首先定义一系列metric相关的interface, IMetric, IReducer, ICombiner (backtype.storm.metric.api)
在task中, 创建一系列builtin-metrics, (backtype.storm.daemon.builtin-metrics), 并注册到topology context里面
task会不断的利用如spout-acked-tuple!的functions去更新这些builtin-metrics
task会定期将builtin-metrics里面的统计数据通过METRICS-STREAM发送给metric-bolt (backtype.storm.metric.MetricsConsumerBolt, 该bolt会创建实现backtype.storm.metric.api.IMetricsConsumer的对象, 用于计算出metrics)
然后如何使用这些metrics?
由于这是builtin metrics, 是不会被外界使用的
如果处理这些metrics, 取决于_metricsConsumer.handleDataPoints, 这里的_metricsConsumer是通过topology's configuration配置的
比如backtype.storm.metric.LoggingMetricsConsumer, 如果使用这个consumer就会将metrics写入log中
1. backtype.storm.metric.api
IMetric
package backtype.storm.metric.api; public interface IMetric { public Object getValueAndReset(); ;;取得当前值并恢复初始状态 }
CountMetric, 计数, reset时清零
AssignableMetric, 赋值, 不用reset
MultiCountMetric, 使用hashmap记录多个count, reset时分别对每个count对象调用getValueAndReset
public class CountMetric implements IMetric { long _value = 0; public CountMetric() { } public void incr() { _value++; } public void incrBy(long incrementBy) { _value += incrementBy; } public Object getValueAndReset() { long ret = _value; _value = 0; return ret; } }
ICombiner
public interface ICombiner<T> { public T identity(); public T combine(T a, T b); }
CombinedMetric, 结合ICombiner和IMetric
public class CombinedMetric implements IMetric { private final ICombiner _combiner; private Object _value; public CombinedMetric(ICombiner combiner) { _combiner = combiner; _value = _combiner.identity(); } public void update(Object value) { _value = _combiner.combine(_value, value); } public Object getValueAndReset() { Object ret = _value; _value = _combiner.identity(); return ret; } }
IReducer
public interface IReducer<T> {
T init();
T reduce(T accumulator, Object input);
Object extractResult(T accumulator);
}
实现IReducer接口, 实现平均数Reducer, reduce里面累加和计数, extractResult里面acc/count求平均数
class MeanReducerState { public int count = 0; public double sum = 0.0; } public class MeanReducer implements IReducer<MeanReducerState> { public MeanReducerState init() { return new MeanReducerState(); } public MeanReducerState reduce(MeanReducerState acc, Object input) { acc.count++; if(input instanceof Double) { acc.sum += (Double)input; } else if(input instanceof Long) { acc.sum += ((Long)input).doubleValue(); } else if(input instanceof Integer) { acc.sum += ((Integer)input).doubleValue(); } else { throw new RuntimeException( "MeanReducer::reduce called with unsupported input type `" + input.getClass() + "`. Supported types are Double, Long, Integer."); } return acc; } public Object extractResult(MeanReducerState acc) { if(acc.count > 0) { return new Double(acc.sum / (double)acc.count); } else { return null; } } }
ReducedMetric
结合IReducer和IMetric
public class ReducedMetric implements IMetric { private final IReducer _reducer; private Object _accumulator; public ReducedMetric(IReducer reducer) { _reducer = reducer; _accumulator = _reducer.init(); } public void update(Object value) { _accumulator = _reducer.reduce(_accumulator, value); } public Object getValueAndReset() { Object ret = _reducer.extractResult(_accumulator); _accumulator = _reducer.init(); return ret; } }
IMetricsConsumer
这个interface, 内嵌TaskInfo和DataPoint类
handleDataPoints, 添加逻辑以处理task对应的一系列DataPoint
public interface IMetricsConsumer { public static class TaskInfo { public TaskInfo() {} public TaskInfo(String srcWorkerHost, int srcWorkerPort, String srcComponentId, int srcTaskId, long timestamp, int updateIntervalSecs) { this.srcWorkerHost = srcWorkerHost; this.srcWorkerPort = srcWorkerPort; this.srcComponentId = srcComponentId; this.srcTaskId = srcTaskId; this.timestamp = timestamp; this.updateIntervalSecs = updateIntervalSecs; } public String srcWorkerHost; public int srcWorkerPort; public String srcComponentId; public int srcTaskId; public long timestamp; public int updateIntervalSecs; } public static class DataPoint { public DataPoint() {} public DataPoint(String name, Object value) { this.name = name; this.value = value; } @Override public String toString() { return "[" + name + " = " + value + "]"; } public String name; public Object value; } void prepare(Map stormConf, Object registrationArgument, TopologyContext context, IErrorReporter errorReporter); void handleDataPoints(TaskInfo taskInfo, Collection<DataPoint> dataPoints); void cleanup(); }
2. backtype.storm.daemon.builtin-metrics
定义Spout和Bolt所需要的一些metric, 主要两个record, BuiltinSpoutMetrics和BuiltinBoltMetrics, [metric-name, metric-object]的hashmap
(defrecord BuiltinSpoutMetrics [^MultiCountMetric ack-count
^MultiReducedMetric complete-latency
^MultiCountMetric fail-count
^MultiCountMetric emit-count
^MultiCountMetric transfer-count])
(defrecord BuiltinBoltMetrics [^MultiCountMetric ack-count
^MultiReducedMetric process-latency
^MultiCountMetric fail-count
^MultiCountMetric execute-count
^MultiReducedMetric execute-latency
^MultiCountMetric emit-count
^MultiCountMetric transfer-count])
(defn make-data [executor-type]
(condp = executor-type
:spout (BuiltinSpoutMetrics. (MultiCountMetric.)
(MultiReducedMetric. (MeanReducer.))
(MultiCountMetric.)
(MultiCountMetric.)
(MultiCountMetric.))
:bolt (BuiltinBoltMetrics. (MultiCountMetric.)
(MultiReducedMetric. (MeanReducer.))
(MultiCountMetric.)
(MultiCountMetric.)
(MultiReducedMetric. (MeanReducer.))
(MultiCountMetric.)
(MultiCountMetric.))))
(defn register-all [builtin-metrics storm-conf topology-context]
(doseq [[kw imetric] builtin-metrics]
(.registerMetric topology-context (str "__" (name kw)) imetric
(int (get storm-conf Config/TOPOLOGY_BUILTIN_METRICS_BUCKET_SIZE_SECS)))))
在mk-task-data的时候, 调用make-data来创建相应的metrics,
:builtin-metrics (builtin-metrics/make-data (:type executor-data))
并在executor的mk-threads中, 会将这些builtin-metrics注册到topologycontext中去,
(builtin-metrics/register-all (:builtin-metrics task-data) storm-conf (:user-context task-data))
上面完成的builtin-metrics的创建和注册, 接着定义了一系列用于更新metrics的functions,
以spout-acked-tuple!为例, 需要更新MultiCountMetric ack-count和MultiReducedMetric complete-latency
.scope从MultiCountMetric取出某个CountMetric, 然后incrBy来将stats的rate增加到count上
(defn spout-acked-tuple! [^BuiltinSpoutMetrics m stats stream latency-ms] (-> m .ack-count (.scope stream) (.incrBy (stats-rate stats))) (-> m .complete-latency (.scope stream) (.update latency-ms)))
3. backtype.storm.metric
MetricsConsumerBolt
创建实现IMetricsConsumer的对象, 并在execute里面调用handleDataPoints
package backtype.storm.metric; public class MetricsConsumerBolt implements IBolt { IMetricsConsumer _metricsConsumer; String _consumerClassName; OutputCollector _collector; Object _registrationArgument; public MetricsConsumerBolt(String consumerClassName, Object registrationArgument) { _consumerClassName = consumerClassName; _registrationArgument = registrationArgument; } @Override public void prepare(Map stormConf, TopologyContext context, OutputCollector collector) { try { _metricsConsumer = (IMetricsConsumer)Class.forName(_consumerClassName).newInstance(); } catch (Exception e) { throw new RuntimeException("Could not instantiate a class listed in config under section " + Config.TOPOLOGY_METRICS_CONSUMER_REGISTER + " with fully qualified name " + _consumerClassName, e); } _metricsConsumer.prepare(stormConf, _registrationArgument, context, (IErrorReporter)collector); _collector = collector; } @Override public void execute(Tuple input) { _metricsConsumer.handleDataPoints((IMetricsConsumer.TaskInfo)input.getValue(0), (Collection)input.getValue(1)); _collector.ack(input); } @Override public void cleanup() { _metricsConsumer.cleanup(); } }
SystemBolt
SystemBolt, 根据comments里面说的, 每个worker都有一个, taskid=-1
定义些system相关的metric, 并注册到topologycontext里面
需要使用Java调用clojure, 所以需要import下面的package
import clojure.lang.AFn; import clojure.lang.IFn; //funtion import clojure.lang.RT; //run-time
并且用到些用于监控memory和JVM的java package
java.lang.management.MemoryUsage, 表示内存使用量快照的MemoryUsage对象
java.lang.management.GarbageCollectorMXBean, 用于Java虚拟机的垃圾回收的管理接口, 比如发生的回收的总次数, 和累计回收时间
java.lang.management.RuntimeMXBean, 用于Java 虚拟机的运行时系统的管理接口
这个bolt的特点是, 只有prepare实现了逻辑, 并且通过_prepareWasCalled保证prepare只被执行一次
prepare中的逻辑, 主要就是定义各种metric, 并且通过registerMetric注册到TopologyContext中
metic包含, JVM的运行时间, 开始时间, memory情况, 和每个GarbageCollector的情况
注册的这些system metrics也会一起被发送到MetricsConsumerBolt进行处理
这应该用spout实现, 为啥用bolt实现?
