1、cal files Column: a key in the k/v container inside a row Timestamp: long milliseconds, sorted descending Value: a time-versioned value in the k/v container What data types are stored in key/value pairs? Row keys, column names, values: arbitrary bytes HBase Data Model: Brief Recap Table: design-time
2、 namespace, has many rows. Row: atomic key/value container, with one row key Column Family: divide columns into physical files Column: a key in the k/v container inside a row Timestamp: long milliseconds, sorted descending Value: a time-versioned value in the k/v container What data types are stored
3、 in key/value pairs? Row keys, column names, values: arbitrary bytes Table and column family names: printable characters HBase Data Model: Brief Recap Table: design-time namespace, has many rows. Row: atomic key/value container, with one row key Column Family: divide columns into physical files Colu
4、mn: a key in the k/v container inside a row Timestamp: long milliseconds, sorted descending Value: a time-versioned value in the k/v container What data types are stored in key/value pairs? Row keys, column names, values: arbitrary bytes Table and column family names: printable characters Timestamps
5、: long integers HBase Data Model: Brief Recap One more thing that bears repeating: every cell (i.e. the time-versioned value of one column in one row) is stored fully qualified (with its full rowkey, column family, column name, etc.) on disk. So now you know whats available. Now, how do you model th
6、ings? Lets start with the entity / attribute / relationship modeling paradigm, and see how far we get applying it to HBase. A note about my example: its for clarity, not realism. For bands the important thing is that any byte array in HBase can represent more than one logical attribute. If we want,
7、we can even add types to the schema definition. If we want, we can even add types to the schema definition. ? If we want, we can even add types to the schema definition. ? HBase dont care, but we do (sometimes). If we want, we can even add types. You could also mark things as ASC or DESC, depending
8、on whether you invert the bits. ? This is pretty textbook stuff, but heres where it gets exciting. This is pretty textbook stuff, but heres where it gets exciting. (If youre astute, youll notice we havent talked about relationships yet.) HBase has no foreign keys, or joins, or cross-table transactio
9、ns. This can make representing relationships between entities . tricky. Part of the beauty of the relational model is that you can punt on this question. If you model the entities as fully normalized, then you can write any query you want at run time, and the DB performs joins on tables as optimally
10、 as it can. (This relies heavily on indexing.) In HBase (or any distributed DB) you dont have that luxury. Joins get way more expensive and complicated across a distributed cluster of machines, as do the indexes that make them happen efficiently. HBase has neither joins nor indexes. You have two cho
11、ices, if you need relationships between entities. You have two choices, if you need relationships between entities. Roll your own joins You have two choices, if you need relationships between entities. Roll your own joins Denormalize the data Rolling your own joins is hard. Youll probably get it wro
