The Cloud Bigtable HBase client for Java can collect client-side metrics that enable you to monitor Bigtable's performance. Other Bigtable client libraries do not provide client-side metrics. This page explains how to enable client-side metrics in the HBase client for Java and lists the metrics that are available.
The Cloud Bigtable HBase client for Java uses Dropwizard Metrics to collect and report client-side metrics. Because collecting metrics can add a very small amount of latency (single-digit microseconds) to each request, metrics are not enabled by default. The following sections explain how to enable client-side metrics.
Using the Log4j reporter
The simplest way to enable metrics is to add the following line to your Log4j configuration file:
This configuration setting turns on metrics collection and logs Bigtable metrics with an SLF4J logger.
Using other reporters
You can use other types of reporters by updating your application's code. For example, here is a reporter that sends metrics to a Graphite server:
Graphite pickledGraphite = new PickledGraphite(new InetSocketAddress("graphite.example.com", 2004)); DropwizardMetricRegistry registry = new DropwizardMetricRegistry(); GraphiteReporter reporter = GraphiteReporter.forRegistry(registry.getRegistry()) .convertRatesTo(TimeUnit.SECONDS) .convertDurationsTo(TimeUnit.MILLISECONDS) .filter(MetricFilter.ALL) .build(pickledGraphite); reporter.start(1, TimeUnit.MINUTES); BigtableClientMetrics.setMetricRegistry(registry);
This section describes the metrics that are available when client-side metrics are enabled. Each metric has one of the following types:
- Counter: A cumulative count per Java virtual machine (JVM).
- Meter: Count information plus throughput information (a count per the last minute, 5 minutes, or 15 minutes)
- Timer: Meter information plus latency information (such as median, mean, and 95th percentile)
Only certain metrics are collected for each type of request. See "Example:
Metrics for a
Put request" for an example of the metrics that
are collected during a
||How many lower-level gRPC/Netty channels are opened. Each
General RPC metrics
||The number of remote procedure calls (RPCs) that are currently active.|
Data method metrics
Data method metrics are collected for the following data methods:
ReadRows: Implements gets and scans.
MutateRow: Implements puts and deletes.
MutateRows: Implements bulk writes.
CheckAndMutateRow: Implements HBase's
ReadModifyWrite: Implements HBase's
SampleRowKeys: Retrieves region information that is used for MapReduce operations.
||The length of time that individual operations take. Operations include the total amount of latency of all RPCs that are performed. (Usually only one RPC is performed. Client-side retries can cause the same RPC to be performed more than once if there is a transient error.)|
||The length of time that it takes to receive the first response to a scan request.|
||The number of retries that were performed.|
||The number of non-retryable failures.|
||The number of times that retrying was aborted because too many retries had failed.|
Bulk metrics are provided for methods that return more than one response, such as a bulk write.
||The throughput of individual rows returned by a scan.|
||The throughput of individual mutations added for each
||The number of individual mutations retried over time.|
Bigtable table metrics
Converting Bigtable objects to HBase objects can add to the
latency of a request. The following timers can be correlated with the specified
*.operation.latency timers to measure the cost of the conversion.
||The length of time that individual
||The length of time that individual
Example: Metrics for a
When client-side metrics are enabled, the following metrics are collected for a
Put request that is not retried:
Collecting these metrics adds approximately 1 microsecond (1/1000 of a
millisecond) to the
Put operation. The overall
Put operation could be as
fast as 2 to 3 milliseconds, assuming that the operation includes about 1 KB of