You Can't Sacrifice Partition Tolerance (response to Stonebraker)

http://codahale.com/you-cant-sacrifice-partition-tolerance/#errata10221010

by benblack • 16 years ago

49 13 16 years ago

tbrownaw2• 16 years ago

Every failed Google search means fewer ads served and advertisers charged

No, I just hit refresh a couple times or try again later. Same number of (successful) pageviews, same number of ad impressions.

every item a user can’t add to their shopping cart means fewer items sold

No, I try again later. Just like I do if something is out-of-stock.

every unprocessed credit charge risks a regulatory fine

So what just happened to that 24 hour window you mentioned for banking transactions?

The choice of availability over consistency is a business choice, not a technical one.

Right. Because it's safer for Citibank to approve a transaction that they shouldn't, than for them to process it five minutes late.

Given this economic context, it becomes clear why most practitioners at any interesting scale meet their business needs using highly-available, eventually consistent systems.

I thought it was because consistency usually costs a bit of performance, and at "interesting scale" that translates into significant hardware costs. Plus of course "interesting scale" apparently comes with "rockstar programmers", who prefer eventual consistency because it's a fun exercise in minimalism.

aaronblohowiak• 16 years ago

You do not exhibit the same behavior as users in aggregate. However, the question of Opportunity Cost in development vs marginal cost of constructing ca systems is open for debate and situationally dependent.

johnrice• 16 years ago

Stonebraker says real systems can obviously achieve "CA" when one node is partitioned off. But he slyly switched out the theorem's definition of "available" for a more colloquial definition. The CAP theorem definition requires every live node, including the lone segregated guy, to respond to clients. Real "CA" systems don't do that--they're "available" because the non-partitioned nodes are still responding.

tbrownaw2• 16 years ago

The CAP theorem definition requires every live node, including the lone segregated guy, to respond to clients.

It also models nuked (hardware failure, power outage, whatever) nodes as being in an isolated single-node partition. Which becomes rather interesting in combination with that definition of availability.

RickHull• 16 years ago

> Real "CA" systems don't do that--they're "available" because the non-partitioned nodes are still responding.

This sounds like a fencepost type of error to me. When you have a partition, you are splitting a single set into two subsets. It may be live/dead or live-net1/live-net2. The point is, in the presence of a partition, there are no non-partitioned nodes.

foobarbazetc2• 16 years ago

"Even the canonical example of an isolated transaction—a transfer of funds between bank accounts—happens with a 24-hour window of indeterminacy. In fact, one of the few financial transactions which actually resembles a database transaction is the physical exchange of cash for goods—a totally analog experience."

With all due respect, that's a bunch of bullshit, and is a great example of passing off how you think something works as how it does work.

My bank transfers between accounts happen instantly. Buying stuff online isn't an analog experience, and happens instantly.

Let's not put banking into the same category as a Twitter-like app, because they're nothing alike.

It's perfectly okay that two different solutions exist for two different sets of problems, and there's no real need to argue that your world view applies to all problems.

moonpolysoft3• 16 years ago

What is this wonderful bank you use where transfers between accounts at a different bank happen instantaneously? Every time I've tried it the process takes 1-2 days and most likely has to go through an ACH.

As for credit and debit charges, my bank always has two transaction histories. Cleared transactions and pending transactions. Pending transactions are merely funds that have been reserved against your account and they may or may not clear after a period of time. Sounds like that fits pretty well with eventual consistency.

foobarbazetc2• 16 years ago

Love the down votes -- he didn't say different bank, did he? Funnily enough, money also moves between accounts at the same bank.

My bank transfers to other banks take a couple of hours, but I don't live in the US. The money comes out of the account instantly though.

I think it's a bit of a stretch to say that a bank doesn't need transactions.

As for credit cards, that's great, but that doesn't address what he said, which is that it's an analog experience, which it isn't. Also, it depends on the bank, but there's generally an available balance and a credit limit. Credit card payments come out of the available balance instantly. I'm pretty sure that's running in a transaction. :)

What happens between then and actual acquisition is irrelevant.

Anyway, the point isn't the technical details, or how implementations differ. The point is that the argument presented is flat out wrong (in this case). Banks are a completely different beast to your web 2.0 startup. :)

riffraff2• 16 years ago

did anyone else notice we have 3 posts a week about the cap theorem? Maybe it's about time we stop talking about it and start using some other terminology.

moonpolysoft3• 16 years ago

Your right. All this talk of distributed systems is making hacker news too nerdy. Lets go rate someone's startup and find a technical cofounder. Perhaps we might also watch a video of a prominent investor trying to promote his personal brand.

riffraff2• 16 years ago

you have misunderstood my point, which is that there is in my opinion too much time spent trying to understand what X means when using a permutation of the letters C,A,P,L,whatever, and we are not seeing enough of the nerdy stuff (such as: new coordination protocols, failure detectors, workflow patterns).

You Can't Sacrifice Partition Tolerance (response to Stonebraker)

Related Stories