| Tip |
|---|
The Framework proposed in this space (Alex Xu) is not exactly applied to propose a design : Getting started - a framework to propose... |
Part 2 - KVS in Single Server
Even with the optimisations (data compression, frequently used data in memory and rest on disk), a distributed KV store is required for big data.
Part 3 - KVS in Distributed Environment
Real-world distributed systems
Partitions cannot be avoided. We have to choose between consistency and availability !
If consistency over availability because high consistent requirements = system coud unavailable.
if availability over consistency = system keeps accepting reads even though it returns stale data.
How to store & fetch data ? (data partition)
We can use the Consistent Hashing technic. Using this technic, gives 2 advantages :
Automatic scaling: servers could be added and removed automatically depending on the load.
Heterogeneity: number of virtual nodes (or replicas) for a server is proportional to the server capacity.
How to replicate data ? (data replication)
To achieve HA and reliability, data must be replicated asynchronously over N servers where N is a configurable parameter.
The replication is useful for preserving the data in a node. For replicating a key-value, we'll fetch next 2 (or 3, depending on replication factor / number of copies) distinct server nodes from the hash ring and save the copies into these nodes.
How to handle consistency ?
Since data is replicated, it must be synchronized. Issue is configuration is a typical tradeoff between latency and consistency.
CONFIGURATION OF W (WRITE OPERATION), R (READ OPERATION) AND N SERVERS (REPLICAS OR NODES) | |
|---|---|
if W or R > 1 | better consistency; but slower because coordinator must wait for response from slowest replica or node. |
if W + R > N | strong consistency is guaranteed because one overlapping node has latest data to ensure consistency. |
if R = 1 and W = N | optimized for fast read. |
if W = 1 and R = N | optimized for fats write. |
if W + R <= N | strong consistency is NOT guaranteed |
See the consistency models - page References & Glossary for Key-Value store.
Versioning is the solution to resolve inconsistency. Versioning means treating each data modification as a new immutable version of data.
2 servers have the same value, but changes are performed at the same time in the both servers (conflicting values).
Example of increment when there is a conflicting value, increment with a vector block.
if D3([Sx,2],[Sy,1]) et D4([Sx,2],[Sz,1]), then
then, D5([Sx,3],[Sy,1],[Sz,1]) → Value reconciled and written by Sx.