fix: harden telemetry for 100k MAU — memory caps, IP rate limiting, size guards

P1 fixes:
- Per-request body capped at 512 KiB (was 24 MiB — would OOM Workers)
- Flush reads at most 50 keys capped at 5 MiB combined (was unbounded)
- Dual rate limiting: per-session + per-IP (blocks session_id rotation)
- Content-Length early reject before reading body into memory
- Chunked UTF-8 base64 encoding (avoids Array.from() on full payload)
- Max 100 events per batch (was 500)
- Rate limiter map eviction at 10k entries (prevents isolate memory leak)
- Direct-commit fallback warns loudly when TELEMETRY_KV not bound
- Deployment checklist in header comment (WAF, IP rate limit, KV plan)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
faber
2026-07-10 21:22:04 -04:00
parent 2be5eebb14
commit 0fcc33f9eb

View File

@@ -3,21 +3,29 @@
*
* Receives metadata events from the frontend and buffers them in
* Cloudflare KV. When enough batches accumulate (or enough time
* passes), all pending batches are flushed as a single JSONL commit
* to HuggingFace. This prevents the 10k+ commit buildup that
* passes), pending batches are flushed as JSONL commits to
* HuggingFace. This prevents the 10k+ commit buildup that
* happens with per-request commits.
*
* URL: POST /api/telemetry
*
* Setup (Cloudflare Pages Dashboard → Settings → Environment Variables):
* HF_TOKEN — HuggingFace write token (hf_...)
* HF_DATASET_REPO — Target dataset repo (e.g. "pliny-the-prompter/g0dm0d3")
* HF_TOKEN — HuggingFace write token (hf_...)
* HF_DATASET_REPO — Target dataset repo (e.g. "pliny-the-prompter/g0dm0d3")
* TELEMETRY_FLUSH_SECRET — Bearer token for authenticated cron flush
*
* KV Binding (Cloudflare Pages Dashboard → Settings → Bindings):
* TELEMETRY_KV — KV namespace for buffering (optional; falls back to direct commit)
* TELEMETRY_KV — KV namespace for buffering
*
* File layout in the HF repo:
* telemetry/batch_<timestamp>_<hash>.jsonl
* IMPORTANT: TELEMETRY_KV must be bound in production. Without it,
* every request falls back to a direct HF commit — which will
* recreate the commit-explosion problem at scale.
*
* Cloudflare-level protections (configure in Dashboard → Security):
* - WAF rule: block POST /api/telemetry with Content-Length > 524288
* - Rate limiting rule: 20 req/min per IP to /api/telemetry
* - Restrict Origin header to godmod3.ai (optional)
* - Monthly usage alert on Workers & KV (paid plan required at scale)
*/
interface Env {
@@ -48,22 +56,29 @@ const CORS_HEADERS = {
'Access-Control-Max-Age': '86400',
}
// KV batching thresholds
const FLUSH_BATCH_COUNT = 50
const FLUSH_AGE_MS = 30 * 60 * 1000 // 30 minutes
// ── Size & batching limits ──────────────────────────────────────────
// Workers have 128 MiB memory. Keep individual requests small and
// cap the total data loaded during a flush.
const MAX_REQUEST_BYTES = 512 * 1024 // 512 KiB per inbound POST
const MAX_BATCH_EVENTS = 100 // events per request
const FLUSH_BATCH_COUNT = 50 // keys before triggering flush
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
// ── Rate Limiter (in-memory, per-isolate) ────────────────────────────
// ── Rate Limiter (dual: session + IP, in-memory per-isolate) ────────
const RATE_LIMIT_WINDOW_MS = 60_000
const RATE_LIMIT_MAX = 10
const RATE_LIMIT_MAX_SESSION = 10 // per session_id
const RATE_LIMIT_MAX_IP = 20 // per IP — catches session_id rotation
const rateLimitMap = new Map<string, number[]>()
function isRateLimited(sessionId: string): boolean {
function isRateLimited(key: string, max: number): boolean {
const now = Date.now()
let timestamps = rateLimitMap.get(sessionId)
let timestamps = rateLimitMap.get(key)
if (!timestamps) {
timestamps = []
rateLimitMap.set(sessionId, timestamps)
rateLimitMap.set(key, timestamps)
}
const cutoff = now - RATE_LIMIT_WINDOW_MS
@@ -71,11 +86,22 @@ function isRateLimited(sessionId: string): boolean {
timestamps.shift()
}
if (timestamps.length >= RATE_LIMIT_MAX) {
if (timestamps.length >= max) {
return true
}
timestamps.push(now)
// Prevent unbounded memory growth across isolate lifetime
if (rateLimitMap.size > 10_000) {
const it = rateLimitMap.keys()
for (let i = 0; i < 1_000; i++) {
const k = it.next()
if (k.done) break
rateLimitMap.delete(k.value)
}
}
return false
}
@@ -110,9 +136,6 @@ export const onRequestOptions: PagesFunction<Env> = async () => {
}
// GET ?flush=force — authenticated cron trigger for guaranteed flush.
