Monitoring and Observability¶

You can monitor the health of your feature store pipelines, online serving endpoint, and stream ingestion directly in Snowsight. The monitoring view is available under AI & ML » Features and is organized into three tabs:

• Offline Refresh: Status of offline refresh jobs for all feature views in the feature store.
• Online Serving: Visualize metrics related to online serving, including latency percentiles, query volume, error rates for online feature retrieval.
• Stream Ingest: Visualize metrics related to ingestion for stream feature views, including throughput, latency, and concurrency for streaming feature pipelines.

Prerequisites¶

• A feature store must already exist. For more information, see Creating or connecting to a feature store.
• For Online Serving and Stream Ingest metrics, online service must be created and running. For more information, see Create the online service.
• Online Serving metrics are only shown when online serving is enabled for at least one feature view in the feature store.
• Stream Ingest metrics are only shown when at least one stream feature view exists in the feature store.
• A warehouse is required to run monitoring queries. You can select the warehouse from the warehouse selector at the top of each tab.
• Navigate to the Feature Store monitoring view:
  1. Sign in to Snowsight.
  2. In the navigation menu, select AI & ML » Features.
  3. Select a feature store from the dropdown.
  4. Select a sub-tab: Offline Refresh, Online Serving, or Stream Ingest.

Offline Refresh¶

The Offline Refresh tab lists offline refresh jobs for all feature views in the feature store. Each row represents one refresh job. By default, empty refreshes (refreshes with no changed rows) are hidden. Only refreshes with changed rows are displayed. To display empty refreshes, select Include empty refreshes.

Columns¶

The Offline Refresh tab includes the following columns:

Refresh status values¶

The Offline Refresh tab uses the following status values:

Online Serving¶

The Online Serving tab displays metrics for your online feature retrieval endpoint. Use this tab to monitor request volume, track latency against your SLA targets, and identify errors in serving.

Visualize metrics¶

Total Request

The total number of online feature retrieval requests over the selected time window, across all feature groups or for the selected feature group.

Errors (%)

Failed requests as a percentage of total requests. A value of 0% indicates that all requests are completing without errors. An increase in this value indicates errors in online serving.

Latency (ms)

Server-side latency in milliseconds for online feature retrieval requests, shown as three percentiles: p50, p90, and p99.

Filter by feature group¶

Use the Feature Group selector at the top of the tab to filter all charts to a specific feature group. Feature groups are shown in name$version format, for example user_features$v1. Select All to view aggregate metrics across all feature groups in the feature store.

Stream Ingest¶

The Stream Ingest tab shows the health and throughput of your streaming feature pipelines. Use this tab to verify that the stream ingest API is processing events at the expected rate, meeting latency targets, and handling concurrent request volume without errors.

Visualize metrics¶

Ingest API Responses Per Second

The total number of stream ingest API responses per second, indicating the throughput of incoming streaming events.

Ingest API Responses by HTTP Code

Response volume broken down by HTTP response code. Use this chart to detect elevated error rates or throttling. A spike in 429 responses indicates that the ingest rate is exceeding the provisioned limit.

Ingest API Latency

Server-side latency in milliseconds for stream ingest API requests, shown as four percentiles (p50, p90, p95, p99). This latency is used to measure the amount of time it takes to ingest streaming events into the feature store.

Concurrency

A metric for concurrent ingest load, derived by dividing QPS at the ingest endpoint by the number of provisioned compute nodes. The default number of provisioned compute nodes is 3, and the online service autoscales according to load. Snowflake scales up or scales down the number of nodes when aggregated CPU usage crosses an 80% threshold. For more information, see How CPU-based autoscaling works. Use this chart to identify bursts of concurrent requests that may cause throttling errors.

Diagnosing stream ingest issues¶

Use the following table to interpret common symptoms on the Stream Ingest tab: