etcd

etcd is a distributed key-value store which is highly available, strongly consistent, and watchable for changes. The name "etcd" was from a unix's configuration directory, "etc" and "d"istributed system.

There are two major use cases: concurrency control in the distributed system and application configuration store. For example, CoreOS Container Linux uses etcd to achieve a global semaphore which avoids that all nodes in the cluster go to reboot at the same time. Also, Kubernetes use etcd for their configuration store.

History

CoreOS released etcd in 2013. Originally, the etcd was developed to manage a cluster of CoreOS Container Linux. In 2014, Google launched the Kubernetes project and used etcd for their configuration store. In 2016, CoreOS announced etcd3 and changed their data structure from a tree model to a flat key space. In 2018, RedHat announced to acquire CoreOS, and IBM announced to acquire RedHat in the same year.

Query Compilation

Not Supported

Storage Model

N-ary Storage Model (Row/Record)

etcd physically stores data as a key-value pair. The key is consist of a 3-tuple: major, sub, type. The major contains the revision holding the key. The sub contains the identifier among the revision because a transaction can produce a single revision with multiple keys. The type is an optional and one use case is for a tombstone. The value of the key-value pair contains a delta from a previous version.

Concurrency Control

Multi-version Concurrency Control (MVCC)

etcd uses MVCC for the concurrency control. The etcd uses revision which corresponds to a version of MVCC. The revision works like a logical clock in the etcd cluster and is updated when any data modification happens. Each key-value contains two revisions which respectively represents the time when the key-value was created and the time when the key-value was updated. The etcd cluster maintains the current revision. When the mutative operation has arrived (e.g., Put, DeleteRange, Txn), the etcd assigns the revision to the data related to the operation and updates the current revision.

Checkpoints

Blocking

etcd provides a snapshot to improve the recovery speed and avoid growing logs. The etcd automatically creates a snapshot based on the number of committed transactions from the last snapshot, which is configurable, while a user can create the snapshot anytime via etcdctl command. The etcd acquires a global latch to produce a snapshot, so the high frequency for taking the snapshot degrades the performance of the database operation.

Storage Organization

Log-structured Sorted Files

etcd stores data in a log file (write-ahead log) first and periodically creates a snapshot file to avoid growing the log files. The file contains a key-value pair as the b+ tree structure sorted by a key.

Isolation Levels

Serializable

etcd provides the serializable isolation by MVCC. Since each data contains a revision, the etcd aborts or reties a transaction which contains an older revision than the revision the etcd holds.

For example, an etcd client 1 started a transaction and got {"a": 1} with a revision 1. After that, the etcd cluster updated the data with a revision 2 ({"a": 2}) due to the client 2's request. When the client 1 requests to update the data, the etcd cluster aborts or retries the request because the client 1 tries to modify the data which has an older revision than the cluster has.

Indexes

B+Tree

etcd creates a secondary in-memory btree index for keys to accelerate range operations (e.g. Range and DeleteRange). The key of the btree uses a key of data and the value of the btree is a pointer to a persistent b+tree.

Storage Architecture

Disk-oriented

etcd stores a key-value pair in a persistent disk as the b+ tree structure sorted by a key.

System Architecture

Shared-Nothing

The etcd cluster is composed of shared-nothing nodes. The cluster has one leader node, and other nodes work as followers. The leader node is determined at run-time (Raft algorithm). When the leader node receives a request, the leader takes votes against followers. If the majority of nodes agrees on the request, the leader commits the request and ask followers to commit. An etcd client does not need to know which node is a current leader. Instead, the client can send a request to any node in the cluster, and the node will forward the request to the leader node if the node is a follower.

Stored Procedures

Not Supported

Data Model

Key/Value

Data Model of etcd is a key-value pair and both a key and a value are represented as a binary array. There is no fixed size limit for the key and the value, but since there is the limit for the request size (in default, 1.5 MiB), the acceptable size of the key and the value is determined by the limit.

In addition to a key and a value, each data has following metadata: create_revision, mod_revision, version, and lease. The creation_revision stands for the creation time of the data, and the mod_revision stand for the updated time of the data. A revision works like a global counter which is incremented when any data is changed, while a version works like a local counter which is incremented when the data is changed. The older version of the data can be retrieved by specifying the revision unless the version is not compacted. The lease is used for the data which has a specific lifetime, and after the lease time elapsed, the data will be removed and not be accessible.

Query Interface

Custom API

etcd has own interface and accepts a request via a gRPC remote procedure call. The interface supports the following operations: Range, Put, DeleteRange, Txn, Compact. The Range operation gets a key-value pair or key-value pairs within a range key. It supports a travel query, so the older version data is retrieved unless it is not compacted. The Put operation creates or updates the data. The DeleteRange operation removes a key-value pair or key-value pairs within a range key. The Txn operation provides an atomic operation for multiple operations. The Compact operation removes the older version of data up to a given revision.