// Requires Authorization: Bearer <TELEMETRY_FLUSH_SECRET> header.
// Set TELEMETRY_FLUSH_SECRET in CF Pages env vars, then configure a
// CF Worker scheduled event or external cron to call this every 30 min.
export const onRequestGet: PagesFunction<Env> = async (context) => {
const { request, env } = context
const url = new URL(request.url)
@@ -143,6 +166,12 @@ export const onRequestGet: PagesFunction<Env> = async (context) => {
export const onRequestPost: PagesFunction<Env> = async (context) => {
const { request, env } = context
// ── Early size rejection (before reading body into memory) ────────
const contentLength = parseInt(request.headers.get('Content-Length') || '', 10)
if (contentLength > MAX_REQUEST_BYTES) {
return jsonResponse({ error: 'Request too large', limit: MAX_REQUEST_BYTES }, 413)
}
if (!env.HF_TOKEN || !env.HF_DATASET_REPO) {
const missing = []
if (!env.HF_TOKEN) missing.push('HF_TOKEN')
@@ -151,9 +180,27 @@ export const onRequestPost: PagesFunction<Env> = async (context) => {
return jsonResponse({ error: `Telemetry not configured (missing: ${missing.join(', ')})` }, 503)
}
// ── IP-based rate limiting (catches session_id rotation attacks) ───
const clientIP = request.headers.get('CF-Connecting-IP') || request.headers.get('X-Forwarded-For') || 'unknown'
if (isRateLimited(`ip:${clientIP}`, RATE_LIMIT_MAX_IP)) {
return jsonResponse({ error: 'Rate limit exceeded — try again later' }, 429)
}
// ── Read body with size cap ───────────────────────────────────────
let bodyText: string
try {
bodyText = await request.text()
} catch {
return jsonResponse({ error: 'Failed to read body' }, 400)
}
if (new TextEncoder().encode(bodyText).byteLength > MAX_REQUEST_BYTES) {
return jsonResponse({ error: 'Request too large', limit: MAX_REQUEST_BYTES }, 413)
}
let payload: TelemetryPayload
try {
payload = await request.json() as TelemetryPayload
payload = JSON.parse(bodyText) as TelemetryPayload
} catch {
return jsonResponse({ error: 'Invalid JSON' }, 400)
}
@@ -162,40 +209,33 @@ export const onRequestPost: PagesFunction<Env> = async (context) => {
return jsonResponse({ error: 'No events provided' }, 400)
}
const MAX_BATCH = 500
const events = payload.events.slice(0, MAX_BATCH)
const events = payload.events.slice(0, MAX_BATCH_EVENTS)
const invalid = events.filter(e => !validateEvent(e))
if (invalid.length > 0) {
return jsonResponse({ error: `${invalid.length} event(s) failed schema validation` }, 400)
}
// ── Session-based rate limiting ───────────────────────────────────
const firstEvent = events[0]
const sessionKey = firstEvent.session_id
? firstEvent.session_id
: deriveSessionKey(firstEvent)
if (isRateLimited(sessionKey)) {
if (isRateLimited(`session:${sessionKey}`, RATE_LIMIT_MAX_SESSION)) {
return jsonResponse({ error: 'Rate limit exceeded — try again later' }, 429)
}
const sanitized = events.map(stripPII)
const jsonl = sanitized.map(e => JSON.stringify(e)).join('\n')
// KV values max out at 25 MiB; reject oversized batches cleanly
const KV_VALUE_LIMIT = 24 * 1024 * 1024 // 24 MiB with margin
const jsonlBytes = new TextEncoder().encode(jsonl).byteLength
if (jsonlBytes > KV_VALUE_LIMIT) {
return jsonResponse({ error: 'Batch too large', bytes: jsonlBytes, limit: KV_VALUE_LIMIT }, 413)
}
// ── KV buffered path (preferred) ──────────────────────────────────
if (env.TELEMETRY_KV) {
const now = Date.now()
const kvKey = `batch:${now}:${shortHash(jsonl)}`
await env.TELEMETRY_KV.put(kvKey, jsonl, {
expirationTtl: 604800, // 7 days — long enough for any low-volume period
expirationTtl: 604800, // 7 days
})
// Check flush conditions
@@ -213,10 +253,6 @@ export const onRequestPost: PagesFunction<Env> = async (context) => {
}
if (shouldFlush) {
// Flush is idempotent: concurrent flushes may produce duplicate
// JSONL files on HF, which is harmless (extra telemetry rows).
