Production monitoring and observability

Genkit is built on OpenTelemetry. Every flow execution, model call, and tool invocation creates a span that records input, output, latency, and token usage. In production you export those spans to Google Cloud Trace, Cloud Monitoring, and Cloud Logging using the @genkit-ai/google-cloud plugin (or its language equivalents).

What is automatically traced

Genkit instruments the following operations without any extra code:

Span type	Recorded automatically
Flow execution	Input, output, duration, success/error
Model call	Model name, input/output (redacted in export), token usage, latency
Tool call	Tool name, invocation count
`ai.run()` step	Step name, duration

Each span carries a genkit:type attribute (flow, action, model, tool) and a genkit:path that shows the full call chain, for example /menuSuggestion/googleai/gemini-2.0-flash.

By default, model input and output content is redacted before export to Google Cloud (genkit:input and genkit:output are replaced with <redacted>). To retain them, set disableLoggingInputAndOutput: false (it defaults to false, meaning redaction is on during export). Check the option description in GcpTelemetryConfigOptions for the exact semantics before enabling this in production.

Node.js — `@genkit-ai/google-cloud`

Installation

npm install @genkit-ai/google-cloud

Configuration

import { enableGoogleCloudTelemetry } from '@genkit-ai/google-cloud';
import { genkit } from 'genkit';

// Call before constructing the Genkit instance.
await enableGoogleCloudTelemetry({
  // projectId is auto-detected from ADC / GOOGLE_CLOUD_PROJECT when omitted.
  projectId: 'my-gcp-project',

  // Sampling — AlwaysOnSampler is the default.
  // sampler: new TraceIdRatioBasedSampler(0.1), // 10 % sampling

  // Disable individual signals if not needed:
  // disableMetrics: false,
  // disableTraces: false,

  // Redact model I/O before exporting to Cloud Trace.
  // disableLoggingInputAndOutput: false,

  // Export intervals (Google Cloud requires >= 5000 ms).
  metricExportIntervalMillis: 60_000,
});

const ai = genkit({ plugins: [...] });

The plugin reads credentials from Application Default Credentials (ADC). On Cloud Run, GKE, or Cloud Functions no additional setup is needed. Outside of GCP, run gcloud auth application-default login or provide a credentials option.

Available options

Option	Type	Default	Description
`projectId`	`string`	ADC / env	Google Cloud project to export to.
`credentials`	`JWTInput`	ADC	Explicit credentials for non-GCP environments.
`sampler`	`Sampler`	AlwaysOn	OpenTelemetry trace sampler.
`autoInstrumentation`	`boolean`	`true`	Enable OTel auto-instrumentation for popular libraries.
`metricExportIntervalMillis`	`number`	`60000`	Metric export frequency (minimum 5000 ms).
`disableMetrics`	`boolean`	`false`	Skip metric export.
`disableTraces`	`boolean`	`false`	Skip trace export.
`disableLoggingInputAndOutput`	`boolean`	`false`	Redact model I/O in exported spans and logs.
`forceDevExport`	`boolean`	`false`	Export even when `GENKIT_ENV=dev` (for local testing).

Python — `genkit-google-cloud`

Install:

pip install genkit-google-cloud

Enable telemetry:

from genkit.plugins.google_cloud import GoogleCloud
from genkit import Genkit

ai = Genkit(
    plugins=[
        GoogleCloud(),   # reads GOOGLE_CLOUD_PROJECT from env
    ],
)

The Python google-cloud plugin follows the same trace/metric/log pipeline as the Node.js plugin.

Third-party telemetry backends (Python)

The genkit-observability community plugin provides exporters for Datadog, Honeycomb, Sentry, and other OpenTelemetry-compatible backends:

from genkit.plugins.observability.datadog import DatadogTelemetry

ai = Genkit(
    plugins=[
        DatadogTelemetry(api_key='DD_API_KEY'),
    ],
)

See the plugin README for the full list of supported backends.

Go — `googlecloud` plugin

Import:

import "github.com/firebase/genkit/go/plugins/googlecloud"

Enable telemetry before calling genkit.Init:

package main

import (
    "context"

    "github.com/firebase/genkit/go/genkit"
    "github.com/firebase/genkit/go/plugins/googlecloud"
    "github.com/firebase/genkit/go/plugins/googlegenai"
)

func main() {
    // Zero-config: project ID auto-detected from GOOGLE_CLOUD_PROJECT or ADC.
    googlecloud.EnableGoogleCloudTelemetry(nil)

    // With explicit configuration:
    // googlecloud.EnableGoogleCloudTelemetry(&googlecloud.GoogleCloudTelemetryOptions{
    //     ProjectID:      "my-project",
    //     DisableMetrics: false,
    //     DisableTraces:  false,
    // })

    ctx := context.Background()
    g := genkit.Init(ctx, genkit.WithPlugins(&googlegenai.GoogleAI{}))
    // ... define flows and start server
}

Go telemetry options

Option	Type	Default	Description
`ProjectID`	`string`	env	Google Cloud project.
`Credentials`	`*google.Credentials`	ADC	Explicit credentials.
`Sampler`	`sdktrace.Sampler`	AlwaysOn	Trace sampler.
`MetricExportIntervalMillis`	`*int`	300000 (prod)	Export frequency.
`DisableMetrics`	`bool`	`false`	Skip metric export.
`DisableTraces`	`bool`	`false`	Skip trace export.
`DisableLoggingInputAndOutput`	`bool`	`false`	Redact model I/O.
`ForceDevExport`	`bool`	`false`	Export in `GENKIT_ENV=dev`.

Firebase telemetry

For Firebase Cloud Functions deployments, use enableFirebaseTelemetry from @genkit-ai/firebase. It wraps enableGoogleCloudTelemetry and exports to the same Google Cloud Observability suite:

import { enableFirebaseTelemetry } from '@genkit-ai/firebase';

enableFirebaseTelemetry({
  // Accepts the same GcpTelemetryConfigOptions as enableGoogleCloudTelemetry.
  disableLoggingInputAndOutput: true,
});

Viewing traces in Google Cloud Console

Open Cloud Trace

Go to console.cloud.google.com/traces and select your project.

Filter by Genkit flow

In the Trace Explorer, filter by the genkit/feature attribute to find traces for a specific flow. Root spans carry genkit/isRoot: true.

Inspect model calls

Expand a trace to see child spans. Spans with genkit/metadata/subtype: model represent model calls. They include the model name (genkit/model) and — if disableLoggingInputAndOutput is false — the prompt and response.

View logs

Open Cloud Logging and filter by logName: genkit_log. Log entries are linked to their parent trace via the logging.googleapis.com/trace field.

Span attributes reference

The following genkit:* attributes are set on every span and normalised to genkit/* before export:

Attribute	Description
`genkit/type`	Span type: `flow`, `action`, `model`, `tool`, `flowStep`
`genkit/name`	Action or flow name
`genkit/path`	Full call path, e.g. `/myFlow/googleai/gemini-2.0-flash`
`genkit/isRoot`	`true` for the outermost flow span
`genkit/state`	`success` or `error`
`genkit/feature`	Feature name (set on root spans)
`genkit/model`	Model name (set on model spans)
`genkit/input`	Flow or model input (redacted during export)
`genkit/output`	Flow or model output (redacted during export)

Structured logging

The @genkit-ai/google-cloud plugin replaces the default logger with a Winston-based logger that:

Emits structured JSON to Cloud Logging (logName: genkit_log).
Attaches logging.googleapis.com/trace, trace_sampled, and spanId fields so log entries are correlated with traces in Cloud Trace.
Uses Pino and Winston OpenTelemetry instrumentations when autoInstrumentation: true.

import { logger } from 'genkit/logging';

logger.info('Processing request', { userId: 'u123' });
logger.warn('Rate limit approaching', { remaining: 10 });
logger.error('Model call failed', { model: 'gemini-2.0-flash', err });

Production monitoring best practices

Sample traces

Use TraceIdRatioBasedSampler at 10–20 % to control costs while retaining statistical signal. Always sample errors (ParentBasedSampler with an error filter).

Set up alerting

Create Cloud Monitoring alerting policies on:

genkit/feature/requests/count with genkit/status != success
genkit/generate/requests/count with high latency
Error rate above a threshold

Redact sensitive data

Set disableLoggingInputAndOutput: true (or leave at default) for flows that handle PII, financial data, or other sensitive information. Review what gets exported before enabling full input/output logging.

Flush before exit

For short-lived processes or Cloud Functions, flush pending spans before returning. The Go plugin calls googlecloud.FlushMetrics(ctx). In Node.js the BatchSpanProcessor flushes on SIGTERM.

Next steps

Deploy to Firebase

Firebase deployment guide with telemetry wiring.

Deploy to Cloud Run

Cloud Run deployment guide.

Plugins overview

Explore all official Genkit plugins.

Get Started

Core Concepts

Guides

Plugins

Deployment

Production monitoring and observability

What is automatically traced

Node.js — `@genkit-ai/google-cloud`

Installation

Configuration

Available options

Python — `genkit-google-cloud`

Third-party telemetry backends (Python)

Go — `googlecloud` plugin

Go telemetry options

Firebase telemetry

Viewing traces in Google Cloud Console

Span attributes reference

Structured logging

Production monitoring best practices

Sample traces

Set up alerting

Redact sensitive data

Flush before exit

Next steps

Deploy to Firebase

Deploy to Cloud Run

Plugins overview

Build docs developers (and LLMs) love

Get Started

Core Concepts

Guides

Plugins

Deployment

Documentation Index

​What is automatically traced

​Node.js — @genkit-ai/google-cloud

​Installation

​Configuration

​Available options

​Python — genkit-google-cloud

​Third-party telemetry backends (Python)

​Go — googlecloud plugin

​Go telemetry options

​Firebase telemetry

​Viewing traces in Google Cloud Console

​Span attributes reference

​Structured logging

​Production monitoring best practices

Sample traces

Set up alerting

Redact sensitive data

Flush before exit

​Next steps

Deploy to Firebase

Deploy to Cloud Run

Plugins overview

Build docs developers (and LLMs) love

What is automatically traced

Node.js — `@genkit-ai/google-cloud`

Installation

Configuration

Available options

Python — `genkit-google-cloud`

Third-party telemetry backends (Python)

Go — `googlecloud` plugin

Go telemetry options

Firebase telemetry

Viewing traces in Google Cloud Console

Span attributes reference

Structured logging

Production monitoring best practices

Next steps