6 ms·

That's an excellent explanation, except for couple of points which I think can be misinterpreted. A distributed system with consensus will in practice provide h

by oggy 8y ago

That's an excellent explanation, except for couple of points which I think can be misinterpreted. A distributed system with consensus will in practice provide higher availability than a single-node system, because it provides fault-tolerance. In fact, fault-tolerance is the primary point of using (non-Byzantine) consensus. But you are absolutely right that a distributed system using consensus has worse availability than a distributed system with no coordination. Also, when using a system "with consensus" there is often no need to actually invoke consensus on the read side of the system, in which case you don't have to pay the throughput and latency penalties. I know you've sort of said this already, but it might be helpful to mention explicitly.

zzzcpan 8y ago
> when using a system "with consensus" there is often no need to actually invoke consensus on the read side of the system, in which case you don't have to pay the throughput and latency penalties. Doesn't matter if the system is designed for faster reads. There is still coordination, you still pay coordination overhead, including throughput and latency penalties. Without coordination, for example, you can have some part of the database relevant to the client stored directly on the client, not doing remote reads at all.
- oggy 8y ago
  You're comparing apples to oranges. If read coordination can be avoided (and it often can, even in systems with consensus), whether you stick a cache on the client or not is completely orthogonal to whether the system uses consensus for write operations.
  - zzzcpan 8y ago
    Think about it, if you read from a replica that is partitioned from the rest of the system and there is no coordination, how would the replica or the client know that the value replica returns is too old and therefore breaks strong consistency guarantee?
    - oggy 8y ago
      No, you don't get linearizability of all operations, you might not even read your earlier writes - but the whole point is that you sometimes don't need these guarantees for reads. You get a consistent snapshot read, and that's often good enough. You can get an idea of how recent the snapshot is based on timestamps, but "recent" is hard to define in a distributed system.
      - zzzcpan 8y ago
        > but the whole point is that you sometimes don't need these guarantees for reads. If that was your point, than sure. If you drop consistency, you can drop coordination too. But typically people expect reads to be consistent in consensus based systems, which requires coordination.
sagichmal 8y ago
> A distributed system with consensus will in practice provide higher availability than a single-node system, because it provides fault-tolerance. I'm not sure this is true. It protects against one class of fault (node failure) but opens you up to another (network failure). As a distsys engineer I am increasingly convinced that fault tolerance is not a good selling point for distribution, the fallacies of distributed computing are real and difficult to accommodate.
- oggy 8y ago
  That's not really true. With a single, non-replicated server, you're also very much exposed to network failures. If the server's connection goes down, you're screwed. Compare this to a Google replication setup of 2 East Cost, 2 West Coast, and 1 central USA server. The client must only be able to reach two out of three data centers (and they have to be able to communicate with each other). That sounds much more resilient to me - and I guess Google agrees, since they deploy the setup.

by oggy 8y ago

zzzcpan 8y ago

> when using a system "with consensus" there is often no need to actually invoke consensus on the read side of the system, in which case you don't have to pay the throughput and latency penalties. Doesn't matter if the system is designed for faster reads. There is still coordination, you still pay coordination overhead, including throughput and latency penalties. Without coordination, for example, you can have some part of the database relevant to the client stored directly on the client, not doing remote reads at all.

oggy 8y ago
You're comparing apples to oranges. If read coordination can be avoided (and it often can, even in systems with consensus), whether you stick a cache on the client or not is completely orthogonal to whether the system uses consensus for write operations.
- zzzcpan 8y ago
  Think about it, if you read from a replica that is partitioned from the rest of the system and there is no coordination, how would the replica or the client know that the value replica returns is too old and therefore breaks strong consistency guarantee?
  - oggy 8y ago
    No, you don't get linearizability of all operations, you might not even read your earlier writes - but the whole point is that you sometimes don't need these guarantees for reads. You get a consistent snapshot read, and that's often good enough. You can get an idea of how recent the snapshot is based on timestamps, but "recent" is hard to define in a distributed system.
    - zzzcpan 8y ago
      > but the whole point is that you sometimes don't need these guarantees for reads. If that was your point, than sure. If you drop consistency, you can drop coordination too. But typically people expect reads to be consistent in consensus based systems, which requires coordination.

sagichmal 8y ago

> A distributed system with consensus will in practice provide higher availability than a single-node system, because it provides fault-tolerance. I'm not sure this is true. It protects against one class of fault (node failure) but opens you up to another (network failure). As a distsys engineer I am increasingly convinced that fault tolerance is not a good selling point for distribution, the fallacies of distributed computing are real and difficult to accommodate.

oggy 8y ago
That's not really true. With a single, non-replicated server, you're also very much exposed to network failures. If the server's connection goes down, you're screwed. Compare this to a Google replication setup of 2 East Cost, 2 West Coast, and 1 central USA server. The client must only be able to reach two out of three data centers (and they have to be able to communicate with each other). That sounds much more resilient to me - and I guess Google agrees, since they deploy the setup.