Atlas — the telemetry plane
Every coworker container and session in this platform reports dimensional metrics into a
Netflix Atlas standalone instance at
localhost:7101 — Atlas's standalone mode is an in-memory time-series store, which is
the point: it is the hot, queryable working set for cross-container/cross-session telemetry, with durable
history landing in the postgres 18 warehouse the data-engineering coworker owns (see /team).
The metric contract
Two real metric families, both instrumented via Netflix Spectator semantics (counter/timer/gauge with tag maps):
1. conductor.api.calls — the Conductor worker poll/complete loop
Language-specific coworker workers (the coworkers-{py,ts,ja,rs,go} container images) interact
with the Conductor server via a polling model, and every API call tracks the same dimensional contract:
| tag | values |
|---|---|
application | coworkers |
language | py | ts | ja | rs | go |
api_method | poll | update | complete |
status | success | failure |
2. coworkers.api.* — the coworkers API itself
From packages/coworkers/typescript/src/metrics.ts, which adopts
@opencoworkers/telemetry (the repo's nflx-spectator facade + AtlasPublishWriter,
defaulting to http://localhost:7101 with 60s buckets). Publishing is explicit opt-in
(ATLAS_PUBLISH=1); with no reachable Atlas the meters still record in-memory and the API keeps
serving — degraded, never crashed.
| metric | type | tags |
|---|---|---|
coworkers.api.requests | counter | route, status |
coworkers.api.latency | timer | route |
coworkers.ooda.tokens | counter | phase |
coworkers.workflow.runs | counter | status |
Atlas Stack Language (ASL) examples
Atlas graphs are queried with a stack language against GET /api/v1/graph?q=... on the
standalone instance. Three real queries against the contract above:
# conductor API call rate, one line per language
GET http://localhost:7101/api/v1/graph?q=name,conductor.api.calls,:eq,(,language,),:by
# failures only, still split by language
GET http://localhost:7101/api/v1/graph?q=name,conductor.api.calls,:eq,status,failure,:eq,:and,(,language,),:by
# total call volume, single summed line
GET http://localhost:7101/api/v1/graph?q=name,conductor.api.calls,:eq,:sumThe same shapes apply to the coworkers.api.* family — e.g.
name,coworkers.api.requests,:eq,(,route,),:by for per-route request rates.
localhost:7101, in-memory) and this Worker does not proxy live
Atlas graphs — wiring a read-only graph proxy (so ASL queries render here) is a documented next step, not
something this page pretends already works.