Refresh rates and SLOs

QuantDesk treats the Postgres hub as the canonical cache boundary. APIs read the latest written state, and websockets fan out incremental updates as they arrive.

Expected cadences

Surface	Typical cadence	Notes
`market-stats` hub row	Event-driven, usually sub-30s	Updated by mark, trade, OI, and funding writers
`candles`	1m base writes with rollups	Higher resolutions are rolled up on read
`trade` fanout	Near real-time	Derived from `place_order_v2` fills
`news`	Poll-driven	Depends on RSS / API provider cadence
`reference/perps`	~15s default	Configurable collector interval
`data-health`	Snapshot at request time	Summarizes collector freshness

Staleness thresholds

Asset class	Default threshold	Why
Crypto	`30s`	24/7 market; prolonged silence is suspicious
TradFi index synthetics	Configurable, session-aware	Off-hours behavior follows market session policy

When a symbol exceeds its freshness threshold, /api/v2/data-health reports it as stale instead of hiding the condition. For TradFi index synthetics, stale evaluation is session-aware: off-hours US500-PERP / US100-PERP rows are not flagged stale by age alone when the cash session is closed.

Client pattern

Load a snapshot with REST.
Subscribe to market_stats_update over Socket.IO.
Reconcile on reconnect by refetching the REST snapshot.

That model avoids using websockets as ledger truth while still keeping the UI current.

Operational notes

Collectors write heartbeats into hub_collector_status.
Backend health summaries aggregate collector status plus latest market-stat timestamps.
Dataset exports capture storage_format, storage_path, checksum_sha256, row count, and export timestamp for reproducibility.