In distributed systems, the CAP theorem provides a framework for thinking about the consistency, availability, and partition tolerance guarantees a system can provide. In their paper "FIT, a distributed database performance trandeoff", Faleiro and Abadi present a similar framework for thinking about distributed database performance.
The authors start with some intuition about distributed transactions: ones that rely on data that sits in different nodes in a distributed system. For the distributed transaction to guarantee atomicity, coordination between the participating nodes is required, The coordination offers systems designers a tradeoff choice between throughput and strong isolation. Guaranteeing strong isolation impacts the system throughput, and increasing throughput would imply allowing transactions to execute concurrently in spite of the presence of conflicts.
The authors introduce another variable, fairness, that interplays with the tradeoffs between strong isolation and throughput. The idea is that when the system is given license to selectively prioritize or delay transactions, it can improve throughput while still guaranteeing strong isolation. Instead of thinking about the tradeoff between strong isolation and throughput, the authors present the three way tradeoff between fairness, isolation, and throughput "FIT", and postulate that a system that forgoes one of them can guarantee the other two.
The authors provide some of examples of fairness play, such as "group commit" for in-memory databases, where the transaction cost is small, but the cost of writing the logs to durable storage is high and limits the throughput. In "group commit", the database accumulates log records from multiple transactions, and writes them to disk in one batch, working around the disk write bottleneck and increasing the system throughput at the cost of decreasing fairness, since the transactions can't commit until their buffered log records are flushed to disk.
Another example the authors provide is "lazy evaluation", where transactions are deferred to ensure that data dependent transactions are executed together, to amortize the cost of bringing the affected data into the processor cache and main memory across the transactions, improving throughput but decreasing fairness.
The authors categorize systems according to the interplay between fairness, isolation, and throughput, and present three classes of systems, with practical examples of each class:
The authors start with some intuition about distributed transactions: ones that rely on data that sits in different nodes in a distributed system. For the distributed transaction to guarantee atomicity, coordination between the participating nodes is required, The coordination offers systems designers a tradeoff choice between throughput and strong isolation. Guaranteeing strong isolation impacts the system throughput, and increasing throughput would imply allowing transactions to execute concurrently in spite of the presence of conflicts.
The authors introduce another variable, fairness, that interplays with the tradeoffs between strong isolation and throughput. The idea is that when the system is given license to selectively prioritize or delay transactions, it can improve throughput while still guaranteeing strong isolation. Instead of thinking about the tradeoff between strong isolation and throughput, the authors present the three way tradeoff between fairness, isolation, and throughput "FIT", and postulate that a system that forgoes one of them can guarantee the other two.
The authors provide some of examples of fairness play, such as "group commit" for in-memory databases, where the transaction cost is small, but the cost of writing the logs to durable storage is high and limits the throughput. In "group commit", the database accumulates log records from multiple transactions, and writes them to disk in one batch, working around the disk write bottleneck and increasing the system throughput at the cost of decreasing fairness, since the transactions can't commit until their buffered log records are flushed to disk.
Another example the authors provide is "lazy evaluation", where transactions are deferred to ensure that data dependent transactions are executed together, to amortize the cost of bringing the affected data into the processor cache and main memory across the transactions, improving throughput but decreasing fairness.
The authors categorize systems according to the interplay between fairness, isolation, and throughput, and present three classes of systems, with practical examples of each class:
- Ones that guarantee strong isolation and fairness at the expense of throughput
- Spanner--Google's geo-scale distributed database
- Ones that guarantee strong isolation and good throughput at the expense of fairness
- G-Store--a key value store with support for multi-key transactions
- Calvin--a database system designed to reduce the impact of coordination in distributed transactions through imposing a total order on the transactions
- Ones that guarantee good throughput and fairness at the expense of strong isolation
- Eventually consistent systems--Cassandra for example
- RAMP systems--read atomic multi partition transactions
Comments
Post a Comment