================== + +- A user is provided with a single authentication URL to the Identity + (Keystone) service. Using that URL, the user authenticates with Keystone by +requesting a token typically using username/password credentials. The keystone +server validates the credentials, possibly with an external LDAP/AD server and +returns a token to the user. With token type UUID/Fernet, the user request the +service catalog. With PKI tokens the service catalog is included in the token. +The user sends a request to a service in a selected region including the token. +Now the service in the region, say Nova needs to validate the token. Nova uses +its configured keystone endpoint and service credentials to request token +validation from Keystone. The Keystone token validation should preferably be +done in the same region as Nova itself. Now Keystone has to validate the token +that also (always?) includes a project ID in order to make sure the user is +authorized to use Nova. The project ID is stored in the assignment backend - +tables in the Keystone SQL database. For this project ID validation the +assignment backend database needs to have the same content as the keystone who +issued the token. + +- So either 1) services in all regions are configured with a central keystone + endpoint through which all token validations will happen. or 2) the Keystone +assignment backend database is replicated and thus available to Keystone +instances locally in each region. + + Alt 2) is obviously the only scalable solution that produce no inter region +traffic for normal service usage. Only when data in the assignment backend is +changed, replication traffic will be sent between regions. Assignment data +includes domains, projects, roles and role assignments. + +Keystone deployment: + + - Centralized: a single Keystone service installed in some location, either + in a "master" region or totally external as a service to OpenStack + regions. + - Distributed: a Keystone service is deployed in each region + +Token types: + + - UUID: tokens are persistently stored and creates a lot of database + traffic, the persistence of token is for the revoke purpose. UUID tokens +are online validated by Keystone, each API calling to service will ask token +validation from KeyStone. Keystone can become a bottleneck in a large system +due to this. UUID token type is not suitable for use in multi region clouds at +all, no matter the solution used for the Keystone database replication (or +not). UUID tokens have a fixed size. + + - PKI: tokens are non persistent cryptographic based tokens and offline + validated (not by the Keystone service) by Keystone middleware +which is part of other services such as Nova. Since PKI tokens include endpoint +for all services in all regions, the token size can become big.There are +several ways to reduce the token size, no catalog policy, endpoint filter to +make a project binding with limited endpoints, and compressed PKI token - PKIZ, +but the size of token is still predictable, make it difficult to manage. If no +catalog applied, that means the user can access all regions, in some scenario, +it's not allowed to do like this. + + - Fernet: tokens are non persistent cryptographic based tokens and online + validated by the Keystone service. Fernet tokens are more lightweigth +then PKI tokens and have a fixed size. + + PKI (offline validated) are needed with a centralized Keystone to avoid +inter region traffic. PKI tokens do produce Keystone traffic for revocation +lists. + + Fernet tokens requires Keystone deployed in a distributed manner, again to +avoid inter region traffic. + + Cryptographic tokens brings new (compared to UUID tokens) issues/use-cases +like key rotation, certificate revocation. Key management is out of scope of +this use case. + +Database deployment: + + Database replication: + -Master/slave asynchronous: supported by the database server itself +(mysql/mariadb etc), works over WAN, it's more scalable + -Multi master synchronous: Galera(others like percona), not so scalable, +for multi-master writing, and need more parameter tunning for WAN latency. + -Symmetrical/asymmetrical: data replicated to all regions or a subset, +in the latter case it means some regions needs to access Keystone in another +region. + + Database server sharing: + In an OpenStack controller normally many databases from different +services are provided from the same database server instance. For HA reasons, +the database server is usually synchronously replicated to a few other nodes +(controllers) to form a cluster. Note that _all_ database are replicated in +this case, for example when Galera sync repl is used. + + Only the Keystone database can be replicated to other sites. Replicating +databases for other services will cause those services to get of out sync and +malfunction. + + Since only the Keystone database is to be sync replicated to another +region/site, it's better to deploy Keystone database into its own +database server with extra networking requirement, cluster or replication +configuration. How to support this by installer is out of scope. + + The database server can be shared when async master/slave repl is used, if +global