// There is one task inside one executor for each worker of the topology. // TaskID is always -1, therefore you can only send-unanchored tuples to co-located SystemBolt. // This bolt was conceived to export worker stats via metrics api. public class SystemBolt implements IBolt { private static Logger LOG = LoggerFactory.getLogger(SystemBolt.class); private static boolean _prepareWasCalled = false; private static class MemoryUsageMetric implements IMetric { IFn _getUsage; public MemoryUsageMetric(IFn getUsage) { _getUsage = getUsage; } @Override public Object getValueAndReset() { MemoryUsage memUsage = (MemoryUsage)_getUsage.invoke(); HashMap m = new HashMap(); m.put("maxBytes", memUsage.getMax()); m.put("committedBytes", memUsage.getCommitted()); m.put("initBytes", memUsage.getInit()); m.put("usedBytes", memUsage.getUsed()); m.put("virtualFreeBytes", memUsage.getMax() - memUsage.getUsed()); m.put("unusedBytes", memUsage.getCommitted() - memUsage.getUsed()); return m; } } // canonically the metrics data exported is time bucketed when doing counts. // convert the absolute values here into time buckets. private static class GarbageCollectorMetric implements IMetric { GarbageCollectorMXBean _gcBean; Long _collectionCount; Long _collectionTime; public GarbageCollectorMetric(GarbageCollectorMXBean gcBean) { _gcBean = gcBean; } @Override public Object getValueAndReset() { Long collectionCountP = _gcBean.getCollectionCount(); Long collectionTimeP = _gcBean.getCollectionTime(); Map ret = null; if(_collectionCount!=null && _collectionTime!=null) { ret = new HashMap(); ret.put("count", collectionCountP - _collectionCount); ret.put("timeMs", collectionTimeP - _collectionTime); } _collectionCount = collectionCountP; _collectionTime = collectionTimeP; return ret; } } @Override public void prepare(final Map stormConf, TopologyContext context, OutputCollector collector) { if(_prepareWasCalled && stormConf.get(Config.STORM_CLUSTER_MODE) != "local") { throw new RuntimeException("A single worker should have 1 SystemBolt instance."); } _prepareWasCalled = true; int bucketSize = RT.intCast(stormConf.get(Config.TOPOLOGY_BUILTIN_METRICS_BUCKET_SIZE_SECS)); final RuntimeMXBean jvmRT = ManagementFactory.getRuntimeMXBean(); context.registerMetric("uptimeSecs", new IMetric() { @Override public Object getValueAndReset() { return jvmRT.getUptime()/1000.0; } }, bucketSize); context.registerMetric("startTimeSecs", new IMetric() { @Override public Object getValueAndReset() { return jvmRT.getStartTime()/1000.0; } }, bucketSize); context.registerMetric("newWorkerEvent", new IMetric() { boolean doEvent = true; @Override public Object getValueAndReset() { if (doEvent) { doEvent = false; return 1; } else return 0; } }, bucketSize); final MemoryMXBean jvmMemRT = ManagementFactory.getMemoryMXBean(); context.registerMetric("memory/heap", new MemoryUsageMetric(new AFn() { public Object invoke() { return jvmMemRT.getHeapMemoryUsage(); } }), bucketSize); context.registerMetric("memory/nonHeap", new MemoryUsageMetric(new AFn() { public Object invoke() { return jvmMemRT.getNonHeapMemoryUsage(); } }), bucketSize); for(GarbageCollectorMXBean b : ManagementFactory.getGarbageCollectorMXBeans()) { context.registerMetric("GC/" + b.getName().replaceAll("\\W", ""), new GarbageCollectorMetric(b), bucketSize); } } @Override public void execute(Tuple input) { throw new RuntimeException("Non-system tuples should never be sent to __system bolt."); } @Override public void cleanup() { } }
4. system-topology!
