There Is More Consensus in Egalitarian Parliaments

• p. 12 Section 7.3, paragraph on EPaxos vs Mencius

  with the notable exception of many conflicts for more than three replicas

• p. 12 Figure 5, 3 Replicas panel

  EPaxos,100%

• p. 12 Section 7.3, right column

  so its throughput is generally higher—with the notable exception of many conflicts for more than three replicas, when EPaxos executes an extra round per command

• p. 12 Figure 5, 3 Replicas panel

  EPaxos,25%

Judge's note: Read off the Figure 5 crop, 3 replicas (approximate): EPaxos 0% about 53k, 2% about 46k, 25% about 44k, 100% about 45.5k, Mencius about 47.5k. Mencius is at or above EPaxos for 2%, 25% and 100% with three replicas. The 25% error bar appears not to overlap the Mencius one. The stated exception ("more than three replicas") therefore does not cover everything the figure shows. The text is hedged with "generally" and the 3-replica differences are small, so this is a wording point only. Values are read from bars, not from printed numbers.

Second check: another model agreed

• p. 7 Section 4.3.3, Proposition 1, proof sketch

  only if replica Q has started Phase 1 for γ at instance Q.i

• p. 7 Figure 3, line 37

  start Phase 1 (at line 2) for no-op at L.i, avoid fast

Only the authors can say what is right here.

Judge's note: Reworked against Figures 2 and 3 with N=3 (Q, A, B). Q runs lines 1-4 for γ at Q.1 and fails before any PreAccept is delivered. A runs Explicit Prepare for Q.1 and gets ⌊3/2⌋+1 = 2 replies (A and B), neither holding a tuple for Q.1. Lines 28, 30, 32 and 34 are all false, so line 37 applies: A starts Phase 1 for no-op at Q.1 without the fast path, runs Paxos-Accept and commits no-op. A command is then committed at Q.1 for which Q never started Phase 1, so the premise is false as printed. The paper treats no-op as a command ("Read is a no-op command"; Sec. 4.6 "will commit a no-op to finalize the instance"). Uniqueness between γ and no-op needs the ballot/quorum argument, which the sketch does not give. The sketch is labelled informal and the formal proof is deferred to technical report [25], which is not among the files, so the proposition itself presumably stands.

Second check: another model agreed

• p. 7 Section 4.3.3, first paragraph after the opening

  Phase 1 is only exe- cuted for commands proposed by clients.

• p. 7 Figure 3, line 37

  start Phase 1 (at line 2) for no-op at L.i

Only the authors can say what is right here.

Judge's note: Both passages are on page 7. Line 37 runs Phase 1 for a no-op generated by the recovering replica, not proposed by a client, and Sec. 4.6 says that no-op is committed. As worded, the justification (and the Sec. 4.2 property "Any command committed by any replica must have been proposed by a client") has an unstated no-op exception. Excluding no-ops is a common implicit convention, so this is a wording gap for the authors to clarify.

Second check: another model agreed

• p. 13 Section 7.5, first paragraph

  EPaxos has lower execution latency than either Multi-Paxos or Mencius, regardless of interfering commands

• p. 13 Figure 8, bottom (99%ile) graph

  Figure 8: Latency vs. throughput for 3 replicas.

Only the authors can say what is right here.

Judge's note: Viewed page 13 and the Figure 8 crop. In the 99%ile panel the Multi-Paxos line runs at roughly 13-24 ms up to about 9,000 req/s, while the EPaxos points sit at roughly 17-36 ms over the same range. In the median panel all curves overlap at 2-3 ms at low load. EPaxos 100% and 25% track Mencius closely across the range; only EPaxos 0% is clearly lower, and only at higher load. The surrounding text of Section 7.5 carries no qualifier such as 'near saturation' or 'at high load'. Values are read off a line plot and are approximate, so this is an overstatement of wording rather than a wrong number.

Second check: another model agreed

• p. 6 Figure 2, lines 9 and 19

  reply PreAcceptOK(γ,seqγ,depsγ,L.i) to L

• p. 8 Section 4.3.3, proof sketch of Lemma 1

  its PreAcceptReplies are taken into account

• p. 8 Section 4.4, corner-case paragraph

  Accept and AcceptReply messages

Judge's note: Figure 2 on page 6 (lines 9, 19 and the leader boxes) uses PreAcceptOK and AcceptOK. Page 8 prints PreAcceptReplies and AcceptReply in the italic used for message names. The meaning is clear, so this is only a naming inconsistency between pseudocode and prose.

Second check: none (judge only)

• p. 8 Section 4.4, second paragraph

  If a un-committed conflict exists

Judge's note: The page 8 image, Sec. 4.4 second paragraph, prints "If a un-committed conflict exists, Q defers recovery".

Second check: none (judge only)

• p. 8 Section 4.4, last paragraph

  another implication of the new fast-path quorums size

Judge's note: The page 8 image, last paragraph of Sec. 4.4, prints "fast-path quo-rums size" across a line break, with the 's' present.

Second check: none (judge only)

• p. 10 Footnote 5

  The typical RTT in an EC2 cluster is 0.4 ms

Judge's note: The crop and the page 10 image show no period after "0.4 ms". Footnote 4 on the same page does end with a period.

Second check: none (judge only)

• p. 11 Section 7.1, last paragraph

  (e.g., “EPaxos 0%”’)

• p. 11 Section 7.1, last sentence

  “EPaxos 0%”’)

• p. 11 Section 7.1

  “EPaxos 0%”’

Judge's note: The crop shows the closing double quote followed by an additional single closing quote before the parenthesis. It is printed in the PDF, not an extraction artefact.

Second check: none (judge only)

• p. 12 Section 7.3, last paragraph on the page

  its throughout is mostly unaffected

• p. 12 Section 7.3, last paragraph

  otherwise its throughout is mostly unaffected

Judge's note: The page 12 image, last line of the right column, prints "its throughout is mostly unaffected". The sentence is about throughput.

Second check: none (judge only)

• p. 15 References, entry [1]

  London, UK, UK, 2000. Springer-Verlag.

• p. 15 Reference [1]

  London, UK, UK, 2000

Judge's note: Viewed page 15 and the crop. The page image shows 'pages 268–282, London, UK, UK, 2000. Springer-Verlag.' The duplicated 'UK' is in the printed PDF, not an extraction artefact. Cosmetic.

Second check: none (judge only)

• p. 15 References, entry [13]

  M. Kaptritsos, Y. Wang, V. Quema

• p. 15 Reference [13]

  M. Kaptritsos

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: that the first author of 'Eve: Execute-verify replication for multi-core servers' (OSDI 2012) is spelled Kapritsos. The page image and crop do show 'M. Kaptritsos, Y. Wang, V. Quema, A. Clement, L. Alvisi, and M. Dahlin', so the quoted text is on the page. The name occurs nowhere else in the files, so the correct spelling cannot be checked from them.

Second check: none (judge only)