phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-16 22:32:38 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

feat(telemetry:major): End to End Testing, Metric Capture, SQL Alchemy Injection

Browse source
This commit is contained in:
Emilio Garcia 2025-10-03 12:17:41 -04:00
parent e815738936
commit 7e3cf1fb20
26 changed files with 2075 additions and 1006 deletions
Download patch file Download diff file Expand all files Collapse all files

5
tests/integration/telemetry/__init__.py Normal file
View file

@ -0,0 +1,5 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.

149
tests/integration/telemetry/mocking/README.md Normal file
View file

@ -0,0 +1,149 @@
# Mock Server Infrastructure
This directory contains mock servers for E2E telemetry testing.
## Structure
```
mocking/
├── README.md ← You are here
├── __init__.py ← Module exports
├── mock_base.py ← Pydantic base class for all mocks
├── servers.py ← Mock server implementations
└── harness.py ← Async startup harness
```
## Files
### `mock_base.py` - Base Class
Pydantic base model that all mock servers must inherit from.
**Contract:**
```python
class MockServerBase(BaseModel):
async def await_start(self):
# Start server and wait until ready
...
def stop(self):
# Stop server and cleanup
...
```
### `servers.py` - Mock Implementations
Contains:
- **MockOTLPCollector** - Receives OTLP telemetry (port 4318)
- **MockVLLMServer** - Simulates vLLM inference API (port 8000)
### `harness.py` - Startup Orchestration
Provides:
- **MockServerConfig** - Pydantic config for server registration
- **start_mock_servers_async()** - Starts servers in parallel
- **stop_mock_servers()** - Stops all servers
## Creating a New Mock Server
### Step 1: Implement the Server
Add to `servers.py`:
```python
class MockRedisServer(MockServerBase):
"""Mock Redis server."""
port: int = Field(default=6379)
# Non-Pydantic fields
server: Any = Field(default=None, exclude=True)
def model_post_init(self, __context):
self.server = None
async def await_start(self):
"""Start Redis mock and wait until ready."""
# Start your server
self.server = create_redis_server(self.port)
self.server.start()
# Wait for port to be listening
for _ in range(10):
sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
if sock.connect_ex(("localhost", self.port)) == 0:
sock.close()
return # Ready!
await asyncio.sleep(0.1)
def stop(self):
if self.server:
self.server.stop()
```
### Step 2: Register in Test
In `test_otel_e2e.py`, add to MOCK_SERVERS list:
```python
MOCK_SERVERS = [
# ... existing servers ...
MockServerConfig(
name="Mock Redis",
server_class=MockRedisServer,
init_kwargs={"port": 6379},
),
]
```
### Step 3: Done!
The harness automatically:
- Creates the server instance
- Calls `await_start()` in parallel with other servers
- Returns when all are ready
- Stops all servers on teardown
## Benefits
✅ **Parallel Startup** - All servers start simultaneously
✅ **Type-Safe** - Pydantic validation
✅ **Simple** - Just implement 2 methods
✅ **Fast** - No HTTP polling, direct port checking
✅ **Clean** - Async/await pattern
## Usage in Tests
```python
@pytest.fixture(scope="module")
def mock_servers():
servers = asyncio.run(start_mock_servers_async(MOCK_SERVERS))
yield servers
stop_mock_servers(servers)
# Access specific servers
@pytest.fixture(scope="module")
def mock_redis(mock_servers):
return mock_servers["Mock Redis"]
```
## Key Design Decisions
### Why Pydantic?
- Type safety for server configuration
- Built-in validation
- Clear interface contract
### Why `await_start()` instead of HTTP `/ready`?
- Faster (no HTTP round-trip)
- Simpler (direct port checking)
- More reliable (internal state, not external endpoint)
### Why separate harness?
- Reusable across different test files
- Easy to add new servers
- Centralized error handling
## Examples
See `test_otel_e2e.py` for real-world usage:
- Line ~200: MOCK_SERVERS configuration
- Line ~230: Convenience fixtures
- Line ~240: Using servers in tests

28
tests/integration/telemetry/mocking/__init__.py Normal file
View file

@ -0,0 +1,28 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
"""
Mock server infrastructure for telemetry E2E testing.
This module provides:
- MockServerBase: Pydantic base class for all mock servers
- MockOTLPCollector: Mock OTLP telemetry collector
- MockVLLMServer: Mock vLLM inference server
- Mock server harness for parallel async startup
"""
from .harness import MockServerConfig, start_mock_servers_async, stop_mock_servers
from .mock_base import MockServerBase
from .servers import MockOTLPCollector, MockVLLMServer
__all__ = [
"MockServerBase",
"MockOTLPCollector",
"MockVLLMServer",
"MockServerConfig",
"start_mock_servers_async",
"stop_mock_servers",
]

106
tests/integration/telemetry/mocking/harness.py Normal file
View file

@ -0,0 +1,106 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
"""
Mock server startup harness for parallel initialization.
HOW TO ADD A NEW MOCK SERVER:
1. Import your mock server class
2. Add it to MOCK_SERVERS list with configuration
3. Done! It will start in parallel with others.
"""
import asyncio
from typing import Any
from pydantic import BaseModel, Field
from .mock_base import MockServerBase
class MockServerConfig(BaseModel):
"""
Configuration for a mock server to start.
**TO ADD A NEW MOCK SERVER:**
Just create a MockServerConfig instance with your server class.
Example:
MockServerConfig(
name="Mock MyService",
server_class=MockMyService,
init_kwargs={"port": 9000, "config_param": "value"},
)
"""
model_config = {"arbitrary_types_allowed": True}
name: str = Field(description="Display name for logging")
server_class: type = Field(description="Mock server class (must inherit from MockServerBase)")
init_kwargs: dict[str, Any] = Field(default_factory=dict, description="Kwargs to pass to server constructor")
async def start_mock_servers_async(mock_servers_config: list[MockServerConfig]) -> dict[str, MockServerBase]:
"""
Start all mock servers in parallel and wait for them to be ready.
**HOW IT WORKS:**
1. Creates all server instances
2. Calls await_start() on all servers in parallel
3. Returns when all are ready
**SIMPLE TO USE:**
servers = await start_mock_servers_async([config1, config2, ...])
Args:
mock_servers_config: List of mock server configurations
Returns:
Dict mapping server name to server instance
"""
servers = {}
start_tasks = []
# Create all servers and prepare start tasks
for config in mock_servers_config:
server = config.server_class(**config.init_kwargs)
servers[config.name] = server
start_tasks.append(server.await_start())
# Start all servers in parallel
try:
await asyncio.gather(*start_tasks)
# Print readiness confirmation
for name in servers.keys():
print(f"[INFO] {name} ready")
except Exception as e:
# If any server fails, stop all servers
for server in servers.values():
try:
server.stop()
except Exception:
pass
raise RuntimeError(f"Failed to start mock servers: {e}") from None
return servers
def stop_mock_servers(servers: dict[str, Any]):
"""
Stop all mock servers.
Args:
servers: Dict of server instances from start_mock_servers_async()
"""
for name, server in servers.items():
try:
if hasattr(server, "get_request_count"):
print(f"\n[INFO] {name} received {server.get_request_count()} requests")
server.stop()
except Exception as e:
print(f"[WARN] Error stopping {name}: {e}")

68
tests/integration/telemetry/mocking/mock_base.py Normal file
View file

@ -0,0 +1,68 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
"""
Base class for mock servers with async startup support.
All mock servers should inherit from MockServerBase and implement await_start().
"""
from abc import abstractmethod
from pydantic import BaseModel
class MockServerBase(BaseModel):
"""
Pydantic base model for mock servers.
**TO CREATE A NEW MOCK SERVER:**
1. Inherit from this class
2. Implement async def await_start(self)
3. Implement def stop(self)
4. Done!
Example:
class MyMockServer(MockServerBase):
port: int = 8080
async def await_start(self):
# Start your server
self.server = create_server()
self.server.start()
# Wait until ready (can check internal state, no HTTP needed)
while not self.server.is_listening():
await asyncio.sleep(0.1)
def stop(self):
if self.server:
self.server.stop()
"""
model_config = {"arbitrary_types_allowed": True}
@abstractmethod
async def await_start(self):
"""
Start the server and wait until it's ready.
This method should:
1. Start the server (synchronous or async)
2. Wait until the server is fully ready to accept requests
3. Return when ready
Subclasses can check internal state directly - no HTTP polling needed!
"""
...
@abstractmethod
def stop(self):
"""
Stop the server and clean up resources.
This method should gracefully shut down the server.
"""
...

