Skip to content
Ameen Abozaid
SRB-003

Distributed data

  • 7 Lessons
  • 3.5 h practice (estimate)
On this page 9 topics

Overview

Topics

  • Consensus concepts
  • quorum
  • replication
  • consistency models

Learn first

3.5 h of practice (estimate)

Learning outcomes

  1. Explain leader-based, multi-leader, and leaderless replication and their failure behavior.
  2. Explain consensus (Raft) at the level needed to evaluate systems that use it (etcd, Kafka KRaft, distributed SQL).
  3. Reason about quorums and tunable consistency.
  4. Evaluate a datastore's consistency claims against published analyses (e.g., Jepsen).

Lessons

  1. 01Replication strategies
  2. 02Consensus and Raft
  3. 03Quorums
  4. 04Consistency models, revisited
  5. 05Reading a Jepsen report
  6. 06Demo: partitioning a three-node cluster and watching it choose
  7. 07Demo: a quorum read returning stale data, explained

Lab and exercises

Lab

SRB-003-LAB1 · Datastore evaluation

Outcomes: 1, 2, 3, 4 · Time: ~3.5 h

  1. State FulfillCo's stock ledger requirements (no oversell, multi-warehouse, failure tolerance).
  2. Pick two candidate datastores and document their replication and consistency models.
  3. Gather evidence: documentation, Jepsen findings, or a small partition experiment.
  4. Write a comparison and recommendation with caveats.

Acceptance: every consistency claim backed by documentation, Jepsen findings, or experiment; recommendation stated.

Exercises

IDOutcomeTask
SRB-003-EX011A leader fails with asynchronous replication. Describe what is lost and how to limit it.
SRB-003-EX022A 5-node cluster splits 3/2. Which side accepts writes and why?
SRB-003-EX033N=5. Choose W and R for a stock ledger and justify, then list one failure scenario to test.
SRB-003-EX043Show that W=1, R=1 with N=3 can return stale data.
SRB-003-EX054Find the version tested in a Jepsen report for a store you know and list the findings relevant to you.
SRB-003-EX064A vendor claims 'strong consistency, always available'. Write three questions to test it.
Show solutions
  • EX01 Writes not yet replicated; reduce with synchronous or semi-synchronous replication at a latency cost.
  • EX02 The 3-node side holds a majority; the 2-node side cannot commit.
  • EX03 W=3, R=3 (overlap); test slow node plus sloppy quorum or concurrent writers.
  • EX04 The write lands on replica A, the read hits replica B which has not received it.
  • EX05 Answers depend on the report; must cite version, configuration, anomaly, and whether it is fixed.
  • EX06 Under which partition? Which configuration? Which published analysis or test supports it? Also: what happens to minority-side clients?

Workplace evidence: a datastore decision you informed with consistency analysis.

Quiz

Outcomes: O1 replication · O2 consensus · O3 quorums · O4 evaluating claims.

A self-check for you: nothing is graded, saved or sent anywhere. Choose an answer, then check it. You can try again as often as you like.

  1. 1Multi-leader replication mainly adds the problem of:O1
  2. 2Asynchronous leader replication risks:O1
  3. 3Leaderless stores keep replicas close using:O1
  4. 4Raft commits an entry when:O2
  5. 5A 3-node Raft cluster tolerates:O2
  6. 6R + W > N means:O3
  7. 7Sloppy quorums can cause:O3
  8. 8Which model is strongest?O4
  9. 9A Jepsen result applies to:O4
  10. 10Absence of a Jepsen report means:O4

Practical part

Scenario quiz plus a short written evaluation of the lab's two datastores. Pass: 70% on the quiz and every claim in the evaluation backed by evidence.

Common mistakes

Trusting marketing consistency claims; quorum math without failure scenarios.

Assessment

Scenario quiz.

Workplace evidence

A datastore decision you informed with consistency analysis.

Career relevance

Multi-region trade-offs (SRA-005).

Type to search levels, phases, modules and lessons.