is hyper-converged architecture really perfect?

The earliest is probably by Nutanix know "hyper-fusion" concept, originally thought this is the manufacturer's gimmick and speculation, who knows otherwise, hyper-integration of the two-year development in full swing. The 2015 was recognized as the ultra-fusion of the years, driven by software-defined storage SDS, hyper-converged architecture is leading a huge transformation into a software-defined data center SDDC future technology development trend.

The advantages of hyper-converged architecture and customer value this has not been questioned, the manufacturers and the media has been successful to help customers brainwashed, and customer exchanges do not mention hyper-fusion is considered outdated, here do not want to repeat. From a dialectical point of view, anything can not be perfect, hyper-integration is there are some problems or limitations? Hyper-Fusion has a suitable scenario, but certainly not universally applicable. Therefore, this article I would like to change a point of view hyper-fusion, find fault, comb the hyper-fusion, but also for everyone to choose a super-converged architecture solution provides a reference. Container so fire, do not talk as if it is not good, the last simple Hu Yu container Super Fusion.

1 What is hyper-fusion?
Hyper-Fusion (hyper-converged) first by who proposed this concept we do not go to research, but there is no strict standard definition, each manufacturer and agency have their own definition of the label, which also proves that hyper-integration is still in the rapid development and evolution, has not formed a unified standard specification.

What does "super" mean in hyper-fusion? The first thought was Superman super, the result almost make a joke. In fact, "super" corresponds to "Hyper" in English "hyper-converged", especially virtualization, corresponding to the virtual computing architecture, such as Esxi/kvm/xen/hyper-v. This concept was first derived from storage startups such as Nutanix to use the compute Storage Fusion architecture used by internet vendors such as Google/facebook for virtualized environments, providing enterprise customers with a computing storage fusion product or solution based on X86 hardware platform. It is not difficult to see that the most fundamental change in hyper-converged architecture is storage, from the original centralized shared storage (San/nas) to software-defined storage, especially distributed storage (including Object/block/file storage), such as Ndfs/vsan/scaleio/ssan. Therefore, based on this "super", the database all-in-one machine and big data integration machine can not be classified as hyper-fusion category, unless Rac/hadoop and other applications run on the virtualization. Also, in hyper-converged software-defined storage is typically distributed storage, ZFS is a SDS-based, but a ZFS-built compute-storage fusion system is not strictly a hyper-converged architecture.

What does it mean to see "Convergence" again? The concept of hyper-fusion now seems to be deified. To put it simply, fusion is the combination of two or more components into a single unit, which can be either hardware or software. As far as virtualization and private cloud are concerned, I personally divide the fusion into physical fusion and hyper-fusion, according to whether it is completely virtualized and centralized. Hyper-convergence is a subset of convergence, where compute and storage are deployed on the same node, which is equivalent to multiple components deployed in a single system, while providing compute and storage capabilities. In a physical fusion system, computing and storage can still be two independent components without direct interdependencies. For example, Ssan+ovirt scheme, a node of the Redhat/centos system Ssan and Ovirt physical fusion, sharing the physical resources of the host. The difference between hyper-fusion and physical fusion is that the focus is on virtualization computing, where compute and storage are tightly related, and storage is controlled by virtual machines rather than physical machine CVM (Controller VMs), and distributed storage resources are formed into a unified storage pool, which is then provided to hypervisor for creating application virtual machines. such as the Nutanix and Ssan+vshpere Hyper-integration scheme. The narrow definition here is the real hyper-convergence, Nutanix first proposed this architecture and applied for a patent. As defined here, Openstack+ceph is a physical fusion rather than a hyper-fusion. It is worth noting that, for performance reasons, hyper-converged architectures usually need to pass the host physical device (pass Through) to the control VM CVM.

