伸缩性、可用性、稳定性(Scalability, Availability & Stability Patterns)

    技术2022-06-12  62

    一 自我有要求的读者应该提出问题:(研习:掌握层次:)能力级别:不会(了解)——领会(理解)——熟练——精(why)——通(融汇贯通) 1.1 什么是Scalability, Availability&Stability Patterns ? 1.2 以上各个模式都说了些什么?   1.2.1 Scalability Patterns 从State和Behavior都说了些什么? 是简单介绍还是有一定深度呢?   1.2.2 Availability  Patterns 都说了些什么?   1.2.3 Stability Patterns ?都说了些什么?   该PPT只是比较全面、轻轻点水般介绍了一下当前架构思想,只能增加架构设计的视野,要想能够很好的使用这些思想,必须选一两个感兴趣并有前景的点深入下去才行。 1.3 (3)这本书说得有道理吗? 是全部有道理,还是部分有道理?why?     作者最终的目标,加上他建议的达成目标的方法--这要看你认为追求的是什么,以及什么才是最好的追求方法而定。   这些Pattern在什么情况下用?如何用? 可以解决那些问题?而不解决那些问题? 1.4 (4)赞同一本实用性的书之后,确实需要你采取行动。 照着作者希望你做的方式来行动。How   行动:为达到某种目的而进行的活动。行动目标,行动方法,行动开始时间,结束时间,行动人,行动地点,行动方式。   在架构设计时,考虑这些因素会提高系统Scalability,Availability&Stability等? 二 研读过程中应该努力寻找问题的答案,对问题的思考越深入,收获也就越多: 2.1 什么是Scalability, Availability&Stability Patterns ?http://www.jdon.com/jivejdon/thread/38928   2.1.1 Scalability(伸缩性、可扩展性):(研习:1 掌握层次:理解) Scale up/Scale out     可伸缩性就是通过增加资源使服务容量产生线性(理想情况下)增长的能力。可伸缩应用程序的主要特点是:附加负载只需要增加资源,而不需要对应用程序本身进行大量修改。     在一些大的系统中,预测最终用户的数量和行为是非常困难的,伸缩性是指系统适应不断增长的用户数的能力。提高这种并发会话能力的一种最直观的方式就增加资源(CPU,内存,硬盘等),集群是解决这个问题的另一种方式,它允许一组服务器组在一起,像单个服务器一样分担处理一个繁重的任务。     尽管原始性能对于确定应用程序所能支持的用户数很重要,但可伸缩性和性能是两个单独的实体。事实上,性能结果有时可能与可伸缩性结果是对立的。       可伸缩性Scalable高性能系统设计:http://www.jdon.com/jivejdon/thread/40668       可伸缩性最佳实战: http://www.jdon.com/jivejdon/thread/37793       CAP理论以及Eventually Consistent (最终一致性)解析:http://www.jdon.com/jivejdon/thread/37999       BASE(Basically Availability、Soft state、Eventually consistent)       你真的明白什么是可伸缩性吗?http://developer.51cto.com/art/200710/57496.htm   2.1.2 Availability(可用性、有效性) (研习:1 掌握层次:理解)     ISO9241/11中的定义是:一个产品可以被特定的用户在特定的上下文中,有效、高效并且满意得达成特定目标的程度(The extent to which a product can be used by specified users to achieve specified goals with effectiveness, efficiency and satisfaction in a specified context of use.)。     GB/T3187-97对可用性的定义:在要求的外部资源得到保证的前提下,产品在规定的条件下和规定的时刻或时间区间内处于可执行规定功能状态的能力。它是产品可靠性、维修性和维修保障性的综合反映。     实际比较常用Shakel(1991)对可用性的定义:可用性是指技术的“能力(按照人的功能特性),它很容易有效地被特定范围的用户使用,经过特定培训和用户支持,在特定的环境情景中,去完成特定范围的任务。”     单一服务器的解决方案并不是一个健壮方式,因为容易出现单点失效。像银行、账单处理这样一些关键的应用程序是不能容忍哪怕是几分钟的死机。它们需要这样一些服务在任何时间都可以访问并在可预期的合理的时间周期内有响应。集群方案通过在集群中增加的冗余的服务器,使得在其中一台服务器失效后仍能提供服务,从而获得高的可用性。        可用性:http://baike.baidu.com/view/1436.htm   2.1.3 Stability(稳定性、稳定度) (研习:1 掌握层次:理解)     软件的稳定性,指软件在持续操作时间内出错的概率,例如一天时间内会出错1次或几次。具体来定义它是否属不属于稳定,根据软件的具体要求来定义。     软件的稳定性应该和软件的可靠性是不同的。软件的稳定性从软件开发的角度出发,强调软件架构的稳定,也就是说需求、代码等的变更对软件系统的影响尽可能地小,这也是架构设计要解决的首要任务。     这需要作边缘测试来检验,而边缘测试的定义和实施都是需要很多经验来支持的,这对于新手来说是无法做到的。     边缘测试,举个例子:在压力测试中,可以在压力的最大值、最小值附近取值进行测试,甚至考虑超过最大值和最小值的方式进行测试。这就属于边缘测试。     平均无故障时间等指标是说明系统的可靠性的。系统的稳定性应该是指系统的一些边缘故障。比如系统运行一直很好,只是偶尔出现一些奇怪的问题,但是找不到原因,经过重启或者重装之后就恢复正常,这就在考验系统的稳定性。从系统本身来说,没有无缘无故的爱也没有无缘无故的恨,出现问题一定是在某个方面有缺陷,而且问题往往是出在设计上。如果要从设计角度去保障软件的稳定性就需要设计人员充分的考虑系统各个模块之间的关系,减少耦合度,是问题隔离起来。很多问题都是出在模块之间的调用上的。模块内部也是一样,最大的问题就出在内存的使用上,不过这就到编码的问题了。总之,稳定的系统需要专业的有丰富经验的设计人员,合理的划分系统,详细设计做到足够细,避免在开发阶段出现问题。        稳定性:http://baike.baidu.com/view/251942.htm        什么是软件的稳定性:http://topic.csdn.net/t/20051220/22/4471364.html 2.2 以上各个模式都说了些什么?   2.2.1A Scalability Patterns 从State和方面都说了些什么? 是简单介绍还是有一定深度呢?     2.2.1.1 Scalability解决什么问题?Managing Overload (研习:1 掌握层次:理解)     2.2.1.2 Scalability有哪两种扩展方式? Scale up vs Scale out (研习:1 掌握层次:理解)     2.2.1.3 关于Scalability一般建议是什么? (研习:1 掌握层次:理解)            *Immutability as the default:不变作为一个缺省            *Referential Transparency(FP) :参考透明性            *Laziness: 懒惰            *Think about your data:               *Different data need different guarantees     2.2.1.4 伸缩扩展系统时,有哪些因素中权衡? (研习:1 掌握层次:理解)            没有免费的午餐,扩展系统时有代价的。        (1)Performance       vs  Scalability ?            How do I know if I have a performance problem? If your system is slow for a single user.            How do I know if I have a scalability problem? If your system is fast for a single user but slow under heavy load        (2)Latency(等待时间) vs  Throughtput(容量/吞吐率) ?            You should strive(力求) for maximal throughput with acceptable latency(等待时间)        (3)Availability      vs  Consistency ?            Brewster's CAP theorem: CAP(Consistency/Availability/Partition tolerance) You can only pick 2            *对于Centralized system 对CAP权衡的因素?              In a centralized system(RDBMS etc.) we don't have network partitions, e.g. P no in CAP               So you get both: Availability and Consistency  如ACID(Atomic/Consistent/Isolated/Durable)            *对于Distributed system 对CAP权衡的因素?             In a distributed syste we (will) have network partitions, e.g. P in CAP             So you get to only pick one: Availability or Consistency            *如何利用CAP理论指导实践呢?             there are only two types of systems:               1. CA == CP (they are equivalent)               2. AP             there is only one choice to make. In case of a network partition, what do you sacrifice?               1. C:Consistency               2. A:Availabilty               权衡的选择:BASE(Basically Avaialable/Soft state/Eventually consistent)               Eventual Consistency is an interesting trade-off   2.2.1.5 Scalability Patterns:State都说了些什么?     2.2.1.5.1 Partitioning:分区技术 (研习:1 掌握层次:理解)     2.2.1.5.2 HTTP Caching (研习:1 掌握层次:理解)           Reverse Proxy(反向代理)软件:Varnish/Squid/rack-cache/Pound/Nginx/Apache mod_proxy           CDN(Content Delivery Network 内容分发网络):使用户可就近取得所需内容           Generate Static Content:             Precompute content:              *Homegrown + cron or Quartz              *Spring Batch              *Gearman              *Hadoop              *Google Data Protocol              *Amazon Elastic MapReduce     2.2.1.5.3 RDBMS Sharding (研习:1 掌握层次:理解)    (属于SoR:Service of Record):分片:水平扩展(Scale Out,亦或横向扩展、向外扩展) Sharding:http://baike.baidu.com/view/3126360.htm           How to scale out RDBMS?          2.2.1.5.3.1 Partitioning:例子如:把User[A-C]放到DB1,把User[D-F]放到DB2...把User[X-Z]放到DBn          2.2.1.5.3.2 Replication :把User[A-C]User[D-F]放到DB1,把User[D-F]User[A-C]放到DB2...把User[N1-N2]User[M1-M2]放到DBn          2.2.1.5.3.3 anti-pattern(反模式):ORM+rich domain model ,Attempt: Read an object from DB, Result:You sit with your whole database in your lap     2.2.1.5.4 NOSQL:Not Only SQL (属于SoR:Service of Record) (研习:1 掌握层次:理解)        2.2.1.5.4.1 Think about your data?            When do you need ACID?            When is Eventually Consistent a better fit?            Different kinds of data has different needs            When is a RMDBS not good enough?            Scaling reads to a RDBMS is hard!            Scaling writes to a RDBMS is impossible            Do we really need a RDBMS? But many times we don't.            Who's ACID?              Relational DBs(MySQL,Oracle, Postgres)/Object DBs(Gemstone, db4o)/Clustering products(Coherence, Terracotta)/Most caching products(ehcache)            Who's BASE?              Distributed databases: Cassandra/Riak/Voldemort/Dynomite/SimpleDB/etc.            