What PALM SOUQ sends to Datadog
PALM SOUQ pushes metrics to Datadog's metrics intake using an API key you provide, from the gateway region that observed the traffic. The metric set covers tokens by type (input, cached input, output, reasoning), cost in USD, request count, breaker trips, fallback decisions, quota state and gateway latency. Each point is tagged with tenant, tier, model, provider and feature.
Breaker trips and quota changes are also sent as Datadog events, so they appear as overlays on any dashboard and can trigger monitors. An on-call engineer looking at an error spike sees the breaker event that caused it.
Tag cardinality, handled
A SaaS with thousands of tenants can create a lot of custom metric series if every metric is tagged by tenant. PALM SOUQ lets you choose: tag every tenant, tag only the top N tenants by spend and roll the rest into an other bucket, or tag tenants on cost metrics only. The console shows the estimated series count before you save.
Per-tenant detail is always available in PALM SOUQ itself, so you can keep Datadog lean and use it for alerting and correlation while finance-grade attribution lives in the gateway.
Alerts that act, not just notify
A Datadog monitor on LLM spend tells you something went wrong. A PALM SOUQ breaker stops it. The two are complementary: PALM SOUQ enforces limits in the request path in under 50 ms, and Datadog gives you the wider picture across services and the alert routing your team already uses.
You can also go the other way. A Datadog webhook can call the PALM SOUQ policy API to tighten a tenant's quota during an incident, for example when a downstream dependency is failing and agents are retrying more than usual.
Rolling it out safely
Start with top-N tenant tagging and a small N, then review the custom metrics usage page in Datadog after a day. Raise N only if your team actually filters by tenants beyond the top group; per-tenant detail for the long tail is always available in PALM SOUQ without adding series.
Build one dashboard first: cost per hour by model, breaker trips as an event overlay, and gateway latency at p99. That combination answers most incident questions about LLM spend in a single view.
Set monitors on the PALM SOUQ metrics that predict cost rather than report it: token velocity per tenant, breaker trips per hour and the share of requests served by fallback. Those move before the dollar figure does, and they route through the alerting your team already watches.
