yugabyte/yugabyte-db

Remove a peer from the Raft quorum only if the load balancer would allow a new one to be added.

Open

#1218 aperta il 17 apr 2019

Vedi su GitHub
 (2 commenti) (0 reazioni) (0 assegnatari)C (1003 fork)batch import
area/docdbhelp wantedkind/enhancementpriority/medium

Metriche repository

Star
 (8229 star)
Metriche merge PR
 (Merge medio 17g 21h) (92 PR mergiate in 30 g)

Descrizione

Jira Link: DB-2435 Currently, Raft groups independently kick out a failed peer from the quorum independently. However, adding new peers into the quorum is controlled by the master/load-balancer and throttled to limit the number of concurrent remote bootstraps.

We should make this symmetric so that a failed peer is kicked out of the quorum only when the load balancer actually has the bandwidth to add some other node to the quorum.

So, instead of a failed peer losing all the tablets that he had after 5 mins of absence, the new behavior will be to remove the failed peer from the tablets he owns in a throttled fashion. So, if he were to come back before the system has moved away all his tablets, he can start off from where he left on those tablets.

(This may require #1217 for this to be effective.)

Guida contributor