600
tests/integration/telemetry/mocking/servers.py Normal file
View file

@ -0,0 +1,600 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
"""
Mock servers for OpenTelemetry E2E testing.
This module provides mock servers for testing telemetry:
- MockOTLPCollector: Receives and stores OTLP telemetry exports
- MockVLLMServer: Simulates vLLM inference API with valid OpenAI responses
These mocks allow E2E testing without external dependencies.
"""
import asyncio
import http.server
import json
import socket
import threading
import time
from collections import defaultdict
from typing import Any
from opentelemetry.proto.collector.metrics.v1.metrics_service_pb2 import (
ExportMetricsServiceRequest,
)
from opentelemetry.proto.collector.trace.v1.trace_service_pb2 import (
ExportTraceServiceRequest,
)
from pydantic import Field
from .mock_base import MockServerBase
class MockOTLPCollector(MockServerBase):
"""
Mock OTLP collector HTTP server.
Receives real OTLP exports from Llama Stack and stores them for verification.
Runs on localhost:4318 (standard OTLP HTTP port).
Usage:
collector = MockOTLPCollector()
await collector.await_start()
# ... run tests ...
print(f"Received {collector.get_trace_count()} traces")
collector.stop()
"""
port: int = Field(default=4318, description="Port to run collector on")
# Non-Pydantic fields (set after initialization)
traces: list[dict] = Field(default_factory=list, exclude=True)
metrics: list[dict] = Field(default_factory=list, exclude=True)
all_http_requests: list[dict] = Field(default_factory=list, exclude=True) # Track ALL HTTP requests for debugging
server: Any = Field(default=None, exclude=True)
server_thread: Any = Field(default=None, exclude=True)
def model_post_init(self, __context):
"""Initialize after Pydantic validation."""
self.traces = []
self.metrics = []
self.server = None
self.server_thread = None
def _create_handler_class(self):
"""Create the HTTP handler class for this collector instance."""
collector_self = self
class OTLPHandler(http.server.BaseHTTPRequestHandler):
"""HTTP request handler for OTLP requests."""
def log_message(self, format, *args):
"""Suppress HTTP server logs."""
pass
def do_GET(self): # noqa: N802
"""Handle GET requests."""
# No readiness endpoint needed - using await_start() instead
self.send_response(404)
self.end_headers()
def do_POST(self): # noqa: N802
"""Handle OTLP POST requests."""
content_length = int(self.headers.get("Content-Length", 0))
body = self.rfile.read(content_length) if content_length > 0 else b""
# Track ALL requests for debugging
collector_self.all_http_requests.append(
{
"method": "POST",
"path": self.path,
"timestamp": time.time(),
"body_length": len(body),
}
)
# Store the export request
if "/v1/traces" in self.path:
collector_self.traces.append(
{
"body": body,
"timestamp": time.time(),
}
)
elif "/v1/metrics" in self.path:
collector_self.metrics.append(
{
"body": body,
"timestamp": time.time(),
}
)
# Always return success (200 OK)
self.send_response(200)
self.send_header("Content-Type", "application/json")
self.end_headers()
self.wfile.write(b"{}")
return OTLPHandler
async def await_start(self):
"""
Start the OTLP collector and wait until ready.
This method is async and can be awaited to ensure the server is ready.
"""
# Create handler and start the HTTP server
handler_class = self._create_handler_class()
self.server = http.server.HTTPServer(("localhost", self.port), handler_class)
self.server_thread = threading.Thread(target=self.server.serve_forever, daemon=True)
self.server_thread.start()
# Wait for server to be listening on the port
for _ in range(10):
try:
sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
result = sock.connect_ex(("localhost", self.port))
sock.close()
if result == 0:
# Port is listening
return
except Exception:
pass
await asyncio.sleep(0.1)
raise RuntimeError(f"OTLP collector failed to start on port {self.port}")
def stop(self):
"""Stop the OTLP collector server."""
if self.server:
self.server.shutdown()
self.server.server_close()
def clear(self):
"""Clear all captured telemetry data."""
self.traces = []
self.metrics = []
def get_trace_count(self) -> int:
"""Get number of trace export requests received."""
return len(self.traces)
def get_metric_count(self) -> int:
"""Get number of metric export requests received."""
return len(self.metrics)
def get_all_traces(self) -> list[dict]:
"""Get all captured trace exports."""
return self.traces
def get_all_metrics(self) -> list[dict]:
"""Get all captured metric exports."""
return self.metrics
# -----------------------------
# Trace parsing helpers
# -----------------------------
def parse_traces(self) -> dict[str, list[dict]]:
"""
Parse protobuf trace data and return spans grouped by trace ID.
Returns:
Dict mapping trace_id (hex) -> list of span dicts
"""
trace_id_to_spans: dict[str, list[dict]] = {}
for export in self.traces:
request = ExportTraceServiceRequest()
body = export.get("body", b"")
try:
request.ParseFromString(body)
except Exception as e:
raise RuntimeError(f"Failed to parse OTLP traces export (len={len(body)}): {e}") from e
for resource_span in request.resource_spans:
for scope_span in resource_span.scope_spans:
for span in scope_span.spans:
# span.trace_id is bytes; convert to hex string
trace_id = (
span.trace_id.hex() if isinstance(span.trace_id, bytes | bytearray) else str(span.trace_id)
)
span_entry = {
"name": span.name,
"span_id": span.span_id.hex()
if isinstance(span.span_id, bytes | bytearray)
else str(span.span_id),
"start_time_unix_nano": int(getattr(span, "start_time_unix_nano", 0)),
"end_time_unix_nano": int(getattr(span, "end_time_unix_nano", 0)),
}
trace_id_to_spans.setdefault(trace_id, []).append(span_entry)
return trace_id_to_spans
def get_all_trace_ids(self) -> set[str]:
"""Return set of all trace IDs seen so far."""
return set(self.parse_traces().keys())
def get_trace_span_counts(self) -> dict[str, int]:
"""Return span counts per trace ID."""
grouped = self.parse_traces()
return {tid: len(spans) for tid, spans in grouped.items()}
def get_new_trace_ids(self, prior_ids: set[str]) -> set[str]:
"""Return trace IDs that appeared after prior_ids snapshot."""
return self.get_all_trace_ids() - set(prior_ids)
def parse_metrics(self) -> dict[str, list[Any]]:
"""
Parse protobuf metric data and return metrics by name.
Returns:
Dict mapping metric names to list of metric data points
"""
metrics_by_name = defaultdict(list)
for export in self.metrics:
# Parse the protobuf body
request = ExportMetricsServiceRequest()
body = export.get("body", b"")
try:
request.ParseFromString(body)
except Exception as e:
raise RuntimeError(f"Failed to parse OTLP metrics export (len={len(body)}): {e}") from e
# Extract metrics from the request
for resource_metric in request.resource_metrics:
for scope_metric in resource_metric.scope_metrics:
for metric in scope_metric.metrics:
metric_name = metric.name
# Extract data points based on metric type
data_points = []
if metric.HasField("gauge"):
data_points = list(metric.gauge.data_points)
elif metric.HasField("sum"):
data_points = list(metric.sum.data_points)
elif metric.HasField("histogram"):
data_points = list(metric.histogram.data_points)
elif metric.HasField("summary"):
data_points = list(metric.summary.data_points)
metrics_by_name[metric_name].extend(data_points)
return dict(metrics_by_name)
def get_metric_by_name(self, metric_name: str) -> list[Any]:
"""
Get all data points for a specific metric by name.
Args:
metric_name: The name of the metric to retrieve
Returns:
List of data points for the metric, or empty list if not found
"""
metrics = self.parse_metrics()
return metrics.get(metric_name, [])
def has_metric(self, metric_name: str) -> bool:
"""
Check if a metric with the given name has been captured.
Args:
metric_name: The name of the metric to check
Returns:
True if the metric exists and has data points, False otherwise
"""
data_points = self.get_metric_by_name(metric_name)
return len(data_points) > 0
