Lab 3: Fault-tolerant Key/Value Service
Due: Apr 11 at 11:59pm
Introduction
In this lab you will build a fault-tolerant key/value storage service using your Raft library from lab 2. You key/value service will be a replicated state machine, consisting of several key/value servers that use Raft to maintain 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.
The service supports three operations: Put(key, value), Append(key, arg), and Get(key). It maintains a simple database of key/value pairs. 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. An Append to a non-existant 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 applications 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 garbage collect old log entries. Please submit each part by the respective deadline.
- This lab doesn’t require you to write much code, but you will most likely spend a substantial amount of time thinking and staring at debugging logs to figure out why your implementation doesn’t work. Debugging will be more challenging than in the Raft lab because there are more components that work asynchronously of each other. Start early.
- You should reread the extended Raft paper, in particular Sections 7 and 8. For a wider perspective, have a look at Chubby, Raft Made Live, Spanner, Zookeeper, Harp, Viewstamped Replication, and Bolosky et al.
- You are allowed to add fields to the Raft ApplyMsg, and to add fields to Raft RPCs such as AppendEntries. But be sure that your code continues to pass the Lab 2 tests.
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:
$ cd src/kvraft
$ GOPATH=<your-repo-directory>
$ export GOPATH
$ go test
...
Part A: Key/value service without log compaction
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 the Raft log.
- After calling Start(), your kvservers will need to wait for Raft to complete agreement. Commands that have been agreed upon arrive on the applyCh. You should think carefully about how to arrange your code so that it will keep reading applyCh, while PutAppend() and Get() handlers submit commands to the Raft log using Start(). It is easy to achieve deadlock between the kvserver and its Raft library.
- Your solution needs to handle the case in which a leader has called Start() for a Clerk’s RPC, but loses its leadership before the request is committed to the log. In this case you should arrange for the Clerk to re-send the request to other servers until it finds the new leader. One way to do this is for the server to detect that it has lost leadership, by noticing that a different request has appeared at the index returned by Start(), or that Raft’s term has changed. If the ex-leader is partitioned by itself, it won’t know about new leaders; but any client in the same partition won’t be able to talk to a new leader either, so it’s OK in this case for the server and client to wait indefinitely until the partition heals.
- You will probably have to modify your Clerk to remember which server turned out to be the leader for the last RPC, and send the next RPC to that server first. This will avoid wasting time searching for the leader on every RPC, which may help you pass some of the tests quickly enough.
- A kvserver should not complete a Get() RPC if it is not part of a majority (so that it does not serve stale data). A simple solution is to enter every Get() (as well as each Put() and Append()) in the Raft log. You don’t have to implement the optimization for read-only operations that is described in Section 8.
- It’s best to add locking from the start because the need to avoid deadlocks sometimes affects overall code design. Check that your code is race-free using go test -race.
In the face of unreliable connections and server failures, a Clerk may 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 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 always execute just once. Your code should pass the go test -run 3A tests.
- You will need to uniquely identify client operations to ensure that the key/value service executes each one just once.
- Your scheme for duplicate detection should free server memory quickly, for example by having each RPC imply that the client has seen the reply for its previous RPC. It’s OK to assume that a client will make only one call into a Clerk at a time.
Your code should now pass the Lab 3A tests, like this:
$ go test -run 3A Test: one client (3A) ... ... Passed -- 15.1 5 12882 2587 Test: many clients (3A) ... ... Passed -- 15.3 5 9678 3666 Test: unreliable net, many clients (3A) ... ... Passed -- 17.1 5 4306 1002 Test: concurrent append to same key, unreliable (3A) ... ... Passed -- 0.8 3 128 52 Test: progress in majority (3A) ... ... Passed -- 0.9 5 58 2 Test: no progress in minority (3A) ... ... Passed -- 1.0 5 54 3 Test: completion after heal (3A) ... ... Passed -- 1.0 5 59 3 Test: partitions, one client (3A) ... ... Passed -- 22.6 5 10576 2548 Test: partitions, many clients (3A) ... ... Passed -- 22.4 5 8404 3291 Test: restarts, one client (3A) ... ... Passed -- 19.7 5 13978 2821 Test: restarts, many clients (3A) ... ... Passed -- 19.2 5 10498 4027 Test: unreliable net, restarts, many clients (3A) ... ... Passed -- 20.5 5 4618 997 Test: restarts, partitions, many clients (3A) ... ... Passed -- 26.2 5 9816 3907 Test: unreliable net, restarts, partitions, many clients (3A) ... ... Passed -- 29.0 5 3641 708 Test: unreliable net, restarts, partitions, many clients, linearizability checks (3A) ... ... Passed -- 26.5 7 10199 997 PASS ok kvraft 237.352s
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 log compaction
As things stand now with your lab 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 Raft and kvserver to cooperate to save space: from time to time kvserver will persistently store a “snapshot” of its current state, and Raft will discard log entries that precede the snapshot. When a server restarts (or falls far behind the leader and must catch up), the server first installs a snapshot and then replays log entries from after the point at which the snapshot was created. Section 7 of the extended Raft paper outlines the scheme; you will have to design the details.
