January 18, 2006

How Many Transactions?

Often people expect your transaction systems to be scaleable, and fast and capable of performing prodigous feats. Usually these claims reduce down to "if you're not using my favourite widget" then you are somehow of minimal stature. But even when exposed as a microdroid sales tactic, it is very hard to argue against someone that is convinced that a system just must use the ACME backend, or it's not worth spit.

One issue is transaction flow. How many transactions must you handle? Here's some numbers, suitable for business plans or just slapping on the table as if your plan depended on it. Please note that this is a fairly boring post - only read it when you need some numbers for your b-plan. I'll update it as new info comes in.

In Tokyo, 17th Jan 2006:

Earlier Wednesday, the Tokyo exchange had issued a warning it would stop trading if the system capacity limit of 4 million transactions was reached. As it reached 3.5 million about an hour before the session's close, it announced it would stop trading 20 minutes early.

According to the 2004 fourth-quarter report issued by Western Union's parent company, First Data Corp., an estimated 420 Western Union transactions occur every minute each day -- amounting to an average of seven transfers every second of the year.

Craig's stellar plotting system for e-gold's transaction flow gives us loads of 34k per day last I looked. Also Fee Income.

eBay's annual report for 2005 reported that "PayPal processed an average of approximately 929,000 transactions per day during 2004." There are contradictory numbers here and here: Total number of payments grew to 117.4 million, up 41 percent year over year and 4 percent vs. the prior quarter. This would imply 320k transactions per day, but we don't expect Paypal to be accurate in filings to the SEC.

Payment News reports:

Robin Sidel reports for the Wall St. Journal on how credit card issuers are now pursuing the market for smaller payments less than $5:

....The market for transactions valued at less than $5 accounted for $1.32 trillion in consumer spending in 2003, representing more than 400 billion transactions, according to research by TowerGroup, a unit of MasterCard International Inc. By comparison, Visa processes a total of about $2 trillion of global transactions each year.

And right from Visa's mouth:

During the busiest hour on December 23 [2005], Visa processed an average of 6,363 transaction messages each second. That's a 14 percent increase over the average of 5,546 transaction messages per second Visa processed during the peak hour on December 24, 2004. Consider that Visa's payment network, VisaNet, can process more transactions over the course of a coffee break than all the stock exchanges of the world in an entire day.

Nice quip! I'll check that out next time I get my exchange running. In closing, let's let Lynn Wheeler have the last word. He reports that the old white elephant of transaction processing, SET, performed like this:

...If even a small percentage of the 2000 transactions/sec that typically go on were to make the transition to SET, the backend processing institution would have to increase their legacy computational processing capability by three orders of magnitude. The only way that SET could ever be succesful was if it totally failed, since the backend processing couldn't build out to meet the SET computational requirements. It was somewhat satisfying to see the number of people that the thot stopped them in their tracks.

The best case demo of SET at a show a year later was on an emulated processing environment with totally idle dedicated boxes. The fastest that they could get was 30 seconds elapsed time, with essentially all of that being RSA crypto computational processing. Now imagine a real-world asymmetric environment that is getting 1000 of those a second. My statement was that a realistic backend processing configuration would require on the order of 30,000 dedicated machines just doing the RSA crypto processing.

There were then remarks from the workstation vendors on the prospect of selling that many machines, and somebody from RSA about selling that many licenses. From then on my description that SET was purely an academic toy demo, since nobody could ever take it as a serious proposal with operational characteristics like that.

private email, 14th December 2005. Note how he's talking about the mid 90's, using 10 year old figures for credit card processing networks like Visa.

Posted by iang at January 18, 2006 10:45 AM | TrackBack

Interesting. Paypal and Western Union both make about 10 transactions a second, while Visa handles two orders of magnitude more.

Posted by: Daniel A. Nagy at January 18, 2006 01:39 PM

Computational transaction costs are usually dwarfed by mental transaction costs:



Posted by: nick at January 20, 2006 04:41 PM
Post a comment

Remember personal info?

Hit preview to see your comment as it would be displayed.