Mirror of Apache Kafka

7.4 KiB

Raw Blame History

Trogdor

Trogdor is a test framework for Apache Kafka.

Trogdor can run benchmarks and other workloads. Trogdor can also inject faults in order to stress test the system.

Quickstart

First, we want to start a single-node Kafka cluster with a ZooKeeper and a broker.

Running ZooKeeper:

> ./bin/zookeeper-server-start.sh ./config/zookeeper.properties &> /tmp/zookeeper.log &

Running Kafka:

> ./bin/kafka-server-start.sh ./config/server.properties &> /tmp/kafka.log &

Then, we want to run a Trogdor Agent, plus a Trogdor broker.

To run the Trogdor Agent:

> ./bin/trogdor.sh agent -c ./config/trogdor.conf -n node0 &> /tmp/trogdor-agent.log &

To run the Trogdor Coordinator:

> ./bin/trogdor.sh coordinator -c ./config/trogdor.conf -n node0 &> /tmp/trogdor-coordinator.log &

Let's confirm that all of the daemons are running:

> jps
116212 Coordinator
115188 QuorumPeerMain
116571 Jps
115420 Kafka
115694 Agent

Now, we can submit a test job to Trogdor. Here's an example of a short bash script which makes it easier.

> ./tests/bin/trogdor-run-produce-bench.sh
Sent CreateTaskRequest for task produce_bench_21634.$TASK_ID = produce_bench_21634

To get the test results, we run --show-tasks:

./bin/trogdor.sh client --show-tasks localhost:8889
Got coordinator tasks: {
  "tasks" : {
    "produce_bench_21634" : {
      "state" : "DONE",
      "spec" : {
        "class" : "org.apache.kafka.trogdor.workload.ProduceBenchSpec",
        "startMs" : 0,
        "durationMs" : 10000000,
        "producerNode" : "node0",
        "bootstrapServers" : "localhost:9092",
        "targetMessagesPerSec" : 10000,
        "maxMessages" : 50000,
        "keyGenerator" : {
          "type" : "sequential",
          "size" : 4,
          "startOffset" : 0
        },
        "valueGenerator" : {
          "type" : "constant",
          "size" : 512,
          "value" : "AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA="
        },
        "activeTopics" : {
          "foo[1-3]" : {
            "numPartitions" : 10,
            "replicationFactor" : 1
          }
        },
        "inactiveTopics" : {
          "foo[4-5]" : {
            "numPartitions" : 10,
            "replicationFactor" : 1
          }
        }
      },
      "startedMs" : 1541435949784,
      "doneMs" : 1541435955803,
      "cancelled" : false,
      "status" : {
        "totalSent" : 50000,
        "averageLatencyMs" : 11.0293,
        "p50LatencyMs" : 9,
        "p95LatencyMs" : 27,
        "p99LatencyMs" : 39
      }
    }
  }
}

Trogdor Architecture

Trogdor has a single coordinator process which manages multiple agent processes. Each agent process is responsible for a single cluster node.

The Trogdor coordinator manages tasks. A task is anything we might want to do on a cluster, such as running a benchmark, injecting a fault, or running a workload. In order to implement each task, the coordinator creates workers on one or more agent nodes.

The Trogdor agent process implements the tasks. For example, when running a workload, the agent process is the process which produces and consumes messages.

Both the coordinator and the agent expose a REST interface that accepts objects serialized via JSON. There is also a command-line program which makes it easy to send messages to either one without manually crafting the JSON message body.

All Trogdor RPCs are idempotent except the shutdown requests. Sending an idempotent RPC twice in a row has the same effect as sending the RPC once.

Tasks

Tasks are described by specifications containing:

A "class" field describing the task type. This contains a full Java class name.
A "startMs" field describing when the task should start. This is given in terms of milliseconds since the UNIX epoch.
A "durationMs" field describing how long the task should last. This is given in terms of milliseconds.
Other fields which are task-specific.

The task specification is usually written as JSON. For example, this task specification describes a network partition between nodes 1 and 2, and 3:

{
    "class": "org.apache.kafka.trogdor.fault.NetworkPartitionFaultSpec",
    "startMs": 1000,
    "durationMs": 30000,
    "partitions": [["node1", "node2"], ["node3"]]
}

Tasks are submitted to the coordinator. Once the coordinator determines that it is time for the task to start, it creates workers on agent processes. The workers run until the task is done.

Task specifications are immutable; they do not change after the task has been created.

Tasks can be in several states:

PENDING, when task is waiting to execute,
RUNNING, when the task is running,
STOPPING, when the task is in the process of stopping,
DONE, when the task is done.

Tasks that are DONE also have an error field which will be set if the task failed.

Workloads

Trogdor can run several workloads. Workloads perform operations on the cluster and measure their performance. Workloads fail when the operations cannot be performed.

ProduceBench

ProduceBench starts a Kafka producer on a single agent node, producing to several partitions. The workload measures the average produce latency, as well as the median, 95th percentile, and 99th percentile latency.

RoundTripWorkload

RoundTripWorkload tests both production and consumption. The workload starts a Kafka producer and consumer on a single node. The consumer will read back the messages that were produced by the producer.

ConsumeBench

ConsumeBench starts one or more Kafka consumers on a single agent node. Depending on the passed in configuration (see ConsumeBenchSpec), the consumers either subscribe to a set of topics (leveraging consumer group functionality and dynamic partition assignment) or manually assign partitions to themselves. The workload measures the average produce latency, as well as the median, 95th percentile, and 99th percentile latency.

Faults

Trogdor can run several faults which deliberately break something in the cluster.

ProcessStopFault

ProcessStopFault stops a process by sending it a SIGSTOP signal. When the fault ends, the process is resumed with SIGCONT.

NetworkPartitionFault

NetworkPartitionFault sets up an artificial network partition between one or more sets of nodes. Currently, this is implemented using iptables. The iptables rules are set up on the outbound traffic from the affected nodes. Therefore, the affected nodes should still be reachable from outside the cluster.

7.4 KiB Raw Blame History