12、ng if you try to do it casually. Ever implemented a multi-way merge sort by hand? Denormalizing is fun! But it also sucks. Well see why. The basic concept of denormalization is simple: two logical entities share one physical representation. The basic concept of denormalization is simple: two logical
13、 entities share one physical representation. Of course, there are lots of ways to skin a cat. Lets look at standard relational database denormalization techniques first. In a relational database, if the normalized physical model is this: You could start with the Band, and give it extra columns to ho
14、ld the shows: You could start with the Band, and give it extra columns to hold the shows: Thats pretty obviously a bad idea, because bands can play a lot of shows. WTF? You could also just give it an unstructured blob of text for all the shows. You could also just give it an unstructured blob of tex
15、t for all the shows. But then youve given up the integrity of your data. (Which might be fine. If so, stop here.) You get similar problems if you try to bring all the info into the Show table. Another solution is to focus on the junction table. And pull in copies of the info in the other tables: Lea
16、ving you with one table, with one row per band/show combination: The cons to this should be self-evident. Updating and querying are more complicated, because you have to deal with many representations of the same data, which could get out of sync, etc. But the pros are that with huge data, answering
17、 some questions is much faster. (If you happen to be asking the query in the same way you structured the denormalization, that is.) So back to HBase. How do you denormalize in an HBase schema? HBase columns can be defined at runtime. A column doesnt have to represent a pre-defined attribute. HBase c
18、olumns can be defined at runtime. A column doesnt have to represent a pre-defined attribute. In fact, it doesnt have to be an attribute at all. A set of dynamically named columns can represent another entity! If you put data into the column name, and expect many such columns in the same row, then lo
19、gically, youve created a nested entity. You can scan over columns. See: hadoop- So we can store shows inside bands. Which is like the denormalization we say earlier, except without the relational DB kludges. HBase dont care. Its just a matter of how your app treats the columns. If you put repeating
20、info in column names, youre doing this. Note! Why is this so difficult for most relational database devs to grok? Because relational databases have no concept of nested entities! Youd make it a separate table in an RDBMS, which is more flexible but much more difficult to optimize. Its very difficult
21、 to talk about HBase schemas if you dont acknowledge this. You dont have to represent it this way (with a nested box) but you have to at least acknowledge it. And maybe name it. But, once you do acknowledge it, you can do some neat things. Nested entities can have attributes, some of which are ident
22、ifying. Identifying attributes make up the column qualifier (just like row keys, it could be multiple attributes mashed together) Nested entities can have attributes, some of which are identifying. Nested entities can have attributes, some of which are identifying. Non-identifying attributes are hel
23、d in the value (again, you could mash many attributes in here) Identifying attributes make up the column qualifier (just like row keys, it could be multiple attributes mashed together) Shows is nested in Band show_id is the column qualifier Other attributes are mashed into the value The column quali