def get_all_metric_names(self) -> list[str]:
"""
Get all unique metric names that have been captured.
Returns:
List of metric names
"""
return list(self.parse_metrics().keys())
class MockVLLMServer(MockServerBase):
"""
Mock vLLM inference server with OpenAI-compatible API.
Returns valid OpenAI Python client response objects for:
- Chat completions (/v1/chat/completions)
- Text completions (/v1/completions)
- Model listing (/v1/models)
Runs on localhost:8000 (standard vLLM port).
Usage:
server = MockVLLMServer(models=["my-model"])
await server.await_start()
# ... make inference calls ...
print(f"Handled {server.get_request_count()} requests")
server.stop()
"""
port: int = Field(default=8000, description="Port to run server on")
models: list[str] = Field(
default_factory=lambda: ["meta-llama/Llama-3.2-1B-Instruct"], description="List of model IDs to serve"
)
# Non-Pydantic fields
requests_received: list[dict] = Field(default_factory=list, exclude=True)
server: Any = Field(default=None, exclude=True)
server_thread: Any = Field(default=None, exclude=True)
def model_post_init(self, __context):
"""Initialize after Pydantic validation."""
self.requests_received = []
self.server = None
self.server_thread = None
def _create_handler_class(self):
"""Create the HTTP handler class for this vLLM instance."""
server_self = self
class VLLMHandler(http.server.BaseHTTPRequestHandler):
"""HTTP request handler for vLLM API."""
def log_message(self, format, *args):
"""Suppress HTTP server logs."""
pass
def log_request(self, code="-", size="-"):
"""Log incoming requests for debugging."""
print(f"[DEBUG] Mock vLLM received: {self.command} {self.path} -> {code}")
def do_GET(self): # noqa: N802
"""Handle GET requests (models list, health check)."""
# Log GET requests too
server_self.requests_received.append(
{
"path": self.path,
"method": "GET",
"timestamp": time.time(),
}
)
if self.path == "/v1/models":
response = self._create_models_list_response()
self._send_json_response(200, response)
elif self.path == "/health" or self.path == "/v1/health":
self._send_json_response(200, {"status": "healthy"})
else:
self.send_response(404)
self.end_headers()
def do_POST(self): # noqa: N802
"""Handle POST requests (chat/text completions)."""
content_length = int(self.headers.get("Content-Length", 0))
body = self.rfile.read(content_length) if content_length > 0 else b"{}"
try:
request_data = json.loads(body)
except Exception:
request_data = {}
# Log the request
server_self.requests_received.append(
{
"path": self.path,
"body": request_data,
"timestamp": time.time(),
}
)
# Route to appropriate handler
if "/chat/completions" in self.path:
response = self._create_chat_completion_response(request_data)
if response is not None: # None means already sent (streaming)
self._send_json_response(200, response)
elif "/completions" in self.path:
response = self._create_text_completion_response(request_data)
self._send_json_response(200, response)
else:
self._send_json_response(200, {"status": "ok"})
# ----------------------------------------------------------------
# Response Generators
# **TO MODIFY RESPONSES:** Edit these methods
# ----------------------------------------------------------------
def _create_models_list_response(self) -> dict:
"""Create OpenAI models list response with configured models."""
return {
"object": "list",
"data": [
{
"id": model_id,
"object": "model",
"created": int(time.time()),
"owned_by": "meta",
}
for model_id in server_self.models
],
}
def _create_chat_completion_response(self, request_data: dict) -> dict | None:
"""
Create OpenAI ChatCompletion response.
Returns a valid response matching openai.types.ChatCompletion.
Supports both regular and streaming responses.
Returns None for streaming responses (already sent via SSE).
"""
# Check if streaming is requested
is_streaming = request_data.get("stream", False)
if is_streaming:
# Return SSE streaming response
self.send_response(200)
self.send_header("Content-Type", "text/event-stream")
self.send_header("Cache-Control", "no-cache")
self.send_header("Connection", "keep-alive")
self.end_headers()
# Send streaming chunks
chunks = [
{
"id": "chatcmpl-test",
"object": "chat.completion.chunk",
"created": int(time.time()),
"model": request_data.get("model", "test"),
"choices": [{"index": 0, "delta": {"role": "assistant"}, "finish_reason": None}],
},
{
"id": "chatcmpl-test",
"object": "chat.completion.chunk",
"created": int(time.time()),
"model": request_data.get("model", "test"),
"choices": [{"index": 0, "delta": {"content": "Test "}, "finish_reason": None}],
},
{
"id": "chatcmpl-test",
"object": "chat.completion.chunk",
"created": int(time.time()),
"model": request_data.get("model", "test"),
"choices": [{"index": 0, "delta": {"content": "streaming "}, "finish_reason": None}],
},
{
"id": "chatcmpl-test",
"object": "chat.completion.chunk",
"created": int(time.time()),
"model": request_data.get("model", "test"),
"choices": [{"index": 0, "delta": {"content": "response"}, "finish_reason": None}],
},
{
"id": "chatcmpl-test",
"object": "chat.completion.chunk",
"created": int(time.time()),
"model": request_data.get("model", "test"),
"choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}],
},
]
for chunk in chunks:
self.wfile.write(f"data: {json.dumps(chunk)}\n\n".encode())
self.wfile.write(b"data: [DONE]\n\n")
return None # Already sent response
# Regular response
return {
"id": "chatcmpl-test123",
"object": "chat.completion",
"created": int(time.time()),
"model": request_data.get("model", "meta-llama/Llama-3.2-1B-Instruct"),
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "This is a test response from mock vLLM server.",
"tool_calls": None,
},
"logprobs": None,
"finish_reason": "stop",
}
],
"usage": {
"prompt_tokens": 25,
"completion_tokens": 15,
"total_tokens": 40,
"completion_tokens_details": None,
},
"system_fingerprint": None,
"service_tier": None,
}
def _create_text_completion_response(self, request_data: dict) -> dict:
"""
Create OpenAI Completion response.
Returns a valid response matching openai.types.Completion
"""
return {
"id": "cmpl-test123",
"object": "text_completion",
"created": int(time.time()),
"model": request_data.get("model", "meta-llama/Llama-3.2-1B-Instruct"),
"choices": [
{
"text": "This is a test completion.",
"index": 0,
"logprobs": None,
"finish_reason": "stop",
}
],
"usage": {
"prompt_tokens": 10,
"completion_tokens": 8,
"total_tokens": 18,
"completion_tokens_details": None,
},
"system_fingerprint": None,
}
def _send_json_response(self, status_code: int, data: dict):
"""Helper to send JSON response."""
self.send_response(status_code)
self.send_header("Content-Type", "application/json")
self.end_headers()
self.wfile.write(json.dumps(data).encode())
return VLLMHandler
async def await_start(self):
"""
Start the vLLM server and wait until ready.
This method is async and can be awaited to ensure the server is ready.
"""
# Create handler and start the HTTP server
handler_class = self._create_handler_class()
self.server = http.server.HTTPServer(("localhost", self.port), handler_class)
self.server_thread = threading.Thread(target=self.server.serve_forever, daemon=True)
self.server_thread.start()
# Wait for server to be listening on the port
for _ in range(10):
try:
sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
result = sock.connect_ex(("localhost", self.port))
sock.close()
if result == 0:
# Port is listening
return
except Exception:
pass
await asyncio.sleep(0.1)
raise RuntimeError(f"vLLM server failed to start on port {self.port}")
def stop(self):
"""Stop the vLLM server."""
if self.server:
self.server.shutdown()
self.server.server_close()
def clear(self):
"""Clear request history."""
self.requests_received = []
def get_request_count(self) -> int:
"""Get number of requests received."""
return len(self.requests_received)
def get_all_requests(self) -> list[dict]:
"""Get all received requests with their bodies."""
return self.requests_received

