Version: 1.2.3

Storm Metrics

Storm exposes a metrics interface to report summary statistics across the full topology. The numbers you see on the UI come from some of these built in metrics, but are reported through the worker heartbeats instead of through the IMetricsConsumer described below.

Metric Types

Metrics have to implement IMetric which contains just one method, getValueAndReset -- do any remaining work to find the summary value, and reset back to an initial state. For example, the MeanReducer divides the running total by its running count to find the mean, then initializes both values back to zero.

Storm gives you these metric types:

  • AssignableMetric -- set the metric to the explicit value you supply. Useful if it's an external value or in the case that you are already calculating the summary statistic yourself.
  • CombinedMetric -- generic interface for metrics that can be updated associatively.
  • CountMetric -- a running total of the supplied values. Call incr() to increment by one, incrBy(n) to add/subtract the given number.
  • ReducedMetric
    • MeanReducer -- track a running average of values given to its reduce() method. (It accepts Double, Integer or Long values, and maintains the internal average as a Double.) Despite his reputation, the MeanReducer is actually a pretty nice guy in person.
    • MultiReducedMetric -- a hashmap of reduced metrics.

Be aware that even though getValueAndReset can return an object returning any object makes it very difficult for an IMetricsConsumer to know how to translate it into something usable. Also note that because it is sent to the IMetricsConsumer as a part of a tuple the values returned need to be able to be serialized by your topology.

Metrics Consumer

You can listen and handle the topology metrics via registering Metrics Consumer to your topology.

To register metrics consumer to your topology, add to your topology's configuration like:

conf.registerMetricsConsumer(org.apache.storm.metric.LoggingMetricsConsumer.class, 1);

You can refer Config#registerMetricsConsumer and overloaded methods from javadoc.

Otherwise edit the storm.yaml config file:

topology.metrics.consumer.register:
  - class: "org.apache.storm.metric.LoggingMetricsConsumer"
    parallelism.hint: 1
  - class: "org.apache.storm.metric.HttpForwardingMetricsConsumer"
    parallelism.hint: 1
    argument: "http://example.com:8080/metrics/my-topology/"

Storm adds a MetricsConsumerBolt to your topolology for each class in the topology.metrics.consumer.register list. Each MetricsConsumerBolt subscribes to receive metrics from all tasks in the topology. The parallelism for each Bolt is set to parallelism.hint and component id for that Bolt is set to __metrics_<metrics consumer class name>. If you register the same class name more than once, postfix #<sequence number> is appended to component id.

Storm provides some built-in metrics consumers for you to try out to see which metrics are provided in your topology.

  • LoggingMetricsConsumer -- listens for all metrics and dumps them to log file with TSV (Tab Separated Values).
  • HttpForwardingMetricsConsumer -- listens for all metrics and POSTs them serialized to a configured URL via HTTP. Storm also provides HttpForwardingMetricsServer as abstract class so you can extend this class and run as a HTTP server, and handle metrics sent by HttpForwardingMetricsConsumer.

Also, Storm exposes the interface IMetricsConsumer for implementing Metrics Consumer so you can create custom metrics consumers and attach to their topologies, or use other great implementation of Metrics Consumers provided by Storm community. Some of examples are versign/storm-graphite, and storm-metrics-statsd.

When you implement your own metrics consumer, argument is passed to Object when IMetricsConsumer#prepare is called, so you need to infer the Java type of configured value on yaml, and do explicit type casting.

Please keep in mind that MetricsConsumerBolt is just a kind of Bolt, so whole throughput of the topology will go down when registered metrics consumers cannot keep up handling incoming metrics, so you may want to take care of those Bolts like normal Bolts. One of idea to avoid this is making your implementation of Metrics Consumer as non-blocking fashion.

Build your own metric (task level)

You can measure your own metric by registering IMetric to Metric Registry.

Suppose we would like to measure execution count of Bolt#execute. Let's start with defining metric instance. CountMetric seems to fit our use case.

private transient CountMetric countMetric;

Notice we define it as transient. IMertic is not Serializable so we defined as transient to avoid any serialization issues.

Next, let's initialize and register the metric instance.

@Override
public void prepare(Map conf, TopologyContext context, OutputCollector collector) {
    // other initialization here.
    countMetric = new CountMetric();
    context.registerMetric("execute_count", countMetric, 60);
}

The meaning of first and second parameters are straightforward, metric name and instance of IMetric. Third parameter of TopologyContext#registerMetric is the period (seconds) to publish and reset the metric.

Last, let's increment the value when Bolt.execute() is executed.

public void execute(Tuple input) {
    countMetric.incr();
    // handle tuple here.   
}

Note that sample rate for topology metrics is not applied to custom metrics since we're calling incr() ourselves.

Done! countMetric.getValueAndReset() is called every 60 seconds as we registered as period, and pair of ("execute_count", value) will be pushed to MetricsConsumer.

Build your own metrics (worker level)

You can register your own worker level metrics by adding them to Config.WORKER_METRICS for all workers in cluster, or Config.TOPOLOGY_WORKER_METRICS for all workers in specific topology.

