6.824: Distributed Systems Spring 2020

课程链接：http://nil.csail.mit.edu/6.824/2020

GitHub链接：https://github.com/fkuner/6.824

Lecture

Lecture1 MapReduce

Lecture2 RPC and Threads

1. I/O Concurrency
2. Parallelism
3. Convenience

race、coordination

web crawler execrise

Lecture3 GFS

Master Data:

1. filename -> array of chunk handles
2. handle ->list of chunk servers
   ->version
   ->primary
   ->lease expiration
3. log, checkpoint

READ：

1. name, offset-master
2. M sends the chunk handle, list of servers which is cached by the client
3. client- CS and CS return data to the client

WRITE:

NO PRIMARY -> find up-to-date replicas -> pick primary, secondary -> increment v -> tells p, s -> master writes VA to disk

Lecture4 Primary-Backup Replication

Failures:

1. fail-stop faults √
2. bugs ×

State transfer (memory)
Replicated State Machine (operation)

Non-def events:

1. Inputs-packet-data + interrupt
2. Weird instructions
3. Multicore

Log entry: instruction #, type, data

Lecture5 Go Threads and Raft

condition variables:

mu.Lock()
// do something that might affect the condition
cond.Braodcast()
mu.Unlock()

---

mu.Lock()
while condition == false {
	cond.Wait()
}
// now condition is true, and we have the lock
mu.Unlock()

死锁：

s0.CallRequestVote, acquire the lock
s0.CallRequestVote, send RPC to s1
s1.CallRequestVote, acquire the lock
s1.CallRequestVote, send RPC to s0
s0.Handler, s1.Handler trying to acquire lock

Lecture6 Fault Tolerance Raft (1)

Leader -> 高效

ELECTION TIMER -> START ELECTION

term++, requestVote

Lecture7 Fault Tolerance Raft (2)

linearizable system ? ? ?

Lecture8 ZooKeeper

WHY ZK?

1. API general-purpose coord service, not library as raft
2. N Servers can get N performance?

ZK GUARANTEES

1. Linearizable writes: all requests that update the state of ZooKeeper are serializable and respect precedence;
2. FIFO client order: all requests from a given client are executed in the order that they were sent by the client.

底层基于ZAB

Lecture9 More Replication CRAQ

Chain Replication

Lecture10 Cloud Replicated DB Aurora

EC2、EBS、RDS

Quorum

Papers

MapReduce: Simplified Data Processing on Large Clusters

The Google File System

The Design of a Practical System for Fault-Tolerant Virtual Machines

通过主从复制，在不同的物理服务器上运行一个备份虚拟机，通过和主机执行相同的操作来保持同步。备份虚拟机被建模为确定性状态机，主从都执行相同的动作，对于不确定性操作，还必须记录事件发生的确切指令。primary VM 接收的所有输入都会通过logging channel传输给backup VM，backup VM根据日志信息进行replay，因此可以和primary VM有相同的执行，从而达到备份的目的。

In Search of an Understandable Consensus Algorithm (Extended Version)

中文翻译

Raft官网

ZooKeeper: Wait-free coordination for Internet-scale systems

the abstraction of a set of data nodes (znodes), organized according to a hierarchical name space.

znode types:

1. Regular: Clients manipulate regular znodes by creating and deleting them explicitly;
2. Clients create such znodes, and they either delete them explicitly, or let the system remove them automatically when the session that creates them terminates (deliberately or due to a failure).

Client API:

create(path, data, flags)
delete(path, version)
exists(path, watch)
getData(path, watch)
setData(path, data, version)
getChildren(path, watch)
sync(path)

ZooKeeper guarantees

1. Linearizable writes: all requests that update the state of ZooKeeper are serializable and respect prece-
   dence
2. FIFO client order: all requests from a given client are executed in the order that they were sent by the
   client.

ZooKeeper Implementation

1. Request Processor
2. Atomic Broadcast
3. Replicated Database
4. Client-Server Interactions

Object Storage on CRAQ

基于对象存储的系统提供两个原语：

1. write(objID, V): The write (update) operation stores the value V associated with object identifier objID
2. V := read(objID): The read (query) operation retrieves the value V associated with object id objID.

普通的Chain Replication（CR）维护一个链表，头结点处理写请求，尾结点处理读请求，头结点收到写请求后，会传播到下一个结点，以此类推，直到到达尾节点，此时该操作为committed，所有只有committed的值才能被读到，因此保证了强一致性

CR存在一个问题，只能在尾结点读，吞吐性不高

CRAQ相比于CR拓展了：

1. 每一个结点包含一个自增的版本号和额外的属性，指示版本号是clean还是dirty，所有版本号属性初始为clean
2. 当一个结点收到一个对象的新版本号（写操作）时，插入最新的版本号到列表中，如果该结点不是尾结点，标记这个版本号为dirty，然后向后传播；否则，标记为clean，这时这个写操作为committed，然后尾结点发送ACK给所有之前的结点
3. 当一个结点收到一个对象的ACK，该结点标记这个对象的版本号为clean，并且删除所有之前的版本号
4. 当一个结点收到一个对象的读请求时，如果对象的最新版本号是clean, 返回值；否则，该结点向尾结点请求last committed版本号，并且返回这个版本号对应的值

在Read-Mostly Workloads和Write-Heavy Workloads场景都能提供比CR更高的吞吐

CRAQ提供了3种一致性：

1. Strong Consistency
2. Eventual Consistency
3. Eventual Consistency with Maximum-Bounded Inconsistency

CRAQ采取跟CR一样的容错

除此之外，为了提升效率，Node再给后续的节点发送消息时可以用多播协议，来并发处理，缩短了线性处理的时间

Amazon Aurora: Design Considerations for High Throughput Cloud-Native Relational Databases

为了减少网络压力，提升数据库性能，Aurora将redo log下推到分布式存储层，采用Quorum协议，提供可扩展的高可用存储

Lab

Lab1、Lab2、Lab3代码在GitHub上，Lab4肝不动了，有机会再写吧！