The Fuzzylog: A Partially Ordered Shared Log (Osdi'18)

Tango newspaper from SOSP'13, upon which they construct here, argued for the simplicity benefits of such a system. Recently Jay Kreps wrote a prissy in addition to accessible explanation of the benefits of using globally shared logs for edifice distributed systems.

Despite the simplicity it provides, maintaining a shared log alongside a system-wide total guild is impractical, because it is

1. expensive: At large scale, the unmarried sequencer volition last a bottleneck for the throughput. Also having a unmarried sequencer inwards broad expanse networks (WANs) adds to the latency prohibitively.
2. often impossible: A network sectionalization tin cutting off clients from the sequencer or a required quorum of the servers implementing the log.
3. and typically unnecessary: updates to disjoint information (e.g., dissimilar keys inwards a map) produce non take to last ordered, piece updates that demeanour upon the same information may commute because the application requires weak consistency guarantees (e.g., causal consistency).

To address this problem, this newspaper explores how to render the simplicity of a shared log without imposing a total order.

FuzzyLog

The FuzzyLog uses an expressive partial ordering API. 

1. Applications sectionalization their set down across logical information shards, such that updates against dissimilar shards are processed concurrently. 
2. When deployed across geographical regions, applications weaken consistency to avoid synchronous cross-region coordination; updates across regions fifty-fifty to the same logical information sectionalization tin come about concurrently.

Clients interact solely alongside their ain region's local re-create of the DAG; they tin modify this re-create past times appending to their ain region's chain for a color. The customer tin synchronize alongside a unmarried color, playing frontwards novel nodes inwards the local region's re-create of that color inwards a contrary topological kind guild of the DAG. A node tin last appended atomically to multiple colors, representing a transactional update across information shards.

To realize the FuzzyLog API over a collection of in-memory storage servers, the newspaper presents Dapple. Dapple scales throughput linearly past times storing each color on a dissimilar replica gear upward of servers, thence that appends to a unmarried color execute inwards a unmarried phase, piece appends that bridge colors execute inwards 2 phases (in the absence of failures) that solely involve the respective replica sets. Dapple achieves this via a fault-tolerant ordering algorithm that builds on Skeen's algorithm in addition to adds recovery protocols to it. The algorithm  provides linear scaling for single-color appends, serializable isolation for multi-color appends, in addition to failure atomicity.

In Figure 5, customer C1 sends dorsum a max timestamp of 2.2 to servers S1 in addition to S2. When a chainserver receives this message, it moves the multi-append from the pending queue to a delivery queue; it in addition to thence waits until at that spot is no other multi-append inwards the pending queue alongside a lower returned timestamp, or inwards the delivery queue alongside a lower max timestamp (i.e., no other multi-append that could conceivably last assigned a lower max timestamp). Once this status is true, the multi-append is removed from the delivery queue in addition to processed. In Figure 5, server S1 receives a stage 2 message alongside a max timestamp of 3.1 from customer C2, only does non response instantly since it previously responded to a stage 1 message from customer C1 alongside a timestamp of 2.1. Once C1 sends a stage 2 message alongside a max timestamp of 2.2, S1 knows the ordering for both outstanding multi-appends in addition to tin response to both C1 in addition to C2.

Evaluation

Published numbers for sequencers inwards fully functional systems include: roughly 200K ops/sec (CORFU), 250K ops/sec (NOPaxos), in addition to 600K ops/sec (Tango). 

They implement FuzzyLog inwards C++ in addition to evaluate to compare alongside Tango. (It seems similar at that spot is an implementation available here.) 

Discussion

The newspaper does non reference in addition to hash out a comparing alongside Kafka, only Kafka is real relevant, in addition to it is prior piece of employment past times several years, in addition to used inwards manufacture extensively. Kafka also provides a total guild per shard. I approximate 1 departure is FuzzyLog uses Skeen's algorithm to enable updates across shards. This algorithm shouldn't last likewise difficult to implement inwards Kafka either for cross shard updates. 

The utilization of Skeen's algorithm was beneficial for getting total guild implemented over replicas inwards a lightweight manner,  by employing the clients to piece of employment every bit component of the system. The master algorithm has many assumptions, similar clients don't die, replicas don't die, all the same it provides a lightweight agency of getting updates linearized across multiple shards. But in addition to thence the enquiry becomes, are nosotros punting likewise much piece of employment in addition to responsibleness to the clients. 

Skeen's algorithm is an quondam algorithm. This goes dorsum to the famous causally in addition to totally ordered communication debate from 1993. Cheriton in addition to Skeen (from Stanford University) argued for a lightweight organisation that punts the damage of causal in addition to total guild to the clients who are interested inwards it. 

Ken Birman (Cornell university) took this every bit an gear upward on to the virtual synchrony approach he has been implementing inwards his systems. He wrote a response.

Then Cheriton in addition to Skeen wrote a response to the response, in addition to Birman wrote a response to the response to the response. This postal service summarizes the starting fourth dimension 2 exchanges in addition to cuts to the gist of the issue. 

This seems founded inwards a to a greater extent than full general debate --should systems developers aim for efficiency in addition to functioning first, giving application developers total control, only leaving them to layer security accordingly, or should they apply an unknown damage to all users, making rigid semantics an indelible component of the system?

The FuzzyLog clearly takes the Cheriton & Skeen's side.

Share this post