For example, we can add worker.metrics to storm.yaml in cluster,

worker.metrics: 
  metricA: "aaa.bbb.ccc.ddd.MetricA"
  metricB: "aaa.bbb.ccc.ddd.MetricB"
  ...

or put Map<String, String> (metric name, metric class name) with key Config.TOPOLOGY_WORKER_METRICS to config map.

There are some restrictions for worker level metric instances:

A) Metrics for worker level should be kind of gauge since it is initialized and registered from SystemBolt and not exposed to user tasks.

B) Metrics will be initialized with default constructor, and no injection for configuration or object will be performed.

C) Bucket size (seconds) for metrics is fixed to Config.TOPOLOGY_BUILTIN_METRICS_BUCKET_SIZE_SECS.

Builtin Metrics

The builtin metrics instrument Storm itself.

builtin_metrics.clj sets up data structures for the built-in metrics, and facade methods that the other framework components can use to update them. The metrics themselves are calculated in the calling code -- see for example ack-spout-msg in clj/b/s/daemon/daemon/executor.clj

Reporting Rate

The rate at which built in metrics are reported is configurable through the topology.builtin.metrics.bucket.size.secs config. If you set this too low it can overload the consumers, so please use caution when modifying it.

Tuple Counting Metrics

There are several different metrics related to counting what a bolt or spout does to a tuple. These include things like emitting, transferring, acking, and failing of tuples.

In general all of these tuple count metrics are randomly sub-sampled unless otherwise stated. This means that the counts you see both on the UI and from the built in metrics are not necessarily exact. In fact by default we sample only 5% of the events and estimate the total number of events from that. The sampling percentage is configurable per topology through the topology.stats.sample.rate config. Setting it to 1.0 will make the counts exact, but be aware that the more events we sample the slower your topology will run (as the metrics are counted in the same code path as tuples are processed). This is why we have a 5% sample rate as the default.

The tuple counting metrics are generally reported to the metrics consumers as maps unless explicitly stated otherwise. They break down each count for finer grained reporting. The keys to these maps fall into two categories "${stream_name}" or "${upstream_component}:${stream_name}". The former is used for all spout metrics and for outgoing bolt metrics (__emit-count and __transfer-count). The latter is used for bolt metrics that deal with incoming tuples.

So for a word count topology the count bolt might show something like the following for the __ack-count metric

{
    "split:default": 80080
}

But the spout instead would show something like the following for the __ack-count metric.

{
    "default": 12500
}
__ack-count

For bolts it is the number of incoming tuples that had the ack method called on them. For spouts it is the number of tuples trees that were fully acked. See Guaranteeing Message Processing for more information about what a tuple tree is. If acking is disabled this metric is still reported, but it is not really meaningful.

__fail-count

For bolts this is the number of incoming tuples that had the fail method called on them. For spouts this is the number of tuple trees that failed. Tuple trees may fail from timing out or because a bolt called fail on it. The two are not separated out by this metric.

__emit-count

This is the total number of times the emit method was called to send a tuple. This is the same for both bolts and spouts.

__transfer-count

This is the total number of tuples transferred to a downstream bolt/spout for processing. This number will not always match __emit_count. If nothing is registered to receive a tuple down stream the number will be 0 even if tuples were emitted. Similarly if there are multiple down stream consumers it may be a multiple of the number emitted. The grouping also can play a role if it sends the tuple to multiple instances of a single bolt down stream.

__execute-count

This count metric is bolt specific. It counts the number of times that a bolt's execute method was called.

Tuple Latency Metrics

Similar to the tuple counting metrics storm also collects average latency metrics for bolts and spouts. These follow the same structure as the bolt/spout maps and are sub-sampled in the same way as well. In all cases the latency is measured in milliseconds.

__complete-latency

The complete latency is just for spouts. It is the average amount of time it took for ack or fail to be called for a tuple after it was emitted. If acking is disabled this metric is likely to be blank or 0 for all values, and should be ignored.

__execute-latency

This is just for bolts. It is the average amount of time that the bolt spent in the call to the execute method. The higher this gets, the lower the throughput of tuples per bolt instance.

__process-latency

This is also just for bolts. It is the average amount of time between when execute was called to start processing a tuple, to when it was acked or failed by the bolt. If your bolt is a very simple bolt and the processing is synchronous then __process-latency and __execute-latency should be very close to one another, with process latency being slightly smaller. If you are doing a join or have asynchronous processing then it may take a while for a tuple to be acked so the process latency would be higher than the execute latency.

__skipped-max-spout-ms

This metric records how much time a spout was idle because more tuples than topology.max.spout.pending were still outstanding. This is the total time in milliseconds, not the average amount of time and is not sub-sampled.

__skipped-throttle-ms

This metric records how much time a spout was idle because back-pressure indicated that downstream queues in the topology were too full. This is the total time in milliseconds, not the average amount of time and is not sub-sampled.

skipped-inactive-ms

This metric records how much time a spout was idle because the topology was deactivated. This is the total time in milliseconds, not the average amount of time and is not sub-sampled.

Queue Metrics

Each bolt or spout instance in a topology has a receive queue and a send queue. Each worker also has a queue for sending messages to other workers. All of these have metrics that are reported.

