This article will introduce and analyze the design and implementation of the Raft Log module in etcd's Raft, starting from the log in the Raft consensus algorithm. The goal is to help readers better understand the implementation of etcd's Raft and provide a possible approach for implementing similar scenarios.
The Raft consensus algorithm is essentially a replicated state machine, with the goal of replicating a series of logs in the same way across a server cluster. These logs enable the servers in the cluster to reach a consistent state.
In this context, the logs refer to the Raft Log. Each node in the cluster has its own Raft Log, which is composed of a series of log entries. A log entry typically contains three fields:
It's important to note that the index of the Raft Log starts at 1, and only the leader node can create and replicate the Raft Log to follower nodes.
When a log entry is persistently stored on the majority of nodes in the cluster (e.g., 2/3, 3/5, 4/7), it is considered committed.
When a log entry is applied to the state machine, it is considered applied.
etcd raft is a Raft algorithm library written in Go, widely used in systems like etcd, Kubernetes, CockroachDB, and others.
The primary characteristic of etcd raft is that it only implements the core part of the Raft algorithm. Users must implement network transmission, disk storage, and other components involved in the Raft process themselves (although etcd provides default implementations).
Interacting with the etcd raft library is somewhat straightforward: it tells you which data needs to be persisted and which messages need to be sent to other nodes. Your responsibility is to handle the storage and network transmission processes and inform it accordingly. It doesn’t concern itself with the details of how you implement these operations; it simply processes the data you submit and, based on the Raft algorithm, tells you the next steps.
In the implementation of etcd raft’s code, this interaction model is seamlessly combined with Go's unique channel feature, making the etcd raft library truly distinctive.
In etcd raft, the main implementation of Raft Log is located in the log.go and log_unstable.go files, with the primary structures being raftLog and unstable. The unstable structure is also a field within raftLog.
etcd raft manages the logs within the algorithm by coordinating raftLog and unstable.
To simplify the discussion, this article will focus only on the processing logic of log entries, without addressing snapshot handling in etcd raft.
type raftLog struct { storage Storage unstable unstable committed uint64 applying uint64 applied uint64 }
Core fields of raftLog:
type unstable struct { entries []pb.Entry offset uint64 offsetInProgress uint64 }
Core fields of unstable:
The core fields in raftLog are straightforward and can easily be related to the implementation in the Raft paper. However, the fields in unstable might seem more abstract. The following example aims to help clarify these concepts.
Assume we already have 5 log entries persisted in our Raft Log. Now, we have 3 log entries stored in unstable, and these 3 log entries are currently being persisted. The situation is as shown below:
offset=6 indicates that the log entries at positions 0, 1, and 2 in unstable.entries correspond to positions 6 (0 6), 7 (1 6), and 8 (2 6) in the actual Raft Log. With offsetInProgress=9, we know that unstable.entries[:9-6], which includes the three log entries at positions 0, 1, and 2, are all being persisted.
The reason offset and offsetInProgress are used in unstable is that unstable does not store all Raft Log entries.
Since we are focusing only on the Raft Log processing logic, "when to interact" here refers to when etcd raft will pass the log entries that need to be persisted by the user.
etcd raft interacts with the user primarily through methods in the Node interface. The Ready method returns a channel that allows the user to receive data or instructions from etcd raft.
type raftLog struct { storage Storage unstable unstable committed uint64 applying uint64 applied uint64 }
The Ready struct received from this channel contains log entries that need processing, messages that should be sent to other nodes, the current state of the node, and more.
For our discussion on Raft Log, we only need to focus on the Entries and CommittedEntries fields:
type unstable struct { entries []pb.Entry offset uint64 offsetInProgress uint64 }
After processing the logs, messages, and other data passed through Ready, we can call the Advance method in the Node interface to inform etcd raft that we have completed its instructions, allowing it to receive and process the next Ready.
etcd raft offers an AsyncStorageWrites option, which can enhance node performance to some extent. However, we are not considering this option here.
On the user side, the focus is on handling the data in the received Ready struct. On the etcd raft side, the focus is on determining when to pass a Ready struct to the user and what actions to take afterward.
I have summarized the main methods involved in this process in the following diagram, which shows the general sequence of method calls (note that this only represents the approximate order of calls):
You can see that the entire process is a loop. Here, we'll outline the general function of these methods, and in the subsequent write-flow analysis, we’ll delve into how these methods operate on the core fields of raftLog and unstable.
There are two important points to consider here:
1. Persisted ≠ Committed
As defined initially, a log entry is considered committed only when it has been persisted by the majority of nodes in the Raft cluster. So even if we persist the Entries returned by etcd raft through Ready, these entries cannot yet be marked as committed.