NOSQL in the wild:              Google:Bigtable/ Amazon:Dynamo / Amazon:SimpleDB/Yahoo:HBase/Microsoft:Dynomite/Facebook:Cassandra/LinkedIn:Voldemort        2.2.1.5.4.2 Chord(和弦) & Pastry(糕点):              Distributed Hash Tables(DHT) / Scalable/ Partitioned/ Fault-tolerant/ Decentralized/ Peer to Peer/ Popularized(Node ring/Consistent Hashing)            Bigtable              How can we build a DB on top of Google File System?              Paper:Bigtable: A distributed storage system for structured data, 2006              Rich data-model, structured storage              Clones: HBase|Hypertable|Nepture            Dynamo:              Hoe can we build a distributed has table for the data center?              Paper:Dynamo:Amazon's highly available key-value store, 2007              Focus:partitioning,replication and availability              Eventually Consistent              Clones:Voldemort|Dynomite        2.2.1.5.4.3 Types of NOSQL stores              Key-Value databases(Voldemort,Dynomite)              Column databases(Cassandra,Vertica)              Document databases(MongoDB,CouchDB)              Graph databases(Neo4J,AllegroGraph)              Datastructure databases(Redis,Hazelcast)       2.2.1.5.5 Distributed Caching (研习:1 掌握层次:理解)            2.2.1.5.5.1 Write-through            2.2.1.5.5.2 Write-behind            2.2.1.5.5.3 Eviction Policies               TTL(time to live)               Bounded FIFO(first in first out)               Bounded LIFO(last  in first out)               Explicit cache invalidation            2.2.1.5.5.4 Replication            2.2.1.5.5.5 Peer-To-Peer(P2P)               Decentralized:分散               No"special" or "blessed" nodes               Nodes can join and leave as they please            2.2.1.5.5.6 Distributed Caching Products:               EHCache               JBoss Cache               OSCache               memcached: Simple/Key-Value(string->binary)/Clients for most languages/Distributed/Not replicated - so I/N chance for local access in cluster            2.2.1.5.6 Data Grids/Custering:数据网格  Parallel data storage (研习:1 掌握层次:了解)               Data replication               Data partitioning               Continuous availability               Data invalidation               Fail-over               C+A in CAP            Data Grids/Custering Products:               Coherence/Terracotta/GigaSpaces/GemStone/Hazelcast/Infinispan      2.2.1.5.7 Concurrency:并发          2.2.1.5.7.1 Shared-State Concurrency (研习:1 掌握层次:理解)             Every one can access anything anytime             Totally indeterministic             Introduce determinism at well-defined places...             ... using locks           Problems with locks:             Locks do not compose             Taking too few   locks             Taking too many  locks             Taking the wrong locks             Taking locks in the wrong order             Error recovery is hard           Please use java.util.concurrent.*:             ConcurrentHashMap/BlockingQueue/concurrentQueue/ExecutorService/ReentrantReadWriteLock/ParallelArray/and much much more..          2.2.1.5.7.2 Message-Passing Concurrency (研习:1 掌握层次:理解)           *Actors: erlang万物皆Actor, Actor之间只有发送消息这一种通信方式              Implemented in Erlang, Occam,Oz              Encapsulates state and behavior              Closer to the definition of OO than classes              Share NOTHING              Isolated lightweight processes              Communicates through messages              Asynchronous and non-blocking              No shared state   ... hence, nothing to synchronize.              