这里会动态的往topology里面, 加入metric-component (MetricsConsumerBolt) 和system-component (SystemBolt), 以及相应的steam信息
system-topology!会往topology加上些东西
1. acker, 后面再说
2. metric-bolt, input是所有component的tasks发来的METRICS-STREAM, 没有output
3. system-bolt, 没有input, output是两个TICK-STREAM
4. 给所有component, 增加额外的输出metrics-stream, system-stream
(defn system-topology! [storm-conf ^StormTopology topology]
(validate-basic! topology)
(let [ret (.deepCopy topology)]
(add-acker! storm-conf ret)
(add-metric-components! storm-conf ret)
(add-system-components! storm-conf ret)
(add-metric-streams! ret)
(add-system-streams! ret)
(validate-structure! ret)
ret
))
4.1 增加component
看下thrift中的定义, 往topology里面增加一个blot component, 其实就是往hashmap中增加一组[string, Bolt]
关键就是看看如何使用thrift/mk-bolt-spec*来创建blot spec
struct StormTopology {
1: required map<string, SpoutSpec> spouts;
2: required map<string, Bolt> bolts;
3: required map<string, StateSpoutSpec> state_spouts;
}
struct Bolt {
1: required ComponentObject bolt_object;
2: required ComponentCommon common;
}
struct ComponentCommon {
1: required map<GlobalStreamId, Grouping> inputs;
2: required map<string, StreamInfo> streams; //key is stream id, outputs
3: optional i32 parallelism_hint; //how many threads across the cluster should be dedicated to this component
4: optional string json_conf;
}
struct StreamInfo {
1: required list<string> output_fields;
2: required bool direct;
}
(defn add-metric-components! [storm-conf ^StormTopology topology]
(doseq [[comp-id bolt-spec] (metrics-consumer-bolt-specs storm-conf topology)] ;;从metrics-consumer-bolt-specs中可以看出该bolt会以METRICS-STREAM-ID为输入, 且没有输出
(.put_to_bolts topology comp-id bolt-spec)))
(defn add-system-components! [conf ^StormTopology topology]
(let [system-bolt-spec (thrift/mk-bolt-spec*
{} ;;input为空, 没有输入
(SystemBolt.) ;;object
{SYSTEM-TICK-STREAM-ID (thrift/output-fields ["rate_secs"])
METRICS-TICK-STREAM-ID (thrift/output-fields ["interval"])} ;;output, 定义两个output streams, 但代码中并没有emit
:p 0
:conf {TOPOLOGY-TASKS 0})]
(.put_to_bolts topology SYSTEM-COMPONENT-ID system-bolt-spec)))
metric-components
首先, topology里面所有的component(包含system component), 都需要往metics-bolt发送统计数据, 所以component-ids-that-emit-metrics就是all-components-ids+SYSTEM-COMPONENT-ID
那么对于任意一个comp, 都会对metics-bolt产生如下输入, {[comp-id METRICS-STREAM-ID] :shuffle} (采用:suffle grouping方式)
然后, 用thrift/mk-bolt-spec*来定义创建bolt的fn, mk-bolt-spec
最后, 调用mk-bolt-spec来创建metics-bolt的spec, 参考上面的定义
关键就是, 创建MetricsConsumerBolt对象, 需要从storm-conf里面读出, MetricsConsumer的实现类和参数
这个bolt负责, 将从各个task接收到的数据, 调用handleDataPoints生成metircs, 参考前面的定义
(defn metrics-consumer-bolt-specs [storm-conf topology]
(let [component-ids-that-emit-metrics (cons SYSTEM-COMPONENT-ID (keys (all-components topology)))
inputs (->> (for [comp-id component-ids-that-emit-metrics]
{[comp-id METRICS-STREAM-ID] :shuffle})
(into {}))
mk-bolt-spec (fn [class arg p]
(thrift/mk-bolt-spec*
inputs ;;inputs集合
(backtype.storm.metric.MetricsConsumerBolt. class arg) ;;object
{} ;;output为空
:p p :conf {TOPOLOGY-TASKS p}))]
(map
(fn [component-id register]
[component-id (mk-bolt-spec (get register "class")
(get register "argument")
(or (get register "parallelism.hint") 1))])
(metrics-consumer-register-ids storm-conf)
(get storm-conf TOPOLOGY-METRICS-CONSUMER-REGISTER))))
4.2 增加stream
给每个component增加两个output stream
METRICS-STREAM-ID, 发送给metric-blot, 数据结构为output-fields ["task-info" "data-points"]
SYSTEM-STREAM-ID, ,数据结构为output-fields ["event"]
(defn add-metric-streams! [^StormTopology topology]
(doseq [[_ component] (all-components topology)
:let [common (.get_common component)]]
(.put_to_streams common METRICS-STREAM-ID
(thrift/output-fields ["task-info" "data-points"]))))
(defn add-system-streams! [^StormTopology topology]
(doseq [[_ component] (all-components topology)
:let [common (.get_common component)]]
(.put_to_streams common SYSTEM-STREAM-ID (thrift/output-fields ["event"]))))
本文章摘自博客园，原文发布日期：2013-07-30