Let's look at a few more representative hyper-fusion definitions:
NUTANIX: Hyper-converged architecture ("HCI") means not only resources and technologies such as compute, network, storage and server virtualization, but also backup software, snapshot technology, deduplication, online data compression and other elements in the same set of unit devices, And multiple sets of units can be aggregated through the network to achieve a modular seamless scale-out, the formation of a unified pool of resources. HCI is the ultimate technical way to achieve a "software-defined data center."

Gartner : HCI is a software-centric architecture that tightly integrates compute, storage, networking, and virtualization resources (and possibly other technologies) into a single vendor-provided hardware device.

IDC : Hyper-converged systems are an emerging, integrated system that integrates core storage, compute, and storage networking capabilities into a single software solution or device. This definition differs from the products that integrate infrastructure and platforms that are owned by a vendor or reseller at the factory, and that integrate the storage and network systems.

Summarize the above definition of hyper-convergence: hyper-converged architecture is a standard-based hardware platform that enables computing, storage, and network convergence through software definitions to realize the technical architecture of a software-defined data center that is virtualized as a center. There are several keywords in this: the Universal hardware platform, software definition, and virtualization, where software-defined distributed storage is the core. How to judge whether a system is hyper-converged? I boldly give a simple standard.
(1) full software definition . Independent of the hardware, the use of commercial common standard hardware platform (such as X86), fully using software to achieve computing, storage, network and other functions.
(2) full virtualization . The computing, storage, and network are managed and dispatched centrally by the virtualization engine, and the software-defined storage is managed by the virtual machine controller CVM.
(3) fully distributed . Scale-out distributed system, computing, storage, network on-demand dynamic expansion, the system does not exist any single point of failure, the use of distributed storage.

2 What is the relationship between hyper-fusion and SDS?
What is software-defined storage SDS? Software definition, from the beginning of Sdn to computing and storage, software definition has been considered a technological trend. Like hyper-convergence, SDS does not yet have a uniform standard definition, and Snia/idc/gartner/vmware has its own definition. As I understand it, software-defined storage is a traditional storage system that uses proprietary hardware, a common standard open hardware platform, and software for all storage functions, including data plane and control plane. In addition to implementing storage capabilities, SDS has a natural advantage in terms of scalability, availability, flexibility, simplified management, and lower total costs. The typical representative of SDS is Zfs/ndfs/vsan/scaleio/ssan/ceph, which is mostly Serversan, which is the most important product form of SDS. Wikibon defines Serversan: A pool of storage resources composed of multiple standalone server stores that provide storage services on an IP network in a clustered manner.

Typical features of Serversan include :

(1) Pure software definition, independent of hardware;
(2) completely distributed, there is no single point of failure;
(3) system autonomy, automatic fault self-healing and data balance;
(4) The fusion system, storage and computing are deployed on the same hardware;
(5) Horizontal expansion, minimum deployment, on-demand capacity;

According to Wikibon's findings, server San will surpass traditional enterprise storage after five years, become the mainstream of the market, a decade after the traditional storage only 10% share, the server San has become the new darling of cloud computing era. Why would SDS, represented by the server SAN, be so hot in the market? The reason is due to the time when SDS was born.
(1) Explosion of storage demand: explosive growth of application data;
(2) Hardware Moore's Law: X86 hardware, software definition becomes possible;
(3) High-speed network development: the elimination of network bottlenecks, distributed become possible;
(4) Innovative Flash: eliminates the huge gap between computing and storage;
(5) Cloud Data Center: elasticity, performance, integration, management;
(6) The pursuit of TCO: higher cost-effective, lower opex;

