Transcript
Replication Enhancements inRavenDB 3.5
Jonathan MatheusRavenHQ
jmatheus@ravenhq.com
What we doWe’re the official cloud hosting provider for RavenDB• Simplify getting RavenDB up and running in
production in AWS & Azure• Commoditize dev-ops and support for
RavenDB• Partnership between Nuvem Nine Consulting
& Hibernating Rhino’s founders• Tight-knitted with Raven product team
How Replication Works• All replication is 1-way• 2-way replication is 2 x 1-way replication
setup• Configure replication on each node• PITA for large number of nodes
• Each node tracks it’s replication to destinations
3.0 Failover• Primary goes down... Boo!• Reads will failover to a Secondary… FTW!• Wait, what about writes?
• Optionally enable writes to Secondary• docStore.Conventions.FailoverBehavior• Changes will be replicated to Primary when it
recovers
3.0 Failover - Recovery • Primary comes back up• Whoo-hoo!
• Let’s use it again!• Write against pre-failover
data• Conflicts!
• “Recovery Jitter”
3.0 Failover - Recovery • Jitter likely to happen when:• High rate of change• Long outages
Clustering - Goals• Reduce Jitter during failover recovery• Introduce Leader Election using Raft
• Simplify Ops Setup & Management• Shared setup and configuration of dbs and
replication
Clustering - Replication• Clustering does not change how replication
works• Clustering changes how failover works• Write only to the leader by default
Clustering - Leader Election• Write only to the Leader in the best case
scenario• Old Leader goes down, choose a new Leader• Old leader recovers & becomes a follower
Clustering - How’s A Leader Elected?• Use Raft!• Consensus Algorithm• Majority vote• At least 3 members
demo: http://raft.github.io
Clustering - Things to know• Leader election is server-wide• Not by database
• ClusterBehavior• Read - Leader or All• Write - Leader or Leader w/ Failover
Selective Replication• Pick collections & transform• More like an ETL Export• Not used for failover• No conflict detection
Selective Replication• Prefer fewer replication “endpoints”• Perhaps combine transform files for an endpoint
during deployments• Use conventions to replace properties (lodash)
Failover SLAs• Failover when responses are slow• Weighted• For reads only
top related