A single retryable delivery failure aborted checkpoints for the whole poll batch, so fully-delivered events were re-collected and re-sent on the next poll, and an unrelated failure froze all change detection. Attach post-delivery checkpoints to each event and run them as soon as that event fully delivers; a failed event defers only its own checkpoints for retry. A subscriber skipped after an earlier retryable failure now counts as the event's failure so the event is not marked delivered to a subscriber that never received it.
Swap the Redis datastore for MongoDB via mongo-driver v2. New
internal/mongostore mirrors the old redisstore type/constant names so
consumers change only the import qualifier; document model collapses the
multi-key subscriber/settings and per-event delivery sets into single-doc
operations and a TTL-indexed delivery collection.
Config now requires MONGODB_URI (Atlas) with MONGODB_DATABASE selecting
dev vs prod on one cluster; REDIS_URL removed. Compose files run bot-only
against Atlas. Default tests stay Docker-free (poller/bot use fakes,
mongostore unit tests are pure-logic); a gated //go:build integration
suite covers the real backend via testcontainers.
tiennm99