Hyper-converged HCI and software-defined storage SDS are two of the most popular technologies in the world, and they are the core components of the next generation of software-defined data center SDDC. From the above definition of hyper-fusion, HCI realizes the fusion of computing, storage and network, and SDS is the core component of hyper-converged architecture. HCI is typically extended by an SDS solution, and the value of hyper-converged architecture is that it abandons dedicated shared storage systems, breaks the limits of traditional storage, and meets the new storage requirements of cloud computing. In other words, hyper-convergence relies on SDS, and the hyper-converged architecture is incomplete without SDS providing storage capabilities. Instead, SDS does not rely on hyper-convergence, which can be used as a standalone software-defined storage system, replacing traditional shared storage systems, which we call traditional EMC-represented storage in the IoE. The advantage of SDS is that it is not only in the hyper-converged form of HCI, but also as an isolated storage system that provides a viable storage alternative for enterprise applications, while maintaining flexible scalability while reducing storage costs, and leveraging the customer's existing storage architecture to Give full play to the functions of dedicated storage networks and storage computing.

The hyper-converged solution focuses on hypervisor and SDS, ideally hypervisor supports different SDS, and SDS can support different hypervisor. Hypervisor and SDS manufacturers in their respective areas of expertise to do a good job, through the depth of integration to provide customers with better ultra-integration solutions, to achieve ecological cooperation and mutual benefit. The reality is that hypervisor manufacturers also want to do SDS,SDS manufacturers also to do hypervisor, resulting in short board impact Hyper-integration program competitiveness, cooperation evolved into a more brutal direct competition. Typical case is, initially Nutanix and VMware deep happy cooperation, after VMware self-developed Vsan Nutanix abandoned, Nutanix had to based on KVM self-developed Acropolis hypervisor, so as to enter the direct competition between the two sides. Therefore, to fully build a full stack hyper-converged solution, in addition to technical factors and deep business factors, this is a very debatable issue, for the SDS or cloud computing in the field of startups to be more cautious.

3 Why is hyper-convergence popular?
With the next 5-10 years of new data center infrastructure moving toward software definition and hyper-convergence, san/nas storage is increasingly being replaced by software-defined storage, and hyper-converged architectures have been widely accepted by the market and customers as the core of the data center infrastructure. EMC, the world's largest storage vendor, sold itself to Dell, which also confirms the decline in traditional san/nas storage. The hyper-converged architecture becomes the enterprise-class customer choice, accelerating the transformation of business systems from traditional architectures to cloud computing architectures. According to Wikibon's analysis, 2016 will be the beginning of real cloud computing, hyper-integration and Serversan future market space will exceed 60 billion U.S. dollars, the annual compound growth rate of more than 20%. In stark contrast, traditional storage is being phased out, with a market share of only 10% per cent in the future. Hyper-converged market has been initially formed, traditional IT vendors data center solutions have lagged behind the new technology manufacturers, the potential market space in the future is huge.

The global hyper-converged market will reach $800 million in 2015, according to IDC's forecast analysis. If the Chinese market is based on a 10% share, it should be $80 million. IDC China has just released the "2015H1 China ultra-fusion market manufacturer Share Report", the report shows that in the first half of 2015, China's ultra-converged market size reached 39.7 million U.S. dollars, almost 2014 year 1.5 times times. The 2015 year-round data is not yet released and is expected to reach $100 million a year, roughly the same as global share. 2015 can be regarded as the Chinese super-fusion market, 2016 is likely to exceed 200 million U.S. dollars.

Why is hyper-convergence so popular? This also boils down to the hyper-converged architecture, which has a significant advantage in bringing high customer value. Hyper-converged HCI, like Google, Facebook and other Internet data centers, a large-scale infrastructure model, the realization of computing, storage, network and other resources of the unified management and scheduling, with more flexible scale-out capability, can provide the data center with the best efficiency, flexibility, scale, cost and data protection. The use of an integrated platform for computing storage hyper-convergence replaces the traditional server plus centralized storage architecture, making the entire architecture clearer and simpler, greatly simplifying the design of complex IT systems. Relative to the traditional it architecture, hyper-converged architecture has a natural advantage, the following descriptions are quoted from Nutanix:
(1) on-demand procurement: Change the procurement model, no one-time large-scale procurement, on-demand procurement, protection of existing investment, extend to the cloud computing architecture;
(2) Fast delivery: It can be delivered in 30 minutes from the top rack and quickly deployed;
(3) Simplified management: Single interface, unified management of computing, storage, virtualization and other resources, operations and maintenance management simplification;
(4) Elastic expansion: Distributed architecture, linear expansion, no node limit, no single point of failure, built-in local backup, the same city and remote disaster capacity;
(5) Single support: Single vendor guarantees all hardware and software, including compute, storage and virtualization support;

