this repo has no description
Editing overview
+9
-11
+9
-11
internals/overview.md
+9
-11
internals/overview.md
···
3
3
Hobbes is a key/value database with a service architecture inspired by FoundationDB.
4
4
The database is unbundled into a number of independent roles which communicate and interoperate.
5
5
6
-
These roles can be organized into three main subsystems:
6
+
These roles can be organized into three subsystems:
7
7
8
8
- The **Control Plane**, which performs distributed consensus (i.e. Paxos)
9
9
- The **Transaction Plane**, which provides transactional support to the database
10
-
- The **Storage Plane**, which stores key/value data for the cluster and serves multiversion reads
10
+
- The **Storage Plane**, which stores key/value data for the cluster and serves versioned reads
11
11
12
12
Each plane is responsible for managing the next.
13
13
14
-
The Control Plane consists of a consensus cluster of Coordinator servers.
14
+
The Control Plane consists of a number of Coordinator servers.
15
15
The Coordinators have known, static names and serve as the entry point of the cluster.
16
-
They perform fault-tolerant distributed consensus, specifically Viewstamped Replication (similar to MultiPaxos or Raft).
16
+
They execute fault-tolerant distributed consensus, specifically Viewstamped Replication (similar to MultiPaxos or Raft).
17
17
18
-
The Transaction Plane consists of several unique roles which work together to provide transactional guarantees like consistency and atomicity.
18
+
The Transaction Plane consists of a number of servers across several unique roles which work together to provide transactional guarantees like consistency and atomicity.
19
19
The Transaction Plane is generational:
20
-
if *any* member fails or is unresponsive, the entire system is destroyed and a new generation is created to replace it.
21
-
The creation of a new generation is kicked off by the Control Plane;
22
-
as the Control Plane is implicitly fault-tolerant (distributed consensus), it therefore confers that fault-tolerance onto the Transaction Plane.
20
+
if *any* member fails or is unresponsive, all of its servers are destroyed and replaced with a new generation in a process called **recovery**.
21
+
The Control Plane is responsible for kicking off the recovery process; it therefore confers its fault tolerance onto the Transaction Plane.
23
22
24
-
The Storage Plane consists of a number of Storage servers which store the actual key/value data.
23
+
The Storage Plane consists of a number of Storage servers which store key/value data and serve versioned reads.
25
24
Data is split into contiguous (ordered by key) shards and distributed (and replicated) amongst Storage servers.
26
25
If a Storage server fails or becomes unresponsive,
27
-
the Transaction Plane will issue commands to re-replicate its shards to restore fault-tolerance.
28
-
Because the Transaction Plane is fault-tolerant as explained above, it therefore confers that fault-tolerance into the Storage Plane.
26
+
the Transaction Plane will issue commands to re-replicate its shards to restore fault tolerance.