Endpoint-level metrics failed to generate baseline.

[km-smith986]

I have configured:

baseline.fetch.metrics=service_cpm,service_percentile,service_resp_time,endpoint_cpm,endpoint_percentile,endpoint_resp_time

While service_cpm, service_percentile, and service_resp_time can generate baseline data, endpoint_cpm, endpoint_percentile, and endpoint_resp_time cannot.

fetcher logs：

2025-12-26 11:24:03,826 - baseline.fetcher - INFO - Fetched 9 data points for service_percentile(service: test-gateway) from 2025122519 to 2025122603, original query time range(HOUR): 2025-12-26 03 to 2025-12-26 11
2025-12-26 11:24:03,826 - baseline.fetcher - INFO - Total fetched 89 data points for service_percentile(service: test-gateway)
2025-12-26 11:24:03,826 - baseline.fetcher - INFO - Total fetched 0 data points for endpoint_percentile(service: test-manager-gateway)
2025-12-26 11:24:03,830 - baseline.fetcher - INFO - Total fetched 0 data points for endpoint_resp_time(service: test-manager-gateway)
2025-12-26 11:24:03,831 - baseline.fetcher - INFO - Total fetched 0 data points for endpoint_cpm(service: test-manager-gateway)

The logs above show that the predictor did not obtain endpoint-level metric data from the OAP service.

How can I resolve this issue?

[wu-sheng]

@mrproliu do we support for endpoint baseline?

[mrproliu]

Nope, it only support service level metrics.

[wu-sheng]

@km-smith986 It is not supported.
We had concerns to generate all endpoints could overload the system easily.

[km-smith986]

Thanks for the clarification 👍
I understand that SkyPredictor currently only supports service-level baseline, and endpoint-level metrics are intentionally not included at the moment.

However, I’d like to share why endpoint-level baseline could be valuable in some real-world scenarios, and see whether this might be considered as a future enhancement.

Why endpoint-level baseline matters

In large-scale or gateway-style services, a single service often contains:

• dozens or even hundreds of endpoints
• very different traffic patterns and latency distributions per endpoint

In these cases, service-level baseline is often too coarse, and anomalies at critical endpoints (e.g. /pay, /withdraw, /callback) may be hidden by the overall service behavior.

Technical challenges (fully understood)

From a technical perspective, I understand the challenges very well:

• endpoint cardinality explosion
• sparse or zero-value time series
• higher query and compute cost

Possible constrained approaches

That said, endpoint baseline might still be feasible if it is:

• explicitly opt-in
• limited to top-N endpoints by traffic
• or scoped to manually configured endpoints

If endpoint baseline is out of scope

If endpoint-level baseline is intentionally out of scope for SkyPredictor, it would also be helpful to:

• document this limitation clearly
• or expose an extension point / plugin mechanism for custom predictors

Even a simple “won’t support endpoint baseline” note in the docs would help avoid confusion for new users like me.

Thanks again for the clarification and for maintaining SkyPredictor!

[wu-sheng]

If we want to do something like this, we are better to support explict API list of endpoint metrics.