Swarm / parallel / networked architecture — no orchestrator
shared queue (durable event bus in production)
T1 (slow) · T2 (fast) · T3 (slow) · T4 (fast) · T5 (fast) · ...
Worker A
pull task
process
write result
Worker B
pull task
process
write result
Worker C
pull task
process
write result
Worker D
pull task
process
write result
results pool
idempotent, durable, per-task trace IDs
tradeoffs
✓ variable-duration balancing · ✓ scales to thousands · ✓ fault-tolerant with durable queue
✗ no central trace · ✗ ordered workflows break · ✗ debugging needs per-task IDs
failure modes: starvation (aging queue), hot-spotting (worker specialization), back-pressure (bounded queue)