[jira] [Updated] (HBASE-4145) Provide metrics for hbase client

["Ming Ma (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/HBASE-4145?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Ming Ma updated HBASE-4145:
---------------------------

    Description: 
Sometimes it is useful to get some metrics from hbase client point of view. This will help understand the metrics for scan/TableInputFormat map job scenario.

What to capture, for example, for each ResultScanner object,

1. The number of RPC calls to RSs.
2. The delta time between consecutive RPC calls in the current serialized scan implementation.
3. The number of RPC retry to RSs.
4. The number of NotServingRegionException got.
5. The number of remote RPC calls. This excludes those call that hbase client calls the RS on the same machine.
6. The number of regions accessed.


How to capture

1. Metrics framework works for a fixed number of metrics. It doesn't fit this scenario.

2. Use some TBD solution in HBase to capture such dynamic metrics. If we assume there is a solution in HBase that HBase client can use to log such kind of metrics, TableInputFormat can pass in mapreduce task ID as application scan ID to HBase client as small addition to existing scan API; and HBase client can log metrics accordingly with such ID. That will allow query, analysis later on the metrics data for specific map reduce job.

3. Expose via MapReduce counter. It lacks certain features, for example, there is no good way to access the metrics on per map instance; the MapReduce framework only performs sum on the counter values so it is tricky to find the max of certain metrics in all mapper instances. However, it might be good enough for now. With this approach, the metrics value will be available via MapReduce counter.
a) Have ResultScanner return a new ResultScannerMetrics interface.
b) TableInputFormat will access data from ResultScannerMetrics and populate MapReduce counters accordingly.







  was:
Sometimes it is useful to get some metrics from hbase client point of view.

1. scan.next() call latency. Useful for regular map job scenario, where we want to see the latency from client side that includes RPC.
2. region server retry count. Some spike could indicate some issues in network or RS availability.

In the map reduce scenario, hbase client will run the same machines as RSs. So if the MetricsContext is usually available for hbase client. 




> Provide metrics for hbase client
> --------------------------------
>
>                 Key: HBASE-4145
>                 URL: https://issues.apache.org/jira/browse/HBASE-4145
>             Project: HBase
>          Issue Type: Improvement
>            Reporter: Ming Ma
>            Assignee: Ming Ma
>
> Sometimes it is useful to get some metrics from hbase client point of view. This will help understand the metrics for scan/TableInputFormat map job scenario.
> What to capture, for example, for each ResultScanner object,
> 1. The number of RPC calls to RSs.
> 2. The delta time between consecutive RPC calls in the current serialized scan implementation.
> 3. The number of RPC retry to RSs.
> 4. The number of NotServingRegionException got.
> 5. The number of remote RPC calls. This excludes those call that hbase client calls the RS on the same machine.
> 6. The number of regions accessed.
> How to capture
> 1. Metrics framework works for a fixed number of metrics. It doesn't fit this scenario.
> 2. Use some TBD solution in HBase to capture such dynamic metrics. If we assume there is a solution in HBase that HBase client can use to log such kind of metrics, TableInputFormat can pass in mapreduce task ID as application scan ID to HBase client as small addition to existing scan API; and HBase client can log metrics accordingly with such ID. That will allow query, analysis later on the metrics data for specific map reduce job.
> 3. Expose via MapReduce counter. It lacks certain features, for example, there is no good way to access the metrics on per map instance; the MapReduce framework only performs sum on the counter values so it is tricky to find the max of certain metrics in all mapper instances. However, it might be good enough for now. With this approach, the metrics value will be available via MapReduce counter.
> a) Have ResultScanner return a new ResultScannerMetrics interface.
> b) TableInputFormat will access data from ResultScannerMetrics and populate MapReduce counters accordingly.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira