Documentation preview for f7e2f52 is available at:

• https://pr-18345--mlflow-docs-preview.netlify.app/docs/latest/

@joelrobin18 we might want to think about creating a v2 autologging module for the scope of the breaking changes to simplify things here. We could do version validation handling (there are other autologging integrations where this has been done) to prevent having to complicate the maintainability with embedding large amounts of try/catch logic or conditional logic within a single implementation.

Hi @BenWilson2 Thank you for the feedbacks. Im trying to refactor the code to use lesser try/catch blocks as well as address the above comments as well.

Very much appreatiated if we could fully integrate v2.
imho it not worth trying to be compatible with v1 agno is updating very fast :)

Hi @ashdam Could you please add the below code at the top of your agent and check?

from opentelemetry import trace
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from openinference.instrumentation.agno import AgnoInstrumentor

# Configure OTLP to export to MLflow
exporter = OTLPSpanExporter(
    endpoint="http://localhost:5000/v1/traces",
    headers={"x-mlflow-experiment-id": "0"}
)

tracer_provider = TracerProvider()
tracer_provider.add_span_processor(BatchSpanProcessor(exporter))
trace.set_tracer_provider(tracer_provider)

AgnoInstrumentor().instrument()

@joelrobin18 i have tested it.

Not really an expert on MLFlow but currenly im using MLflow 3.6.0 OSS with PostgreSQL backend with Agno v2.2.6

Its agno is capturing calls but it raises the following error

i got this error in agno-os:

{"TimeStamp": "2025-11-12T16:55:42.7841868+00:00", "Log": "error Not Implemented encountered while exporting span batch, retrying in 0.96s."}
{"TimeStamp": "2025-11-12T16:55:43.74478+00:00", "Log": "error Not Implemented encountered while exporting span batch, retrying in 1.63s."}
{"TimeStamp": "2025-11-12T16:55:45.3754366+00:00", "Log": "error Not Implemented encountered while exporting span batch, retrying in 3.62s."}
{"TimeStamp": "2025-11-12T16:55:49.0009375+00:00", "Log": "to export span batch code: 501, reason: {\"detail\":\"REST OTLP span logging is not supported by FileStore\"}"}

It looks like we are using FileStore as backend here. Can you share a simple repo code for the same?

Hi @joelrobin18,

Thanks for looking into this! Yes, I'm definitely using PostgreSQL as the backend store, not FileStore.
I would like to add that im not python expert and I am using Sonnet 4.5 to help me out :P

Here's the configuration:

MLflow Server Setup

Deployment: Azure Container Apps running MLflow v3.6.0
Start Command:

mlflow server \
  --host 0.0.0.0 \
  --port 5000 \
  --backend-store-uri postgresql://agnoadmin:****@psql-agno-storage.postgres.database.azure.com:5432/mlflow_db?sslmode=require \
  --default-artifact-root wasbs://mlflow-artifacts@saagenticaifinancedemo.blob.core.windows.net/ \
  --serve-artifacts

Verified Backend Configuration:

$ az containerapp show --name mlflow-server --query "properties.template.containers[0].env"
[
  {
    "name": "MLFLOW_BACKEND_STORE_URI",
    "secretRef": "postgres-uri"  # Points to PostgreSQL connection string
  },
  ...
]

Agno v2.2.6 Integration Code

Following your recommendation, I'm using OpenInference AgnoInstrumentor:

from opentelemetry import trace
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from openinference.instrumentation.agno import AgnoInstrumentor

mlflow_tracking_uri = os.getenv("MLFLOW_TRACKING_URI")
mlflow_experiment_id = os.getenv("MLFLOW_EXPERIMENT_ID", "0")

exporter = OTLPSpanExporter(
    endpoint=f"{mlflow_tracking_uri}/v1/traces",
    headers={"x-mlflow-experiment-id": mlflow_experiment_id},
)

tracer_provider = TracerProvider()
tracer_provider.add_span_processor(BatchSpanProcessor(exporter))
trace.set_tracer_provider(tracer_provider)

AgnoInstrumentor().instrument()

Dependencies:

