eBay thinks MPP DBMS clobber MapReduce
I talked with Oliver Ratzesberger and his team at eBay last week, who I already knew to be MapReduce non-fans. This time I added more detail.
Oliver believes that, on the whole, MapReduce is 6-8X slower than native functionality in an MPP DBMS, and hence should only be used sporadically. This view is based on part on simulations eBay ran of the Terasort benchmark. On 72 Teradata nodes or 96 lower-powered nodes running another (currently unnamed, as per yet another of my PR fire drills) MPP DBMS, a simulation of Terasort executed in 78 and 120 secs respectively, which is very comparable to the times Google and Yahoo got on 1000 nodes or more.
And by the way, if you use many fewer nodes, you also consume much less floor space or electric power.
Comments
11 Responses to “eBay thinks MPP DBMS clobber MapReduce”
Leave a Reply
Yet you don’t see eBay or Teradata publishing a Terasort benchmark of their own for some reason.
Also, while mapreduce might be 6-8 times slower, if it ends up being a hundred times cheaper, and it scales linearly as you add hardware, then it starts to not matter don’t you think? Expect as you said for floor and rack space.
Would be interesting to see some solid, unbiased numbers on how much a Hadoop node costs, TOC, as comapred to a Teradata node. I’m not sure what eBay is paying for their Teradata nodes, but Teradata list price is easily 100 times marked up from base hardware cost.
I think there are really three interplaying factors here:
1: Cost of actual hardware/footprint as it compares to software licensing cost
2: Pros/cons of distributed db’s as comapred to map/reduce infrastructures
3: ease of use, target market, target user
Purchase and maintenance prices for hardware are no longer dispositive if viewed in isolation from power and floor space costs.
eBay talked about that to me, and it’s also in the link above.
Putting a big markup on hardware because you add your software doesn’t add to the power consumption.
Perfectly possible to back in to the break even point given the inputs.
Assume parity of functionality between the two systems and linear scalability.
Assume that the open source Map/Reduce system (Hadoop) requires 8 times the compute resources on identical hardware to solve the same problem as the proprietary (Teradata)
Let N = #nodes
Let C = software cost
Let K = fully loaded cost to operate a node for a year (power, rack maintenance )
Parity is solved by
C * N + KN = 8KN, N’s reduce
C+K=7K
Plug in $200,000 a node for Teradata
The cost to operate a machine has to be around $28,000/year before the teradata solution is cost effective.
If you assume a fully loaded total cost to operate commodity hardware of around $1000/year (I’m not sure this is correct but it is likely ballpark)
then Teradata needs to come down to about $7000/node to break even with the open source solution.
Whoops left out the hardware purchase cost
Assume parity of functionality between the two systems and linear scalability.
Assume that the open source Map/Reduce system (Hadoop) requires 8 times the compute resources on identical hardware to solve the same problem as the proprietary (Teradata)
Assume nodes cost $3000
Let N = #nodes in the Teradata cluster
Let C = software cost
Let K = fully loaded cost to operate a node for a year (power, rack maintenance )
Parity is solved by
C * N + KN = 8KN + $3000*8N
N’s reduce
C = 7K +$24,000
Plug in $200,000 a node for Teradata
The cost to operate a machine has to be around $25,000/year before the teradata solution is cost effective.
If you assume a fully loaded total cost to operate commodity hardware of around $1000/year (I’m not sure this is correct but it is likely ballpark)
then Teradata needs to come down to about $31,000/node to break even with the open source solution.
@UnHolyGuy – stop it with all of the facts. I prefer to read that MapReduce is bad because a big Teradata customer says so.
If I understood correctly, eBay told me that it costs $10-20K/year to operate a node. One data point was that power tends to run 3X what you’d think it would just looking at a device’s power rating.
That said, I think the real competitor at eBay for MapReduce isn’t Teradata but rather a much cheaper MPP database management system alternative, whose vendor’s executives I’m quite displeased with at the moment because of the stupid games they’re playing around yea or nay having their name mentioned.
CAM
@Curt – have vendors such as Teradata become more open about the cost of their systems (per node, per GB, …) given the availability of open-source alternatives?
Mark,
I think the driver was doing business over the Web. Large vendors started publishing very clear price sheets, just in case somebody wanted to place an online order. Perhaps in response to this transparency, other vendors did the same thing, even if they didn’t take online orders. I’ve certainly posted references to detailed price sheets from Oracle and Teradata.
Smaller vendors are more circumspect. In some cases I’ve had to beat them up pretty hard to get any clarity.
It’s also a little bit academic, as there are huge quantity discounts for much of this stuff. I’m glad Vertica (posted a while ago) and Greenplum (coming soon) gave me permission to post their list prices, but frankly those are not very precise guides to the actual cost of buying a Vertica or Greenplum system.
Unless you are running your servers on uranium power packs, power does not cost $20K a year for a $3K server.
However, the fully loaded cost might indeed come close to that.
“Fully loaded” is such a hard thing to come to grips with however. Certainly 1000 exactly idenitical, low availabilty, very simple hadoop servers are going to be a lot less expensive then 1000 on-offs.
I think another one of the hidden values to the map/reduce architecture is the extreme simplicity of the whole stack, and lack of requirement for any kind of uptime requirements on an individual server. Something goes wrong, kill the server, replace with a freshly minted copy, at your leisure.
However, in all fairness, current map/reduce infrastrucutres offer NOTHING like the ease of use and interoperability that something like a distributed database offers.
My opinion as follows:
1: Map/Reduce and distributed databases are synergistic not competitive technologies and will converge
2: Th price point on a distributed database node can not stay above $30K, anymore then Sun could continue to charge $200K for an 8proc E4500
3: Pricing per core is kind of dumb in general, price point should drive off data under management.
[…] segments. (For the record, I probably put more weight on that reason than Aster itself does.) eBay seems even more negative on MapReduce, if that is possible, than it previously was. Also, I gathered I should talk with Hadoop-centric […]
[…] eBay’s prior skepticism about MapReduce, it is quoted saying nice things in a Cloudera press release, and has apparently become quite a […]