622
tests/integration/telemetry/test_otel_e2e.py Normal file
View file

@ -0,0 +1,622 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
"""
End-to-end tests for the OpenTelemetry inline provider.
What this does:
- Boots mock OTLP and mock vLLM
- Starts a real Llama Stack with inline OTel
- Calls real HTTP APIs
- Verifies traces, metrics, and custom metric names (non-empty)
"""
# ============================================================================
# IMPORTS
# ============================================================================
import os
import socket
import subprocess
import time
from typing import Any
import pytest
import requests
import yaml
from pydantic import BaseModel, Field
# Mock servers are in the mocking/ subdirectory
from .mocking import (
MockOTLPCollector,
MockServerConfig,
MockVLLMServer,
start_mock_servers_async,
stop_mock_servers,
)
# ============================================================================
# DATA MODELS
# ============================================================================
class TelemetryTestCase(BaseModel):
"""
Pydantic model defining expected telemetry for an API call.
**TO ADD A NEW TEST CASE:** Add to TEST_CASES list below.
"""
name: str = Field(description="Unique test case identifier")
http_method: str = Field(description="HTTP method (GET, POST, etc.)")
api_path: str = Field(description="API path (e.g., '/v1/models')")
request_body: dict[str, Any] | None = Field(default=None)
expected_http_status: int = Field(default=200)
expected_trace_exports: int = Field(default=1, description="Minimum number of trace exports expected")
expected_metric_exports: int = Field(default=0, description="Minimum number of metric exports expected")
should_have_error_span: bool = Field(default=False)
expected_metrics: list[str] = Field(
default_factory=list, description="List of metric names that should be captured"
)
expected_min_spans: int | None = Field(
default=None, description="If set, minimum number of spans expected in the new trace(s) generated by this test"
)
# ============================================================================
# TEST CONFIGURATION
# **TO ADD NEW TESTS:** Add TelemetryTestCase instances here
# ============================================================================
# Custom metric names (defined in llama_stack/providers/inline/telemetry/otel/otel.py)
CUSTOM_METRICS_BASE = [
"http.server.request.duration",
"http.server.request.count",
]
CUSTOM_METRICS_STREAMING = [
"http.server.streaming.duration",
"http.server.streaming.count",
]
TEST_CASES = [
TelemetryTestCase(
name="models_list",
http_method="GET",
api_path="/v1/models",
expected_trace_exports=1, # Single trace with 2-3 spans (GET, http send)
expected_metric_exports=1, # Metrics export periodically, but we'll wait for them
expected_metrics=[], # First request: middleware may not be initialized yet
expected_min_spans=2,
),
TelemetryTestCase(
name="chat_completion",
http_method="POST",
api_path="/v1/chat/completions",
request_body={
"model": "meta-llama/Llama-3.2-1B-Instruct",
"messages": [{"role": "user", "content": "Hello!"}],
},
expected_trace_exports=1, # Single trace with 4 spans (POST, http receive, 2x http send)
expected_metric_exports=1, # Metrics export periodically
expected_metrics=CUSTOM_METRICS_BASE,
expected_min_spans=3,
),
TelemetryTestCase(
name="chat_completion_streaming",
http_method="POST",
api_path="/v1/chat/completions",
request_body={
"model": "meta-llama/Llama-3.2-1B-Instruct",
"messages": [{"role": "user", "content": "Streaming test"}],
"stream": True, # Enable streaming response
},
expected_trace_exports=1, # Single trace with streaming spans
expected_metric_exports=1, # Metrics export periodically
# Validate both base and streaming metrics with polling
expected_metrics=CUSTOM_METRICS_BASE + CUSTOM_METRICS_STREAMING,
expected_min_spans=4,
),
]
# ============================================================================
# TEST INFRASTRUCTURE
# ============================================================================
class TelemetryTestRunner:
"""
Executes TelemetryTestCase instances against real Llama Stack.
**HOW IT WORKS:**
1. Makes real HTTP request to the stack
2. Waits for telemetry export
3. Verifies exports were sent to mock collector
4. Validates custom metrics by name (if expected_metrics is specified)
5. Ensures metrics have non-empty data points
"""
def __init__(
self,
base_url: str,
collector: MockOTLPCollector,
poll_timeout_seconds: float = 8.0,
poll_interval_seconds: float = 0.1,
):
self.base_url = base_url
self.collector = collector
self.poll_timeout_seconds = poll_timeout_seconds # how long to wait for telemetry to be exported
self.poll_interval_seconds = poll_interval_seconds # how often to poll for telemetry
def run_test_case(self, test_case: TelemetryTestCase, verbose: bool = False) -> bool:
"""Execute a single test case and verify telemetry."""
initial_traces = self.collector.get_trace_count()
prior_trace_ids = self.collector.get_all_trace_ids()
initial_metrics = self.collector.get_metric_count()
if verbose:
print(f"\n--- {test_case.name} ---")
print(f" {test_case.http_method} {test_case.api_path}")
if test_case.expected_metrics:
print(f" Expected custom metrics: {', '.join(test_case.expected_metrics)}")
# Make real HTTP request to Llama Stack
is_streaming_test = test_case.request_body and test_case.request_body.get("stream", False)
try:
url = f"{self.base_url}{test_case.api_path}"
# Streaming requests need longer timeout to complete
timeout = 10 if is_streaming_test else 5
if test_case.http_method == "GET":
response = requests.get(url, timeout=timeout)
elif test_case.http_method == "POST":
response = requests.post(url, json=test_case.request_body or {}, timeout=timeout)
else:
response = requests.request(test_case.http_method, url, timeout=timeout)
if verbose:
print(f" HTTP Response: {response.status_code}")
status_match = response.status_code == test_case.expected_http_status
except requests.exceptions.RequestException as e:
if verbose:
print(f" Request exception: {type(e).__name__}")
# For streaming requests, exceptions are expected due to mock server behavior
# The important part is whether telemetry metrics were captured
status_match = is_streaming_test # Pass streaming tests, fail non-streaming
# Poll until all telemetry expectations are met or timeout (single loop for speed)
missing_metrics: list[str] = []
empty_metrics: list[str] = []
