CS 451/651 - Spring 2021

CS 451/651 Lab 3: Fault-tolerant Key/Value Service

Due Part A: Friday April 9 23:59

Due Part B: Friday Apri 23 23:59


Introduction

In this lab you will build a fault-tolerant key/value storage service using your Raft library from lab 2. Your key/value service will be a replicated state machine, consisting of several key/value servers that use Raft for replication. Your key/value service should continue to process client requests as long as a majority of the servers are alive and can communicate, in spite of other failures or network partitions. After Lab3, you will have implemented all parts (Clerk, Service, and Raft) shown in the diagram of Raft interactions.

The service supports three operations: Put(key, value), Append(key, arg), and Get(key). It maintains a simple database of key/value pairs. Keys and values are strings. Put() replaces the value for a particular key in the database, Append(key, arg) appends arg to key's value, and Get() fetches the current value for a key. A Get for a non-existent key should return an empty string. An Append to a non-existent key should act like Put. Each client talks to the service through a Clerk with Put/Append/Get methods. A Clerk manages RPC interactions with the servers.

Your service must provide strong consistency to application calls to the Clerk Get/Put/Append methods. Here's what we mean by strong consistency. If called one at a time, the Get/Put/Append methods should act as if the system had only one copy of its state, and each call should observe the modifications to the state implied by the preceding sequence of calls. For concurrent calls, the return values and final state must be the same as if the operations had executed one at a time in some order. Calls are concurrent if they overlap in time, for example if client X calls Clerk.Put(), then client Y calls Clerk.Append(), and then client X's call returns. Furthermore, a call must observe the effects of all calls that have completed before the call starts (so we are technically asking for linearizability).

Strong consistency is convenient for applications because it means that, informally, all clients see the same state and they all see the latest state. Providing strong consistency is relatively easy for a single server. It is harder if the service is replicated, since all servers must choose the same execution order for concurrent requests, and must avoid replying to clients using state that isn't up to date.

This lab has two parts. In part A, you will implement the service without worrying that the Raft log can grow without bound. In part B, you will implement snapshots (Section 7 in the paper), which will allow Raft to discard old log entries. Please submit each part by the respective deadline.

You should reread the extended Raft paper, in particular Sections 7 and 8. For a wider perspective, have a look at Chubby, Paxos Made Live, Spanner, Zookeeper, Harp, Viewstamped Replication, and Bolosky et al.

Start early.

Getting Started

We supply you with skeleton code and tests in src/kvraft. You will need to modify kvraft/client.go, kvraft/server.go, and perhaps kvraft/common.go.

To get up and running, execute the following commands. Don't forget the git pull to get the latest software.

$ cd ~/6.824
$ git pull
...
$ cd src/kvraft
$ go test -race
...
$

Part A: Key/value service without snapshots

Each of your key/value servers ("kvservers") will have an associated Raft peer. Clerks send Put(), Append(), and Get() RPCs to the kvserver whose associated Raft is the leader. The kvserver code submits the Put/Append/Get operation to Raft, so that the Raft log holds a sequence of Put/Append/Get operations. All of the kvservers execute operations from the Raft log in order, applying the operations to their key/value databases; the intent is for the servers to maintain identical replicas of the key/value database.

A Clerk sometimes doesn't know which kvserver is the Raft leader. If the Clerk sends an RPC to the wrong kvserver, or if it cannot reach the kvserver, the Clerk should re-try by sending to a different kvserver. If the key/value service commits the operation to its Raft log (and hence applies the operation to the key/value state machine), the leader reports the result to the Clerk by responding to its RPC. If the operation failed to commit (for example, if the leader was replaced), the server reports an error, and the Clerk retries with a different server.

Your kvservers should not directly communicate; they should only interact with each other through Raft.

Your first task is to implement a solution that works when there are no dropped messages, and no failed servers.

You'll need to add RPC-sending code to the Clerk Put/Append/Get methods in client.go, and implement PutAppend() and Get() RPC handlers in server.go. These handlers should enter an Op in the Raft log using Start(); you should fill in the Op struct definition in server.go so that it describes a Put/Append/Get operation. Each server should execute Op commands as Raft commits them, i.e. as they appear on the applyCh. An RPC handler should notice when Raft commits its Op, and then reply to the RPC.

You have completed this task when you reliably pass the first test in the test suite: "One client".