transaction identifiers GTID is enabled. + + +Candidate solution analysis +------------------------------------ + +- KeyStone service (Distributed) with Fernet token + + Fernet token is a very new format, and just introduced recently,the biggest +gain for this token format is :1) lightweight, size is small to be carried in +the API request, not like PKI token( as the sites increased, the endpoint-list +will grows and the token size is too long to carry in the API request) 2) no +token persistence, this also make the DB not changed too much and with light +weight data size (just project. User, domain, endpoint etc). The drawback for +the Fernet token is that token has to be validated by KeyStone for each API +request. + + This makes that the DB of KeyStone can work as a cluster in multisite (for +example, using MySQL galera cluster). That means install KeyStone API server in +each site, but share the same the backend DB cluster.Because the DB cluster +will synchronize data in real time to multisite, all KeyStone server can see +the same data. + + Because each site with KeyStone installed, and all data kept same, +therefore all token validation could be done locally in the same site. + + The challenge for this solution is how many sites the DB cluster can +support. Question is aksed to MySQL galera developers, their answer is that no +number/distance/network latency limitation in the code. But in the practice, +they have seen a case to use MySQL cluster in 5 data centers, each data centers +with 3 nodes. + + This solution will be very good for limited sites which the DB cluster can +cover very well. + +- KeyStone service(Distributed) with Fernet token + Async replication ( + multi-cluster mode). + + We may have several KeyStone cluster with Fernet token, for example, +cluster1 ( site1, site2, … site 10 ), cluster 2 ( site11, site 12,..,site 20). +Then do the DB async replication among different cluster asynchronously. + + A prototype of this has been down on this. In some blogs they call it +"hybridreplication". Architecturally you have a master region where you do +keystone writes. The other regions is read-only. +http://severalnines.com/blog/deploy-asynchronous-slave-galera-mysql-easy-way +http://severalnines.com/blog/replicate-mysql-server-galera-cluster + + Only one DB cluster (the master DB cluster) is allowed to write(but still +multisite, not all sites), other clusters waiting for replication. Inside the +master cluster, "write" is allowed in multiple region for the distributed lock +in the DB. But please notice the challenge of key distribution and rotation for +Fernet token, you can refer to these two blogs: http://lbragstad.com/?p=133, +http://lbragstad.com/?p=156 + +- KeyStone service(Distributed) with Fernet token + Async replication ( + star-mode). + + one master KeyStone cluster with Fernet token in two sites (for site level +high availability purpose), other sites will be installed with at least 2 slave +nodes where the node is configured with DB async replication from the master +cluster members, and one slave’s mater node in site1, another slave’s master +node in site 2. + + Only the master cluster nodes are allowed to write, other slave nodes +waiting for replication from the master cluster ( very little delay) member. +But the chanllenge of key distribution and rotation for Fernet token should be +settled, you can refer to these two blogs: http://lbragstad.com/?p=133, +http://lbragstad.com/?p=156 + + Pros. + Why cluster in the master sites? There are lots of master nodes in the +cluster, in order to provide more slaves could be done with async. replication +in parallel. Why two sites for the master cluster? to provide higher +reliability (site level) for writing request. + Why using multi-slaves in other sites. Slave has no knowledge of other +slaves, so easy to manage multi-slaves in one site than a cluster, and +multi-slaves work independently but provide multi-instance redundancy(like a +cluster, but independent). + + Cons. The distribution/rotation of key management. + +- KeyStone service(Distributed) with PKI token + + The PKI token has one great advantage is that the token validation can be +done locally, without sending token validation request toKeyStone server. The +drawback of PKI token is 1) the endpoint list size in the token. If a project +will be only spread in very limited site number(region number), then we can use +the endpoint filter to reduce the token size, make it workable even a lot of +sites in the cloud. 