谷歌Jeff Dean的分布式系统设计模式
IKS workshop: semantic technology Parsing a large JSON file efficiently and easily » In a recent keynote at SOCC, Jeff Dean of Google listed a number of design patterns for system design and a number of challenges for the future. I wrote them down, and thought I might as well share them.
You can find the presentation here (it begins with something else, just skip forward a few slides, Linux users: install moonlight).
He starts off noting a shift that happened over the last 5 to 10 years: (small) devices interact with services that are provided by large data centers. This allows clients to use large bursts of computational power, such as in the case of a single Google search which runs across 1000's of servers.
Then he goes on with the typical introductions to MapReduce (with a map tile generation example) and BigTable (mentioning what's new since the paper). He also mentions Colossus (next-gen GFS) and talks about Spanner, a cross data center storage and computing system.
And then we get to the system design experiences and design patterns. Some of these are very generic, others are specifically for distributed systems. The below is just a tight summary, the actual slides and talk go into more detail on each of these.
1.Break large complex systems down into many services, with few dependencies.
Easy to test and deploy, allows lots of experiments, reimplement without affecting clients, small teams can work independently.
A single google.com search touches over 100 services.
Personal note: this is important in any design, distributed or not. It is the purpose of the module system in our Daisy/Kauri Runtime system.
2.Protocol description language is a must.
See protocol buffers.
Servers ignore tags they don't understand, but pass the information through.
Personal note: in the XML world, this is also known as the "must ignore" pattern.
3.Ability to estimate performance of a system design, without actually having to build it: do 'back of the envelope' calculations. See slide on numbers everyone should know, shown here to the right. Know your basic building blocks (understand their implementation at a high level).
4.Designing and building infrastructure: important not to try to be all things to all people, don't build infrastructure just for its own sake: identify common needs and address them.
5.Design for growth, but don't design to scale infinitely: 5 to 50 times growth good to consider, 1000 times probably requires rethink and rewrite.
6.Single master, 1000's of servers. Master orchestrates global operation of the system, but client interaction with the master is fairly minimal. Often: hot standby of master. Simpler to reason about, but scales less (1000's of workers, not 100,000's).
7.Canary requests: odd requests sometimes crash server process. When sending same request to many servers, all the servers might crash. Therefore: first send the request to one server.
8.Tree distribution of requests, to avoid many outgoing RCP requests from one server.
9.Use backup requests to minimize latency. This avoids waiting on a few slow machines when request is sent to 1000's of machines.
10.Use multiple smaller units per machine, to minimize recovery time when a machine crashes, and to have fine-grained load balancing. See the many tablets per tablet server in BigTable.
Personal note: I found this a key insight in understanding how scalable stores or indexes work in contrast to say a more traditional partitioned RDBMS setup (see earlier blog). Besides BigTable/HBase, this idea is also applied in Elastic Search and Katta.
11.Range distribution of data, not hash. Allows users to reason about, and control, locality across keys.
12.Elastic systems. Avoid overcapacity and undercapacity. Design to shrink & grow capacity. Do something reasonable in case of overload, e.g. disable certain features (reduce size of index searched, disable spelling correction tip, ...)
13.One interface, multiple implementations. E.g. in search the combination of freshness & massive size is rather impossible, therefore partition into subproblems.
14.Add sufficient monitoring/status/debugging hooks.
He ends with some challenges for the future:
1.Adaptivity in world-wide systems. Challenge: automatic, dynamic world-wide placement of data & computation to minimize latency and/or cost.
2.Building applications on top of weakly consistent storage systems. Challenge: general model of consistency choice, explained and codified. Challenge: easy to use abstractions for resolving conflicting updates to multiple versions of a piece of state.
3.Distributed system abstractions. Cf. MapReduce: are there unifying abstractions for other kinds of distributed systems problems?
You can find the presentation here (it begins with something else, just skip forward a few slides, Linux users: install moonlight).
He starts off noting a shift that happened over the last 5 to 10 years: (small) devices interact with services that are provided by large data centers. This allows clients to use large bursts of computational power, such as in the case of a single Google search which runs across 1000's of servers.
Then he goes on with the typical introductions to MapReduce (with a map tile generation example) and BigTable (mentioning what's new since the paper). He also mentions Colossus (next-gen GFS) and talks about Spanner, a cross data center storage and computing system.
And then we get to the system design experiences and design patterns. Some of these are very generic, others are specifically for distributed systems. The below is just a tight summary, the actual slides and talk go into more detail on each of these.
1.Break large complex systems down into many services, with few dependencies.
Easy to test and deploy, allows lots of experiments, reimplement without affecting clients, small teams can work independently.
A single google.com search touches over 100 services.
Personal note: this is important in any design, distributed or not. It is the purpose of the module system in our Daisy/Kauri Runtime system.
2.Protocol description language is a must.
See protocol buffers.
Servers ignore tags they don't understand, but pass the information through.
Personal note: in the XML world, this is also known as the "must ignore" pattern.
3.Ability to estimate performance of a system design, without actually having to build it: do 'back of the envelope' calculations. See slide on numbers everyone should know, shown here to the right. Know your basic building blocks (understand their implementation at a high level).
4.Designing and building infrastructure: important not to try to be all things to all people, don't build infrastructure just for its own sake: identify common needs and address them.
5.Design for growth, but don't design to scale infinitely: 5 to 50 times growth good to consider, 1000 times probably requires rethink and rewrite.
6.Single master, 1000's of servers. Master orchestrates global operation of the system, but client interaction with the master is fairly minimal. Often: hot standby of master. Simpler to reason about, but scales less (1000's of workers, not 100,000's).
7.Canary requests: odd requests sometimes crash server process. When sending same request to many servers, all the servers might crash. Therefore: first send the request to one server.
8.Tree distribution of requests, to avoid many outgoing RCP requests from one server.
9.Use backup requests to minimize latency. This avoids waiting on a few slow machines when request is sent to 1000's of machines.
10.Use multiple smaller units per machine, to minimize recovery time when a machine crashes, and to have fine-grained load balancing. See the many tablets per tablet server in BigTable.
Personal note: I found this a key insight in understanding how scalable stores or indexes work in contrast to say a more traditional partitioned RDBMS setup (see earlier blog). Besides BigTable/HBase, this idea is also applied in Elastic Search and Katta.
11.Range distribution of data, not hash. Allows users to reason about, and control, locality across keys.
12.Elastic systems. Avoid overcapacity and undercapacity. Design to shrink & grow capacity. Do something reasonable in case of overload, e.g. disable certain features (reduce size of index searched, disable spelling correction tip, ...)
13.One interface, multiple implementations. E.g. in search the combination of freshness & massive size is rather impossible, therefore partition into subproblems.
14.Add sufficient monitoring/status/debugging hooks.
He ends with some challenges for the future:
1.Adaptivity in world-wide systems. Challenge: automatic, dynamic world-wide placement of data & computation to minimize latency and/or cost.
2.Building applications on top of weakly consistent storage systems. Challenge: general model of consistency choice, explained and codified. Challenge: easy to use abstractions for resolving conflicting updates to multiple versions of a piece of state.
3.Distributed system abstractions. Cf. MapReduce: are there unifying abstractions for other kinds of distributed systems problems?