new_trace_ids: set[str] = set()
def compute_status() -> tuple[bool, bool, bool, bool]:
traces_ok_local = (self.collector.get_trace_count() - initial_traces) >= test_case.expected_trace_exports
metrics_count_ok_local = (
self.collector.get_metric_count() - initial_metrics
) >= test_case.expected_metric_exports
metrics_ok_local = True
if test_case.expected_metrics:
missing_metrics.clear()
empty_metrics.clear()
for metric_name in test_case.expected_metrics:
if not self.collector.has_metric(metric_name):
missing_metrics.append(metric_name)
else:
data_points = self.collector.get_metric_by_name(metric_name)
if len(data_points) == 0:
empty_metrics.append(metric_name)
metrics_ok_local = len(missing_metrics) == 0 and len(empty_metrics) == 0
spans_ok_local = True
if test_case.expected_min_spans is not None:
nonlocal new_trace_ids
new_trace_ids = self.collector.get_new_trace_ids(prior_trace_ids)
if not new_trace_ids:
spans_ok_local = False
else:
counts = self.collector.get_trace_span_counts()
min_spans: int = int(test_case.expected_min_spans or 0)
spans_ok_local = all(counts.get(tid, 0) >= min_spans for tid in new_trace_ids)
return traces_ok_local, metrics_count_ok_local, metrics_ok_local, spans_ok_local
# Poll until all telemetry expectations are met or timeout (single loop for speed)
start = time.time()
traces_ok, metrics_count_ok, metrics_by_name_validated, spans_ok = compute_status()
while time.time() - start < self.poll_timeout_seconds:
if traces_ok and metrics_count_ok and metrics_by_name_validated and spans_ok:
break
time.sleep(self.poll_interval_seconds)
traces_ok, metrics_count_ok, metrics_by_name_validated, spans_ok = compute_status()
if verbose:
total_http_requests = len(getattr(self.collector, "all_http_requests", []))
print(f" [DEBUG] OTLP POST requests: {total_http_requests}")
print(
f" Expected: >={test_case.expected_trace_exports} traces, >={test_case.expected_metric_exports} metrics"
)
print(
f" Actual: {self.collector.get_trace_count() - initial_traces} traces, {self.collector.get_metric_count() - initial_metrics} metrics"
)
if test_case.expected_metrics:
print(" Custom metrics:")
for metric_name in test_case.expected_metrics:
n = len(self.collector.get_metric_by_name(metric_name))
status = "✓" if n > 0 else "✗"
print(f" {status} {metric_name}: {n}")
if missing_metrics:
print(f" Missing: {missing_metrics}")
if empty_metrics:
print(f" Empty: {empty_metrics}")
if test_case.expected_min_spans is not None:
counts = self.collector.get_trace_span_counts()
span_counts = {tid: counts[tid] for tid in new_trace_ids}
print(f" New trace IDs: {sorted(new_trace_ids)}")
print(f" Span counts: {span_counts}")
result = bool(
(status_match or is_streaming_test)
and traces_ok
and metrics_count_ok
and metrics_by_name_validated
and spans_ok
)
print(f" Result: {'PASS' if result else 'FAIL'}")
return bool(
(status_match or is_streaming_test)
and traces_ok
and metrics_count_ok
and metrics_by_name_validated
and spans_ok
)
def run_all_test_cases(self, test_cases: list[TelemetryTestCase], verbose: bool = True) -> dict[str, bool]:
"""Run all test cases and return results."""
results = {}
for test_case in test_cases:
results[test_case.name] = self.run_test_case(test_case, verbose=verbose)
return results
# ============================================================================
# HELPER FUNCTIONS
# ============================================================================
def is_port_available(port: int) -> bool:
"""Check if a TCP port is available for binding."""
try:
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
sock.bind(("localhost", port))
return True
except OSError:
return False
# ============================================================================
# PYTEST FIXTURES
# ============================================================================
@pytest.fixture(scope="module")
def mock_servers():
"""
Fixture: Start all mock servers in parallel using async harness.
**TO ADD A NEW MOCK SERVER:**
Just add a MockServerConfig to the MOCK_SERVERS list below.
"""
import asyncio
# ========================================================================
# MOCK SERVER CONFIGURATION
# **TO ADD A NEW MOCK:** Just add a MockServerConfig instance below
#
# Example:
# MockServerConfig(
# name="Mock MyService",
# server_class=MockMyService, # Must inherit from MockServerBase
# init_kwargs={"port": 9000, "param": "value"},
# ),
# ========================================================================
mock_servers_config = [
MockServerConfig(
name="Mock OTLP Collector",
server_class=MockOTLPCollector,
init_kwargs={"port": 4318},
),
MockServerConfig(
name="Mock vLLM Server",
server_class=MockVLLMServer,
init_kwargs={
"port": 8000,
"models": ["meta-llama/Llama-3.2-1B-Instruct"],
},
),
# Add more mock servers here - they will start in parallel automatically!
]
# Start all servers in parallel
servers = asyncio.run(start_mock_servers_async(mock_servers_config))
# Verify vLLM models
models_response = requests.get("http://localhost:8000/v1/models", timeout=1)
models_data = models_response.json()
print(f"[INFO] Mock vLLM serving {len(models_data['data'])} models: {[m['id'] for m in models_data['data']]}")
yield servers
# Stop all servers
stop_mock_servers(servers)
@pytest.fixture(scope="module")
def mock_otlp_collector(mock_servers):
"""Convenience fixture to get OTLP collector from mock_servers."""
return mock_servers["Mock OTLP Collector"]
@pytest.fixture(scope="module")
def mock_vllm_server(mock_servers):
"""Convenience fixture to get vLLM server from mock_servers."""
return mock_servers["Mock vLLM Server"]
@pytest.fixture(scope="module")
def llama_stack_server(tmp_path_factory, mock_otlp_collector, mock_vllm_server):
"""
Fixture: Start real Llama Stack server with inline OTel provider.
**THIS IS THE MAIN FIXTURE** - it runs:
opentelemetry-instrument llama stack run --config run.yaml
**TO MODIFY STACK CONFIG:** Edit run_config dict below
"""
config_dir = tmp_path_factory.mktemp("otel-stack-config")
# Ensure mock vLLM is ready and accessible before starting Llama Stack
print("\n[INFO] Verifying mock vLLM is accessible at http://localhost:8000...")
try:
vllm_models = requests.get("http://localhost:8000/v1/models", timeout=2)
print(f"[INFO] Mock vLLM models endpoint response: {vllm_models.status_code}")
except Exception as e:
pytest.fail(f"Mock vLLM not accessible before starting Llama Stack: {e}")
# Create run.yaml with inference and telemetry providers
# **TO ADD MORE PROVIDERS:** Add to providers dict
run_config = {
