In this post, I’m going to take a look at a sample application that uses the Couchbase Server Multi-Cluster Aware (MCA) Java client. This client goes hand-in-hand with Couchbase’s Cross-Data Center Replication (XDCR) capabilities.
XDCR supports flexible replication of data between different clusters. XDCR is an important feature for large scale, multi-data center deployments. The references below include several on XDCR. The MCA client can automatically redirect traffic to these separate clusters based on library configuration, relieving the application developer of managing much of the complexity involved.
These and other features allow Couchbase deployments to support a range of strong high availability and disaster recovery strategies. To find out more before diving deeper into client specifics, take a look at these resources:
Intro to Couchbase HA/DR – High Availability/Disaster Recover documentation
Couchbase XDCR – XDCR documentation
XDCR/MCA webinar (including technical demo) – HA/DR webinar
Webinar tutorial – Setting up clusters step-by-step, with XDCR
A deeper look at XDCR – Deep Dive on Cross Data Center Replication (XDCR) by Denis Rosa
Setting up XDCR using Docker – Replicate NoSQL Data Between Datacenters with Couchbase XDCR by Nick Raboy
Short video on setting up XDCR using Docker – Use XDCR to Replicate NoSQL Data Between Couchbase Docker Containers – Video Tutorial by Nick Raboy
MCA Client Details in Brief
The MCA Java SDK builds on the standard Couchbase Java SDK. The core API mostly mimics that of the standard client.
- The ability to prioritize a list of clusters.
- Ways to supply rules on what constitutes a failure.
- APIs to coordinate client instances.
- A management interface.
We’ll discuss the first two using the code sample. Management happens programmatically, or via a JMX interface. This allows administrators to switch clusters in the priority list. It falls outside the scope of this post.
The Sample application
To illustrate use of the MCA client, we have a straightforward sample application. The app runs a configurable number of threads in groups. One group simply reads from the database, one writes new, random records, one reads documents, modifies them, then writes them back, and one performs a simple N1QL query. The idea is to provide a variety of loads to apply against a cluster. In this webinar, we used this app to illustrate some simple inter-cluster failover behavior.
You can find the complete code, along with build instructions and other materials on GitHub. Here’s an example invocation running against two 3-node clusters with 10 threads updating documents.
|
1 |
java -jar mca-1.0.jar -c 172.23.123.4,172.23.123.5,172.23.123.6:172.23.122.55,172.23.122.56,172.23.122.57 -b my_bucket -i user -p password -u 10 -l statistics |
Now, let’s go through the main parts of the code.
Code Walkthrough
The application code, in this case, consists of one main class and some simple helper classes. Once we have the MCA client instance, it’s used almost exactly like the regular client. I’ll just focus, then, on the client set up. The code comes from the ClusterBasics class.
Cluster Nodes and Groupings
Clusters are specified on the command line as groups of nodes separated by colons (:). Within each group, the nodes names or IP addresses are separated by commas. Implicitly, the clusters are specified in priority order. This code block breaks down that simple format into sets and creates a cluster spec for each cluster. (A cluster spec can include an id for easy reference. It wasn’t necessary here, but worthwhile to point out.)
|
1 2 3 4 5 6 7 8 |
String[] clusters = options.stringValueOf("clusters").split(":"); List<ClusterSpec> specs = new ArrayList<>(clusters.length); for (String cluster: clusters) { Set<String> nodes = Arrays.stream(cluster.split(",")).collect(Collectors.toSet()); specs.add(ClusterSpec.create(nodes)); } |
Couchbase Service Types
We’ll talk about coordination and failure monitoring in a moment. You can tune both by service type. In this example, we’re using direct key/value retrieval (ServiceType.BINARY) and N1QL queries (ServiceType.QUERY). These next two lines prepare a set of service types for use later.
|
1 2 |
serviceTypes.add(ServiceType.BINARY); serviceTypes.add(ServiceType.QUERY); |
Inter-client Coordination
Next we start to see the real difference in capabilities of this client. Coordinators manage orchestration of behavior across client instances. Currently, only isolated coordinators have been implemented, but that’s appropriate for this case anyway. More sophisticated ones are planned for the future. Still, looking at all the options, you can see even an isolated coordinator allows quite a bit of control. Here’s how to create one with some non-default options.
