Shaped SWE 2026

Assignments - Phase 3

4 min read

📝 Assignments: Phase 3 (Distributed Systems: the durable engine)

W10 - Durable Execution v1 (Go) 🔴

🎯 A10 Turn orbit-dag into orbit-engine (Orbit v2 - Durable Workflows)

  • An append-only run_events table; run state = replay(events); snapshots every N events
  • Task queue: SELECT … FOR UPDATE SKIP LOCKED LIMIT n; leases (leased_until) + heartbeats; expired leases are re-queued
  • Fencing token = lease epoch; CompleteTask(taskId, epoch) is rejected if the epoch is stale
  • Idempotency key per step attempt passed to tools; an idempotent sink to verify effects
  • Acceptance (chaos test): 1,000 runs × 5 steps while a script kill -9s random workers every 2 s and SIGSTOPs one for 30 s (a “GC pause”) → all runs complete, 0 duplicate effects, 0 stale completions accepted

🔬 Lab: MIT 6.5840 Lab 1 MapReduce 🧠 Cognitive: First-principles: write the exact interleaving where a paused worker corrupts state without fencing, then show how the epoch prevents it 🌀 Open: Postgres queue vs Kafka vs Redis Streams for task dispatch at 100 / 10k / 1M tasks/s: where does each break? Score: __/28

W11 - Kafka Backbone, Triggers & Webhooks

🎯 A11

  • Transactional outbox in orbit-api + engine → Debezium (or a Go relay) → run.events, usage.events (Protobuf + Schema Registry, key = runId/tenantId)
  • orbit-hooks (Go): inbound webhooks (HMAC verification, dedupe by delivery ID), cron triggers (next-fire index), Kafka-topic triggers → StartRun
  • Outbound webhooks: HMAC signatures, exponential backoff, retry topics (1m/10m/1h), DLQ + replay CLI
  • UC9: meeting transcript webhook → summarize → extract action items (schema) → create tasks via tool → idempotent on redelivery
  • Acceptance: send each inbound webhook 3× → exactly 1 run; kill the relay mid-stream → no lost events

🔬 Lab: Gossip Glomers 1–3 (echo, unique IDs, broadcast 3a–3e) (Kafka Internals for the log challenge later) 🧠 Cognitive: Predict: ordering guarantees for keyed events across 6 partitions during a consumer rebalance → verify with an experiment 🌀 Open: Choreography vs orchestration for Orbit’s internal flows (billing, notifications, search indexing) Score: __/28

W12 - Sagas, HITL, Versioning & Resilience

🎯 A12

  • human step: approval inbox API, signals, timeout → escalate timers
  • Compensation handlers + reverse-order compensation on failure
  • UC3 refund agent: lookup order (tool) → policy check → propose a refund → approval → payment tool (idempotent) → notify; on a payment failure → compensate
  • Workflow versioning: in-flight runs stay on v1 while v2 publishes; a migration policy for long waits
  • Circuit breakers per LLM provider (gobreaker) + fallback provider; bulkheads per step type
  • Acceptance: 100 UC3 runs with random worker kills + a 20% payment tool failure rate → 0 double refunds; every refund has an audit trail

🔬 Lab: MIT 6.5840 Lab 2 (KV server); run one sample in Temporal (Go SDK) 🧠 Cognitive: Pre-mortem of UC3; map each failure to its compensation 🌀 Open: ADR: “Our engine vs Temporal”: when would you migrate? What do you lose and gain? Score: __/28

W13 - Raft & Multi-Tenant Fairness

🎯 A13

  • Per-tenant queues + weighted fair scheduling (deficit round robin) + per-tenant concurrency caps + priority tiers
  • Admission control: 429 + Retry-After when a tenant’s backlog exceeds its quota
  • Run search: OpenSearch projection from run.events (CQRS), rebuildable from scratch
  • Acceptance: tenant A floods 100k runs; tenants B–E (100 runs each) finish with p95 queue wait < 2 s

🔬 Lab (⚫): MIT 6.5840 Lab 3A (leader election) + 3B (log replication) 🧠 Cognitive: Feynman: explain Raft election safety + log matching in 5 minutes 🌀 Open: Noisy neighbors at 10,000 tenants: shared pools vs cells vs dedicated workers for enterprise tiers Score: __/28

W14 - v2 Ship & Chaos Day

🎯 A14 Ship Orbit v2 - Durable Workflows: ADRs, sequence diagrams (happy, retry, lease expiry, compensation), README update 🔬 Lab: Gossip Glomers 4 (grow-only counter) + 5 (Kafka-style log); Raft 3C if you have time

⚫ Phase boss fight: Chaos Day. 5,000 runs across UC1/UC3/UC9 while randomly: killing engine and worker pods, restarting a Kafka broker, failing over Postgres, injecting 2 s of latency into the LLM mock (toxiproxy)

  • Acceptance: 0 lost runs, 0 duplicate side effects, all SLO breaches explained. Write it as a post-mortem (timeline, impact, root causes, action items)

🧠 Cognitive: Constraint flip: redesign the engine for 1M runs/s. What changes first? Score: __/28

Graph View

Backlinks