this repo has no description
Add implementation plan
+14
+14
rfds/0004_tlog_copysets.md
+14
rfds/0004_tlog_copysets.md
···
100
100
a log will only be GC'd once all storage servers in the team have popped past a particular version.
101
101
102
102
Copysets also naturally solve the fault-tolerance issue, by their nature.
103
+
104
+
## Implementation
105
+
106
+
- In manager, always spawn TLogs in sets of `replication_factor`.
107
+
The TLog copysets are stored implicitly as each N TLogs in the list:
108
+
`[1, 2, 3, 4, 5, 6]` -> `[{1, 2, 3}, {4, 5, 6}]`
109
+
- In the Manager/Distributor, shards point to a team id directly instead of to a set of storage server ids
110
+
- In the CommitBuffer, mutations are tagged with a team id as above and then sent to the corresponding TLog team.
111
+
The TLog team for a storage team can be computed using the same mod trick, i.e. `rem(storage_team_id, length(tlog_teams))`
112
+
- Storage servers can now safely read from any TLog in their corresponding team for a given generation.
113
+
- During recovery, the Manager needs to hear from at least one TLog from every copyset in order to proceed.
114
+
- The Distributor can *only* move shards between storage teams.
115
+
Storage servers can not be added to or removed from teams; teams can only be created and destroyed.
116
+
(Note that the current implementation does not do this anyway, because this was planned.)