dubin555 opened a new pull request, #27724:
URL: https://github.com/apache/flink/pull/27724
## What is the purpose of the change
The `flink-models` module (both `flink-model-triton` and
`flink-model-openai`) currently has no metric instrumentation. Users running
model inference in production have no visibility into request rates, error
rates, or latency — making it impossible to set up monitoring or alerting for
inference degradation.
This PR adds unified inference metrics to both Triton and OpenAI model
function base classes, following the same `MetricGroup` patterns used elsewhere
in Flink (e.g., `CachingAsyncLookupFunction`, `AsyncMLPredictRunner`).
Four metrics are registered under the `model_inference` group:
| Metric | Type | Description |
|--------|------|-------------|
| `inference_requests` | Counter | Total inference requests initiated |
| `inference_requests_success` | Counter | Successful inference completions |
| `inference_requests_failure` | Counter | Failed requests (network, HTTP,
parse errors) |
| `inference_latency_ms` | Gauge | Last inference round-trip time in ms |
## Brief change log
- Added metric fields and `registerMetrics()` call in
`AbstractTritonModelFunction.open()` so all Triton subclasses automatically get
metrics
- Instrumented all success/failure paths in
`TritonInferenceModelFunction.asyncPredict()` with counter increments and
latency tracking
- Added metric fields, registration, and `whenComplete()` instrumentation in
`AbstractOpenAIModelFunction.open()` / `asyncPredict()`
- Null inputs and context-overflow-skipped inputs in OpenAI are filtered
before incrementing request counts to avoid inflation
- Used a `volatile long` gauge for latency rather than histogram, since
`DescriptiveStatisticsHistogram` lives in `flink-runtime` which is not
available as a dependency in `flink-models`
## Verifying this change
This change added tests and can be verified as follows:
- Added `TritonInferenceMetricsTest` — integration test using MockWebServer
that verifies metrics are correctly registered and updated after successful
inference calls
- Added `OpenAIInferenceMetricsTest` — integration test using MockWebServer
that verifies chat inference metrics and null-input skip behavior
- Existing tests pass unchanged (no behavioral changes to inference logic)
## Does this pull request potentially affect one of the following parts:
- Dependencies (does it add or upgrade a dependency): no
- The public API, i.e., is any changed class annotated with
`@Public(Evolving)`: no
- The serializers: no
- The runtime per-record code paths (performance sensitive): no
- Anything that affects deployment or recovery: JobManager (and its
components), Checkpointing, Kubernetes/Yarn, ZooKeeper: no
- The S3 file system connector: no
## Documentation
- Does this pull request introduce a new feature? yes
- If yes, how is the feature documented? JavaDocs
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
