Motivation

Currently Paimon runs without monitoring and metrics out of box. In the production environment, users need to know how their Paimon Table behaves like what is the commit duration, how many files each commit added or deleted, the status of compaction operation, the duration of scan query, lag of streaming reading and so on. So we need to introduce a metrics system and more metrics for Paimon.  

...

Histogram is a type of metric interface measure the statistical distribution of a set of values including the min, max, mean, standard deviation and quantilespercentile.

/** The histogram allows to record values, get the current count of recorded values and create
 * histogram statistics for the currently seen elements.*/
@Public
public interface Histogram extends Metric {

    /**
     * Update the histogram with the given value.
     *
     * @param value Value to update the histogram with
     */
    void update(long value);

    /**
     * Get the count of seen elements.
     *
     * @return Count of seen elements
     */
    long getCount();

    /**
     * Create statistics for the currently recorded elements.
     *
     * @return Statistics about the currently recorded elements
     */
    HistogramStatistics getStatistics();
}

Metrics

Class Metrics  is the core of metrics system, there are  MetricRegistry  and  MetricsReporter  container in it. When the Metrics  instance is initiating, the MetricRegistry  is instantiated and metrics reporters are started. 

Metrics reporters are configurable, users can use custom reporters, Paimon will provide a default metrics reporter of JMX metrics reporter. 

public class Metrics {

    /** The registry that holds the metrics. */
 	private final MetricRegistry registry

	interface HistogramStatistics {
    	/**
     	* Return the number of values that the statistics computation is based on.
     	*/
    	int size();

    	/** The metrics reporters container Returns the mean value of the histogram observations. */
  	private final List<MetricsReporter> reporters	double mean();

    	/** RegisterReturns metricsthe tostandard MetricRegistry. 
 		@param name The name of metric.
		@param metric The metric to register.	
	*/
	public void registerMetrics(String name, Metric metric) {}
}

MetricRegistry

MetricRegistry is a class responsible for metrics registering, there is a metrics container in it. It provides register method for each type of measurable metric, registering metrics will put metrics to the metrics container.

public class MetricRegistry {

	/** Map of gauge metrics. */
    private final Map<String, Gauge<?>> gauges = new HashMap<>();

    /** Map of counter metrics. */
    private final Map<String, Counter> counters = new HashMap<>();

	/** Register gauge metric. */
 	public void gauge(String name, Gauge gauge) {}

	/** Register counter metric. */
	public void counter(String name, Counter counter) {}
}

MetricsReporter

MetricsReporter  is used to report metrics to external backend, Paimon will implement an out of box reporter as JMX `MetricsReporter`.

public interface MetricsReporter {
	/** Configure reporter after instantiating it.*/
     void open();

    /** Closes this reporter. */
    void close();

	/** Report the current measurements. This method is called periodically by the Metrics. */
	void report();
}

Proposed Changes

Architecture

...

deviation of the histogram distribution. */
    	double stdDev();

    	/** Returns the maximum value of the histogram observations. */
    	long max();

    	/** Returns the minimum value of the histogram observations. */
    	long min();

    	/**
     	* Returns the percentile value based on the histogram statistics.
     	*
     	* @param percentile percentile point in double. E.g., 0.75 means 75 percentile. It is up to the
     	*     implementation to decide what valid percentile points are supported.
     	* @return Value for the given percentile
     	*/
    	long percentile(double percentile);
  } 
}

Metrics

Class Metrics  is the core of metrics system, there are  MetricGroup  and  MetricsReporter  container in it. When the Metrics  instance is initiating, the MetricGroup  is instantiated and metrics reporters are started. 

Metrics reporters are configurable, users can use custom reporters, Paimon will provide a default metrics reporter of JMX metrics reporter. 

public class Metrics {

    /** The metrics reporters container. */
  	private final List<MetricsReporter> reporters = new ArrayList<>();

	/** The metrics group. */  
	private final List<MetricGroup> metricGroups = new ArrayList<>();   

	/** Add a metrics reporter. */
	public void addReporter(MetricsReporter reporter) {};

	/** Add a metric group. */
	public void addGroup(MetricGroup group) {};

	/** Get metrics reporters. */
	public List<MetricsReporter> getReporters() {};

	/** Get metric groups. */
	public List<MetricGroup> getMetricGroups() {};
}

MetricGroup

MetricGroup is a class responsible for metrics registering and tagging, there is a metrics container in it. It provides register method for each type of measurable metric, registering metrics will put metrics to the metrics container.

public interface MetricGroup {
   /** Register gauge metric. */
 	public void gauge(String name, Gauge gauge) {}

	/** Register counter metric. */
	public void counter(String name, Counter counter) {}

	/** Register histogram metric. */
	public void histogram(String name, Histogram counter) {}
}

Paimon support to add tags for metric groups, tags is likely to scopes in Flink's metric group. The tagged metrics group will report metric value with tags as a part of prefix of the metric name. 

Table partition and bucket can be added as tag for some metric groups. Like the number of files monitoring in committing / scan / compaction, users can monitor the files of different buckets to check the data skew, and by checking the compaction duration of different buckets to see which unexpected bucket caused the long compaction duration.

MetricsReporter

MetricsReporter  is used to report metrics to external backend, Paimon will implement an out of box reporter as JMX MetricsReporter .

public interface MetricsReporter {
	/** Configure reporter after instantiating it.*/
     void open();

    /** Closes this reporter. */
    void close();

	/** Report the current measurements. This method is called periodically by the Metrics. */
	void report();
}

Proposed Changes

Architecture

Metrics Registering

Take CommitMetrics as example, the CommitMetrics will be instantiated by FileStoreCommitImpl, then commit related metrics will be registered by the corresponding MetricGroup in the Metrics instance. 