The above seems to be hyper-fusion and SDS manufacturer's self-hi, for marketing promotion or control of bidding parameters, actual real customer ideas? If you are a customer, why choose a hyper-converged architecture? Here are a few real reasons, probably beyond your imagination, and the reality is so brutal.
(1) reason 1: Business need
is really a business need, and traditional it architecture is not enough or expensive, it may be the size, data availability, business continuity, performance and other requirements. This type of customer is highly valued but less capable of proactively seeking and accepting hyper-converged architectures. With performance as an example, applications that require more than 100,000 iops are not so much imagined, although hyper-converged architectures can easily reach hundreds of thousands of IOPS. There is a project that ssan+vsphere all-flash hyper-converged, running on Oracle RAC applications, performance leverage, and the use of common San arrays is difficult to meet demand.

(2) reason 2: Cost
This is a pseudo-demand, traditional IT architecture can be satisfied, but the cost is too high, want to save costs, this kind of customer accounted for a large proportion. This can not be the greatest embodiment of the advantages of hyper-integration, but often the most impress customers, after all, can save money is the most affordable benefits. There are a few real customers who use Emc/netapp storage before, which is because of the high cost problem and go to adopt ssan+vsphere/ovirt hyper-fusion or physical fusion scheme.

(3) reason 3: Lee-old
This kind of demand is actually a bit of a deviation, originally not hyper-integration should do things, but the reality is real. Enterprise procurement of servers, storage and other IT resources are fixed assets, customers often want to retire or over-insured these resources for the old, so as to achieve the purpose of protecting investment. Recently implemented a project, the customer will be eliminated a bunch of IBM servers, the Cpu/mem/disk resources pieced together, simply spell out 4 of the server can be run properly, and then deploy a set of Ssan as a separate SDS, for some enterprise business to provide storage services.

Is the 4 hyper-fusion really perfect?
The benefits of hyper-converged architecture and customer value this has not been questioned, the global and domestic markets have been initially formed, HCI is the next 5-10 years of new data center infrastructure of choice. But is hyper-fusion really perfect? I personally prefer the dialectical view of the problem, hyper-integration is not there are some limitations or problems? Here we will find fault, but also the choice of hyper-converged architecture to consider factors.
(1) new islands of information
Almost all hyper-converged solutions do not support external storage in the data center, and most businesses cannot replace the entire data center infrastructure in the short term, resulting in a split-off data center into two separate, disparate infrastructure silos. For large data centers, there is a high likelihood that different hyper-converged architectures will be deployed at the same time for different business needs and balances, and the result is that there are several new silos of information that are not integrated and interoperable between HCI. The new information Island brings the problem of resource utilization efficiency and unified management.

(2) performance consistency issues
The performance of storage in the data center is critical, and the expected performance is predictable and consistent, including latency, IOPS, and bandwidth, which is especially critical for core business systems. This is a big challenge for hyper-converged architectures. The main reasons are two points, one is hyper-converged architecture "share everything." Compute and storage compete for physical resources such as cpu/memory/network, and compute and storage are interdependent, and once one resource demand rises, the other side is depleted, impacting performance and creating ripple effects throughout the infrastructure. Although Cgroup or container technology can be used for resource isolation limitations, the effects of non-hyper-converged architectures are different. Second, hyper-converged architecture "all distributed and software definition", cluster scale, network, hard disk, server failure probability will increase, data deduplication/compression/encryption/erasure code and other functions are implemented with software, fault self-repair and data function implementation will consume certain system resources, resulting in performance degradation and jitter. Self-healing flow control, data function bypass to hardware module processing, these methods can alleviate the performance consistency problem, but it seems to deviate from the concept of hyper-fusion.

