The ingest pipeline

There is a single ingest path for every kind of source — worker push, file upload and manual create all funnel through it (result_router.go):

ingest (IngestActivityFromWorker)   identity dedup + persist source/stream (atomic)
  → ReclusterBucket                 fuzzy-match + union-find group same-workout sources
  → RecomputeActivityFromSources    domain.Merge over all non-detached sources
  → follow-ups                      best-efforts, segments (+ranks), training load, PR notifs

Durability rule

The ingest is durable; follow-ups never fail it. Follow-ups run nil-tolerant and log on failure. This means a transient hiccup computing, say, training load can never lose your activity — it just gets recomputed later. Any new follow-up stage must hold to this pattern.

Async via NATS

Workers and core communicate over NATS. Subjects are cairn.<kind>.<job_type>.<provider>, with the provider as the last token:

• cairn.jobs.<type>.<provider> — work for a provider on the CAIRN_JOBS work-queue stream: fetch_source, reconcile, parse_blob, import_metrics, backfill. The subject is exact, e.g. cairn.jobs.fetch_source.strava; workers also publish fetch_source sub-jobs here from reconcile and webhooks.
• cairn.results.<type>.<provider> — JobResult events back to core on the CAIRN_RESULTS stream; reconcile results are account-suffixed, cairn.results.reconcile.<provider>.<external_account_id>.
• cairn.discover.<provider> — request/reply listing of importable items, served by the worker.
• cairn.tokens.<provider>.* / cairn.creds.<provider>.get — token and credential exchange (request/reply, served by core).
• cairn.blobs.presign_upload.<provider> / ..presign_download.<provider> — presigned S3 URLs for claim-checked results and raw blobs.
• cairn_worker_presence (KV) — capability-manifest heartbeats.

Core is the job producer + result consumer; a worker is the job consumer + result producer. The exact subjects and their semantics are in the provider contract.

Stream or core publish

Some subjects are stored in JetStream streams (cairn.jobs.>, cairn.results.>, cairn.events.>, cairn.dlq.>, cairn.webhooks.>), others are plain request/reply subjects with a core NATS subscriber. Core's bus, and the Strava worker's, decide per subject before publishing: they look up whether a stream covers the subject ($JS.API.STREAM.NAMES with a subject filter), cache the answer, and use a JetStream publish (with Nats-Msg-Id dedup) for stream subjects and a core NATS publish for the rest. A miss is re-checked after a while, so a stream created later is picked up.

Before core v0.3.0 / Strava v0.2.6 a JetStream publish to a core-only subject waited five seconds for an ack that never came and delivered the message twice on retry.