3 ms·

You cannot have at-least-once broadcast

11 points by rumcajz 11y ago

nemothekid 11y ago
While I understand the intent, I feel that is much less rigorous than the 2 problems mentioned ("You Cannot Have Exactly-Once Delivery", and CAP), both deal with the fact the network is unreliable and only that - even if you had unlimited RAM and disk, the two proofs would still hold. Here a possible solution is proposed "Send the message to Bob even if Carol is offline and store the message to Carol for later delivery.", and is really just handwaved away by saying no one has unlimited disk space (or a 24/7 ops team) (also doesn't Amazon sort of solve this problem?). I'm sure that is a practical limitation, but for the sake of building reliable systems I'm unsure its strong one. If thats the case, then one could also argue Raft also isn't linearizable, because the replicated log file could get really large and no one has unlimited disk space.
- StavrosK 11y ago
  It's just silly. You can't have any delivery to anyone, because if they're offline your message buffer might get too large for your disk.
- rumcajz 11y ago
  Why should BAP theorem take into account 24/7 human support and CAP theorem shouldn't?
brongondwana 11y ago
Also the earth could be demolished to make way for a hyperspace bypass, so you can't even have at-least-once delivery to a single recipient. It's amazing how much you can get done without perfect guarantees though.
PeterisP 11y ago
The argument is a bit ridiculous - simple logging of all messages sent also requires the same storage, and that is commonly done whenever each individual message and it's delivery is considered important. As long as the storage requirements are bounded by number of messages times number of recipients times some reasonable constant, it is perfectly feasible to assume that you will have sufficient storage; the storage requirements are finite and predictable, and it's only a matter of your choice and priorities. Acknowledging that doing X will require Y petabytes of storage and cost Z dollars may mean that you won't do it, but it's very, very different from stating that it's impossible.

11 points by rumcajz 11y ago

While I understand the intent, I feel that is much less rigorous than the 2 problems mentioned ("You Cannot Have Exactly-Once Delivery", and CAP), both deal with the fact the network is unreliable and only that - even if you had unlimited RAM and disk, the two proofs would still hold. Here a possible solution is proposed "Send the message to Bob even if Carol is offline and store the message to Carol for later delivery.", and is really just handwaved away by saying no one has unlimited disk space (or a 24/7 ops team) (also doesn't Amazon sort of solve this problem?). I'm sure that is a practical limitation, but for the sake of building reliable systems I'm unsure its strong one. If thats the case, then one could also argue Raft also isn't linearizable, because the replicated log file could get really large and no one has unlimited disk space.

StavrosK 11y ago
It's just silly. You can't have any delivery to anyone, because if they're offline your message buffer might get too large for your disk.
rumcajz 11y ago
Why should BAP theorem take into account 24/7 human support and CAP theorem shouldn't?

brongondwana 11y ago

Also the earth could be demolished to make way for a hyperspace bypass, so you can't even have at-least-once delivery to a single recipient. It's amazing how much you can get done without perfect guarantees though.

PeterisP 11y ago

The argument is a bit ridiculous - simple logging of all messages sent also requires the same storage, and that is commonly done whenever each individual message and it's delivery is considered important. As long as the storage requirements are bounded by number of messages times number of recipients times some reasonable constant, it is perfectly feasible to assume that you will have sufficient storage; the storage requirements are finite and predictable, and it's only a matter of your choice and priorities. Acknowledging that doing X will require Y petabytes of storage and cost Z dollars may mean that you won't do it, but it's very, very different from stating that it's impossible.