undefined | Better HN

0 pointsdadkins15y ago0 comments

Not quite. What I'm saying is that a dropped packet or a failed node are not partitions as far as the CAP theorem is considered.

A distributed system is considered available if "every request received by a non-failing node [results] in a response." It does not mean you cannot retransmit or retry.

Similarly, the consistency guarantee only requires that there exist a total order on operations. Failures are ok, as we're allowed to retransmit, retry, and otherwise tolerate faults. There is no inconsistency, nor is anything eventual.

My point is that a "temporary" partition is just a fault, and as long as the fault is shorter than the allowed response time of the system, it doesn't make a difference.

0 comments

8 comments · 2 top-level

HenryR15y ago· 4 in thread

No, dropped packets are partitions. They really are. A partitionable network is modelled as one which may fail to deliver any subset of sent messages between nodes. The Gilbert and Lynch paper makes this explicit.

The consistency guarantee requires that RW histories are compatible with some sequentially consistent history on a non-concurrent RW register. Defining a total order on operations is sufficient, I believe, but not necessary (does it matter what order two consecutive reads happened in?).

dadkinsOP15y ago

How do you explain Paxos, then? How does a dropped packet prevent the system from responding to queries? How about if I broadcast every response 10 times to everyone I know? How many packets must be dropped for the system to be considered unavailable?

HenryR15y ago

Depends on the protocol, in general.

Paxos is, fundamentally, a quorum-based system that deals with reordering of messages. It sacrifices liveness for correctness - if the proposer does not hear back from a majority of nodes (in the case of, e.g. a partition), the protocol will not complete (availability is sacrificed).

My point is not that there is a 'vital packet' in every protocol, the omission of which will cause either a lack of availability or consistency (although I can certainly design protocols that way!) - it's that for every protocol there is a network partition which causes it to be either unavailable or inconsistent. That network partition might be dropping ten messages, or just one. Retransmitting would make sense, but in real life message failures are often highly temporally correlated :(

The proof of this, by the way, is in a very famous paper by Fischer, Lynch and Patterson called "The Impossibility of Distributed Consensus With One Faulty Process". One take away is that one slow-running process can take down any protocol. It may take a few missed messages, but only a single node...

1 more reply

dadkinsOP15y ago

Actually, according to your own blog post on the subject, Gilbert and Lynch define a partition tolerance as:

    “The network will be allowed to lose arbitrarily many messages sent from one node to another”

There's a huge world of difference between the network losing arbitrarily many messages, and the definition you use elsewhere in this thread: namely, any subset of packets dropped, no matter how small, counts as a network partition.

HenryR15y ago

No, arbitrarily many doesn't automatically mean a huge amount :) This definition covers permanent partitions, but also encompasses temporary partitions which are effectively one dropped message or more. There exist protocols which will be broken by the loss of a single message. Paxos may not be one of them, but there is also a pattern of loss which will break that as well.

The theory behind all this really does hold this point up. I have another blog post with much more detail on the theory here: http://the-paper-trail.org/blog/?p=49, but I warn you it may be heavy going.

allertonm15y ago· 2 in thread

Yes, a system is available if one node doesn't respond and you can contact another. But that node will be unable to guarantee consistency.

If a node you can contact is required to guarantee consistency, there will be some times that it will have to refuse your request because other nodes are not contactable.

The author's point was that in any distributed system there is a non-zero probability of a network failure. While both clients and server nodes can retry connections, there is a non-zero probability that the problem will persist longer than your "availability agreement" allows. In that case, you have a choice - return potentially inconstent data or refuse the request.

What you seem to be arguing is that the probabilities of failure - in particular of repeated failure - while non-zero, are effectively zero. The author would disagree (as they point out, the probabilities combine exponentially as the number of nodes increase.) I think he's right and that you are wrong.

dadkinsOP15y ago

Paxos is the quintessential example of a highly available, consistent system. It is available as long as more than half of the nodes are up and able to communicate with each other. It remains consistent, regardless of the failure pattern. You really do only have to worry about a true network partition. This isn't a probabilistic argument in any sense.

allertonm15y ago

As you have yourself pointed out, Paxos will in some cases (when less than half the nodes are up) become unavailable, but remain consistent in all cases where it is available.

So it is tolerant of partition and it sacrifices availability in favour of consistency. So it is CP, not CA.

2 more replies

j / k navigate · click thread line to collapse

0 comments

8 comments · 2 top-level

HenryR15y ago· 4 in thread

dadkinsOP15y ago

HenryR15y ago

Depends on the protocol, in general.

1 more reply

dadkinsOP15y ago

Actually, according to your own blog post on the subject, Gilbert and Lynch define a partition tolerance as:

    “The network will be allowed to lose arbitrarily many messages sent from one node to another”

HenryR15y ago

allertonm15y ago· 2 in thread

Yes, a system is available if one node doesn't respond and you can contact another. But that node will be unable to guarantee consistency.

If a node you can contact is required to guarantee consistency, there will be some times that it will have to refuse your request because other nodes are not contactable.

dadkinsOP15y ago

allertonm15y ago

As you have yourself pointed out, Paxos will in some cases (when less than half the nodes are up) become unavailable, but remain consistent in all cases where it is available.

So it is tolerant of partition and it sacrifices availability in favour of consistency. So it is CP, not CA.

2 more replies

j / k navigate · click thread line to collapse