mirror of
https://github.com/kanidm/kanidm.git
synced 2025-02-23 20:47:01 +01:00
23cc2e7745
Fixes two major issues with replication. The first was related to server refreshes. When a server was refreshed it would retain it's server unique id. If the server had lagged and was disconnected from replication and administrator would naturally then refresh it's database. This meant that on next tombstone purge of the server, it's RUV would jump ahead causing it's refresh-supplier to now believe it was lagging (which was not the case). In the situation where a server is refreshed, we reset the servers unique replication ID which avoids the RUV having "jumps". The second issue was related to RUV trimming. A server which had older RUV entries (say from servers that have been trimmed) would "taint" and re-supply those server ID's back to nodes that wanted to trim them. This also meant that on a restart of the server, that if the node had correctly trimmed the server ID, it would be re-added in memory. This improves RUV trimming by limiting what what compare and check as a supplier to only CID's that are within the valid changelog window. This itself presented challenges with "how to determine if a server should be removed from the RUV". To achieve this we now check for "overlap" of the RUVS. If overlap isn't occurring it indicates split brain or node isolation, and replication is stopped in these cases. |
||
|---|---|---|
| .. | ||
| benches | ||
| src | ||
| build.rs | ||
| Cargo.toml | ||