mirror of
https://github.com/elder-plinius/G0DM0D3.git
synced 2026-11-16 07:02:25 +00:00
fix: in-memory isolate buffer replaces direct-commit fallback
When TELEMETRY_KV is not bound, telemetry events now accumulate in the Worker isolate's module-level buffer (same persistence model as the rate limiter) and flush to HF when the buffer reaches 50 events, 256 KiB, or 15 minutes — whichever comes first. This eliminates the one-commit-per-request explosion without requiring any Cloudflare Dashboard configuration. KV is still preferred (durable across isolate evictions), but the fallback now produces ~1 HF commit per 50 requests instead of 50. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
@@ -66,6 +66,20 @@ const FLUSH_MAX_KEYS = 50 // max keys consumed per flush cycle
|
||||
const FLUSH_MAX_BYTES = 5 * 1024 * 1024 // 5 MiB combined JSONL per HF commit
|
||||
const FLUSH_AGE_MS = 30 * 60 * 1000 // 30 minutes
|
||||
|
||||
// ── In-memory isolate buffer (fallback when KV not bound) ───────────
|
||||
// Same persistence model as the rate limiter: module-level state
|
||||
// survives across requests within the same Worker isolate. We
|
||||
// accumulate JSONL lines and flush to HF when the buffer is large
|
||||
// or old enough. Isolate eviction loses buffered data — acceptable
|
||||
// for best-effort telemetry, not acceptable for the commit explosion.
|
||||
const MEM_FLUSH_COUNT = 50 // events before flush
|
||||
const MEM_FLUSH_BYTES = 256 * 1024 // 256 KiB before flush
|
||||
const MEM_FLUSH_AGE_MS = 15 * 60 * 1000 // 15 minutes before flush
|
||||
const MEM_MAX_BYTES = 2 * 1024 * 1024 // 2 MiB hard cap — drop oldest if exceeded
|
||||
const memBuffer: string[] = []
|
||||
let memBufferBytes = 0
|
||||
let memBufferFirstTs = 0
|
||||
|
||||
// ── Rate Limiter (dual: session + IP, in-memory per-isolate) ────────
|
||||
const RATE_LIMIT_WINDOW_MS = 60_000
|
||||
const RATE_LIMIT_MAX_SESSION = 10 // per session_id
|
||||
@@ -264,23 +278,40 @@ export const onRequestPost: PagesFunction<Env> = async (context) => {
|
||||
}, 200)
|
||||
}
|
||||
|
||||
// ── Direct commit fallback (KV not configured) ────────────────────
|
||||
// WARNING: At scale this creates one HF commit per request.
|
||||
// Bind TELEMETRY_KV in production to avoid commit explosion.
|
||||
console.warn('[Telemetry] TELEMETRY_KV not bound — falling back to direct HF commit')
|
||||
// ── In-memory buffer fallback (KV not configured) ──────────────────
|
||||
// Accumulates JSONL in the isolate's module scope and flushes to HF
|
||||
// when the buffer is large or old enough. Zero config required.
|
||||
// Data loss on isolate eviction is acceptable for telemetry.
|
||||
const now = Date.now()
|
||||
if (memBufferFirstTs === 0) memBufferFirstTs = now
|
||||
|
||||
const ts = new Date().toISOString().replace(/[:.]/g, '-')
|
||||
const hash = shortHash(jsonl)
|
||||
const filePath = `telemetry/batch_${ts}_${hash}.jsonl`
|
||||
|
||||
const branch = env.HF_DATASET_BRANCH || 'main'
|
||||
const ok = await commitToHF(env.HF_TOKEN, env.HF_DATASET_REPO, branch, filePath, jsonl)
|
||||
|
||||
if (ok) {
|
||||
return jsonResponse({ accepted: sanitized.length, file: filePath }, 200)
|
||||
// Drop oldest lines if hard cap would be exceeded
|
||||
const lineBytes = new TextEncoder().encode(jsonl).byteLength
|
||||
while (memBufferBytes + lineBytes > MEM_MAX_BYTES && memBuffer.length > 0) {
|
||||
const dropped = memBuffer.shift()!
|
||||
memBufferBytes -= new TextEncoder().encode(dropped).byteLength
|
||||
}
|
||||
|
||||
return jsonResponse({ error: 'Failed to publish to HuggingFace' }, 502)
|
||||
memBuffer.push(jsonl)
|
||||
memBufferBytes += lineBytes
|
||||
|
||||
const age = now - memBufferFirstTs
|
||||
const shouldFlush =
|
||||
memBuffer.length >= MEM_FLUSH_COUNT ||
|
||||
memBufferBytes >= MEM_FLUSH_BYTES ||
|
||||
age >= MEM_FLUSH_AGE_MS
|
||||
|
||||
if (shouldFlush) {
|
||||
context.waitUntil(flushMemToHF(env))
|
||||
}
|
||||
|
||||
return jsonResponse({
|
||||
accepted: sanitized.length,
|
||||
buffered: true,
|
||||
pending: memBuffer.length,
|
||||
flushing: shouldFlush,
|
||||
storage: 'memory',
|
||||
}, 200)
|
||||
}
|
||||
|
||||
// ── KV Flush: read up to FLUSH_MAX_KEYS, cap at FLUSH_MAX_BYTES ────
|
||||
@@ -343,6 +374,39 @@ async function flushKVToHF(env: Env): Promise<number> {
|
||||
return 0
|
||||
}
|
||||
|
||||
// ── In-memory flush: drain buffer → one HF commit ──────────────────
|
||||
|
||||
async function flushMemToHF(env: Env): Promise<number> {
|
||||
if (memBuffer.length === 0) return 0
|
||||
|
||||
// Snapshot and clear before the async commit so new requests
|
||||
// accumulate into a fresh buffer, not into the one we're flushing.
|
||||
const snapshot = memBuffer.splice(0, memBuffer.length)
|
||||
const snapshotBytes = memBufferBytes
|
||||
memBufferBytes = 0
|
||||
memBufferFirstTs = 0
|
||||
|
||||
const combined = snapshot.join('\n')
|
||||
const digest = await sha256Hex(combined)
|
||||
const filePath = `telemetry/membuf_${digest}.jsonl`
|
||||
const branch = env.HF_DATASET_BRANCH || 'main'
|
||||
|
||||
const ok = await commitToHF(env.HF_TOKEN, env.HF_DATASET_REPO, branch, filePath, combined)
|
||||
|
||||
if (ok) {
|
||||
console.log(`[Telemetry] Memory flush: ${snapshot.length} batches (${(snapshotBytes / 1024).toFixed(0)} KiB) → ${filePath}`)
|
||||
return snapshot.length
|
||||
}
|
||||
|
||||
// Commit failed — push lines back so the next request retries.
|
||||
// Prepend so ordering is roughly preserved.
|
||||
memBuffer.unshift(...snapshot)
|
||||
memBufferBytes += snapshotBytes
|
||||
if (memBufferFirstTs === 0) memBufferFirstTs = Date.now()
|
||||
console.error(`[Telemetry] Memory flush failed — ${snapshot.length} batches returned to buffer`)
|
||||
return 0
|
||||
}
|
||||
|
||||
// ── HuggingFace Hub Commit ───────────────────────────────────────────
|
||||
|
||||
async function commitToHF(
|
||||
|
||||
Reference in New Issue
Block a user