The idea behind client that controls location of a data is performance, to avoid unnecessary network round-trips between nodes and unnecessary caching of backup ranges. All of this mostly is true for reads at CL.ONE and RF>1.
How it works (in our case): Our client uses describe_ring that returns ring for specified Keyspace with token ranges and replica endpoints for each range. First node in the list for a token range is a kind of a primary, others are backup replicas. The client for most requests for a single key calculates a token and connects to node that is a primary node for this token. If primary is down, next endpoint from the list of endpoints for that token range is used. This way the network load between nodes is much lower. In our case, when load balancing just rotated all nodes we've seen a 100Mbps load on a node, while with an approach above 25Mbps only. Caches on a single node is filled up with data of a primary range of that node, avoiding of caching replica ranges that also belongs to this node by RF. The downside is when primary node is not accessible, backup node has no cache for a range we're switched to. 2012/1/11 Andreas Rudolph <andreas.rudo...@spontech-spine.com> > Hi! > > ... > Again, it's probably a bad idea. > > I agree on that, now. > > Thank you. > > > Cheers > > ----------------- > Aaron Morton > Freelance Developer > @aaronmorton > http://www.thelastpickle.com > > On 11/01/2012, at 4:56 AM, Roland Gude wrote: > > ** ** > Each node in the cluster is assigned a token (can be done automatically – > but usually should not)**** > The token of a node is the start token of the partition it is responsible > for (and the token of the next node is the end token of the current tokens > partition)**** > ** ** > Assume you have the following nodes/tokens (which are usually numbers but > for the example I will use letters)**** > ** ** > N1/A**** > N2/D**** > N3/M**** > N4/X**** > ** ** > This means that N1 is responsible (primary) for [A-D)**** > N2 for [D-M)**** > N3 for [M-X)**** > And N4 for [X-A)**** > ** ** > If you have a replication factor of 1 data will go on the nodes like this: > **** > ** ** > B -> N1**** > E->N2**** > X->N4**** > ** ** > And so on**** > If you have a higher replication factor, the placement strategy decides > which node will take replicas of which partition (becoming secondary node > for that partition)**** > Simple strategy will just put the replica on the next node in the ring**** > So same example as above but RF of 2 and simple strategy:**** > ** ** > B-> N1 and N2**** > E -> N2 and N3**** > X -> N4 and N1**** > ** ** > Other strategies can factor in things like “put data in another > datacenter” or “put data in another rack” or such things.**** > ** ** > Even though the terms primary and secondary imply some means of quality or > consistency, this is not the case. If a node is responsible for a piece of > data, it will store it.**** > ** ** > ** ** > But placement of the replicas is usually only relevant for availability > reasons (i.e. disaster recovery etc.)**** > Actual location should mean nothing to most applications as you can ask > any node for the data you want and it will provide it to you (fetching it > from the responsible nodes).**** > This should be sufficient in almost all cases.**** > ** ** > So in the above example again, you can ask N3 “what data is available” and > it will tell you: B, E and X, or you could ask it “give me X” and it will > fetch it from N4 or N1 or both of them depending on consistency > configuration and return the data to you.**** > ** ** > ** ** > So actually if you use Cassandra – for the application the actual storage > location of the data should not matter. It will be available anywhere in > the cluster if it is stored on any reachable node.**** > ** ** > *Von:* Andreas Rudolph [mailto:andreas.rudo...@spontech-spine.com] > *Gesendet:* Dienstag, 10. Januar 2012 15:06 > *An:* user@cassandra.apache.org > *Betreff:* Re: AW: How to control location of data?**** > ** ** > Hi!**** > ** ** > Thank you for your last reply. I'm still wondering if I got you right...** > ** > ** ** > > ... **** > A partitioner decides into which partition a piece of data belongs**** > > Does your statement imply that the partitioner does not take any decisions > at all on the (physical) storage location? Or put another way: What do you > mean with "partition"?**** > ** ** > To quote http://wiki.apache.org/cassandra/ArchitectureInternals: "... > AbstractReplicationStrategy > controls what nodes get secondary, tertiary, etc. replicas of each key > range. Primary replica is always determined by the token ring (...)"**** > > > **** > ... **** > You can select different placement strategies and partitioners for > different keyspaces, thereby choosing known data to be stored on known > hosts.**** > This is however discouraged for various reasons – i.e. you need a lot of > knowledge about your data to keep the cluster balanced. What is your > usecase for this requirement? there is probably a more suitable solution.* > *** > **** > What we want is to partition the cluster with respect to key spaces.**** > That is we want to establish an association between nodes and key spaces > so that a node of the cluster holds data from a key space if and only if > that node is a *member* of that key space.**** > ** ** > To our knowledge Cassandra has no built-in way to specify such a > membership-relation. Therefore we thought of implementing our own replica > placement strategy until we started to assume that the partitioner had to > be replaced, too, to accomplish the task.**** > ** ** > Do you have any ideas?**** > ** ** > > > **** > *Von:* Andreas Rudolph [mailto:andreas.rudo...@spontech-spine.com] > *Gesendet:* Dienstag, 10. Januar 2012 09:53 > *An:* user@cassandra.apache.org > *Betreff:* How to control location of data?**** > **** > Hi!**** > **** > > We're evaluating Cassandra for our storage needs. One of the key benefits > we see is the online replication of the data, that is an easy way to share > data across nodes. But we have the need to precisely control on what node > group specific parts of a key space (columns/column families) are stored > on. Now we're having trouble understanding the documentation. Could anyone > help us with to find some answers to our questions?**** > > · What does the term "replica" mean: If a key is stored on exactly three > nodes in a cluster, is it correct then to say that there are three replicas > of that key or are there just two replicas (copies) and one original?**** > > · What is the relation between the Cassandra concepts "Partitioner" and > "Replica Placement Strategy"? According to documentation found on DataStax > web site and architecture internals from the Cassandra Wiki the first > storage location of a key (and its associated data) is determined by the > "Partitioner" whereas additional storage locations are defined by "Replica > Placement Strategy". I'm wondering if I could completely redefine the way > how nodes are selected to store a key by just implementing my own subclass > of AbstractReplicationStrategy and configuring that subclass into the key > space.**** > > · How can I suppress that the "Partitioner" is consulted at all to > determine what node stores a key first?**** > > · Is a key space always distributed across the whole cluster? Is it > possible to configure Cassandra in such a way that more or less freely > chosen parts of a key space (columns) are stored on arbitrarily chosen > nodes?**** > **** > Any tips would be very appreciated :-)**** > > > > >