"image_name": "test-otel-e2e",
"apis": ["inference"],
"providers": {
"inference": [
{
"provider_id": "vllm",
"provider_type": "remote::vllm",
"config": {
"url": "http://localhost:8000/v1",
},
},
],
"telemetry": [
{
"provider_id": "otel",
"provider_type": "inline::otel",
"config": {
"service_name": "llama-stack-e2e-test",
"span_processor": "simple",
},
},
],
},
"models": [
{
"model_id": "meta-llama/Llama-3.2-1B-Instruct",
"provider_id": "vllm",
}
],
}
config_file = config_dir / "run.yaml"
with open(config_file, "w") as f:
yaml.dump(run_config, f)
# Find available port for Llama Stack
port = 5555
while not is_port_available(port) and port < 5600:
port += 1
if port >= 5600:
pytest.skip("No available ports for test server")
# Set environment variables for OTel instrumentation
# NOTE: These only affect the subprocess, not other tests
env = os.environ.copy()
env["OTEL_EXPORTER_OTLP_ENDPOINT"] = "http://localhost:4318"
env["OTEL_EXPORTER_OTLP_PROTOCOL"] = "http/protobuf" # Ensure correct protocol
env["OTEL_SERVICE_NAME"] = "llama-stack-e2e-test"
env["OTEL_SPAN_PROCESSOR"] = "simple" # Force simple processor for immediate export
env["LLAMA_STACK_PORT"] = str(port)
env["OTEL_PYTHON_LOGGING_AUTO_INSTRUMENTATION_ENABLED"] = "true"
# Configure fast metric export for testing (default is 60 seconds)
# This makes metrics export every 500ms instead of every 60 seconds
env["OTEL_METRIC_EXPORT_INTERVAL"] = "500" # milliseconds
env["OTEL_METRIC_EXPORT_TIMEOUT"] = "1000" # milliseconds
# Disable inference recording to ensure real requests to our mock vLLM
# This is critical - without this, Llama Stack replays cached responses
# Safe to remove here as it only affects the subprocess environment
if "LLAMA_STACK_TEST_INFERENCE_MODE" in env:
del env["LLAMA_STACK_TEST_INFERENCE_MODE"]
# Start server with automatic instrumentation
cmd = [
"opentelemetry-instrument", # ← Automatic instrumentation wrapper
"llama",
"stack",
"run",
str(config_file),
"--port",
str(port),
]
print(f"\n[INFO] Starting Llama Stack with OTel instrumentation on port {port}")
print(f"[INFO] Command: {' '.join(cmd)}")
process = subprocess.Popen(
cmd,
env=env,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT, # Merge stderr into stdout
text=True,
)
# Wait for server to start
max_wait = 30
base_url = f"http://localhost:{port}"
startup_output = []
for i in range(max_wait):
# Collect server output non-blocking
import select
if process.stdout and select.select([process.stdout], [], [], 0)[0]:
line = process.stdout.readline()
if line:
startup_output.append(line)
try:
response = requests.get(f"{base_url}/v1/health", timeout=1)
if response.status_code == 200:
print(f"[INFO] Server ready at {base_url}")
# Print relevant initialization logs
print(f"[DEBUG] Captured {len(startup_output)} lines of server output")
relevant_logs = [
line
for line in startup_output
if any(keyword in line.lower() for keyword in ["telemetry", "otel", "provider", "error creating"])
]
if relevant_logs:
print("[DEBUG] Relevant server logs:")
for log in relevant_logs[-10:]: # Last 10 relevant lines
print(f" {log.strip()}")
time.sleep(0.5)
break
except requests.exceptions.RequestException:
if i == max_wait - 1:
process.terminate()
stdout, _ = process.communicate(timeout=5)
pytest.fail(f"Server failed to start.\nOutput: {stdout}")
time.sleep(1)
yield {
"base_url": base_url,
"port": port,
"collector": mock_otlp_collector,
"vllm_server": mock_vllm_server,
}
# Cleanup
print("\n[INFO] Stopping Llama Stack server")
process.terminate()
try:
process.wait(timeout=5)
except subprocess.TimeoutExpired:
process.kill()
# ============================================================================
# TESTS: End-to-End with Real Stack
# **THESE RUN SLOW** - marked with @pytest.mark.slow
# **TO ADD NEW E2E TESTS:** Add methods to this class
# ============================================================================
@pytest.mark.slow
class TestOTelE2E:
"""
End-to-end tests with real Llama Stack server.
These tests verify the complete flow:
- Real Llama Stack with inline OTel provider
- Real API calls
- Automatic trace and metric collection
- Mock OTLP collector captures exports
"""
def test_server_starts_with_auto_instrumentation(self, llama_stack_server):
"""Verify server starts successfully with inline OTel provider."""
base_url = llama_stack_server["base_url"]
# Try different health check endpoints
health_endpoints = ["/health", "/v1/health", "/"]
server_responding = False
for endpoint in health_endpoints:
try:
response = requests.get(f"{base_url}{endpoint}", timeout=5)
print(f"\n[DEBUG] {endpoint} -> {response.status_code}")
if response.status_code == 200:
server_responding = True
break
except Exception as e:
print(f"[DEBUG] {endpoint} failed: {e}")
assert server_responding, f"Server not responding on any endpoint at {base_url}"
print(f"\n[PASS] Llama Stack running with OTel at {base_url}")
def test_all_test_cases_via_runner(self, llama_stack_server):
"""
**MAIN TEST:** Run all TelemetryTestCase instances with custom metrics validation.
This executes all test cases defined in TEST_CASES list and validates:
1. Traces are exported to the collector
2. Metrics are exported to the collector
3. Custom metrics (defined in CUSTOM_METRICS_BASE, CUSTOM_METRICS_STREAMING)
are captured by name with non-empty data points
Each test case specifies which metrics to validate via expected_metrics field.
**TO ADD MORE TESTS:**
- Add TelemetryTestCase to TEST_CASES (line ~132)
- Reference CUSTOM_METRICS_BASE or CUSTOM_METRICS_STREAMING in expected_metrics
- See examples in existing test cases
**TO ADD NEW METRICS:**
- Add metric to otel.py
- Add metric name to CUSTOM_METRICS_BASE or CUSTOM_METRICS_STREAMING (line ~122)
- Update test cases that should validate it
"""
base_url = llama_stack_server["base_url"]
collector = llama_stack_server["collector"]
# Create test runner
runner = TelemetryTestRunner(base_url, collector)
# Execute all test cases (set verbose=False for cleaner output)
results = runner.run_all_test_cases(TEST_CASES, verbose=False)
print(f"\n{'=' * 50}\nTEST CASE SUMMARY\n{'=' * 50}")
passed = sum(1 for p in results.values() if p)
total = len(results)
print(f"Passed: {passed}/{total}\n")
failed = [name for name, ok in results.items() if not ok]
for name, ok in results.items():
print(f" {'[PASS]' if ok else '[FAIL]'} {name}")
print(f"{'=' * 50}\n")
assert not failed, f"Some test cases failed: {failed}"