|
1 2 3 4 5 6 7 8 |
Coordinator coordinator = Coordinators.isolated(new IsolatedCoordinator.Options() .clusterSpecs(specs) .activeEntries(specs.size()) .failoverNumNodes(2) .gracePeriod(TIMEOUT) .topologyBehavior(TopologyBehavior.WRAP_AT_END) .serviceTypes(serviceTypes) ); |
Code notes (by line number):
- Create an
IsolatedCoordinatorinstance with the configured options - Specify the nodes for each member of this topology
- Set the number of clusters to consider active
- Set the number of individual failed nodes before a whole cluster fails over completely
- Grace period after failure before cluster switchover
- Select the behavior when the end of a topology is reached. Alternatives include staying at the end cluster, or reversing through the list. The best option will depend on your specific scenario.
- Assign the services governed
Failure Detection
The client can monitor quite a lot of information while operating. Failure detectors make use of this wealth of insight happening under the hood. They use it to signal the coordinator. The coordinator can then decide how to respond.
A TrafficMonitoringFailureDetector watches requests sent to the server. You can configure the kinds of errors (exceptions) to count, the threshold number of exceptions, time interval in which they must occur and more. I mostly use the defaults and just tune the failed operation count and interval.
The client also has visibility into what’s going on with nodes and the cluster directly. This is captured using a NodeHealthFailureDetector. There are some nuances best left for the documentation. I just use the defaults here.
Finally, you can construct complex condition combinations using ConjunctionFailureDetector and DisjunctionFailureDetector. Think of conjunction detectors as performing a logical “and” of the constituent detectors. Disjunction does an “or”. I want to trigger on any failure, so I plug my traffic and health monitoring into a DisjunctionFailureDetector to get the final detector object to hand to the client.
|
1 2 3 4 5 6 7 8 9 10 11 |
TrafficMonitoringFailureDetector.Options trafficOptions = TrafficMonitoringFailureDetector.options() .maxFailedOperations(5) .failureInterval(60); FailureDetectorFactory<TrafficMonitoringFailureDetector> traffic = FailureDetectorFactories.trafficMonitoring(coordinator, trafficOptions); NodeHealthFailureDetector.Options healthOptions = NodeHealthFailureDetector.options(); FailureDetectorFactory<NodeHealthFailureDetector> health = FailureDetectorFactories.nodeHealth(coordinator, healthOptions); DisjunctionFailureDetectorFactory detector = FailureDetectorFactories.disjunction(traffic, health); |
Code notes (by line number):
- Instantiate and configure the
TrafficMonitoringFailureDetectoroptions - Fix the number of failed operations needed to indicate a problem
- Set the time interval over which to bin failures
- Create a
TrafficMonitoringFailureDetectorfactory instance, with access to theCoordinatorto trigger, and our options - Create a
NodeHealthFailureDetectorwith default options - Define the final logic using a
DisjunctionFailureDetector, again via a factory
Client Instantiation
Finally, we instantiate a client, and use it to obtain a bucket instance. For those unfamiliar with Couchbase, buckets are a high-level organizational abstraction. You can think of them as something between a table and a full database.
The bucket handle is all we need to run the loads against our clusters. A multi-cluster client bucket lets you specify timeouts for individual operations. For simplicity, I’ve wrapped the interface in a facade, applying only one fixed timeout to every operation. This makes the interface identical to the standard one.
|
1 2 3 4 |
MultiClusterClient client = new MultiClusterClient(coordinator, detector); client.authenticate(options.stringValueOf("id"), options.stringValueOf("password")); bucket = new BucketFacade(client.openBucket(bucketName, null), TIMEOUT, TimeUnit.MILLISECONDS); |
Wrapping up
That’s it for the new concepts needed to get off the ground with the MCA client. The rest of the code parses options, sets up the requested threads, and just runs them in loops. There are some statistics thrown in. I do make calls out to obtain the random data for writing and updating documents, so this code is not optimized for driving load. It can give some insight into the performance that Couchbase offers, though.
Again, to see this all in action, along with how the clusters behave, check out the demonstration during the second half of this webinar.
Version 1.0 of the MCA client was recently released. It’s currently an Enterprise Edition only feature, as of this writing. To find out more and get access to the client, talk to a Couchbase sales representative (sales@couchbase.com). You can contact me for more information also, as described below.
English