-
1. Re: Partitioning type
rvansa Dec 8, 2016 3:41 AM (in response to vladsz83)1 of 1 people found this helpfulSplit brain is an 'extra feature' on top of regular cache resilience, therefore the situations that can be interpreted as node crash - anytime the cluster does not lose all copies of data - are considered a node crash. An exception is if the cluster loses at least half of its members (even if it's lucky enough that it still contains at least one copy of each segment) - then there's the possibility that this cluster is actually the partition that "crashed".
When the node that was considered crashed rejoins, it wipes out all the (possibly stale) data and gets a fresh copy.
-
2. Re: Partitioning type
vladsz83 Dec 8, 2016 6:06 AM (in response to rvansa)Thanks, Radim!
But if the partition nevertheless happenes and the cluster enters 'degradated' mode, after getting healed, do the nodes only merge their state? In that case no cache copies/segments can be wiped. Am I right?
-
3. Re: Partitioning type
rvansa Dec 8, 2016 6:31 AM (in response to vladsz83)Yes. Upon split-brain, each *partition* either becomes degraded or stays available, and there can be at most one available partition. In degraded partition, no modifications can happen (1) and the partition won't rebalance. When a degraded partition joins available partition, degraded one wipes all its data. If degraded partition joins degraded partition, they merge the state and check if they contain enough nodes & data to become available.
(1) unless this partition contains all copies of given segment
-
4. Re: Partitioning type
vladsz83 Dec 8, 2016 10:18 AM (in response to rvansa)Understood. Thank you.