24、fier is the show id. Everything else gets mashed into the value field. 1 table can have many nested entities, provided your app can tell them apart. How do you tell them apart? With prefixes . qualifier starts with: s + show_id a + album_id m + name Where can you nest entities? Knock yourself out. W
25、here can you nest entities? Knock yourself out. In columns Where can you nest entities? Knock yourself out. In columns Two levels deep in columns! Where can you nest entities? Knock yourself out. In columns Two levels deep in columns! In the row key?! Where can you nest entities? Knock yourself out.
26、 In columns Two levels deep in columns! In the row key?! Using timestamps as a dimension! This is a fundamental modeling property of HBase: nesting entities. Wait, what about Column Families? Theyre just namespaces-additional vertical sections on the same entity. Where column families arent shown ex
27、plicitly, lets assume theres just one. So that brings us to a standard way to show an HBase schema: Table: Top level entity name, fixed at design time. Row key: Can consist of multiple parts, each with a name every Answer relates to a single Applicant (by id). It also relates to a Question (not show
28、n). SQL would have you JOIN them to materialize an applicant and their answers. Example 1: a simple parent/child relationship HBase Schema: Answers By Applicant Answer is contained implicitly in Applicant If you know an applicant_id, you get O(1) access If you know an applicant_id AND question_id, O
29、(1) access Answer.created_date is implicit in the timestamp on value HBase Schema: Answers By Applicant You get answer history for free Applicant.applied_date can be implicit in the timestamp More attributes on applicant directly? Just add them! Answers are atomic and transactional by Applicant! Exa
30、mple of rows in HBase: Before you get too excited, remember that there are cons to denormalization. The cons: Nested entities arent independent any more. No: SELECT avg(value) FROM Answer WHERE question_id = 123; But you can still do that with map/reduce This decomposition only works in one directio
31、n. No relation chains without serious trickery. Timestamps are another dimension, but that counts as trickery. No way to enforce the foreign key to another table. But would you really do that in a big RDBMS? On disk, it repeats the row key for every column You didnt really save anything by having ap
32、plicant_id be implied. Or did you? Compression negates that on disk . . and prefix compression (HBASE-4218) will totally sink this. Relational Schema: Users And Messages Many-to-many relationship. One User sees many Messages, and a single Message can be seen by many Users. We want to do things like
33、show the most recent message by subject (e.g. an inbox view). Example 2: dropping some database science What kind of SQL would you run on this? Say, we want to get the most recent 20 messages for 1 user. SELECT TOP 20 M.subject, M.body FROM User_Message UM INNER JOIN Message M ON UM.message_id = M.m
34、essage_id WHERE UM.user_id = ORDER BY M.sent_date DESC Seems easy, right? Well, the database is doing some stuff behind the scenes for you: Assuming no secondary indexes, it might: Drive the join from the User_Message table For each new record with our given user_id, do a single disk access into the
