Great job sleuthing. Originally repair did not have a -pr. When you run the standard repair the node compares it's data with its neighbours and vice versa. They also send each other updates. Since you are supposed to repair every node < gc_grace submitting a full repair to each node would create duplicated work, since a repair on node A has an effect on node B and node C.
If you want to understand this some more you should run compactionstats and netstats across your cluster while a repair is going on, then you can see what effect the commands have on other nodes. I will try to write up some documentation on it as well because -pr is a nice feature. Many may not even be expressly aware of it. On Sat, Jul 14, 2012 at 2:00 PM, Michael Theroux <mthero...@yahoo.com> wrote: > Hello, > > I'm looking at nodetool repair with the "-pr", vs. non "-pr" option. Looking > around, I'm seeing a lot of conflicting information out there. Almost > universally, the recommendation is to run nodetool repair with the "-pr" for > any day-to-day maintenance. > > This is my understanding of how it works. I appreciate any corrections to my > misinformation. > > nodetool repair -pr > > - This performs a repair on the "primary range" of the node. The primary > range is essentially the part of the ring that the node is responsible for. > When this command is run, synchronization of replicas will occur for the rows > that this node is responsible for. If replicas are missing from that node's > neighbors for those rows, they will be replicated. > > nodetool repair > > - This is where I see a lot of conflicting information. I see a lot of > answers in which there is a suggestion that this command will perform a > repair across the entire cluster. However, I don't believe this is true from > my observations (and some of the items I read seems to agree with this). > Instead, this command performs synchronization of your primary range, but > also for other ranges that this node maybe responsible for in a replica > capacity. The way I'm thinking about it is that the -pr option causes > repairs to push information from its primary range to replicas. Without -pr, > nodetool replair does a push, and pull for its neighbors that this node maybe > a replica for. This makes sense to me, as people recommend running nodetool > repair after a node has been down. This is to allow the downed node to get > any missed information that should have been replicated to it while it was > down. > > I'm sure there lots of flaws to the above understanding as I'm cobbling it > together. I appreciate the feedback, > > -Mike
