Great explanation! For the single partition read, it makes sense to read data from only one replica. Thank you so much Ben! Jun
From: ben.sla...@instaclustr.com Date: Tue, 20 Sep 2016 05:30:43 +0000 Subject: Re: Question about replica and replication factor To: wuxiaomi...@hotmail.com CC: user@cassandra.apache.org “replica” here means “a node that has a copy of the data for a given partition”. The scenario being discussed hear is CL > 1. In this case, rather than using up network and processing capacity sending all the data from all the nodes required to meet the consistency level, Cassandra gets the full data from one replica and checksums from the others. Only if the checksums don’t match the full data does Cassandra need to get full data from all the relevant replicas. I think the other point here is, conceptually, you should think of the coordinator as splitting up any query that hits multiple partitions into a set of queries, one per partition (there might be some optimisations that make this not quite physically correct but conceptually it’s about right). Discussion such as the one you quote above tend to be considering a single partition read (which is the most common kind of read in most uses of Cassandra). CheersBen On Tue, 20 Sep 2016 at 15:18 Jun Wu <wuxiaomi...@hotmail.com> wrote: Yes, I think for my case, at least two nodes need to be contacted to get the full set of data. But another thing comes up about dynamic snitch. It's the wrapped snitch and enabled by default and it'll choose the fastest/closest node to read data from. Another post is about this.http://www.datastax.com/dev/blog/dynamic-snitching-in-cassandra-past-present-and-future The thing is why it's still emphasis only one replica to read data from. Below is from the post: To begin, let’s first answer the most obvious question: what is dynamic snitching? To understand this, we’ll first recall what a snitch does. A snitch’s function is to determine which datacenters and racks are both written to and read from. So, why would that be ‘dynamic?’ This comes into play on the read side only (there’s nothing to be done for writes since we send them all and then block to until the consistency level is achieved.) When doing reads however, Cassandra only asks one node for the actual data, and, depending on consistency level and read repair chance, it asks the remaining replicas for checksums only. This means that it has a choice of however many replicas exist to ask for the actual data, and this is where the dynamic snitch goes to work.Since only one replica is sending the full data we need, we need to chose the best possible replica to ask, since if all we get back is checksums we have nothing useful to return to the user. The dynamic snitch handles this task by monitoring the performance of reads from the various replicas and choosing the best one based on this history. Sent from my iPadOn Sep 20, 2016, at 00:03, Ben Slater <ben.sla...@instaclustr.com> wrote: If your read operation requires data from multiple partitions and the partitions are spread across multiple nodes then the coordinator has the job of contacting the multiple nodes to get the data and return to the client. So, in your scenario, if you did a select * from table (with no where clause) the coordinator would need to contact and execute a read on at least one other node to satisfy the query. CheersBen On Tue, 20 Sep 2016 at 14:50 Jun Wu <wuxiaomi...@hotmail.com> wrote: Hi Ben, Thanks for the quick response. It's clear about the example for single row/partition. However, normally data are not single row. Then for this case, I'm still confused. http://docs.datastax.com/en/cassandra/2.1/cassandra/dml/architectureClientRequestsRead_c.html The link above gives an example of 10 nodes cluster with RF = 3. But the figure and the words in the post shows that the coordinator only contact/read data from one replica, and operate read repair for the left replicas. Also, how could read accross all nodes in the cluster? Thanks! Jun From: ben.sla...@instaclustr.com Date: Tue, 20 Sep 2016 04:18:59 +0000 Subject: Re: Question about replica and replication factor To: user@cassandra.apache.org Each individual read (where a read is a single row or single partition) will read from one node (ignoring read repairs) as each partition will be contained entirely on a single node. To read the full set of data, reads would hit at least two nodes (in practice, reads would likely end up being distributed across all the nodes in your cluster). CheersBen On Tue, 20 Sep 2016 at 14:09 Jun Wu <wuxiaomi...@hotmail.com> wrote: Hi there, I have a question about the replica and replication factor. For example, I have a cluster of 6 nodes in the same data center. Replication factor RF is set to 3 and the consistency level is default 1. According to this calculator http://www.ecyrd.com/cassandracalculator/, every node will store 50% of the data. When I want to read all data from the cluster, how many nodes should I read from, 2 or 1? Is it 2, because each node has half data? But in the calculator it show 1: You are really reading from 1 node every time. Any suggestions? Thanks! Jun -- ————————Ben SlaterChief Product OfficerInstaclustr: Cassandra + Spark - Managed | Consulting | Support+61 437 929 798-- ————————Ben SlaterChief Product OfficerInstaclustr: Cassandra + Spark - Managed | Consulting | Support+61 437 929 798 -- ————————Ben SlaterChief Product OfficerInstaclustr: Cassandra + Spark - Managed | Consulting | Support+61 437 929 798