发帖
雷达卡
- put(key, value)
- get(key)
1 Single Server key-value store
# Problem: can't fit everything in memory
Solutions:
- Compress the data
- Store the data on disk and only put frequently used data in hash table
2 Distributed Key-Value store
#1 CAP theorem
refer to previous notes.
#2 Design Goals
- Ability to store "big data" (> 1T)
- High availability: always respond quickly, even during failures
- Scalability
- Heterogeneity: the work distribution must be proportional to the capability of individual servers.
- Tunable tradeoffs between consistency and latency
- Low latency read and write operations (On average < 10 ms for reads, < 20 ms for writes)
- Comprehensible conflict resolution
- Robust failure detection and resolution techniques
#3 System Architecture
- Data Partition
- Goals:
- Goals:
- minimize data movement when nodes are added or removed.
- Consistent hashing:
- Approach:
- a key is hashed into the same ring, and it is stored in the first server that it encounters while traveling in clockwise direction.
- Advantages for consistent hashing
- Heterogeneity: make the number of virtual nodes proportional to the server capacity
- Data Replication: data replicated asynchronously over N servers (N is a configurable parameter, N < number of servers in the system)
- Consistency: Quorum consensus can be used to guarantee consistency for both read and write operations.
- Failure Scenarios
- Failure Scenarios
#2. learn common failure scenarios and failure resolution strategies
- Failure Detection
#2 decentralized failure detection methods like gossip protocol for inter-node communication.
京公网安备 11010802022788号