The receive queue metrics are reported under the __receive name and send queue metrics are reported under the __sendqueue for the given bolt/spout they are a part of. The metrics for the queue that sends messages to other workers is under the __transfer metric name for the system bolt (__system).

They all have the form.

{
    "arrival_rate_secs": 1229.1195171893523,
    "overflow": 0,
    "read_pos": 103445,
    "write_pos": 103448,
    "sojourn_time_ms": 2.440771591407277,
    "capacity": 1024,
    "population": 19
    "tuple_population": 200
}

In storm we sometimes batch multiple tuples into a single entry in the disruptor queue. This batching is an optimization that has been in storm in some form since the beginning, but the metrics did not always reflect this so be careful with how you interpret the metrics and pay attention to which metrics are for tuples and which metrics are for entries in the disruptor queue. The __receive and __transfer queues can have batching but the __sendqueue should not.

arrival_rate_secs is an estimation of the number of tuples that are inserted into the queue in one second, although it is actually the dequeue rate. The sojourn_time_ms is calculated from the arrival rate and is an estimate of how many milliseconds each tuple sits in the queue before it is processed. Prior to STORM-2621 (v1.1.1, v1.2.0, and v2.0.0) these were the rate of entries, not of tuples.

A disruptor queue has a set maximum number of entries. If the regular queue fills up an overflow queue takes over. The number of tuple batches stored in this overflow section are represented by the overflow metric. Storm also does some micro batching of tuples for performance/efficiency reasons so you may see the overflow with a very small number in it even if the queue is not full.

read_pos and write_pos are internal disruptor accounting numbers. You can think of them almost as the total number of entries written (write_pos) or read (read_pos) since the queue was created. They allow for integer overflow so if you use them please take that into account.

capacity is the maximum number of entries in the disruptor queue. population is the number of entries currently filled in the queue.

tuple_population is the number of tuples currently in the queue as opposed to the number of entries. This was added at the same time as STORM-2621 (v1.1.1, v1.2.0, and v2.0.0)

System Bolt (Worker) Metrics

The System Bolt __system provides lots of metrics for different worker wide things. The one metric not described here is the __transfer queue metric, because it fits with the other disruptor metrics described above.

Be aware that the __system bolt is an actual bolt so regular bolt metrics described above also will be reported for it.

Receive (NettyServer)

__recv-iconnection reports stats for the netty server on the worker. This is what gets messages from other workers. It is of the form

{
    "dequeuedMessages": 0,
    "enqueued": {
      "/127.0.0.1:49952": 389951
    }
}

dequeuedMessages is a throwback to older code where there was an internal queue between the server and the bolts/spouts. That is no longer the case and the value can be ignored. enqueued is a map between the address of the remote worker and the number of tuples that were sent from it to this worker.

Send (Netty Client)

The __send-iconnection metric holds information about all of the clients for this worker. It is of the form

{
    NodeInfo(node:7decee4b-c314-41f4-b362-fd1358c985b3-127.0.01, port:[6701]): {
        "reconnects": 0,
        "src": "/127.0.0.1:49951",
        "pending": 0,
        "dest": "localhost/127.0.0.1:6701",
        "sent": 420779,
        "lostOnSend": 0
    }
}

The value is a map where the key is a NodeInfo class for the downstream worker it is sending messages to. This is the SupervisorId + port. The value is another map with the fields

  • src What host/port this client has used to connect to the receiving worker.
  • dest What host/port this client has connected to.
  • reconnects the number of reconnections that have happened.
  • pending the number of messages that have not been sent. (This corresponds to messages, not tuples)
  • sent the number of messages that have been sent. (This is messages not tuples)
  • lostOnSend. This is the number of messages that were lost because of connection issues. (This is messages not tuples).
JVM Memory

JVM memory usage is reported through memory/nonHeap for off heap memory and memory/heap for on heap memory. These values come from the MemoryUsage mxbean. Each of the metrics are reported as a map with the following keys, and values returned by the corresponding java code.

Key Corresponding Code
maxBytes memUsage.getMax()
committedBytes memUsage.getCommitted()
initBytes memUsage.getInit()
usedBytes memUsage.getUsed()
virtualFreeBytes memUsage.getMax() - memUsage.getUsed()
unusedBytes memUsage.getCommitted() - memUsage.getUsed()
JVM Garbage Collection

The exact GC metric name depends on the garbage collector that your worker uses. The data is all collected from ManagementFactory.getGarbageCollectorMXBeans() and the name of the metrics is "GC/" followed by the name of the returned bean with white space removed. The reported metrics are just

  • count the number of gc events that happened and
  • timeMs the total number of milliseconds that were spent doing gc.

Please refer to the JVM documentation for more details.

JVM Misc
  • threadCount is the number of threads currently in the JVM.
Uptime
  • uptimeSecs reports the number of seconds the worker has been up for
  • newWorkerEvent is 1 when a worker is first started and 0 all other times. This can be used to tell when a worker has crashed and is restarted.
  • startTimeSecs is when the worker started in seconds since the epoch