35、 Message table (i.e. a hash_join) Get the records for *every* message for this user Sort them all Take the top 20 No shortcuts; we cant find the top 20 by date w/o seeing ALL messages. This gets more expensive as a user gets more messages. (But its still pretty fast if a given user has a reasonable
36、number of messages). How could you do this in HBase? Try the same pattern as parent / child? 1. Because of the many-to-many, you now have N copies of the message (one per user). Maybe thats OK (especially if its immutable!). Disk is che 母题：从资源消耗型到环境友好型开发 三亚红沙棕榈滩项目将成为海南旅游度假项目的里程碑。 长久以来，三亚的旅游度假项目一直以“阳
37、光、海滨、沙滩”为 特色。众多的同质化产品消耗了大量的稀缺性公共生态旅游资 源，同时造成了相当程度上的依赖资源型发展的低效低质无序竞 争。这种现实的矛盾随着海南新的国际旅游岛地位的确定愈加突 出。如何思考一种新型的发展理念和发展模式成为当前迫在眉睫 的问题。 三亚红沙棕榈滩项目位于三亚著名的亚龙湾西北侧田独镇附近， 左右两翼有群山环抱，大茅河从基地内部流过，每天海潮都会在 基地内大茅河下游起落。海陆交界的自然环境塑造出非常具有特 色的红树林带和非常活跃的动植物栖息地。通过几次现场踏勘和 生态环境研究，本项目的生态敏感性和独特的生态潜力正逐渐地 被解读和释放出来。 与此同时，本项目内部包含现状田
38、独镇在基地的北部，基地内曾 经是当地村民进行农业生产和盐场的经济收入所在，如何创造新 的工作就业机会，保证社会和谐，通过项目的发展带动当地的现 状改造和再生同样具有非常重要的战略意义和示范作用。 基于对环境和社会的高度责任感，业主在原控制性详细规划基础 上重新委托艾奕康（原易道公司）项目组进行本项目的概念性总 体规划，希望总体规划能够厘清现有的机遇和挑战，同时通过精 细的规划设计，创造项目的特色和与众不同的内在价值。 项目小组认为本项目的独特性在于基地具有整体的山、河、湖、 海自然特征架构下蕴含了非常多样的红树林栖息地，尺度的转换 和变化具有异常旖旎的魅力和特色。虽然滨水风景受到潮水起落 带来
39、的影响，但是感受自然的变化、创造一种新的价值观、生活 方式成为业主和项目小组的共同期待。 为此，项目组从生态环境保护和再生、乐居生活创造、景致可达 性、社会和谐共生、未来项目的市场等综合角度出发，为项目制 定了富有前瞻性的项目愿景： 重塑河湖山海，演绎乐居慢活 针对本项目水位多变、环境敏感、市场同质的项目挑战，首先对 既有的设计条件进行重新提炼和思考，进而制定了量化和可实施的 针对性规划设计策略： - 活水有源，创造滨水半岛 - 生态提升，提升环境价值 - 动静和谐，丰富两岸体验 - 混合开发，多元高尚社区 首先，在此设计策略下，通过对内河、内港的塑造，保证在水位变 化的前提下，能够最大限度地
40、保证近距离的亲水生活。同时通过在 上游设置人工湿地，引入中水，净化大茅河现有的水质，都为本项 目未来的洁净的码头水岸、沙滩、湿地、水上活动奠定了基础，同 时为形成各大面积、更多数量的自然栖息地创造了衍生的条件。各 个片区被外河、内水环绕，滨水半岛的自然地缘感，更加利于社区 不同生活方式和各具特色的个性塑造，吸引来自不同地区的人们到 达并以此为目的地，乐居慢活，充分享受生活的闲适与自然之美。 其次，精心塑造红树林和生态廊道，以创造开放的自然保护区，提 升自然栖息地与高贵生活的交汇。对高尔夫球场的生态设计要求、 对河岸、海滨红树林的保护和恢复，都能使本项目的生态容量大大 增强，寓教于居，重归自然，
41、漫步红树林、观鸟泛舟、挥杆果岭、 自然探索等生活方式，能够为国内外游客和城市居民创造出富有风 情的生态场所。 第三，在东西两岸创造不同的开发内容，形成动静相宜的分区开 发；同时在河西南北塑造老、新城镇，共建和谐的综合发展环境。 另外，对近期建设和启动区进行了更加详细的城市设计和建筑导 引，希望本规划设计方案能立足现实，重点有效的推进开发控制引 导，为项目业主和当地政府提供一系列影响深远的技术依据，以推 动和指导本地区不断走向未来。 相信在业主和当时政府的不懈努力之下，本项目将引领当地旅游度 假开发的发展，其生态、人文的创造性实践，将使自身的特色和魅 力逐渐体现，同时发挥引领作用，逐渐实现与周边
42、环境和城市的整 体融合，自身发展与蜕变，引领当地的各项事业的不断发展，逐渐 成为海南乃至中国最具影响的重要旅游度假目的地之一。 三亚红沙棕榈滩，三亚新的起点。 eXecUTIVe sUMMaRY + 项目概述 conTenTs + 目录 01020304 项目背景 PROJECT BACKGROUND 1.1 项目介绍 Project Introduction 1.2 区域背景 Regional Background 1.3 原有规划回顾 Existing Master Plan Review 1.4 基地分析 Site Analysis 1.5 关键议题 Key Issues 愿景和目标 V
43、ISION AND OBJECTIVES 2.1 项目愿景 Project Vision 2.2 定位和目标 Positioning and Objectives 规划策略 PLANNING STRATEGIES 3.1 水资源管理策略 Water Management Strategies 3.2 生态策略 Ecological Strategies 3.3 规划和发展策略 Planning and Development Strategies 总体框架 OVERALL FRAMEWORK 4.1 绿色基础设施 Green Infrastructure 4.2 发展框架 Developmen
44、t Framework 4.3 分区特征 District Identity 4.4 交通规划 Road System 050607 附录 APPENDIX 7.1 生态系统研究 Ecological Study 7.2 案例研究 Case Studies 7.3 设计过程 Design Process 一期概念设计 PHASE 01 CONCEPTUAL DESIGN 5.1 概念总平面 Concept Master Plan 5.2 开放空间系统 Open Space System 5.3 城市水环境规划策略 Water Sensitive Urban Design Strategies
45、5.4 河岸设计 Riverbank Design 5.5 红树林生态系统修复 Ecological Environment Recreation 5.6 邻里特征 Neighborhood Character 启动区导则 KICK-OFF AREA DESIGN GUIDELINES 6.1 总体设计导则 General Design Guidelines 6.2 建议建筑体量 Suggested Building Massing 6.3 开放空间特征 Open Space Features 6.4 可持续社区导则 Sustainable Community Guidelines 1.1 项目介绍 Project Introduction 1.2 区域背景 Regional Background 1.3 原有规划回顾 Existing Master Plan Review 1.4 基地分析 Site Analysis 1.5 关键议题 Key Issues 项目背景分析 PROJECT BACKGROUND 第一章