• agno==2.2.6
• mlflow==3.6.0
• openinference-instrumentation-agno==0.1.22
• opentelemetry-sdk==1.38.0
• opentelemetry-exporter-otlp-proto-http==1.38.0

The Issue

✅ What's Working:

• OpenInference AgnoInstrumentor successfully hooks into Agno v2.2.6 Team/Model calls
• Spans are being generated and batched
• OTLP exporter attempts to send to MLflow

❌ The Error:

error Not Implemented encountered while exporting span batch, retrying in 0.96s.
error Not Implemented encountered while exporting span batch, retrying in 1.63s.
error Not Implemented encountered while exporting span batch, retrying in 3.62s.
Failed to export span batch code: 501, reason: {"detail":"REST OTLP span logging is not supported by FileStore"}

From Agent Execution Logs:
The instrumentation is definitely working - I can see the OpenInference wrappers in the stack traces:

File "/app/.venv/lib/python3.12/site-packages/openinference/instrumentation/agno/_runs_wrapper.py", line 512, in arun_stream
    async for response in wrapped(*args, **kwargs):
File "/app/.venv/lib/python3.12/site-packages/agno/team/team.py", line 2452, in _arun_stream
    async for event in self._ahandle_model_response_stream(
...
File "/app/.venv/lib/python3.12/site-packages/openinference/instrumentation/agno/_model_wrapper.py", line 493, in arun_stream
    async for chunk in wrapped(*args, **kwargs):

The Paradox

MLflow is configured with PostgreSQL backend (--backend-store-uri postgresql://...), but the 501 error indicates traces are still using FileStore. This suggests MLflow 3.6.0 has a separate trace storage layer that defaults to FileStore even when the main backend is PostgreSQL.

Is this a known limitation, or is there a missing configuration flag for enabling database trace storage?

I'm happy to provide a minimal reproducible repo if that helps debug this further!

Thank you very much for your work @joelrobin18 . this is highly anticipated in my company :)

@BenWilson2 @joelrobin18 any news? :)

Is this a known limitation, or is there a missing configuration flag for enabling database trace storage?

@ashdam I can see Agno traces logged successfuly via otel on my local, could not reproduce the error. Could you double check if the tracking URI points to the correct MLflow instance. The error message indicates the backend is actually file store.

You can also test it locally to see if this is related to your azure app container settings or not.

pip install mlflow==3.6.0
mlflow ui --backend-store-uri sqlite:///mlruns.db 

Yes, we only have 1 server of MLFlow and it had postgreSQL configured. Its really hard (devops + security) for me to replicate everything , including agno, in local tbh :(

@B-Step62 @joelrobin18 Thank you guys for your work :)

@ashdam I still believe what happens inside the app container is that MLflow is started with file store. The error message incldues the class name of the store and

If the store is configured to SQL properly, you should see server logs like this

Registry store URI not provided. Using sqlite:///mlruns.db
2025/11/21 20:56:58 INFO mlflow.store.db.utils: Creating initial MLflow database tables...
2025/11/21 20:56:58 INFO mlflow.store.db.utils: Updating database tables

One common gotcha is that the multi-line commands are not properly formatted in the YAML file and it only runs the first line mlflow ui, which defaults to a file store. If that happens, you should see a server log like this instead:

Backend store URI not provided. Using ./mlruns
Registry store URI not provided. Using ./mlruns
.../server/handlers.py:258: FutureWarning: The filesystem tracking backend (e.g., './mlruns') will be deprecated in February 2026. Consider transitioning to a database backend (e.g., 'sqlite:///mlflow.db') to take advantage of the latest MLflow features. See https://github.com/mlflow/mlflow/issues/18534 for more details and migration guidance.
  return FileStore(store_uri, artifact_uri)

For example, this works

    command:
      - /bin/bash
      - -c
      - |
        mlflow server \
            --backend-store-uri postgresql://... \
            --port 5000

but this does not work (only mlflow server will be executed).

    command: >
      /bin/bash -c "
        mlflow server \
            --backend-store-uri postgresql://... \
            --port 5000
        "

Thank you for the help :) I will test it next week :D thank you!