|
| 1 | +# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. |
| 2 | +# SPDX-License-Identifier: Apache-2.0 |
| 3 | + |
| 4 | +""" |
| 5 | +Prometheus metrics utilities for Dynamo components. |
| 6 | +
|
| 7 | +This module provides shared functionality for collecting and exposing Prometheus metrics |
| 8 | +from backend engines (SGLang, vLLM, etc.) via Dynamo's metrics endpoint. |
| 9 | +
|
| 10 | +Note: Engine metrics take time to appear after engine initialization, |
| 11 | +while Dynamo runtime metrics are available immediately after component creation. |
| 12 | +""" |
| 13 | + |
| 14 | +import logging |
| 15 | +import re |
| 16 | +from typing import TYPE_CHECKING, Optional |
| 17 | + |
| 18 | +from prometheus_client import generate_latest |
| 19 | + |
| 20 | +from dynamo._core import Endpoint |
| 21 | + |
| 22 | +# Import CollectorRegistry only for type hints to avoid importing prometheus_client at module load time. |
| 23 | +# prometheus_client must be imported AFTER set_prometheus_multiproc_dir() is called. |
| 24 | +# See main.py worker() function for detailed explanation. |
| 25 | +if TYPE_CHECKING: |
| 26 | + from prometheus_client import CollectorRegistry |
| 27 | + |
| 28 | + |
| 29 | +def register_engine_metrics_callback( |
| 30 | + endpoint: Endpoint, |
| 31 | + registry: "CollectorRegistry", |
| 32 | + metric_prefix: str, |
| 33 | + engine_name: str, |
| 34 | +) -> None: |
| 35 | + """ |
| 36 | + Register a callback to expose engine Prometheus metrics via Dynamo's metrics endpoint. |
| 37 | +
|
| 38 | + This registers a callback that is invoked when /metrics is scraped, passing through |
| 39 | + engine-specific metrics alongside Dynamo runtime metrics. |
| 40 | +
|
| 41 | + Args: |
| 42 | + endpoint: Dynamo endpoint object with metrics.register_prometheus_expfmt_callback() |
| 43 | + registry: Prometheus registry to collect from (e.g., REGISTRY or CollectorRegistry) |
| 44 | + metric_prefix: Prefix to filter metrics (e.g., "vllm:" or "sglang:") |
| 45 | + engine_name: Name of the engine for logging (e.g., "vLLM" or "SGLang") |
| 46 | +
|
| 47 | + Example: |
| 48 | + from prometheus_client import REGISTRY |
| 49 | + register_engine_metrics_callback( |
| 50 | + generate_endpoint, REGISTRY, "vllm:", "vLLM" |
| 51 | + ) |
| 52 | + """ |
| 53 | + |
| 54 | + def get_expfmt() -> str: |
| 55 | + """Callback to return engine Prometheus metrics in exposition format""" |
| 56 | + return get_prometheus_expfmt(registry, metric_prefix_filter=metric_prefix) |
| 57 | + |
| 58 | + endpoint.metrics.register_prometheus_expfmt_callback(get_expfmt) |
| 59 | + |
| 60 | + |
| 61 | +def get_prometheus_expfmt( |
| 62 | + registry, |
| 63 | + metric_prefix_filter: Optional[str] = None, |
| 64 | +) -> str: |
| 65 | + """ |
| 66 | + Get Prometheus metrics from a registry formatted as text using the standard text encoder. |
| 67 | +
|
| 68 | + Collects all metrics from the registry and returns them in Prometheus text exposition format. |
| 69 | + Optionally filters metrics by prefix. |
| 70 | +
|
| 71 | + Prometheus exposition format consists of: |
| 72 | + - Comment lines starting with # (HELP and TYPE declarations) |
| 73 | + - Metric lines with format: metric_name{label="value"} metric_value timestamp |
| 74 | +
|
| 75 | + Example output format: |
| 76 | + # HELP vllm:request_success_total Number of successful requests |
| 77 | + # TYPE vllm:request_success_total counter |
| 78 | + vllm:request_success_total{model="llama2",endpoint="generate"} 150.0 |
| 79 | + # HELP vllm:time_to_first_token_seconds Time to first token |
| 80 | + # TYPE vllm:time_to_first_token_seconds histogram |
| 81 | + vllm:time_to_first_token_seconds_bucket{model="llama2",le="0.01"} 10.0 |
| 82 | + vllm:time_to_first_token_seconds_bucket{model="llama2",le="0.1"} 45.0 |
| 83 | + vllm:time_to_first_token_seconds_count{model="llama2"} 50.0 |
| 84 | + vllm:time_to_first_token_seconds_sum{model="llama2"} 2.5 |
| 85 | +
|
| 86 | + Args: |
| 87 | + registry: Prometheus registry to collect from. |
| 88 | + Pass CollectorRegistry with MultiProcessCollector for SGLang. |
| 89 | + Pass REGISTRY for vLLM single-process mode. |
| 90 | + metric_prefix_filter: Optional prefix to filter displayed metrics (e.g., "vllm:"). |
| 91 | + If None, returns all metrics. (default: None) |
| 92 | +
|
| 93 | + Returns: |
| 94 | + Formatted metrics text in Prometheus exposition format. Returns empty string on error. |
| 95 | +
|
| 96 | + Example: |
| 97 | + from prometheus_client import REGISTRY |
| 98 | + metrics_text = get_prometheus_expfmt(REGISTRY) |
| 99 | + print(metrics_text) |
| 100 | +
|
| 101 | + # With filter |
| 102 | + vllm_metrics = get_prometheus_expfmt(REGISTRY, metric_prefix_filter="vllm:") |
| 103 | + """ |
| 104 | + try: |
| 105 | + # Generate metrics in Prometheus text format |
| 106 | + metrics_text = generate_latest(registry).decode("utf-8") |
| 107 | + |
| 108 | + if metric_prefix_filter: |
| 109 | + # Filter lines: keep metric lines starting with prefix and their HELP/TYPE comments |
| 110 | + escaped_prefix = re.escape(metric_prefix_filter) |
| 111 | + pattern = rf"^(?:{escaped_prefix}|# (?:HELP|TYPE) {escaped_prefix})" |
| 112 | + filtered_lines = [ |
| 113 | + line for line in metrics_text.split("\n") if re.match(pattern, line) |
| 114 | + ] |
| 115 | + result = "\n".join(filtered_lines) |
| 116 | + if result: |
| 117 | + # Ensure result ends with newline |
| 118 | + if result and not result.endswith("\n"): |
| 119 | + result += "\n" |
| 120 | + return result |
| 121 | + else: |
| 122 | + # Ensure metrics_text ends with newline |
| 123 | + if metrics_text and not metrics_text.endswith("\n"): |
| 124 | + metrics_text += "\n" |
| 125 | + return metrics_text |
| 126 | + |
| 127 | + except Exception as e: |
| 128 | + logging.error(f"Error getting metrics: {e}") |
| 129 | + return "" |
0 commit comments