Why

The counter owns all standard-window counts for each routed hashtag, so geo and topic expansion stays local instead of multiplying network traffic.

Processing

Consume ordered raw partitions on ARR-3, reject offsets already covered by the checkpoint, and assign the event to a minute pane.
Expand each event into geo ancestors times all-plus-topic labels, then update exact pane rings, rolling sums, and velocity baselines.
Append state before the source offset on ARR-17; emit changed and closed-pane candidates on ARR-4.
For ARR-18, replay the latest checkpoint and raw offsets, publishing only after the source vector is complete.

Core logic

def consume(task, rec): # ARR-3 if rec.off ≤ task.ckpt: return pane = minute(rec.time) for scope in expand(rec): state.incr(scope, rec.tag, pane) state.roll(scope, rec.tag, pane) changelog.append(delta(rec), rec.off) # ARR-17 if flush(): emit_candidates() # ARR-4 checkpoint_state_before_offset()

Machine count

ceil(N-QPS-PEAK ÷ (N-COUNTER-CAP × 0.55)) + + one-AZ headroom = 20 nodes

What one more machine buys

ReadsAdds replay bandwidth and reduces recovery contention.

WritesAdds counter capacity after task rebalance, but a single hot hashtag still needs salting.

Arrows

ARR-3/ARR-18 in — raw flow or recovery replay.
ARR-4/ARR-17 out — candidates and durable state.