Why ScambleDB ?

ScambleDB should be a framework to hide the complexity of deployement of complexe database architetures

Simplicity

The stack should be control by very limited set of parameters

• mem_pct memory dedicated to the service per node
• io_pct trigger when an instance reach pct capacity
• cpu_pct trigger when an instance reach pct cpu capacity
• price

Availability

• Election for single ressource in the cluster in case of failure (memcache tarantool)
• Automatic master failover cf MHA
• Upgrade parameter with rolling restart and switch over
• Upgrade software with rolling restart and switch over
• Upgrade vm instances with rolling restart and switch over

Flexibility

• Can force per query routing decision to slaves
• Can force per query routing decision to nosql slaves
• Can force per query routing decision to analitic slaves
• Can force per table group routing decision

Auto scale up

Rules:

• Increase Memory capacity until we reach the instances limit
• Monitor io until we reach instance limit
• Increase Memory by provisioning better instances
• Increase io capacity

Actions:

• Presente a set of tables and queries that can be replicated to dedicated pool of slaves set replication filter to shard the data per table set or per db
• Present a set of tables and queries canditate for horizontal sharding partition by range/hash set a replication with spider storage engine per table route the queries to that pool of slave

Auto scale out

Rules:

• Provision more slaves when cpu_pct is reach

Actions:

• Presente queries that can be routed to memory store Memcache route
• Presente queries that can be routed to persistent store Tarantool route
• Presente queries that can be routed to material views to limit joins (cf sphinx RT indexes) Provision replication to sphinx query rewrite
• Presente queries that can be routed to column store DB Provision replication via delta load cf (tungsten replicator)

Security

• Public cloud identity should be keeped secret and cloud provisioning should be engage only under your own control

Billing metrics

Scrambledb should provide inovation in database invoice framework

What makes a fair price should be monitored first : SQL is a powefull language in few words you can parse billions of records

Number of queries does not matter

• Record size * records reads just make sense as it relate to CPU usage and good modeling
• Number of disque iops as it is directly related to DB size
• Bandwidth send/receive