YouTube Lecture Notes: Multi-Clouds and Consensus

Those are my notes from watching the following video: Google I/O 2009 - Transactions Across Datacenters.

URL: https://www.youtube.com/watch?v=srOgpXECblk

Those notes are not curated, with lots of open questions. Take it for what it is.

Different Kind of Consistencies

None

You can just not care, record the writes. Data to be figured-out by the reader.

Weak Consistency

Weak consistency: best try, “message in a bottle”.

• Have a cache somewhere, could evict any write at any time
• No guarantee that a write “survives”.

Eventual Consistency

Eventual consistency: 1 master (r/w), some replicas (read-only). Called Master Slave Replication (MSR).

• Writes are guaranteed to be consistent eventually. Period where stale information is expected. A write that succeeded could be lost on master crash. e.g: SQL replication

Pros / Cons: - (+) Better read availability / speed for people on other continent - (-) Read lags if done on a replica - (-) If master crash, some data loss possible. - (-) If master goes offline, read OK but write goes down as long as the master is down.

Eventual Consistency: Multi-Master Replication

Accept write from multiple places!

• (-) Eventual consistency Need to merge and reconcile writes
• (+) Compared to MSR, if master goes down can still serve reads ?? impose an ordering across all writes ??

Strong Consistency

Implemented by ??? (Paxos I guess).

Pros / Cons: - (+) Absolutely consistent. - (-) Performance degradation.

How to Achieve Consensus

Two Phase-Commit (2PC)

Write-ahead log. I don’t get how you can completely be sure something makes it in. - (-) Everything needs to be synchronous: poor throughput.

3PC allows async.

Paxos

Paxos is a consensus protocol that is similar to 2PC/3PC, but lighter. Still impacts latency quite a lot.

• No single point of failure, not single master.
• Can run multiple Paxos transactions in parallel
• It’s the only game in town -- every other consensus algorithm is either “Paxos with cruft or broken”.
• Still high latency (5ms for other schemes, 150ms Paxos -- depends how far the different participants are).
• Not that much better than 3PC, throughput is better -- that’s it.

RSS

RSS / Subscribe