File size: 8,042 Bytes
c79d967 e181667 c79d967 e181667 c79d967 e181667 c79d967 e181667 c79d967 e181667 c79d967 e181667 c79d967 e181667 c79d967 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 | """
Telemetry: in-memory counters + HF Dataset persistence.
In-memory — rolling deque(500), per-metric pass counters, avg latency.
Resets on restart. Powers /metrics (live session stats).
Persistent — events flushed as JSONL shards to TELEMETRY_REPO HF dataset
every FLUSH_EVERY queries. Powers /report (accumulated history).
Falls back to in-memory if TELEMETRY_REPO is unset or write fails.
TELEMETRY_REPO is inferred from HF Spaces env vars:
SPACE_AUTHOR_NAME + SPACE_REPO_NAME → {author}/{repo}-telemetry
Override with explicit TELEMETRY_REPO env var.
"""
import json
import logging
import os
import threading
from collections import defaultdict, deque
from datetime import UTC, datetime
from typing import Any
from grader import GradeReport
log = logging.getLogger(__name__)
BUFFER_SIZE = 500
FLUSH_EVERY = 20
_lock = threading.Lock()
_events: deque[dict[str, Any]] = deque(maxlen=BUFFER_SIZE)
_unflushed: list[dict[str, Any]] = []
_counters: dict[str, float] = defaultdict(float)
_space_author = os.environ.get("SPACE_AUTHOR_NAME", "")
_space_repo = os.environ.get("SPACE_REPO_NAME", "ai-response-validator")
TELEMETRY_REPO = os.environ.get(
"TELEMETRY_REPO",
f"{_space_author}/{_space_repo}-telemetry" if _space_author else "",
)
_METRICS = ["pii_leakage", "token_budget", "answer_relevancy", "faithfulness", "chain_terminology"]
def record(
client: str,
domain: str,
query_len: int,
latency_ms: dict[str, float],
report: GradeReport,
docs_retrieved: int,
min_retrieval_score: float,
) -> None:
"""Record one query event. Thread-safe. Flushes to HF dataset in background."""
event = {
"ts": datetime.now(UTC).isoformat(),
"client": client,
"domain": domain,
"query_len": query_len,
"latency_ms": {k: round(v) for k, v in latency_ms.items()},
"metrics": {r.metric: round(r.score, 4) for r in report.results},
"metric_passed": {r.metric: r.passed for r in report.results},
"overall_pass": report.overall,
"docs_retrieved": docs_retrieved,
"min_retrieval_score": round(min_retrieval_score, 4),
}
with _lock:
_events.append(event)
_unflushed.append(event)
_counters["total"] += 1
if report.overall:
_counters["overall_pass"] += 1
for r in report.results:
_counters[f"{r.metric}_total"] += 1
if r.passed:
_counters[f"{r.metric}_pass"] += 1
for stage, ms in latency_ms.items():
_counters[f"lat_{stage}_sum"] += ms
_counters[f"lat_{stage}_n"] += 1
should_flush = len(_unflushed) >= FLUSH_EVERY
if should_flush and TELEMETRY_REPO:
threading.Thread(target=_flush, daemon=True).start()
def live_stats() -> dict[str, Any]:
"""In-memory aggregate for the current session (/metrics endpoint)."""
with _lock:
total = int(_counters.get("total", 0))
if total == 0:
return {"total_queries": 0, "message": "No queries recorded this session."}
metric_stats = {}
for m in _METRICS:
mt = int(_counters.get(f"{m}_total", 0))
mp = int(_counters.get(f"{m}_pass", 0))
metric_stats[m] = {
"pass_rate": round(mp / mt, 3) if mt else None,
"pass_count": mp,
"total": mt,
}
avg_latency = {}
for stage in ("retrieve", "generate", "grade"):
n = _counters.get(f"lat_{stage}_n", 0)
if n:
avg_latency[stage] = round(_counters[f"lat_{stage}_sum"] / n)
return {
"source": "in_memory",
"total_queries": total,
"overall_pass_rate": round(_counters.get("overall_pass", 0) / total, 3),
"metrics": metric_stats,
"avg_latency_ms": avg_latency,
"events_in_buffer": len(_events),
"telemetry_repo": TELEMETRY_REPO or None,
}
def persistent_report() -> dict[str, Any]:
"""Aggregate from HF Dataset shards (/report endpoint). Falls back to live_stats."""