Each actor has a mailbox(message queue)              Easier to reson about              Raised abstraction level              Easier to avoid  -Race conditions -Deadlocks -Starvation -Live locks           *Actor libs for the JVM:              Akka(Java/Scala)/scalaz actors(Scala)/Lift Actors(Scala)/Scala Actors(Scala)/Kilim(Java)/Jetlang(Java)/Actors'Guild(Java)/Actorom(Java)/FunctionalJava(Java)/GPars(Groovy)          2.2.1.5.7.3 Dataflow Concurrency (研习:1 掌握层次:了解)              Declarative               No observable non-determinism              Data-driven - thread block until data is available              On-demand, lazy              No difference between:                  Concurrent&Sequential code              Limitations: can't have side-effects          2.2.1.5.7.4 Software Transactional Memory (研习:1 掌握层次:了解)              See the memory(head and stack )as a transactional dataset              Similar to a database: begin commit  abort/rollback              Transactions are retired automatically upon collision              Rolls back the memory on abort              Transactions can nest              Transactions compose          Transactions restrictions: All operations in scope of a transaction: Need to be idempotent          STM libs for the JVM:              Akka(Java/Scala)              Multiverse(Java)              Clojure STM(Clojure)              CCSTM(Scale)              Deuce STM(Java)    2.2.1B Scalability Patterns:Behavior(行为、性能)都说了些什么?         1.2.1B.1 Event-Driven Architecture (研习:1 掌握层次:了解)             1.2.1.6.1 Domain Events             1.2.1.6.2 Event Sourcing             1.2.1.6.3 Command and Query Responsibility Segregation(CQRS) pattern                     in a nutshell             1.2.1.6.4 Event Stream Processing             1.2.1.6.5 Messaging                  Publish-Subscribe          Point-to-Point          Store-forward          Request-Reply                Standards: AMQP(即Advanced Message Queuing Protocol,高级消息队列协议) 和JMS(Java Messaging Service)                Products: RabbitMQ(AMQP)/ActiveMQ(JMS)/Tibco/MQSeries/etc             1.2.1.6.6 Enterprise Service Bus                products: ServiceMix(Open Source)|Mule(Open Source)|Open ESB(Open Source)|Sonic ESB|WebSphere ESB|Oracle ESB|Tibco|BizTalk Server             1.2.1.6.7 Actors                Fire-forget:Async send            Fire-And-Receive-Eventually:Async send + wait on Future for reply             1.2.1.6.8 Enterprise Integration Architecture(EIA)                参考书《Enterprise Integration Patterns》                Apache Camel: More than 80 endpoints/XML(Spring) DSL/Scala DSL         1.2.1.6.2 Compute Grids(研习:1 掌握层次:了解)            Parallel execution:并行执行                  Automatic provisioning              Load balancing              Fail-over              Topology(拓扑) resolution            Products:          Platform/DataSynapse/Google MapReduce/Hadoop/GigaSpeaces/GridGain         1.2.1.6.3 Load-balancing(研习:1 掌握层次:理解)                Random allocation 随机分配算法                Round robin allocation 循环分配算法            Weighted allocation 负载分配算法            Dynamic load balancing:                   Least connections :连接数最少               Least server CPU  :CPU服务最少                   etc.                