:hedgehog: PostHog is the leading platform for building self-driving products. Our developer tools – AI observability, analytics, session replay, flags, experiments, error tracking, logs, and more – capture all the context agents need to diagnose problems, uncover opportunities, and ship fixes. Steer it all from Slack, web, desktop, or the MCP.
Python ab-testingai-analyticsanalyticscdpdata-warehouseexperimentsfeature-flagsjavascriptproduct-analyticspythonreactsession-replaysurveystypescriptweb-analytics
export function parseList(text: string | undefined): string[] {
if (!text) {
return []
}
return text.split(',').map((item) => item.trim())
}
// parseInt returns NaN on junk, and NaN flows silently into Worker.create / pool sizing,
// so every numeric env var goes through a fallback plus a positivity bound.
export function parsePositiveInt(raw: string | undefined, fallback: number): number {
const n = parseInt(raw ?? '', 10)
return Number.isFinite(n) && n > 0 ? n : fallback
}
export function parseFraction(raw: string | undefined, fallback: number): number {
const n = parseFloat(raw ?? '')
return Number.isFinite(n) && n > 0 && n <= 1 ? n : fallback
}
export const config = {
// Temporal
temporalHost: process.env.TEMPORAL_HOST || '127.0.0.1',
temporalPort: process.env.TEMPORAL_PORT || '7233',
temporalNamespace: process.env.TEMPORAL_NAMESPACE || 'default',
taskQueue: process.env.TEMPORAL_TASK_QUEUE || 'rasterization-task-queue',
temporalClientRootCA: process.env.TEMPORAL_CLIENT_ROOT_CA,
temporalClientCert: process.env.TEMPORAL_CLIENT_CERT,
temporalClientKey: process.env.TEMPORAL_CLIENT_KEY,
// Worker
logLevel: (process.env.LOG_LEVEL || 'info') as 'debug' | 'info' | 'warn' | 'error',
maxConcurrentActivities: parsePositiveInt(process.env.MAX_CONCURRENT_ACTIVITIES, 4),
// Above these fractions of the cgroup limit, the worker takes no activity beyond its first slot.
tunerTargetMemoryUsage: parseFraction(process.env.TUNER_TARGET_MEMORY_USAGE, 0.7),
tunerTargetCpuUsage: parseFraction(process.env.TUNER_TARGET_CPU_USAGE, 0.9),
tunerRampThrottleMs: parsePositiveInt(process.env.TUNER_RAMP_THROTTLE_MS, 10_000),
// On SIGTERM (deploy, scale-down, node replacement) the worker waits this long for in-flight renders before
// the SDK cancels them. The SDK default is 0, which cancels every running render at once. Keep it below the
// pod's terminationGracePeriodSeconds (1800s in prod) so the process still exits before it is killed.
shutdownGraceTimeMs: parsePositiveInt(process.env.SHUTDOWN_GRACE_TIME_MS, 25 * 60 * 1000),
browserRecycleAfter: parsePositiveInt(process.env.BROWSER_RECYCLE_AFTER, 100),
// Browsers held warm beyond this are closed on release: each idle Chromium pins 100-250MB RSS,
// and a pod that once ran at full concurrency would otherwise keep that footprint forever.
maxIdleBrowsers: parsePositiveInt(process.env.MAX_IDLE_BROWSERS, 2),
disableBrowserSecurity: process.env.DISABLE_BROWSER_SECURITY === '1',
captureBrowserLogs: process.env.CAPTURE_BROWSER_LOGS === '1',
screenshotFormat: (process.env.SCREENSHOT_FORMAT || 'jpeg') as 'png' | 'jpeg',
// Hard cap on a single beginFrame; mass-image-decode stalls legitimately run 30-60s, so keep
// this well above that. It only exists to bound a truly wedged compositor.
beginFrameTimeoutMs: parsePositiveInt(process.env.BEGINFRAME_TIMEOUT_MS, 120_000),
screenshotJpegQuality: parsePositiveInt(process.env.SCREENSHOT_JPEG_QUALITY, 80),
metricsPort: parsePositiveInt(process.env.METRICS_PORT, 6738),
// Encryption
secretKey: process.env.TEMPORAL_SECRET_KEY || process.env.SECRET_KEY,
fallbackKeys: parseList(process.env.TEMPORAL_FALLBACK_SECRET_KEYS ?? process.env.SECRET_KEY).filter(Boolean),
// S3
s3Endpoint: process.env.VIDEO_EXPORT_OBJECT_STORAGE_ENDPOINT,
s3Region: process.env.VIDEO_EXPORT_OBJECT_STORAGE_REGION || 'us-east-1',
// `s3://bucket/prefix/` entries a render's source_s3_uri may point into. Empty disables file sources.
sourceS3Prefixes: parseList(process.env.RASTERIZER_SOURCE_S3_PREFIXES ?? '').filter(Boolean),
// Bounds a zstd source's decompressed size, which its compressed size under the gate above does not. The
// body reaches the page as base64 text, so it has to stay well under V8's string limit.
maxSourceDecompressedBytes: parsePositiveInt(process.env.MAX_SOURCE_DECOMPRESSED_BYTES, 256 * 1024 * 1024),
// Recording API. The dev recording-api listens on 6741 (bin/temporal-recording-rasterizer-worker).
recordingApiBaseUrl: process.env.RECORDING_API_BASE_URL || 'http://localhost:6741',
// The listing hits ClickHouse through recording-api, so internalFetch's 3s default aborts it under load.
blockListingTimeoutMs: parsePositiveInt(process.env.BLOCK_LISTING_TIMEOUT_MS, 30_000),
// Renders above this many compressed bytes fail permanently instead of loading the pod into its
// memory limit. Deliberately generous: the every-render byte log is what tightens it over time.
maxRecordingCompressedBytes: parsePositiveInt(process.env.MAX_RECORDING_COMPRESSED_BYTES, 512 * 1024 * 1024),
// Player
siteUrl: process.env.SITE_URL || 'http://localhost:8000',
playerHtmlPath: process.env.PLAYER_HTML_PATH || '/code/common/replay-headless/dist/player.html',
// Serve the player under a script-locking CSP (nonce). On by default: recordings are untrusted,
// and the CSP is the backstop if the replay iframe sandbox ever regresses. ENABLE_PLAYER_CSP=0
// is the escape hatch if a player build ships without the nonce placeholder.
enablePlayerCsp: process.env.ENABLE_PLAYER_CSP !== '0',
}