However, when we call the Advance method to inform etcd raft that we have completed the persistence step, etcd raft will evaluate the persistence status across other nodes in the cluster and mark some log entries as committed. These entries are then provided to us through the CommittedEntries field of the Ready struct so we can apply them to the state machine.
Thus, when using etcd raft, the timing for marking entries as committed is managed internally, and users only need to fulfill the persistence prerequisites.
Internally, commitment is achieved by calling the raftLog.commitTo method, which updates raftLog.committed, corresponding to the commitIndex in the Raft paper.
2. Committed ≠ Applied
After the raftLog.commitTo method is called within etcd raft, the log entries up to the raft.committed index are considered committed. However, entries with indices in the range lastApplied < index <= committedIndex have not yet been applied to the state machine. etcd raft will return these committed but un-applied entries in the CommittedEntries field of Ready, allowing us to apply them to the state machine. Once we call Advance, etcd raft will mark these entries as applied.
The timing for marking entries as applied is also handled internally in etcd raft; users only need to apply the committed entries from Ready to the state machine.
Another subtle point is that, in Raft, only the Leader can commit entries, but all nodes can apply them.
Here, we’ll connect all the previously discussed concepts by analyzing the flow of etcd raft as it handles a write request.
To discuss a more general scenario, we’ll start with a Raft Log that has already committed and applied three log entries.
In the illustration, green represents raftLog fields and the stored log entries in Storage, while red represents unstable fields and the unpersisted log entries stored in entries.
Since we have committed and applied three log entries, both committed and applied are set to 3. The applying field holds the index of the highest log entry from the previous application, which is also 3 in this case.
At this point, no requests have been initiated, so unstable.entries is empty. The next log index in the Raft Log is 4, making offset 4. Since no logs are currently being persisted, offsetInProgress is also set to 4.
Now, we initiate a request to append two log entries to the Raft Log.
As shown in the illustration, the appended log entries are stored in unstable.entries. At this stage, no changes are made to the index values recorded in the core fields.
Remember the HasReady method? HasReady checks if there are unpersisted log entries and, if so, returns true.
The logic for determining the presence of unpersisted log entries is based on whether the length of unstable.entries[offsetInProgress-offset:] is greater than 0. Clearly, in our case:
type raftLog struct { storage Storage unstable unstable committed uint64 applying uint64 applied uint64 }
indicating that there are two unpersisted log entries, so HasReady returns true.
The purpose of readyWithoutAccept is to create the Ready struct to be returned to the user. Since we have two unpersisted log entries, readyWithoutAccept will include these two log entries in the Entries field of the returned Ready.
acceptReady is called after the Ready struct is passed to the user.
acceptReady updates the index of log entries that are in the process of being persisted to 6, meaning that log entries within the range [4, 6) are now marked as being persisted.
After the user persists the Entries in Ready, they call Node.Advance to notify etcd raft. Then, etcd raft can execute the "callback" created in acceptReady.
This "callback" clears the already persisted log entries in unstable.entries, then sets offset to Storage.LastIndex 1, which is 6.
We assume that these two log entries have already been persisted by the majority of nodes in the Raft cluster, so we can mark these two log entries as committed.
Continuing with our loop, HasReady detects the presence of log entries that are committed but not yet applied, so it returns true.
readyWithoutAccept returns a Ready containing log entries (4, 5) that are committed but have not been applied to the state machine.
These entries are calculated as low, high := applying 1, committed 1, in a left-open, right-closed interval.
acceptReady then marks the log entries [4, 5] returned in Ready as being applied to the state machine.
After the user calls Node.Advance, etcd raft executes the "callback" and updates applied to 5, indicating that the log entries at index 5 and earlier have all been applied to the state machine.
This completes the processing flow for a write request. The final state is as shown below, which can be compared to the initial state.
We started with an overview of the Raft Log, gaining an understanding of its basic concepts, followed by an initial look at the etcd raft implementation. We then delved deeper into the core modules of Raft Log within etcd raft and considered important questions. Finally, we tied everything together through a complete analysis of a write request flow.
I hope this approach helps you gain a clear understanding of the etcd raft implementation and develop your own insights into the Raft Log.
That concludes this article. If there are any mistakes or questions, feel free to reach out via private message or leave a comment.
BTW, raft-foiver is a simplified version of etcd raft that I implemented, retaining all the core logic of Raft and optimized according to the process in the Raft paper. I’ll release a separate post introducing this library in the future. If you're interested, feel free to Star, Fork, or PR!
The above is the detailed content of Understanding etcds Raft Implementation: A Deep Dive into Raft Log. For more information, please follow other related articles on the PHP Chinese website!