Consensus Under Partial Synchrony: From Paxos to Raft

Monthly research note. Theme: Distributed Systems Under Failure.

TL;DR

Consensus Under Partial Synchrony: From Paxos to Raft as an engineering constraint: write down assumptions, make invariants executable, and design operational recovery as part of correctness.

Key takeaways

• Treat membership changes and compaction as protocol events—not operational details.
• Mixed-version operation is the default; upgrades must preserve invariants.
• Expose protocol state (epoch/term/commit index) as first-class telemetry.
• Design rollbacks as part of the happy path.
• Make failure modes explicit and observable.

Why this matters

• Most protocol bugs hide in timeouts, retries, and membership changes.
• If your protocol isn’t testable under reordering, it isn’t deployable.
• Global systems fail in correlated ways (regions, dependencies, routing).
• Safety failures are permanent; liveness failures are (sometimes) recoverable.

Key questions

• Which components require determinism for reproducibility?
• What is the unit of ordering (per key, per partition, global)?
• Which safety property is non-negotiable (no double-commit, no forks, no split brain)?
• How do clients discover leaders safely (and what happens during flaps)?
• How do you prevent overload from becoming inconsistency?
• What is the compaction story (snapshots, log truncation, state transfer)?

Assumptions

• Clients retry and amplify load right when the system is weakest.
• Reconfigurations happen mid-incident (the worst time).
• Workload is skewed: hot keys exist and dominate.
• Packets can be duplicated and reordered; acks can be lost.

Non-goals

• Treating membership as static or human-managed only.
• Assuming the network eventually behaves “nicely” under load.

Model & invariants

For quorum-based protocols, the intersection property is the backbone of safety:

Liveness is always conditional: specify when progress is expected and what you do otherwise.

Write down the safety property first. If it’s not written, it’s not implemented.

Security properties

• Downgrade resistance: negotiation can’t silently weaken security posture.
• Least authority: privileges are scoped by purpose and time.
• Replay resistance: duplicated inputs do not change outcomes.
• Evidence: critical actions emit verifiable audit events.

Failure modes

• Timeout ambiguity causing double-apply or partial state transitions.
• Config drift that weakens security posture over time.
• Recovery paths that only work when nothing is broken.
• Observability gaps during incidents (missing evidence).

Design sketch

sequenceDiagram
  participant C as Client
  participant L as Leader
  participant F1 as Follower 1
  participant F2 as Follower 2
  C->>L: propose(cmd)
  L->>F1: appendEntries
  L->>F2: appendEntries
  F1-->>L: ack
  F2-->>L: ack
  L-->>C: commit(result)

Implementation notes

Your protocol is an interface between failures and invariants. Encode both.

Operational invariants to monitor:
- leader_changes_per_minute
- commit_index_monotonic
- snapshot_install_failures
- quorum_acks_latency_p99
- rejected_requests_due_to_admission_control

Verification strategy

• Model checking the smallest core (timeouts, election, reconfiguration).
• Stress + skew tests: hot keys, slow disks, noisy neighbors.
• Linearizability checks for read/write APIs that claim it.
• Upgrade tests: mixed versions and rolling deploy invariants.
• Jepsen-style fault injection: partitions + reordering + client retries.

Operational notes

• Expose protocol state: term/epoch, leader, commit index, config version.
• Make client behavior part of the system: document retry semantics.
• Treat compaction and snapshot install as first-class SLOs.
• Rate-limit retries and apply admission control before saturation.
• Prefer monotonic time sources for leases; alert on clock discontinuities.

What to monitor

• Authz failures and policy denials (unexpected spikes).
• Rollback events and the conditions that triggered them.
• Invariant violation rate (should be ~0).
• Admission-control / rate-limit rejections (by reason).
• Error budget burn + tail latency under load.

Rollback plan

• Keep dual-write / dual-verify windows where appropriate.
• Prefer backward-compatible changes; avoid “flag day” upgrades.
• Define an explicit rollback trigger (metrics + thresholds).
• Use canaries and staged rollout; stop early when signals degrade.
• Preserve evidence (configs, artifacts, audit logs) to reconstruct what changed.

Evidence

• Jepsen (1) — Testing correctness under partitions and faults.
  • Evidence: Turn faults into test cases; prioritize partition and clock-skew scenarios that violate user-visible guarantees.
• Learn TLA+ (2) — Practical entry point for specification and model checking.
  • Evidence: Model the smallest thing that can break; use model checking to validate invariants before optimizing.

Open questions

• Where does your protocol assume synchrony without admitting it?
• How do you prevent “operator fixes” from changing safety properties?
• Which invariants are violated first under overload: latency, availability, or correctness?
• What is the worst-case recovery time after a leader + disk failure?

Checklist

• Assumptions listed and reviewed.
• Failure modes enumerated with mitigations.
• Costs bounded (CPU/memory/bandwidth) under adversarial inputs.
• Safety properties stated as invariants.
• Rollback plan rehearsed and automated.
• Telemetry captures correctness signals.

Further reading

• Jepsen — Testing correctness under partitions and faults.
• In Search of an Understandable Consensus Algorithm (Raft) — Consensus with explicit state machines and practical tradeoffs.
• Paxos Made Simple (Lamport) — Agreement basics and the invariants that matter.
• Time, Clocks, and the Ordering of Events (Lamport) — Causality, ordering, and why clocks are tricky.
• Learn TLA+ — Practical entry point for specification and model checking.
• Site Reliability Engineering (Google) — Error budgets, incident response, and reliability as an engineering discipline.