if not TELEMETRY_REPO:
log.info("TELEMETRY_REPO not set — report from in-memory only")
return {"source": "in_memory", **live_stats()}
try:
from huggingface_hub import HfApi
hf_token = os.environ.get("HF_TOKEN")
api = HfApi(token=hf_token)
files = api.list_repo_files(TELEMETRY_REPO, repo_type="dataset")
shard_paths = [f for f in files if f.startswith("events/") and f.endswith(".jsonl")]
if not shard_paths:
return {"source": "hf_dataset", "repo": TELEMETRY_REPO,
"message": "No shards yet — data accumulates after first flush."}
events = []
for path in shard_paths:
content = api.hf_hub_download(
TELEMETRY_REPO, path, repo_type="dataset", token=hf_token,
)
with open(content) as f:
for line in f:
if line.strip():
events.append(json.loads(line))
if not events:
return {"source": "hf_dataset", "repo": TELEMETRY_REPO, "total_events": 0}
total = len(events)
overall_pass = sum(1 for e in events if e.get("overall_pass"))
metric_stats = {}
for m in _METRICS:
passed = sum(1 for e in events if e.get("metric_passed", {}).get(m))
scores = [e["metrics"][m] for e in events if m in e.get("metrics", {})]
metric_stats[m] = {
"pass_rate": round(passed / total, 3),
"avg_score": round(sum(scores) / len(scores), 3) if scores else None,
}
client_breakdown: dict[str, dict[str, int]] = defaultdict(lambda: {"total": 0, "pass": 0})
for e in events:
c = e.get("client", "unknown")
client_breakdown[c]["total"] += 1
if e.get("overall_pass"):
client_breakdown[c]["pass"] += 1
return {
"source": "hf_dataset",
"repo": TELEMETRY_REPO,
"total_queries": total,
"overall_pass_rate": round(overall_pass / total, 3),
"first_event": min(e["ts"] for e in events),
"last_event": max(e["ts"] for e in events),
"metrics": metric_stats,
"by_client": {
c: {"total": v["total"], "pass_rate": round(v["pass"] / v["total"], 3)}
for c, v in client_breakdown.items()
},
"shards_read": len(shard_paths),
}
except Exception as e:
log.warning("HF Dataset report failed (%s) — falling back to in-memory", e)
return {"source": "in_memory_fallback", **live_stats()}
def _flush() -> None:
"""Upload buffered events to HF Dataset as a JSONL shard. Runs in background thread."""
with _lock:
if not _unflushed:
return
batch = list(_unflushed)
_unflushed.clear()
try:
from huggingface_hub import HfApi
hf_token = os.environ.get("HF_TOKEN")
if not hf_token:
log.warning("HF_TOKEN not set — telemetry flush skipped")
return
api = HfApi(token=hf_token)
try:
api.create_repo(TELEMETRY_REPO, repo_type="dataset", exist_ok=True, private=False)
except Exception:
pass
ts = datetime.now(UTC).strftime("%Y%m%dT%H%M%S%f")
content = "\n".join(json.dumps(e) for e in batch).encode()
api.upload_file(
path_or_fileobj=content,
path_in_repo=f"events/shard_{ts}.jsonl",
repo_id=TELEMETRY_REPO,
repo_type="dataset",
)
log.info("Flushed %d telemetry events to %s", len(batch), TELEMETRY_REPO)
except Exception as e:
log.warning("Telemetry flush failed: %s — events returned to buffer", e)
with _lock:
_unflushed.extend(batch)
|