532
tests/integration/telemetry/test_otel_provider.py
View file

@ -1,532 +0,0 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
"""
Integration tests for OpenTelemetry provider.
These tests verify that the OTel provider correctly:
- Initializes within the Llama Stack
- Captures expected metrics (counters, histograms, up/down counters)
- Captures expected spans/traces
- Exports telemetry data to an OTLP collector (in-memory for testing)
Tests use in-memory exporters to avoid external dependencies and can run in GitHub Actions.
"""
import os
import time
from collections import defaultdict
from unittest.mock import patch
import pytest
from opentelemetry.sdk.metrics.export import InMemoryMetricReader
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
from llama_stack.providers.inline.telemetry.otel.config import OTelTelemetryConfig
from llama_stack.providers.inline.telemetry.otel.otel import OTelTelemetryProvider
@pytest.fixture(scope="module")
def in_memory_span_exporter():
"""Create an in-memory span exporter to capture traces."""
return InMemorySpanExporter()
@pytest.fixture(scope="module")
def in_memory_metric_reader():
"""Create an in-memory metric reader to capture metrics."""
return InMemoryMetricReader()
@pytest.fixture(scope="module")
def otel_provider_with_memory_exporters(in_memory_span_exporter, in_memory_metric_reader):
"""
Create an OTelTelemetryProvider configured with in-memory exporters.
This allows us to capture and verify telemetry data without external services.
Returns a dict with 'provider', 'span_exporter', and 'metric_reader'.
"""
# Set mock environment to avoid warnings
os.environ["OTEL_EXPORTER_OTLP_ENDPOINT"] = "http://localhost:4318"
config = OTelTelemetryConfig(
service_name="test-llama-stack-otel",
service_version="1.0.0-test",
deployment_environment="ci-test",
span_processor="simple",
)
# Patch the provider to use in-memory exporters
with patch.object(
OTelTelemetryProvider,
'model_post_init',
lambda self, _: _init_with_memory_exporters(
self, config, in_memory_span_exporter, in_memory_metric_reader
)
):
provider = OTelTelemetryProvider(config=config)
yield {
'provider': provider,
'span_exporter': in_memory_span_exporter,
'metric_reader': in_memory_metric_reader
}
def _init_with_memory_exporters(provider, config, span_exporter, metric_reader):
"""Helper to initialize provider with in-memory exporters."""
import threading
from opentelemetry import metrics, trace
from opentelemetry.sdk.metrics import MeterProvider
from opentelemetry.sdk.resources import Attributes, Resource
from opentelemetry.sdk.trace import TracerProvider
# Initialize pydantic private attributes
if provider.__pydantic_private__ is None:
provider.__pydantic_private__ = {}
provider._lock = threading.Lock()
provider._counters = {}
provider._up_down_counters = {}
provider._histograms = {}
provider._gauges = {}
# Create resource attributes
attributes: Attributes = {
key: value
for key, value in {
"service.name": config.service_name,
"service.version": config.service_version,
"deployment.environment": config.deployment_environment,
}.items()
if value is not None
}
resource = Resource.create(attributes)
# Configure tracer provider with in-memory exporter
tracer_provider = TracerProvider(resource=resource)
tracer_provider.add_span_processor(SimpleSpanProcessor(span_exporter))
trace.set_tracer_provider(tracer_provider)
# Configure meter provider with in-memory reader
meter_provider = MeterProvider(
resource=resource,
metric_readers=[metric_reader]
)
metrics.set_meter_provider(meter_provider)
class TestOTelProviderInitialization:
"""Test OTel provider initialization within Llama Stack."""
def test_provider_initializes_successfully(self, otel_provider_with_memory_exporters):
"""Test that the OTel provider initializes without errors."""
provider = otel_provider_with_memory_exporters['provider']
span_exporter = otel_provider_with_memory_exporters['span_exporter']
assert provider is not None
assert provider.config.service_name == "test-llama-stack-otel"
assert provider.config.service_version == "1.0.0-test"
assert provider.config.deployment_environment == "ci-test"
def test_provider_has_thread_safety_mechanisms(self, otel_provider_with_memory_exporters):
"""Test that the provider has thread-safety mechanisms in place."""
provider = otel_provider_with_memory_exporters['provider']
assert hasattr(provider, "_lock")
assert provider._lock is not None
assert hasattr(provider, "_counters")
assert hasattr(provider, "_histograms")
assert hasattr(provider, "_up_down_counters")
class TestOTelMetricsCapture:
"""Test that OTel provider captures expected metrics."""
def test_counter_metric_is_captured(self, otel_provider_with_memory_exporters):
"""Test that counter metrics are captured."""
provider = otel_provider_with_memory_exporters['provider']
metric_reader = otel_provider_with_memory_exporters['metric_reader']
# Record counter metrics
provider.record_count("llama.requests.total", 1.0, attributes={"endpoint": "/chat"})
provider.record_count("llama.requests.total", 1.0, attributes={"endpoint": "/chat"})
provider.record_count("llama.requests.total", 1.0, attributes={"endpoint": "/embeddings"})
# Force metric collection - collect() triggers the reader to gather metrics
metric_reader.collect()
metric_reader.collect()
metrics_data = metric_reader.get_metrics_data()
# Verify metrics were captured
assert metrics_data is not None
assert len(metrics_data.resource_metrics) > 0
# Find our counter metric
found_counter = False
for resource_metric in metrics_data.resource_metrics:
for scope_metric in resource_metric.scope_metrics:
for metric in scope_metric.metrics:
if metric.name == "llama.requests.total":
found_counter = True
# Verify it's a counter with data points
assert hasattr(metric.data, "data_points")
assert len(metric.data.data_points) > 0
assert found_counter, "Counter metric 'llama.requests.total' was not captured"
def test_histogram_metric_is_captured(self, otel_provider_with_memory_exporters):
"""Test that histogram metrics are captured."""
provider = otel_provider_with_memory_exporters['provider']
metric_reader = otel_provider_with_memory_exporters['metric_reader']
# Record histogram metrics with various values
latencies = [10.5, 25.3, 50.1, 100.7, 250.2]
for latency in latencies:
provider.record_histogram(
"llama.inference.latency",
latency,
attributes={"model": "llama-3.2"}
)
# Force metric collection
metric_reader.collect()
metrics_data = metric_reader.get_metrics_data()
# Find our histogram metric
found_histogram = False
for resource_metric in metrics_data.resource_metrics:
for scope_metric in resource_metric.scope_metrics:
for metric in scope_metric.metrics:
if metric.name == "llama.inference.latency":
found_histogram = True
# Verify it's a histogram
assert hasattr(metric.data, "data_points")
data_point = metric.data.data_points[0]
# Histograms should have count and sum
assert hasattr(data_point, "count")
assert data_point.count == len(latencies)
assert found_histogram, "Histogram metric 'llama.inference.latency' was not captured"
def test_up_down_counter_metric_is_captured(self, otel_provider_with_memory_exporters):
"""Test that up/down counter metrics are captured."""
provider = otel_provider_with_memory_exporters['provider']
metric_reader = otel_provider_with_memory_exporters['metric_reader']
# Record up/down counter metrics
provider.record_up_down_counter("llama.active.sessions", 5)
provider.record_up_down_counter("llama.active.sessions", 3)
provider.record_up_down_counter("llama.active.sessions", -2)
# Force metric collection
metric_reader.collect()
metrics_data = metric_reader.get_metrics_data()
# Find our up/down counter metric
found_updown = False
for resource_metric in metrics_data.resource_metrics:
for scope_metric in resource_metric.scope_metrics:
for metric in scope_metric.metrics:
if metric.name == "llama.active.sessions":
found_updown = True
assert hasattr(metric.data, "data_points")
assert len(metric.data.data_points) > 0
assert found_updown, "Up/Down counter metric 'llama.active.sessions' was not captured"
def test_metrics_with_attributes_are_captured(self, otel_provider_with_memory_exporters):
"""Test that metric attributes/labels are preserved."""
provider = otel_provider_with_memory_exporters['provider']
metric_reader = otel_provider_with_memory_exporters['metric_reader']
# Record metrics with different attributes
provider.record_count("llama.tokens.generated", 150.0, attributes={
"model": "llama-3.2-1b",
"user": "test-user"
})
# Force metric collection
metric_reader.collect()
metrics_data = metric_reader.get_metrics_data()
# Verify attributes are preserved
found_with_attributes = False
for resource_metric in metrics_data.resource_metrics:
for scope_metric in resource_metric.scope_metrics:
for metric in scope_metric.metrics:
if metric.name == "llama.tokens.generated":
data_point = metric.data.data_points[0]
# Check attributes - they're already a dict in the SDK
attrs = data_point.attributes if isinstance(data_point.attributes, dict) else {}
if "model" in attrs and "user" in attrs:
found_with_attributes = True
assert attrs["model"] == "llama-3.2-1b"
assert attrs["user"] == "test-user"
assert found_with_attributes, "Metrics with attributes were not properly captured"
def test_multiple_metric_types_coexist(self, otel_provider_with_memory_exporters):
"""Test that different metric types can coexist."""
provider = otel_provider_with_memory_exporters['provider']
metric_reader = otel_provider_with_memory_exporters['metric_reader']
# Record various metric types
provider.record_count("test.counter", 1.0)
provider.record_histogram("test.histogram", 42.0)
provider.record_up_down_counter("test.gauge", 10)
# Force metric collection
metric_reader.collect()
metrics_data = metric_reader.get_metrics_data()
# Count unique metrics
metric_names = set()
for resource_metric in metrics_data.resource_metrics:
for scope_metric in resource_metric.scope_metrics:
for metric in scope_metric.metrics:
metric_names.add(metric.name)
# Should have all three metrics
assert "test.counter" in metric_names
assert "test.histogram" in metric_names
assert "test.gauge" in metric_names
class TestOTelSpansCapture:
"""Test that OTel provider captures expected spans/traces."""
def test_basic_span_is_captured(self, otel_provider_with_memory_exporters):
"""Test that basic spans are captured."""
provider = otel_provider_with_memory_exporters['provider']
metric_reader = otel_provider_with_memory_exporters['metric_reader']
span_exporter = otel_provider_with_memory_exporters['span_exporter']
# Create a span
span = provider.custom_trace("llama.inference.request")
span.end()
# Get captured spans
spans = span_exporter.get_finished_spans()
assert len(spans) > 0
assert any(span.name == "llama.inference.request" for span in spans)
def test_span_with_attributes_is_captured(self, otel_provider_with_memory_exporters):
"""Test that span attributes are preserved."""
provider = otel_provider_with_memory_exporters['provider']
span_exporter = otel_provider_with_memory_exporters['span_exporter']
# Create a span with attributes
span = provider.custom_trace(
"llama.chat.completion",
attributes={
"model.id": "llama-3.2-1b",
"user.id": "test-user-123",
"request.id": "req-abc-123"
}
)
span.end()
# Get captured spans
spans = span_exporter.get_finished_spans()
# Find our span
our_span = None
for s in spans:
if s.name == "llama.chat.completion":
our_span = s
break
assert our_span is not None, "Span 'llama.chat.completion' was not captured"
# Verify attributes
attrs = dict(our_span.attributes)
assert attrs.get("model.id") == "llama-3.2-1b"
assert attrs.get("user.id") == "test-user-123"
assert attrs.get("request.id") == "req-abc-123"
def test_multiple_spans_are_captured(self, otel_provider_with_memory_exporters):
"""Test that multiple spans are captured."""
provider = otel_provider_with_memory_exporters['provider']
span_exporter = otel_provider_with_memory_exporters['span_exporter']
# Create multiple spans
span_names = [
"llama.request.validate",
"llama.model.load",
"llama.inference.execute",
"llama.response.format"
]
for name in span_names:
span = provider.custom_trace(name)
time.sleep(0.01) # Small delay to ensure ordering
span.end()
# Get captured spans
spans = span_exporter.get_finished_spans()
captured_names = {span.name for span in spans}
# Verify all spans were captured
for expected_name in span_names:
assert expected_name in captured_names, f"Span '{expected_name}' was not captured"
def test_span_has_service_metadata(self, otel_provider_with_memory_exporters):
"""Test that spans include service metadata."""
provider = otel_provider_with_memory_exporters['provider']
span_exporter = otel_provider_with_memory_exporters['span_exporter']
# Create a span
span = provider.custom_trace("test.span")
span.end()
# Get captured spans
spans = span_exporter.get_finished_spans()
assert len(spans) > 0
# Check resource attributes
span = spans[0]
resource_attrs = dict(span.resource.attributes)
assert resource_attrs.get("service.name") == "test-llama-stack-otel"
assert resource_attrs.get("service.version") == "1.0.0-test"
assert resource_attrs.get("deployment.environment") == "ci-test"
class TestOTelDataExport:
"""Test that telemetry data can be exported to OTLP collector."""
def test_metrics_are_exportable(self, otel_provider_with_memory_exporters):
"""Test that metrics can be exported."""
provider = otel_provider_with_memory_exporters['provider']
metric_reader = otel_provider_with_memory_exporters['metric_reader']
# Record metrics
provider.record_count("export.test.counter", 5.0)
provider.record_histogram("export.test.histogram", 123.45)
# Force export
metric_reader.collect()
metrics_data = metric_reader.get_metrics_data()
# Verify data structure is exportable
assert metrics_data is not None
assert hasattr(metrics_data, "resource_metrics")
assert len(metrics_data.resource_metrics) > 0
# Verify resource attributes are present (needed for OTLP export)
resource = metrics_data.resource_metrics[0].resource
assert resource is not None
assert len(resource.attributes) > 0
def test_spans_are_exportable(self, otel_provider_with_memory_exporters):
"""Test that spans can be exported."""
provider = otel_provider_with_memory_exporters['provider']
span_exporter = otel_provider_with_memory_exporters['span_exporter']
# Create spans
span1 = provider.custom_trace("export.test.span1")
span1.end()
span2 = provider.custom_trace("export.test.span2")
span2.end()
# Get exported spans
spans = span_exporter.get_finished_spans()
# Verify spans have required OTLP fields
assert len(spans) >= 2
for span in spans:
assert span.name is not None
assert span.context is not None
assert span.context.trace_id is not None
assert span.context.span_id is not None
assert span.resource is not None
def test_concurrent_export_is_safe(self, otel_provider_with_memory_exporters):
"""Test that concurrent metric/span recording doesn't break export."""
import concurrent.futures
provider = otel_provider_with_memory_exporters['provider']
metric_reader = otel_provider_with_memory_exporters['metric_reader']
span_exporter = otel_provider_with_memory_exporters['span_exporter']
def record_data(thread_id):
for i in range(10):
provider.record_count(f"concurrent.counter.{thread_id}", 1.0)
span = provider.custom_trace(f"concurrent.span.{thread_id}.{i}")
span.end()
# Record from multiple threads
with concurrent.futures.ThreadPoolExecutor(max_workers=5) as executor:
futures = [executor.submit(record_data, i) for i in range(5)]
concurrent.futures.wait(futures)
# Verify export still works
metric_reader.collect()
metrics_data = metric_reader.get_metrics_data()
spans = span_exporter.get_finished_spans()
assert metrics_data is not None
assert len(spans) >= 50 # 5 threads * 10 spans each
@pytest.mark.integration
class TestOTelProviderIntegration:
"""End-to-end integration tests simulating real usage."""
def test_complete_inference_workflow_telemetry(self, otel_provider_with_memory_exporters):
"""Simulate a complete inference workflow with telemetry."""
provider = otel_provider_with_memory_exporters['provider']
metric_reader = otel_provider_with_memory_exporters['metric_reader']
span_exporter = otel_provider_with_memory_exporters['span_exporter']
# Simulate inference workflow
request_span = provider.custom_trace(
"llama.inference.request",
attributes={"model": "llama-3.2-1b", "user": "test"}
)
# Track metrics during inference
provider.record_count("llama.requests.received", 1.0)
provider.record_up_down_counter("llama.requests.in_flight", 1)
# Simulate processing time
time.sleep(0.01)
provider.record_histogram("llama.request.duration_ms", 10.5)
# Track tokens
provider.record_count("llama.tokens.input", 25.0)
provider.record_count("llama.tokens.output", 150.0)
# End request
provider.record_up_down_counter("llama.requests.in_flight", -1)
provider.record_count("llama.requests.completed", 1.0)
request_span.end()
# Verify all telemetry was captured
metric_reader.collect()
metrics_data = metric_reader.get_metrics_data()
spans = span_exporter.get_finished_spans()
# Check metrics exist
metric_names = set()
for rm in metrics_data.resource_metrics:
for sm in rm.scope_metrics:
for m in sm.metrics:
metric_names.add(m.name)
assert "llama.requests.received" in metric_names
assert "llama.requests.in_flight" in metric_names
assert "llama.request.duration_ms" in metric_names
assert "llama.tokens.input" in metric_names
assert "llama.tokens.output" in metric_names
# Check span exists
assert any(s.name == "llama.inference.request" for s in spans)