Now you should modify your solution to continue in the face of network and server failures. One problem you'll face is that a Clerk may have to send an RPC multiple times until it finds a kvserver that replies positively. If a leader fails just after committing an entry to the Raft log, the Clerk may not receive a reply, and thus may re-send the request to another leader. Each call to Clerk.Put() or Clerk.Append() should result in just a single execution, so you will have to ensure that the re-send doesn't result in the servers executing the request twice.

Add code to handle failures, and to cope with duplicate Clerk requests, including situations where the Clerk sends a request to a kvserver leader in one term, times out waiting for a reply, and re-sends the request to a new leader in another term. The request should execute just once. Your code should pass the go test -run 3A -race tests.

Your code should now pass the Lab 3A tests, like this:

$ go test -run 3A -race
Test: one client (3A) ...
  ... Passed --  15.5  5  4576  903
Test: ops complete fast enough (3A) ...
  ... Passed --  15.7  3  3022    0
Test: many clients (3A) ...
  ... Passed --  15.9  5  5884 1160
Test: unreliable net, many clients (3A) ...
  ... Passed --  19.2  5  3083  441
Test: concurrent append to same key, unreliable (3A) ...
  ... Passed --   2.5  3   218   52
Test: progress in majority (3A) ...
  ... Passed --   1.7  5   103    2
Test: no progress in minority (3A) ...
  ... Passed --   1.0  5   102    3
Test: completion after heal (3A) ...
  ... Passed --   1.2  5    70    3
Test: partitions, one client (3A) ...
  ... Passed --  23.8  5  4501  765
Test: partitions, many clients (3A) ...
  ... Passed --  23.5  5  5692  974
Test: restarts, one client (3A) ...
  ... Passed --  22.2  5  4721  908
Test: restarts, many clients (3A) ...
  ... Passed --  22.5  5  5490 1033
Test: unreliable net, restarts, many clients (3A) ...
  ... Passed --  26.5  5  3532  474
Test: restarts, partitions, many clients (3A) ...
  ... Passed --  29.7  5  6122 1060
Test: unreliable net, restarts, partitions, many clients (3A) ...
  ... Passed --  32.9  5  2967  317
Test: unreliable net, restarts, partitions, random keys, many clients (3A) ...
  ... Passed --  35.0  7  8249  746
PASS
ok  	6.824/kvraft	290.184s

The numbers after each Passed are real time in seconds, number of peers, number of RPCs sent (including client RPCs), and number of key/value operations executed (Clerk Get/Put/Append calls).

Part B: Key/value service with snapshots

As things stand now with your code, a rebooting server replays the complete Raft log in order to restore its state. However, it's not practical for a long-running server to remember the complete Raft log forever. Instead, you'll modify kvserver to cooperate with Raft to save space using Raft's Snapshot() and CondInstallSnapshot from lab 2D.

The tester passes maxraftstate to your StartKVServer(). maxraftstate indicates the maximum allowed size of your persistent Raft state in bytes (including the log, but not including snapshots). You should compare maxraftstate to persister.RaftStateSize(). Whenever your key/value server detects that the Raft state size is approaching this threshold, it should save a snapshot using Snapshot, which in turn uses persister.SaveRaftState(). If maxraftstate is -1, you do not have to snapshot. maxraftstate applies to the GOB-encoded bytes your Raft passes to persister.SaveRaftState().

Modify your kvserver so that it detects when the persisted Raft state grows too large, and then hands a snapshot to Raft. When a kvserver server restarts, it should read the snapshot from persister and restore its state from the snapshot.

Your code should pass the 3B tests (as in the example here) as well as the 3A tests (and your Raft must continue to pass the Lab 2 tests).

$ go test -run 3B -race
Test: InstallSnapshot RPC (3B) ...
  ... Passed --   4.0  3   289   63
Test: snapshot size is reasonable (3B) ...
  ... Passed --   2.6  3  2418  800
Test: ops complete fast enough (3B) ...
  ... Passed --   3.2  3  3025    0
Test: restarts, snapshots, one client (3B) ...
  ... Passed --  21.9  5 29266 5820
Test: restarts, snapshots, many clients (3B) ...
  ... Passed --  21.5  5 33115 6420
Test: unreliable net, snapshots, many clients (3B) ...
  ... Passed --  17.4  5  3233  482
Test: unreliable net, restarts, snapshots, many clients (3B) ...
  ... Passed --  22.7  5  3337  471
Test: unreliable net, restarts, partitions, snapshots, many clients (3B) ...
  ... Passed --  30.4  5  2725  274
Test: unreliable net, restarts, partitions, snapshots, random keys, many clients (3B) ...
  ... Passed --  37.7  7  8378  681
PASS
ok  	6.824/kvraft	161.538s