DNS Round Robin(simplest) : Ask DNS for IP for host/Get a new IP every time                Reverse Proxy(better)            Hardware Load Balancing             Load balancing products:                Reverse Proxies: Apache mod_proxy(OSS)|HAProxy(OSS)|Squid(OSS)|Nginx(OSS)|VLS            Hardware Load Balancers: BIG-IP|Cisco         1.2.1.6.4 Parallel Computing(研习:1 掌握层次:了解)             SPMD Pattern:         Single Program Multiple Data             Very generic pattern, used in many other patterns         Use a single program for all the UEs             Use the UE's ID to select different pathways through the program. F.e:            Branching on ID                Use ID in loop index to split loops             Keep interactions between UEs explicit         Master/Worker Pattern             Good scalability             Automatic load-balancing             How to detect termination? Bag of tasks is empty/ Poison pill             If we bottleneck on single queue? Use multiple work queues/ Work stealing         What about fault tolerance? Use"in-progress" queue         Loop Parallelism Pattern         Fork/Join Pattern         MapReduce Pattern             UE:Unit of Execution: Process/Thread/Coroutine/Actor  2.2.2 Availability  Patterns 都说了些什么?       What do we mean with Availability ?     2.2.2.1 Fail-over:故障切换 (研习:1 掌握层次:理解)             simple  Fail-over             complex Fail-over             Network fail-over     2.2.2.2 Replication (研习:1 掌握层次:理解)         *Active  replication - Push         *Passive replication - Pull            * Data not available, read from peer, then store it locally              Works well with timeout-based caches           Master-Slave           Tree replication           Master-Master           Buddy(伙伴) Replication   2.2.3 Stability Patterns ?都说了些什么? (研习:1 掌握层次:了解)       2.2.3.1 Timeouts:Always use timeouts (if possible):       2.2.3.2 Circuit Breaker:断路开关,断路器       2.2.3.3 Let-it-crash       2.2.3.4 Fail fast       2.2.3.5 Bulkheads       2.2.3.6 Steady State       2.2.3.7 Throttling   2.2.4 Extra material(Client-side consistency|Server-side consistency) (研习:1 掌握层次:理解)      Client-side consistency      Server-side consistency 三、(3)这本书说得有道理吗? 是全部有道理,还是部分有道理?why?     在你不能回答上面两个问题时,无法回答这个问题的    作者最终的目标,加上他建议的达成目标的方法--这要看你认为追求的是什么,以及什么才是最好的追求方法而定。    这些Pattern在什么情况下用?如何用? 可以解决那些问题?而不解决那些问题?       3.1.1 作者最终的目标:让软件架构师及产品经理们,了解当前主流架构模式。    3.1.2 他建议的达成目标的方法? 先了解基本软件架构特性如:Scalability/Avaiability/Stability Pattern,          其次介绍各个Pattern具体技术思想,便于自己在设计软件时思考借鉴,          再其次,介绍各个Pattern具体技术产品(开源),便于在设计软件时做参考等。    3.1.3 这些Pattern在什么情况下用?在软件开发周期:设计阶段(尤其架构设计,粗粒度技术选型时)    3.1.4 如何用? 这个文档指示给你一个引子,后续如何使用,需要研究具体的技术产品(一般产品都有:deom/reference/api help 等)    3.1.5 可以解决那些问题?而不解决那些问题? 可以帮助架构设计技术选型等,帮助提高找到满足设计目标的方法程度。          不解决那些问题: 具体如何设计、如何技术选型、如何研习选型产品、编码。。。。 四、(4)赞同一本实用性的书之后,确实需要你采取行动。 照着作者希望你做的方式来行动。How   行动:为达到某种目的而进行的活动。行动目标,行动方法,行动开始时间,结束时间,行动人,行动地点,行动方式。   在架构设计时,考虑这些因素会提高系统Scalability,Availability&Stability等?    4.1 我看这个技术资料的目的:        1 在关于当前主流架构设计讨论中,能清楚知道别人说的概念,能够讨论一些相关技术,提高自己设计水平        2 在自己产品设计中如何应用这些思想,帮助设计出更好的产品    4.2 我应该采取什么实际行动:        1. 技术交流要能够说相关知识点,及why        2. 大型分布式架构设计、网管架构设计、中心架构设计、。。。能够用上这些Pattern思想提高设计水平

     

    五、 参考:

        5.1  Scalability, Availability & Stability Patterns ( Jonas Bonér)

    http://www.slideshare.net/jboner/scalability-availability-stability-patterns


    最新回复(0)