...
Flink has defined a few standard metrics for jobs, tasks and operators. It also supports custom metrics in various scenarios. However, so far there is no standard or conventional metric definition for the connectors. Each connector defines their own metrics at the moment. This complicates operation and monitoring. Admittedly, different connectors may have different metrics, but some commonly used metrics can probably be standardized. This FLIP proposes a set of standard connector metrics that each connector should emit if applicable. The metrics proposed in this FLIP will serve as a convention for the connector implementations
In the future, the other projects in Flink ecosystem may rely on this metric convention. Therefore, the connector implementations are expected to follow the conventions when reporting the metrics.
Public Interfaces
We propose to introduce a set of conventional / standard metrics for the connectors.
...
Name | Type | Unit | Description |
numBytesIn | Counter | Bytes | The total number of input bytes since the source started |
numBytesInPerSec | Meter | Bytes/Sec | The input bytes per second |
numRecordsIn | Counter | Records | The total number of input records since the source started |
numRecordsInPerSec | Meter | Records/Sec | The input records per second |
numRecordsInErrors | Counter | Records | The total number of record that failed to consume |
recordSize* | Histogram | Bytes | The size of a record. |
currentFetchLatency | Gauge | ms | The latency occurred before Flink fetched the record. This metric is different from fetchLatency in that it is an instantaneous value recorded for the last processed record. This metric is provided because latency histogram could be expensive. The instantaneous latency value is usually a good enough indication of the latency. fetchLatency = FetchTime - EventTime |
currentLatency | Gauge | ms | The latency occurred before the record is emitted by the source connector. This metric is different from latency in that it is an instantaneous value recorded for the last processed record. This metric is provided because latency histogram could be expensive. The instantaneous latency value is usually a good enough indication of the latency. latency = EmitTime - EventTime |
fetchLatency* | Histogram | ms | The latency occurred before Flink fetched the record. fetchLatency = FetchTime - EventTime |
latency* | Histogram | ms | The latency occurred before the record is emitted by the source connector. latency = EmitTime - EventTime |
idleTime | Gauge | ms | The time in milliseconds that the source has not processed any record. idleTime = CurrentTime - LastRecordProcessTime |
...
Name | Type | Unit | Description |
numBytesOut | Counter | Bytes | The total number of output bytes since the source started |
numBytesOutPerSec | Meter | Bytes/Sec | The output bytes per second |
numRecordsOut | Counter | Records | The total number of output records since the source started |
numRecordsOutPerSec | Meter | Records/Sec | The output records per second |
numRecordsOutErrors | Counter | Record | The total number of records failed to send |
recordSize* | Histogram | Bytes | The size of a record |
currentSendTime | Gauge | ms | The time it takes to send the last record. |
sendTime* | Histogram | ms | The time it takes to send a record |
...
- A connector implementation does not have report all the following metrics. But the connectors that do report these metrics should conform to this convention.
- The histogram metrics are usually very expensive. Due to its performance impact, so it is strongly recommended that the connectors do not report them by default. But give the options to the , but allow users to enable opt them in on demand.
Scope
The metric group for each source and sink would be the same as ordinary operator scope, i.e. default to <host>.taskmanager.<tm_id>.<job_name>.<operator_name>.<subtask_index>
...