(3) horizontal expansion of the calamity
One of the key features of hyper-converged architectures is ease of expansion, minimal deployment, and on-demand capacity. Hyper-converged architecture vendors claim that the largest cluster size varies widely from dozens of to thousands of nodes, typically from 3 nodes. In hyper-fusion, the computing power, storage performance and capacity are synchronous expansion, can not meet the expansion of the individual ability in reality, some manufacturers also have requirements for the expansion of the minimum unit, expansion flexibility will be limited. When the cluster reaches a certain scale, the complexity of the system architecture becomes more nonlinear, the cluster management becomes more difficult, and the probability of hardware failure and self-repairing will be greatly increased. Therefore, we are not recommended to build a large cluster, if the business allows to build more than the appropriate size of smaller clusters, or in a large cluster to build fault domain or sub-resource pool, Everbright is not possible. Cluster expansion also faces a tricky problem, namely, capacity balancing. If the storage cluster has a large capacity, equalization is a very long and painful process, and can have a significant impact on normal business load.

(4) system complexity
Hyper-converged architecture simplifies it architecture, dramatically reduces data center design complexity, enables rapid delivery, and greatly simplifies operations management. However, this is all based on the user perspective, in terms of product development, hyper-convergence actually makes the internal software complexity higher. As we have explained earlier, hyper-converged architectures require CVM virtual machine controllers and need to pass host physical devices through to control VMS, increasing the complexity of deployment configuration management. Computing and storage have different requirements for the hardware platform, which in some way increases the complexity of compatibility verification. In hyper-converged architectures, management, compute, storage, and high availability often require a separate virtual network, and the network configuration is more complex. At the same time, the allocation, isolation, and scheduling of shared physical resources is an added complication. Also, if a failure occurs, the problem of tracking debugging and analyzing the diagnosis becomes more difficult.

(5)SSD tiered storage
Flash SSDs are essential elements in hyper-converged architectures, eliminating the huge gap in compute and storage, and addressing the I/O performance bottlenecks, especially the I/O random read and write capability. At present, the price of flash memory is higher than HDD disk, due to cost factors, all-flash hyper-converged solution application is still less, most of the applications are based on SSD hybrid storage configuration, so as to achieve a higher price/performance ratio. Typically, we assume that hot data accounts for 10-20%, configure the appropriate scale of SSD storage, store hotspot data in SSD storage in cache acceleration or tier tiering mode, and once the hotspot data exceeds the pre-set threshold or triggers a migration strategy, The cooler data is migrated back to the HDD disk storage in accordance with the corresponding elimination algorithm, which is expected to achieve an overall balance of performance and capacity. It looks perfect, doesn't it? SSD specializes in random read and write, bandwidth is not its strength, for bandwidth-based applications, SSD performance does not help. This is not a good estimate of hot-spot data ratios, which can get worse if the SSD is not configured enough. What if the scenario is appropriate and the SSD is properly configured? When SSD space is eventually filled with hotspot data, data migration is triggered, and HDD storage remains an I/O performance bottleneck, with normal I/O business load, and overall performance degradation and jitter. In order to alleviate this problem, the SSD cache/tier feature implementation, on the one hand will filter out sequential read/write I/O, on the other hand, the space threshold is set low, early data migration, and select the System idle time execution and flow control. The negative effect is that the SSD performance acceleration is limited, and the physical equipment efficiency is not fully achieved. In addition, the SSD itself is written full of performance will also appear large fluctuations. Therefore, SSD hybrid storage is not an ideal mode, and in practice we recommend a full flash SSD or full disk HDD configuration based on the scenario to achieve consistent performance. If you really can't use SSDs all the time, there's another way to use it, and create a full SSD and a full HDD storage pool that people can assign to different storage pools for performance requirements.

(6) Enterprise-class data functions
Today, most hyper-converged systems and SDS systems have core enterprise-class capabilities, including data redundancy, auto-thin provisioning, snapshots, clones, SSD cache/tier, Data Auto-rebuilds, high-availability/multipath data functions, and some even data deduplication, data encryption, Advanced data functions such as data compression. However, compared to high-end storage systems, there is a big gap between hyper-converged architectures that host core critical applications, including but not limited to QoS control, data protection, data migration, backup disaster recovery, and consistent high performance. The core storage system should follow the ras-p principle, first of all to stabilize the reliability, followed by the enterprise data function completeness, finally is the high performance, this order can not mess, the light has high performance is not good. For example, Ceph, Enterprise-class Data feature list and full, functional specifications are very tempting, but really stable and can actually produce deployment applications are not much. At present, the core business system is not very dare to move to hyper-converged architecture, mainly from the non-core business to start testing, after all, hyper-convergence time is relatively short, need more time and practice to verify the ras-p characteristics. However, the future hyper-integration must be the core of key business of the mainstream architecture.

(7) Physical environment Application
Today's universally accepted scenarios are new applications such as desktop cloud, server virtualization, OpenStack Private cloud, and big data analytics. In theory, hyper-converged systems can be applied to all types of applications in the IT environment, and it is important to note that hyper-converged systems manage virtualized environments while more traditional IT applications are still running on physical servers and traditional storage systems. We can be optimistic that no application can be deployed on a hyper-converged infrastructure, but many IT applications are best kept running in physical hardware architectures, such as database applications, hardware dependencies, and the compatibility of virtualized environments. Real-time control systems and a large number of legacy IT systems.

(8) Heterogeneous virtualized environment
The current hyper-converged scenario typically supports only one virtualized environment, and Nutanix can support a variety of virtualized environments, but for a set of hyper-converged architecture deployments, it actually supports only one virtualized environment. Each virtualized environment has its own advantages, and many enterprises may need to run several virtualized environments at the same time, such as VMware, KVM, Hyper-V, and XEN, because hyper-converged does not support heterogeneous virtualized environments and requires multiple sets of hyper-converged architectures, which are new silos of information. Customers very much want to see hyper-converged architectures that support heterogeneous virtualized environments.

(9) Hyper-converged data sharing
The hyper-converged architecture replaces traditional shared storage with software-defined storage to solve the problem of virtualized storage, where the SDS essentially refers to Serversan, which provides distributed block storage. However, whether it is a virtual machine or a physical machine, the actual IT application has data sharing requirements and requires a distributed file system or NAS storage System. This is currently a common lack of hyper-convergence, and the reality is dependent on externally deployed NAS or clustered NAS storage systems, such as Glusterfs, ZFS. From the technical architecture and implementation, an SDS system is well-integrated to support object/block/file storage, which is very difficult to achieve. For example, Ceph, whose CEPHFS has never met the production environment deployment standard, let alone performance. As a result, two sets of SDS storage can be deployed in the same way in the hyper-converged architecture, providing distributed block storage and file system file sharing storage, such as Ssan and Glusterfs, without requiring distributed unified storage.

(ten) full stack hyper-converged architecture
At present, many manufacturers are pushing hyper-converged architectures, some of which are SDS vendors, some are virtualization vendors, and more are integrators like Vmware/nutanix, who do SDS and virtualization at the same time. SDS and virtualization are two completely different areas, and the technology threshold is very high, a manufacturer at the same time the two pieces are very good is very difficult. Both want to do the manufacturers, nothing but to control costs, self-controllable, expand the market, financing good storytelling. For startups, teams and funds are very limited, and creating a full stack hyper-converged architecture is even more difficult. The correct posture may be, determine the strategic direction to focus on a good piece, and then find another well-done manufacturers to carry out strategic ecological cooperation, strong alliances to create a competitive hyper-converged architecture, whether from the product technology or market angle is a win.

