🌀 Open Assignments (no single right answer)
These are research + build + write challenges. Pick one whenever you finish a week early, and 2 in the W25–W26 buffer. Deliverable: prototype + benchmark/evidence + a 2-page write-up with trade-offs. They make excellent blog posts and interview stories.
Distributed systems
- O1 Mini-Temporal: workflows as code in Go with deterministic replay (record side effects; replay from history). Compare it with your JSON-DSL engine
- O2 Kafka as the source of truth: move run history from Postgres to Kafka (compacted + event topics). Prototype it and argue whether it’s better
- O3 Distributed cron: leader election via etcd leases; exactly-once firing across 3 replicas; clock skew tests
- O4 Consistent hashing: a ring with virtual nodes + bounded loads; simulate node joins/failures; plot key movement and load variance
- O5 Raft KV (MIT Lab 4) + linearizability checking with Porcupine
- O6 CRDTs: G-counter, PN-counter, OR-set, then a collaborative workflow editor over WebSockets
- O7 Multi-region Orbit: active-active control plane, home-region runs, failover drills (on paper + a 2-region local simulation)
- O8 Exactly-once metering: billing-grade usage counting from Kafka to ClickHouse to invoices; prove there’s no double counting
Storage & performance
- O9 LSM engine in Go: memtable, WAL, SSTables, compaction, bloom filters; benchmark against bbolt (B+tree)
- O10 Java vs Go honest comparison: reimplement the LLM gateway in Java (virtual threads) and compare throughput, p99, memory, startup, and dev time
- O11 GraalVM native image for orbit-api: startup, RSS, and throughput trade-offs; what broke?
- O12 Bloom filter + count-min sketch for webhook dedupe and top-K tenants by traffic
Infra & platform
- O13 A mini service mesh sidecar in Go: a TCP/HTTP proxy with mTLS (your own CA), retries, and metrics
- O14 Custom K8s scheduler plugin placing “GPU-ish” workers by a custom resource
- O15 WASM plugin system for custom tools (wazero in Go): isolation, limits, and a host API
- O16 Zanzibar-style authz with OpenFGA: workflow sharing between users/teams/tenants
- O17 Reproduce a famous outage: a retry storm, a cache stampede, or a metastable failure, in your lab. Write the post-mortem
AI
- O18 Fine-tune vs prompt: a LoRA fine-tune of a small open model for UC1 classification vs few-shot prompting vs a cascade. Compare accuracy, cost, and latency
- O19 Semantic cache research: choose embedding model + threshold; measure precision/recall of cache hits on paraphrase sets
- O20 Agent that builds its own tools: generates a tool, tests it in the sandbox, registers it after approval. What guardrails are needed?
- O21 Multi-agent debate vs single agent: eval study on 50 reasoning tasks
- O22 Voice agent: a realtime speech pipeline (STT → LLM → TTS) with barge-in; measure end-to-end latency
- O23 GraphRAG vs hybrid RAG on a multi-hop question set
- O24 Eval-driven prompt optimization: automatic prompt search with your eval suite as the objective
Community
- O25 A merged open-source PR to something you used: Spring AI, an MCP SDK, LangGraph, franz-go, pgx, KEDA, or Envoy Gateway