4 ms·
If the network latency is negligible, this sounds reasonable. However, the mapping of 2x IOPs to 2x cost seems tenuous to me. >Someone who wanted an interestin
by pbailis 15y ago
If the network latency is negligible, this sounds reasonable. However, the mapping of 2x IOPs to 2x cost seems tenuous to me.
>Someone who wanted an interesting research project might be able to pull back the covers on DynamoDB a bit by issuing a large number of requests, very carefully measuring the response latency curves for consistent and inconsistent reads, and looking at what model fits them best.
This is possibly interesting, but why shouldn't Amazon just provide this data? The data would useful to developers who otherwise have to guess.
And finally, why not allow users to place this cost on the write path instead? "W=3" seems like a reasonable option (given that consistent reads may be unavailable anyway).
- cperciva 15y agothe mapping of 2x IOPs to 2x cost seems tenuous to me. Well, it's not just the number of disk I/Os which increases -- you've also got the network traffic and CPU time associated with each copy of the read request you send out. Given that (a) DynamoDB is using SSDs, and (b) Amazon seems to love using slow languages, I'd guess that the CPU time is what ends up dominating their cost. why shouldn't Amazon just provide this data? Amazon is very secretive. And I can't really blame them; after all, why would they want to subsidize their competitors' development? And finally, why not allow users to place this cost on the write path instead? "W=3" seems like a reasonable option I'd guess they wanted to be able to tolerate partitions without sacrificing availability. (At least for the common case where "partition" means that nodes are completely offline and are both unable to communicate with other nodes and unable to receive incoming requests.)