5 Do you want the container to be hyper-fused?
The container is a magical thing, no less fiery than hyper-fusion, it is leading a new cloud data center architecture changes, and the object of revolution is not yet a big line of hyper-converged architecture, many enterprises have or are migrating applications from virtual machines to containers. So what is a container? A simple container is an isolated process on the host that runs in a sandbox, qualifying and isolating the host physical resources used by the Cgroup/namspace technology. Why are containers so heated? The hypervisor abstracts the entire device, usually with high system requirements, and the container simply abstracts the operating system kernel, using a shared operating system, which can use system resources more efficiently, and the same hardware can create 4-6 times as many containers as virtual machines. This can save a significant amount of cost to the data center while quickly building containerized applications that run anywhere and simplifying deployment and management. "Minimal deployment, on-demand capacity" is the elastic scaling issue that cloud computing needs to address. Containers are very lightweight relative to virtual machines, and the fundamental problem it solves is to increase efficiency and speed, allowing for second-level scaling (including scaling). It is these remarkable advantages that the container has the potential to replace the virtual machine as the cloud-computing infrastructure, so hot and natural to understand. It is worth mentioning that not all applications are containerized, the key to see whether the business is suitable for highly elastic computing micro-services, can not be blindly respected.

Container is used to carry the application, its design is for the application of the running environment packaging, start, migration, elastic expansion, the container one of the most important features is stateless, can be dynamically created and destroyed as needed. However, not all applications are stateless, what about stateful containers? The data that needs to be persisted in the container is the state, which cannot be discarded arbitrarily. How to persist data for a container is a problem that has existed since the day of Docker's birth. The general view is that data should not be placed in containers, it is best to ensure that all containers are stateless, but still provide an internal mechanism for preserving state. Docker's only state-related concept is volume, where the container accesses the external application data interface, completely out of Docker control. Volume solves the problem of data persistence storage for a container, but it is only a data interface and the container itself is not responsible for the management of persisted data. This problem is almost ignored by all container manufacturers, mainly rely on external storage to solve, one of the solutions is to persist the container data in reliable distributed storage, such as Glusterfs, Ceph, the administrator no longer consider the migration of container data.

The application of direct running in the container, which is inherently closer to the application than VM VMS, is most capable of dynamically configuring different storage policies through application-aware deep-seated requirements for storage. Therefore, an external storage system that provides state persistence for the container should be an application-oriented storage system that provides granular storage policies for containers of different types of applications and dynamic Intelligent Application awareness. So, does the container COMPUTE + application-aware storage need to be hyper-converged? Here we define container hyper-convergence: distributed storage, fully containerized, and storage controller containerized. The container should naturally be stateless, its function is to achieve agile elastic calculation, if the container is stateful, this advantage will be greatly weakened. Distributed storage is a heavy system, such as Ceph,glusterfs, which needs to manage a large number of disks and network resources that are not inherently containerized, and are better suited to run directly on native physical machine operating systems. Distributed storage has strict status requirements, if the application container and storage controller container is fused in one, the change of storage state will seriously affect the application. Therefore, I personally understand that the container essentially does not need to do so-called hyper-fusion, the container is focused on a flexible cloud computing architecture, distributed storage is focused on container-aware application storage, need to have a state by the independent external professional storage to be responsible for data access through the storage mechanism provided by the container, such as rancher Convoy or Flocker container storage driver. In fact, there is very little demand for stateful containers, which is why container storage is a neglected factor. Finally, the biggest difference between a container and VM VMs is that it's not a VM, and the container hyper-converged is not a bad idea.

(End of full text)

is hyper-converged architecture really perfect?