You should spend some time figuring out what the interface will be between your Raft library and your service so that your Raft library can discard log entries. Think about how your Raft will operate while storing only the tail of the log, and how it will discard old log entries. You should discard them in a way that allows the Go garbage collector to free and re-use the memory; this requires that there be no reachable references (pointers) to the discarded log entries.
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, and tell the Raft library that it has snapshotted, so that Raft can discard old log entries. If maxraftstate is -1, you do not have to snapshot.
Your raft.go probably keeps the entire log in a Go slice. Modify it so that it can be given a log index, discard the entries before that index, and continue operating while storing only log entries after that index. Make sure you pass all the Raft tests after making these changes.
Modify your kvserver so that it detects when the persisted Raft state grows too large, and then hands a snapshot to Raft and tells Raft that it can discard old log entries. Raft should save each snapshot with persister.SaveStateAndSnapshot() (don’t use files). A kvserver instance should restore the snapshot from the persister when it re-starts.
- You can test your Raft and kvserver’s ability to operate with a trimmed log, and its ability to re-start from the combination of a kvserver snapshot and persisted Raft state, by running the Lab 3A tests while artificially setting maxraftstate to 1.
- Think about when a kvserver should snapshot its state and what should be included in the snapshot. Raft must store each snapshot in the persister object using SaveStateAndSnapshot(), along with corresponding Raft state. You can read the latest stored snapshot using ReadSnapshot().
- Your kvserver must be able to detect duplicated operations in the log across checkpoints, so any state you are using to detect them must be included in the snapshots. Remember to capitalize all fields of structures stored in the snapshot.
- You are allowed to add methods to your Raft so that kvserver can manage the process of trimming the Raft log and manage kvserver snapshots.
Modify your Raft leader code to send an InstallSnapshot RPC to a follower when the leader has discarded the log entries the follower needs. When a follower receives an InstallSnapshot RPC, your Raft code will need to send the included snapshot to its kvserver. You can use the applyCh for this purpose, by adding new fields to ApplyMsg. Your solution is complete when it passes all of the Lab 3 tests.
The maxraftstate limit applies to the GOB-encoded bytes your Raft passes to persister.SaveRaftState().
- You should send the entire snapshot in a single InstallSnapshot RPC. You do not have to implement Figure 13’s offset mechanism for splitting up the snapshot.
- Make sure you pass TestSnapshotRPC before moving on to the other Snapshot tests.
- A reasonable amount of time to take for the Lab 3 tests is 400 seconds of real time and 700 seconds of CPU time. Further, go test -run TestSnapshotSize should take less than 20 seconds of real time.
Your code should pass the 3B tests (as in the example here) as well as the 3A tests.
$ go test -run 3B Test: InstallSnapshot RPC (3B) ... ... Passed -- 1.5 3 163 63 Test: snapshot size is reasonable (3B) ... ... Passed -- 0.4 3 2407 800 Test: restarts, snapshots, one client (3B) ... ... Passed -- 19.2 5 123372 24718 Test: restarts, snapshots, many clients (3B) ... ... Passed -- 18.9 5 127387 58305 Test: unreliable net, snapshots, many clients (3B) ... ... Passed -- 16.3 5 4485 1053 Test: unreliable net, restarts, snapshots, many clients (3B) ... ... Passed -- 20.7 5 4802 1005 Test: unreliable net, restarts, partitions, snapshots, many clients (3B) ... ... Passed -- 27.1 5 3281 535 Test: unreliable net, restarts, partitions, snapshots, many clients, linearizability checks (3B) ... ... Passed -- 25.0 7 11344 748 PASS ok kvraft 129.114s
Please post questions on Piazza.