// KV is eventually consistent and can't do atomic locks, so we
// accept rare duplicates rather than pretend we can prevent them.
context.waitUntil(flushKVToHF(env))
}
@@ -229,6 +265,10 @@ export const onRequestPost: PagesFunction<Env> = async (context) => {
}
// ── 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')
const ts = new Date().toISOString().replace(/[:.]/g, '-')
const hash = shortHash(jsonl)
const filePath = `telemetry/batch_${ts}_${hash}.jsonl`
@@ -243,48 +283,63 @@ export const onRequestPost: PagesFunction<Env> = async (context) => {
return jsonResponse({ error: 'Failed to publish to HuggingFace' }, 502)
}
// ── KV Flush: combine all pending batches into one HF commit ────────
// ── KV Flush: read up to FLUSH_MAX_KEYS, cap at FLUSH_MAX_BYTES ────
async function flushKVToHF(env: Env): Promise<number> {
const kv = env.TELEMETRY_KV!
const allKeys: { name: string }[] = []
let cursor: string | undefined
do {
const list = await kv.list({ prefix: 'batch:', cursor })
allKeys.push(...list.keys)
cursor = list.list_complete ? undefined : list.cursor
} while (cursor)
// Only list up to FLUSH_MAX_KEYS — leave the rest for the next cycle.
// This bounds both KV read ops and memory consumption.
const list = await kv.list({ prefix: 'batch:', limit: FLUSH_MAX_KEYS })
const keys = list.keys
if (allKeys.length === 0) return 0
if (keys.length === 0) return 0
// Read values with a cumulative size cap
const flushedKeys: { name: string }[] = []
const chunks: string[] = []
for (const key of allKeys) {
let totalBytes = 0
for (const key of keys) {
const value = await kv.get(key.name)
if (value) chunks.push(value)
if (!value) {
flushedKeys.push(key) // empty key — still delete it
continue
}
const valueBytes = new TextEncoder().encode(value).byteLength
if (totalBytes + valueBytes > FLUSH_MAX_BYTES && chunks.length > 0) {
// Would exceed cap — stop here, remaining keys stay for next flush
break
}
chunks.push(value)
flushedKeys.push(key)
totalBytes += valueBytes
}
if (chunks.length === 0) return 0
if (chunks.length === 0) {
// Only empty keys found — clean them up
if (flushedKeys.length > 0) {
await Promise.all(flushedKeys.map(key => kv.delete(key.name)))
}
return 0
}
const combined = chunks.join('\n')
// Filename is derived entirely from the key set — no timestamp.
// Concurrent flushes reading identical keys produce the exact same
// filename, so HF overwrites idempotently (no duplicate rows).
const keyDigest = await sha256Hex(allKeys.map(k => k.name).sort().join(','))
const keyDigest = await sha256Hex(flushedKeys.map(k => k.name).sort().join(','))
const filePath = `telemetry/flush_${keyDigest}.jsonl`
const branch = env.HF_DATASET_BRANCH || 'main'
const ok = await commitToHF(env.HF_TOKEN, env.HF_DATASET_REPO, branch, filePath, combined)
if (ok) {
// Delete flushed keys. If a concurrent flush already deleted some,
// the delete is a no-op — KV delete on missing keys is safe.
await Promise.all(allKeys.map(key => kv.delete(key.name)))
console.log(`[Telemetry] Flushed ${allKeys.length} batches → ${filePath}`)
return allKeys.length
await Promise.all(flushedKeys.map(key => kv.delete(key.name)))
console.log(`[Telemetry] Flushed ${flushedKeys.length} batches (${(totalBytes / 1024).toFixed(0)} KiB) → ${filePath}`)
return flushedKeys.length
}
console.error(`[Telemetry] Flush failed — ${allKeys.length} batches remain in KV`)
console.error(`[Telemetry] Flush failed — ${flushedKeys.length} batches remain in KV`)
return 0
}
@@ -299,10 +354,7 @@ async function commitToHF(
): Promise<boolean> {
const url = `${HF_API}/datasets/${repo}/commit/${branch}`
// btoa() only handles Latin-1; telemetry JSONL can contain Unicode
const bytes = new TextEncoder().encode(content)
const binStr = Array.from(bytes, b => String.fromCodePoint(b)).join('')
const contentBase64 = btoa(binStr)
const contentBase64 = utf8ToBase64(content)
const ndjson = [
JSON.stringify({ key: 'header', value: { summary: `[telemetry] ${filePath}` } }),
JSON.stringify({ key: 'file', value: { content: contentBase64, path: filePath, encoding: 'base64' } }),
@@ -406,6 +458,21 @@ async function sha256Hex(str: string): Promise<string> {
return [...new Uint8Array(buf)].map(b => b.toString(16).padStart(2, '0')).join('')
}
// Chunked UTF-8-safe base64: processes 32 KiB at a time to avoid
// creating a single enormous string via Array.from() on the full input.
function utf8ToBase64(content: string): string {
const bytes = new TextEncoder().encode(content)
const CHUNK = 32 * 1024 // 32 KiB — 3-byte aligned (32768 % 3 = 0)
let binStr = ''
for (let i = 0; i < bytes.length; i += CHUNK) {
const slice = bytes.subarray(i, Math.min(i + CHUNK, bytes.length))
for (let j = 0; j < slice.length; j++) {
binStr += String.fromCharCode(slice[j])
}
}
return btoa(binStr)
}
function jsonResponse(data: unknown, status: number): Response {
return new Response(JSON.stringify(data), {
status,