Griffin is a library to replicate a blob of data to a cluster of machines spanning across multiple data centers.
Griffin was developed at Turn Inc. where it was used to replicate data across machines in a single data center. At Turn, Griffin was tested in a production setting and was found to be able to replicate data
- Of sizes ranging from a few megabytes to up to a gigabytes (compressed size)
- To a couple of hundred machines in a single data center
- With a latency of couple of minutes (time from data being available at any node to the time when data was available at all the nodes in the destination group)
While Griffin has been designed and well tested (in a docker environment) for replication across multiple data centers it has never been used in a production setting across multiple data centers.
Requires
- Docker 1.8+
- Docker compose (https://docs.docker.com/compose/install/)
A docker-compose.yml to create a full cluster of zookeeper, kafka, and griffin is included in this repository. To
create a cluster, update KAFKA_ADVERTISED_HOST_NAME in docker-compose.yml to the IP address of the host machine
(linux box) and start the cluster as
docker-compose up -d
docker-compose scale kafkalax1=3
docker-compose scale griffinlax1=3This might take some time as it requires docker to download few images.
The above steps starts a zookeeper node, a Apache Kafka cluster of three brokers, and cluster of three nodes running Griffin. To get the blob of data (or files) available locally at each of the Griffin nodes
for i in $(seq 3)
do
curl -i http://$(docker port griffin_griffinlax1_$i 8080)/griffin/localrepo
doneSince we have not pushed any data via griffin yet the output of the above command would be empty
We will now upload a file(build.gradle) to one of the griffin node, namely griffin_griffinlax1_1, and then ask griffin to
replicate it to all the other griffin nodes (dest='.*'). We would like this file to be named gradle
(blobname=gradle) on all the griffin nodes.
curl -i# -F blobname=gradle -F dest='.*' -F [email protected] http://$(docker port griffin_griffinlax1_1 8080)/griffin/localrepoWe can now observe (after a short delay) that the file is available at all griffin nodes
for i in $(seq 3)
do
curl -i http://$(docker port griffin_griffinlax1_$i 8080)/griffin/localrepo
doneList all the files in the local repository of that instance of griffin
Push a given file to the specified set of machines
| Parameter | Type | Required | Description |
|---|---|---|---|
| blobname | string | true | A name by which this blob of bytes will be referred by the repository |
| dest | string | true | A regular expression specifying the destination. This blob will be downloaded by any server whose serverid matches the regex |
| file | form | true | Multi-part content specifying the data bytes |
List all the files available in all the instances of griffin (henceforth called global repository)
Push a given file to the specified set of machines
| Parameter | Type | Required | Description |
|---|---|---|---|
| blobname | string | true | A name by which this blob of bytes will be referred by the repository |
| dest | string | true | A regular expression specifying the destination. This blob will be downloaded by any server whose serverid matches the regex |
| file | form | true | Multi-part content specifying the data bytes |
List all the files in global repository that is supposed to be in the local repository but has not yet been download
TODO: Explain the design space and the design choices
TODO: Component design and description
- Gradle 2.5+
- protoc 2.5+
- Java 1.7
Prepare a single griffin fat jar with
gradle buildFat jar is available at ./build/libs/griffin-*.jar. Include this jar in your application's library folder to use griffin.
Griffin requires a configuration file griffin.conf in the local directory. A sample griffin.conf is included in the
examples directory. Please update the file with the required information.
- Griffin.ZkServers:
LIST_OF_ZKSERVERS - Griffin.KafkaBrokersList:
LIST_OF_KAFKA_BROKERS_WITH_PORTS - Griffin.Error.Email.Recipients:
EMAIL_ID - serverid:
UNIQUE_SERVER_ID - dcname:
DATACENTER_NAME
Griffin requires that Apache Kafka has automatic topic creation and topic deletion enabled, i.e.,
auto.create.topics.enable=true
delete.topic.enable=trueWe will run two instances of griffin. Each instance of griffin is required to have a unique serverid. Zookeeper
cluster specified using ZkServers and Kafka cluster specified using KafkaBrokersList should be same for all griffin
services that share the same dcname
Open two terminals. In first terminal start the first instance of griffin as
cd examples
java -jar ../build/libs/griffin-*.jarIn second terminal change serverid in examples/griffin.conf to griffin-2 and server.port in
examples/application.properties to 8081 and start another instance of griffin as
cd examples
java -jar ../build/libs/griffin-*.jarIn yet another terminal we can interact with these griffin services. We will push a file to one instance of griffin and we will observe it getting replicated to another instance of griffin. First we can observe that both instances for griffin does not contain any file in their local repository
curl -i http://localhost:8080/griffin/localrepo/
curl -i http://localhost:8081/griffin/localrepo/We can push a file to one of the griffin instances as
curl -i -F blobname=gradle -F dest='.*' -F [email protected] http://localhost:8080/griffin/localrepoAfter a short while one can observe the file available in local repository of both griffin services
curl -i http://localhost:8080/griffin/localrepo/
curl -i http://localhost:8081/griffin/localrepo/Check out REST API section for full description of rest calls above.
gradle javadocJavadocs are generated in ./build/docs/javadoc/index.html.
- The first incarnation of Griffin was written for Java 6. So while the current version of Griffin requires Java 7, it does not fully leverage all the new features of Java 7.