The Metrics has instance of MetricGroups and MetricsReporters set. MetricGroup maintains metrics map containers. Metrics registering is a process of putting metrics instances into the metric (gauge, counter) map container. 

Update metrics value

The CommitMetrics values will be updated around commit() operation, for example the commit starting time will be recorded before commit operation, CommitDuration value will be recorded after commit completing.

CompactionMetrics values will be updated around compaction operation, and ScanMetrics will be recorded through the scan planning process.

Report metrics to external backend

Each reporter instance has an timer task fetching metrics from the metrics containers periodically and report them out to the external backends. Paimon will has a default reporter backend with JMX, users can define their own MetricsReporter by implement MetricsReporter  interface. Here we can introduce a core option metrics.reporter  to specify a metrics backend.

Metrics list

We introduce CommitMetrics , CompactionMetrics , ScanMetrics as metrics set to measure the stats of Paimon table committing, compaction and scanning.

Common metrics

CommitMetrics

public class CommitMetrics {
	private final String LAST_COMMIT_DURATION_METRIC = "lastCommitDuration";
	...

    private final Map<Integer, BucketMetricGroup> bucketMetricGroups = new HashMap<>();
    private final Map<BinaryRow, PartitionMetricGroup> partitionMetricGroups = new HashMap<>();
    private final MetricGroup genericMetricGroup;

	private void registerGenericCommitMetrics() {   
		group.gauge(LAST_COMMIT_DURATION_METRIC, new CommitDurationGauge());
		...
	}

 	private void registerTaggedCommitMetrics() {
		if(isNewBucket()) {
			registerTaggedMetrics();
		}

	 	if(isNewBucket()) {
			registerTaggedMetrics();
		}
	}
}

CommitMetrics list includes commit duration, counter of files / records etc.

Metrics Registering

Take CommitMetrics as example, the CommitMetrics will be instantiated by FileStoreCommitImpl, then commit related metrics will be registered by MetricRegistry in singleton Metrics. 

The Metrics has instance of MetricRegistry and MetricsReporters set. MetricRegistry maintains metrics map containers. Metrics registering is a process of putting metrics instances into the metric (gauge, counter) map container. 

Update metrics value

The CommitMetrics values will be updated around commit() operation, for example the commit starting time will be recorded before commit operation, CommitDuration value will be recorded after commit completing.

CompactionMetrics values will be updated around compaction operation, and ScanMetrics will be recorded through the scan planning process.

Report metrics to external backend

Each reporter instance has an timer task fetching metrics from the metrics containers periodically and report them out to the external backends. Paimon will has a default reporter backend with JMX, users can define their own MetricsReporter by implement MetricsReporter  interface. Here we can introduce a core option metrics.reporter  to specify a metrics backend.

Metrics list

We introduce CommitMetrics , CompactionMetrics , ScanMetrics as metrics set to measure the stats of Paimon table committing, compaction and scanning.

Common metrics

CommitMetrics

public class CommitMetrics {
	private Metrics metrics;
	private final String COMMIT_DURATION_METRIC = "lastCommitDuration";
	...
	private void registerCommitMetrics(Metrics metrics) {
		metrics.gauge(COMMIT_DURATION_METRIC, new CommitDurationTimer());
		...
	}
	...
}

CommitMetrics list includes commit duration, counter of files / records etc.

ScanMetrics

public class ScanMetrics {
	private final String LAST_SCAN_SKIPPED_MANIFESTS_METRIC = "lastScanSkippedManifests";
	...
 	
	private final Map<Integer, BucketMetricGroup> bucketMetricGroups = new HashMap<>();
    private final Map<BinaryRow, PartitionMetricGroup> partitionMetricGroups = new HashMap<>();

 

public class ScanMetrics {
	private Metrics metrics;
	private final String SCAN_FILES_METRIC = "lastScanDuration";
	...
	private void registerScanMetricsregisterTaggedScanMetrics(Metrics metrics) {
		if(isNewBucket()) {
		metrics.counter(SCAN_FILES_METRIC, new ScanFilesCounter()	registerTaggedMetrics();
		}

	 	if(isNewPartition()) {
			registerTaggedMetrics();
		...}
	}
	...       
}

ScanMetrics list includes duration, data files and manifest files counter.

CompactionMetrics

public class CompactionMetrics {

	private Metrics metrics;
	private final String COMPACTED_FILES_METRIC = "lastCompactionDuration";
	...
 final String LAST_TABLE_FILES_COMPACTED_BEFORE_METRIC = "lastTableFilesCompactedBefore";
 	...
 	
	private final Map<Integer, BucketMetricGroup> bucketMetricGroups = new HashMap<>();
    private final Map<BinaryRow, PartitionMetricGroup> partitionMetricGroups = new HashMap<>();

 	private void registerCompactionMetricsregisterTaggedCompactionMetrics(Metrics metrics) {
		if(isNewBucket()) {
			metrics.counter(COMPACTED_FILES_METRIC, new CompactedFilesCounterregisterTaggedMetrics();
		}

	 	if(isNewPartition()); {
		...
	registerTaggedMetrics();
		}
	...}
}

CompactionMetrics list includes duration, and counter of files, sorted runs etc.

Flink connector metrics

Implement important source metrics in FLIP-33.

...