I think that map-reduce has broader applicability than just places were you need the sort, but I completely agree that other models are far better than most graph theoretic programs unless you have a problem that is susceptible to spectral methods. This last proviso applies because map-reduce can be used to compute graph spectra (aka singular values) very effectively.
On Fri, Apr 20, 2012 at 11:16 AM, Mike Spreitzer <[email protected]>wrote: > I, too, think there is not one best paradigm for everything. However, it > seems to me that Pregel-for-key-value-data, if I may call it that, > dominates map-reduce except for applications where you really need the > sorting built into map-reduce (and if you think about it, many things for > which map-reduce is being used do not really need that sorting). Aside > from that sorting, map-reduce is just a subset of what you can do with > Pregel-for-key-value-data. And sorting could be added to the richer > programming model, so it would be a strict dominance situation --- for > this particular pair of paradigms. Wouldn't you agree? > > The maintenance problem I cited is one that looks like it would really > demonstrate some of the value of the richer programming model. I would > like to try an actual comparison. To be fair, the map-reduce based > solution should not be stupid. So that's why I am asking how this > maintenance problem can best be solved using map-reduce. > > Thanks, > Mike > > > > From: Sean Owen <[email protected]> > To: [email protected] > Date: 04/20/2012 01:57 PM > Subject: Re: shortest-path maintenance > > > > Having seen both these paradigms "from the inside" I can make two > vague general comments. There is not going to be one best paradigm for > distributed computing. Some paradigms are better for some problems. > > That said, MapReduce has always been really for one type of problem: > analytics / log processing. It's great for anything shaped like that. > It's amazing how it's been abused to do more. But it's weird for a > general paradigm. Hence I generally find these newer paradigms that > are emerging better, since at least they're more purpose-built for > general purposes. > > Hadoop's what we've got easy access to in 2012. I think that will > start to change in 2013. > > Sean > > > >
