[BUG] No easy way to change gRPC port on the client side

[HKozubek]

Describe the bug
Correct me if I'm wrong but currently there doesn't exists an easy way to change gRPC port on the client side of the app. The Collector Endpoint is either treated as HTTP by default, or as gRPC, but the port is forced to be 4317.

To Reproduce
Steps to reproduce the behavior:

1. This one defaults to HTTP, since when passing endpoint like this only the endpoint with port 4317 will be interpreted as gRPC

from phoenix.otel import register
from openinference.instrumentation.llama_index import LlamaIndexInstrumentor

tracer_provider = register(
  endpoint="http://localhost:8188",
)
LlamaIndexInstrumentor().instrument(tracer_provider=tracer_provider)

Check for gRPC port:

phoenix/packages/phoenix-otel/src/phoenix/otel/otel.py

Lines 390 to 393 in 8854a20

	def _maybe_grpc_endpoint(parsed_endpoint: ParseResult) -> bool:
	if not parsed_endpoint.path and parsed_endpoint.port == 4317:
	return True
	return False

2. This one will get the endpoint from env variable and will default to gRPC, but will change the port from 8188 to 4317

import os
os.environ["PHOENIX_COLLECTOR_ENDPOINT"] = "http://localhost:8188"

from phoenix.otel import register
from openinference.instrumentation.llama_index import LlamaIndexInstrumentor

tracer_provider = register()
LlamaIndexInstrumentor().instrument(tracer_provider=tracer_provider)

The code overwrites the port from env variable with _DEFAULT_GRPC_PORT variable

phoenix/packages/phoenix-otel/src/phoenix/otel/otel.py

Lines 415 to 416 in 8854a20

	def _construct_grpc_endpoint(parsed_endpoint: ParseResult) -> ParseResult:
	return parsed_endpoint._replace(netloc=f"{parsed_endpoint.hostname}:{_DEFAULT_GRPC_PORT}")

Expected behavior
The port shouldn't be overwritten, or there should be an easy way to define a port as gRPC using env variables

Environment (please complete the following information):

• OS: Ubuntu 22.04
• Browser: chrome
• Version: 7.5.2

Additional context
Currently the only way I found to change this behavior is to overwrite the "_maybe_grpc_endpoint" function

import os
os.environ["PHOENIX_COLLECTOR_ENDPOINT"] = "http://localhost:8188"

import phoenix.otel.otel

def _new_maybe_grpc_endpoint(parsed_endpoint) -> bool:
    if not parsed_endpoint.path and parsed_endpoint.port == 8188:
        return True
    return False

phoenix.otel.otel._maybe_grpc_endpoint = _new_maybe_grpc_endpoint

from openinference.instrumentation.llama_index import LlamaIndexInstrumentor

tracer_provider = phoenix.otel.otel.register()
LlamaIndexInstrumentor().instrument(tracer_provider=tracer_provider)

This will set the Collector endpoint to http://localhost:8188 and the type to gRPC

[cephalization]

Hey @HKozubek thanks for reaching out! Let me take a look at this and get back to you soon

[cephalization]

I was able to replicate this issue, and it does not quite work how I would expect either.

This is in fact quite suspicious.

phoenix/packages/phoenix-otel/src/phoenix/otel/otel.py

Lines 390 to 393 in 8854a20

	def _maybe_grpc_endpoint(parsed_endpoint: ParseResult) -> bool:
	if not parsed_endpoint.path and parsed_endpoint.port == 4317:
	return True
	return False

Any thoughts @axiomofjoy @RogerHYang ?

Here is a minimal reproduction

docker compose up -d

# docker-compose.yml
services:
  phoenix:
    image: arizephoenix/phoenix:latest
    ports:
      - 6006:6006
      - 4999:4999
    environment:
      - PHOENIX_GRPC_PORT=4999
      - PHOENIX_COLLECTOR_ENDPOINT=http://localhost:4999

export OPENAI_API_KEY="my key"; uv run instrument.py

# instrument.py
# /// script
# requires-python = ">=3.12"
# dependencies = [
#     "arize-phoenix",
#     "arize-phoenix-otel",
#     "llama-index",
#     "openinference-instrumentation-llama-index",
#     "opentelemetry-exporter-otlp",
#     "opentelemetry-proto>=1.12.0",
#     "opentelemetry-sdk",
# ]
# ///
from openinference.instrumentation.llama_index import LlamaIndexInstrumentor
from phoenix.otel import register
import os

# these match the env in the docker-compose.yml
os.environ["PHOENIX_COLLECTOR_ENDPOINT"] = "http://localhost:4999"
os.environ["PHOENIX_GRPC_PORT"] = "4999"

# without specifying the endpoint, it defaults to localhost:4317 despite documentation saying otherwise
tracer_provider = register(endpoint=os.environ["PHOENIX_COLLECTOR_ENDPOINT"])

LlamaIndexInstrumentor().instrument(tracer_provider=tracer_provider)

You can see the tracer provider emit the configured grpc endpoint, but with the transport configured as HTTP.

[RogerHYang]

As a workaround for the time being, you can fall back to raw otel by defining tracer_provider as follows.

import os

from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor

port = os.environ["PHOENIX_GRPC_PORT"]
endpoint = f"http://localhost:{port}"
tracer_provider = TracerProvider()
tracer_provider.add_span_processor(SimpleSpanProcessor(OTLPSpanExporter(endpoint)))

[mikeldking]

@HKozubek agree this is confusing and we should be respecting the GRPC config. We actually designed it so that you can bail out of the defaults and specify a gRPC exporter. https://github.com/Arize-ai/phoenix/blob/main/packages/phoenix-otel/README.md

Something like

from phoenix.otel import TracerProvider, BatchSpanProcessor, GRPCSpanExporter

tracer_provider = TracerProvider()
batch_processor = BatchSpanProcessor(
    span_exporter=GRPCSpanExporter(endpoint="http://custom-endpoint.com:6789")
)
tracer_provider.add_span_processor(batch_processor)

[anticorrelator]

hi @HKozubek thanks for pointing this out! We plan on allowing the endpoint to pick up on the PHOENIX_GRPC_PORT env var soon can you can track it's progress in this PR: #6017