2) KeyStone middleware(the old KeyStone client, which +co-locate in Nova/xxx-API) will have to send the request to the KeyStone server +frequently for the revoke-list, in order to reject some malicious API request, +for example, a user has be deactivated, but use an old token to access +OpenStack service. + + For this solution, except above issues, we need also to provide KeyStone +Active-Active mode across site to reduce the impact of site failure. And the +revoke-list request is very frequently asked, so the performance of the +KeyStone server needs also to be taken care. + + Site level keystone load balance is required to provide site level +redundancy. Otherwise the KeyStone middleware will not switch request to the +health KeyStone server in time. + + This solution can be used for some scenario, especially a project only +spread in limited sites ( regions ). + + And also the cert distribution/revoke to each site / API server for token +validation is required. + +- KeyStone service(Distributed) with UUID token + + Because each token validation will be sent to KeyStone server,and the token +persistence also makes the DB size larger than Fernet token, not so good as the +fernet token to provide a distributed KeyStone service. UUID is a solution +better for small scale and inside one site. + + Cons: UUID tokens are persistently stored so will cause a lot of inter +region replication traffic, tokens will be persisted for authorization and +revoke purpose, the frequent changed database leads to a lot of inter region +replication traffic. + +- KeyStone service(Distributed) with Fernet token + KeyStone federation You + have to accept the drawback of KeyStone federation if you have a lot of +sites/regions. Please refer to KeyStone federation section + +- KeyStone federation + In this solution, we can install KeyStone service in each site and with +its own database. Because we have to make the KeyStone IdP and SP know each +other, therefore the configuration needs to be done accordingly, and setup the +role/domain/group mapping, create regarding region in the pair.As sites +increase, if each user is able to access all sites, then full-meshed +mapping/configuration has to be done. Whenever you add one more site, you have +to do n*(n-1) sites configuration/mapping. The complexity will be great enough +as the sites number increase. + + KeyStone Federation is mainly for different cloud admin to borrow/rent +resources, for example, A company and B company, A private cloud and B public +cloud, and both of them using OpenStack based cloud. Therefore a lot of mapping +and configuration has to be done to make it work. + +- KeyStone service (Centralized)with Fernet token + + cons: inter region traffic for token validation, token validation requests +from all other sites has to be sent to the centralized site. Too frequent inter +region traffic. + +- KeyStone service(Centralized) with PKI token + + cons: inter region traffic for tokenrevocation list management, the token +revocation list request from all other sites has to be sent to the centralized +site. Too frequent inter region traffic. + +- KeyStone service(Centralized) with UUID token + + cons: inter region traffic for token validation, the token validation +request from all other sites has to be sent to the centralized site. Too +frequent inter region traffic. + +Prototype +----------- + A prototype of the candidate solution "KeyStone service(Distributed) with +Fernet token + Async replication ( multi-cluster mode)" has been executed Hans +Feldt and Chaoyi Huang, please refer to https://github.com/hafe/dockers/ . And +one issue was found "Can't specify identity endpoint for token validation among +several keystone servers in keystonemiddleware", please refer to the Gaps +section. + +Gaps +==== + Can't specify identity endpoint for token validation among several keystone +servers in keystonemiddleware. + + +**NAME-THE-MODULE issues:** + +* keystonemiddleware + + * Can't specify identity endpoint for token validation among several keystone + * servers in keystonemiddleware: + * https://bugs.launchpad.net/keystone/+bug/1488347 + +Affected By +----------- + OPNFV multisite cloud. + +Conclusion +----------- + + As the prototype demonstrate the cluster level aysn. replication capability +and fernet token validation in local site is feasible. And the candidate +solution "KeyStone service(Distributed) with Fernet token + Async replication ( +star-mode)" is simplified solution of the prototyped one, it's much more easier +in deployment and maintenance, with better scalability. + + Therefore the candidate solution "KeyStone service(Distributed) with Fernet +token + Async replication ( star-mode)" for multsite OPNFV cloud is +recommended. + +References +========== + + There are 3 format token (UUID, PKI/PKIZ, Fernet) provided byKeyStone, this +blog give a very good description, benchmark and comparation: + http://dolphm.com/the-anatomy-of-openstack-keystone-token-formats/ + http://dolphm.com/benchmarking-openstack-keystone-token-formats/ + + To understand the benefit and shortage of PKI/PKIZ token, pleaserefer to : + https://www.mirantis.com/blog/understanding-openstack-authentication-keystone-pk + + To understand KeyStone federation and how to use it: + http://blog.rodrigods.com/playing-with-keystone-to-keystone-federation/ + + To integrate KeyStone with external enterprise ready authentication system + https://blog-nkinder.rhcloud.com/?p=130. + + Key repliocation used in KeyStone Fernet token + http://lbragstad.com/?p=133, + http://lbragstad.com/?p=156 + + KeyStone revoke + http://specs.openstack.org/openstack/keystone-specs/api/v3/identity-api-v3-os-revoke-ext.html diff --git a/docs/requirements/multisite-vnf-gr-requirement.rst b/docs/requirements/multisite-vnf-gr-requirement.rst new file mode 100644 index 0000000..7e67cd0 --- /dev/null +++ b/docs/requirements/multisite-vnf-gr-requirement.rst @@ -0,0 +1,241 @@ +This work is licensed under a Creative Commons Attribution 3.0 Unported License. +http://creativecommons.org/licenses/by/3.0/legalcode + + +========================================= + Multisite VNF Geo site disaster recovery +========================================= + +Glossary +======== + + +There are serveral concept required to be understood first + **Volume Snapshot** + + **Volume Backup** + + **Volume Replication** + + **VM Snapshot** + +Please refer to reference section for these concept and comparison. + + +Problem description +=================== + +Abstract +------------ + +a VNF (telecom application) should, be able to restore in another site for +catastrophic failures happened. + +Description +------------ +GR is to deal with more catastrophic failures (flood, earthquake, propagating +software fault), and that loss of calls, or even temporary loss of service, +is acceptable. It is also seems more common to accept/expect manual / +administrator intervene into drive the process, not least because you don’t +want to trigger the transfer by mistake. + +In terms of coordination/replication or backup/restore between geographic +sites, discussion often (but not always) seems to focus on limited application +level data/config replication, as opposed to replication backup/restore between +of cloud infrastructure between different sites. + +And finally, the lack of a requirement to do fast media transfer (without +resignalling) generally removes the need for special networking behavior, with +slower DNS-style redirection being acceptable. + +This use case is more concerns about cloud infrastructure level capability to +support VNF geo site redundancy + +Requirement and candidate solutions analysis +============================================ + +For VNF to be restored from the backup site for catastrophic failures, +the VNF's bootable volume and data volumes must be restorable. + +There are three ways of restorable boot and data volumes. Choosing the right +one largely depends on the underlying characteristics and requirements of a +VNF. + +1. Nova Quiesce + Cinder Consistency volume snapshot+ Cinder backup + 1).GR(Geo site disaster recovery )software get the volumes for each VM + in the VNF from Nova + 2).GR software call Nova quiesce API to quarantee quiecing VMs in desired + order + 3).GR software takes snapshots of these volumes in Cinder (NOTE: Because + storage often provides fast snapshot, so the duration between quiece and + unquiece is a short interval) + 4).GR software call Nova unquiece API to unquiece VMs of the VNF in reverse + order + 5).GR software create volumes from the snapshots just taken in Cinder + 6).GR software create backup (incremental) for these volumes to remote + backup storage ( swift or ceph, or.. ) in Cinder + 7).if this site failed, + 7.1)GR software restore these backup volumes in remote Cinder in the + backup site. + 7.2)GR software boot VMs from bootable volumes from the remote Cinder in + the backup site and attach the regarding data volumes. + +Pros: Quiesce / unquiesce api from Nova, make transactional snapshot +of a group of VMs is possible, for example, quiesce VM1, quiesce VM2, +quiesce VM3, snapshot VM1's volumes, snapshot VM2's volumes, snapshot +VM3's volumes, unquiesce VM3, unquiesce VM2, unquiesce VM1. For some +telecom application, the order is very important for a group of VMs +with strong relationship. + +Cons: Need Nova to expose the quiesce / unquiesce, fortunately it's alreay +there in Nova-compute, just to add API layer to expose the functionality. +NOTE: It's up to the DR policy and VNF character. Some VNF may afford short +unavailable for DR purpose, and some other may use the standby of the VNF +or member of the cluster to do disaster recovery replication to not interfere +the service provided by the VNF. For these VNFs which can't be quieced/unquiece +should use the option3 (VNF aware) to do the backup/replication. + +Requirement to OpenStack: Nova needs to expose quiesce / unquiesce api, +which is lack in Nova now. + +Example characteristics and requirements of a VNF: + - VNF requires full data consistency during backup/restore process - + entire data should be replicated. + - VNF's data changes infrequently, which results in less number of volume + snapshots during a given time interval (hour, day, etc.); + - VNF is not highly dynamic, e.g. the number of scaling (in/out) operations + is small. + - VNF is not geo-redundant, does not aware of available cloud replication + mechanisms, has no built-in logic for replication: doesn't pre-select the + minimum replication data required for restarting the VNF in a different + site. + (NOTE: The VNF who can perform such data cherry picking should consider + case 3) + +2. Nova Snapshot + Glance Image + Cinder Snapshot + Cinder Backup + - GR software create VM snapshot in Nova + - Nova quiece the VM internally + (NOTE: The upper level application or GR software should take care of + avoiding infra level outage induced VNF outage) + - Nova create image in Glance + - Nova create a snapshot of the VM, including volumes + - If the VM is volume backed VM, then create volume snapshot in Cinder + - No image uploaded to glance, but add the snapshot in the meta data of the + image in Glance + - GR software to get the snapshot information from the Glance + - GR software create volumes from these snapshots + - GR software create backup (incremental) for these volumes to backup + storage( swift or ceph, or.. ) in Cinder if this site failed, + - GR software restore these backup volumes to Cinder in the backup site. + - GR software boot vm from bootable volume from Cinder in the backup site + and attach the data volumes. + +Pros: 1) Automatically quiesce/unquiesce, and snapshot of volumes of one VM. + +Cons: 1) Impossible to form a transactional group of VMs backup. for example, + quiesce VM1, quiesce VM2, quiesce VM3, snapshot VM1, snapshot VM2, + snapshot VM3, unquiesce VM3, unquiesce VM2, unquiesce VM1. This is + quite important in telecom application in some scenario + 2) not leverage the Cinder consistency group. + 3) One more service Glance involved in the backup. Not only to manage the + increased snapshot in Cinder, but also need to manage the regarding + temporary image in Glance. + +Requirement to OpenStack: None. + +Example: It's suitable for single VM backup/restore, for example, for the small +scale configuration database virtual machine which is running in active/standby +model. There is very rare use case for application that only one VM need to be +taken snapshot for back up. + +3. Selective Replication of Persistent Data + - GR software creates datastore (Block/Cinder, Object/Swift, App Custom + storage) with replication enabled at the relevant scope, for use to + selectively backup/replicate desire data to GR backup site + - Cinder : Various work underway to provide async replication of cinder + volumes for disaster recovery use, including this presentation from + Vancouver http://www.slideshare.net/SeanCohen/dude-wheres-my-volume-open-stack-summit-vancouver-2015 + - Swift : Range of options of using native Swift replicas (at expense of + tighter coupling) to replication using backend plugins or volume + replication + - Custom : A wide range of OpenSource technologies including Cassandra + and Ceph, with fully application level solutions also possible + - GR software get the reference of storage in the remote site storage + - If primary site failed, + - GR software managing recovery in backup site gets references to + relevant storage and passes to new software instances + - Software attaches (or has attached) replicated storage, in the case of + volumes promoting to writable. + +Pros: 1) Replication will be done in the storage level automatically, no need + to create backup regularly, for example, daily. + 2) Application selection of limited amount of data to replicate reduces + risk of replicating failed state and generates less overhear. + 3) Type of replication and model (active/backup, active/active, etc) can + be tailored to application needs + +Cons: 1) Applications need to be designed with support in mind, including both + selection of data to be replicated and consideration of consistency + 2) "Standard" support in Openstack for Disaster Recovery currently + fairly limited, though active work in this area. + +Requirement to OpenStack: save the real ref to volume admin_metadata after it +has been managed by the driver https://review.openstack.org/#/c/182150/. + +Prototype +----------- + None. + +Proposed solution +----------- + + requirements perspective we could recommend all three options for different + sceanrio, that it is an operator choice. + Options 1 & 2 seem to be more about replicating/backing up any VNF, whereas + option 3 is about proving a service to a replication aware application. It + should be noted that HA requirement is not a priority here, HA for VNF + project will handle the specific HA requirement. It should also be noted + that it's up to specific application how to do HA (out of scope here). + For the 3rd option, the app should know which volume has replication + capability, and write regarding data to this volume, and guarantee + consistency by the app itself. Option 3 is preferrable in HA scenario. + + +Gaps +==== + 1) Nova to expose quiesce / unquiesce API: + https://blueprints.launchpad.net/nova/+spec/expose-quiesce-unquiesce-api + 2) Get the real ref to volume admin_metadata in Cinder: + https://review.openstack.org/#/c/182150/ + + +**NAME-THE-MODULE issues:** + +* Nova + +Affected By +----------- + OPNFV multisite cloud. + +References +========== + + Cinder snapshot ( no material/BP about snapshot itself availble from web ) + http://docs.openstack.org/cli-reference/content/cinderclient_commands.html + + + Cinder volume backup + https://blueprints.launchpad.net/cinder/+spec/volume-backups + + Cinder incremtal backup + https://blueprints.launchpad.net/cinder/+spec/incremental-backup + + Cinder volume replication + https://blueprints.launchpad.net/cinder/+spec/volume-replication + + Create VM snapshot with volume backed ( not found better matrial to explain + the volume backed VM snapshot, only code tells ) + https://bugs.launchpad.net/nova/+bug/1322195 + + Cinder consistency group + https://github.com/openstack/cinder-specs/blob/master/specs/juno/consistency-groups.rst |