Pith. sign in

REVIEW 3 major objections 4 minor 17 references

Not All Reads Are Conflicts: A Write-Only Analysis of the Sui Blockchain

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that standard read+write conflict graphs overstate contention on Sui, and that a write-only graph—edges only when two transactions write the same object—gives a lower bound on contention and a practical upper bound on para

desk verdict A useful W-only methodology undermines its own headline by excluding a deterministic dependency class it dismisses as conditional. read the letter →

arxiv 2607.26691 v2 pith:WXQLEZ4R submitted 2026-07-29 cs.CE

Haygen Tsoi , Alberto Sonnino , Philipp Jovanovic This is my paper · ORCID
classification cs.CE
keywords Suiblockchainparalleltransactionexecutionwrite-onlyconflictgraphcontentionmeasurementsystemclockartifactDeepBookmaximalextractablevalueobject-centricdatamodel
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that prior measurements of Sui's execution contention are systematically too pessimistic because they count read-only accesses as conflicts, even though Sui's engine never serializes on reads. Its remedy is a write-only conflict graph, where an edge means two transactions both modify the same object—a guaranteed serialization event. Applied to Sui mainnet data through 2025, this changes the picture: the hub-and-spoke structures reported earlier collapse once the shared system clock is removed, and the remaining topology is clique-like and assortative. Measured this way, Sui's optimal-parallelism headroom at routine load is about 1× rather than 1.2–1.4×, a 30–40% tightening. The paper also uses the graph to show that DeepBook dominates contention by volume but not by special bottleneck logic, and that 10–50% of USD-denominated value flows through sequentially constrained paths.

What carries the argument

The write-only conflict graph: two transactions share an edge iff their write sets intersect, meaning both mutate the same shared object. Each such edge corresponds to a write-serialization event that Sui's engine must preserve, so the graph is a lower bound on true contention (and a practical upper bound on achievable parallelism). The analysis is carried on per-checkpoint graphs reconstructed from transaction traces, using LSP/χ and gas-weighted bounds for parallelism and union-find grouping to isolate application ecosystems.

What would settle it

Instrument Sui's scheduler at runtime on a sample of checkpoints and compare actual execution stalls against W-only graph predictions; if transactions that only read a shared object written earlier in the same checkpoint are observed waiting, the W-only graph undercounts true contention.

Watch

Extended reading notes

Core claim

The central discovery is that the conflict model, not the engine, was inflating Sui's apparent contention. Because Sui's scheduler, version manager, congestion tracker, and checkpoint accounting all key off exclusive mutable access, a read of a shared object is never an execution dependency. The paper builds W-only graphs per checkpoint and shows three consequences: the previously reported star topology is an artifact of every user transaction reading the system clock; the true graph is clique-dominated and highly assortative; and the W-only bound cuts the estimate of achievable speedup by 30–40% at matched load. DeepBook's 87.3% share of application-level conflicts tracks its transaction vo

Load-bearing premise

The lower bound is sound only if Sui's engine genuinely serializes exclusively on mutable shared-object writes—the paper pins this to four source locations—and if no write-to-read version forwarding within a checkpoint ever constrains execution; if either fails, sequentiality can be higher than the W-only graph shows.

Editorial extensions

If this is right

  • Under the W-only model, routine Sui checkpoints have near-zero headroom for extra parallelism (~1× vs. the earlier 1.2–1.4×), so the bottleneck is workload composition, not the engine.
  • The system clock's prologue-to-prologue clique accounts for 68% of W-only conflict events and is housekeeping, not user contention.
  • DeepBook's per-pool and per-balance-manager sharding means contention scales with volume, so a central limit order book need not be a sequential bottleneck on an object-centric chain.
  • 50–90% of USD-denominated value flows through parallel paths currently outside typical reordering-based MEV; 10–50% is order-sensitive and potentially exposed.
  • The W-only methodology transfers naturally to other object- or resource-based chains for cross-platform contention comparisons.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the W-only graph truly captures all guaranteed serialization, then future parallelism improvements on Sui should target write-hot objects rather than read patterns; further sharding of hot objects may yield diminishing returns at current volumes.
  • The 68% system-clock share suggests an architectural alternative—exposing time and randomness as validator-side metadata rather than on-chain mutable objects—could eliminate most structural conflicts without changing the user workload, but would not reduce application contention.
  • A testable extension is to count intra-checkpoint write-to-read forwarding via object_changes; if those conditional edges frequently materialize as stalls, the W-only lower bound could understate true contention and the true headroom would lie between the W-only and R+W estimates.
  • The 10–50% economic-exposure figure describes the observed workload, not a ceiling: a sophisticated MEV actor could inject new conflicts and migrate some currently-parallel value into the sequential burden.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a write-only (W-only) conflict model for Sui mainnet checkpoint data, arguing that because Sui's execution engine serialises only on mutable shared-object accesses, the previously used read+write (R+W) conflict graphs contain spurious edges (most prominently from every transaction reading the system clock). The authors construct per-checkpoint W-only conflict graphs from a deterministic sample of checkpoints, measure graph-theoretic contention metrics, and report three main findings: (1) W-only graphs are clique-dominated and highly assortative rather than hub-and-spoke, tightening the estimate of optimal-parallelism headroom by roughly 30–40% at matched load regimes; (2) DeepBook dominates contention by transaction volume but does not show uniquely sequential logic, consistent with its sharded object design; and (3) 10–50% of USD-denominated value flows through sequential-burden paths, bounding the share currently exposed to reordering-based MEV. The paper provides source-code pins for the serialisation semantics, a reproducibility repository, and an explicit discussion of limitations, including the exclusion of intra-checkpoint write-to-read forwarding.

Significance. If the findings hold, the paper makes a useful methodological contribution to blockchain workload analysis: it gives a principled way to separate engine-imposed serialisation from read-side artefacts in object-centric chains, and it provides concrete quantitative bounds on Sui's residual parallelism and on the economic value flowing through sequentially constrained paths. The paper is honest about its limitations, pins its central semantic claim to four specific Sui source locations, and ships code and data-processing scripts. The DeepBook result is genuinely interesting and empirically grounded, and the W-only/R+W bracketing idea is a natural and potentially transferable analytical framing. The main value is in the measurements and the proposed methodology, not in any theoretical derivation.

major comments (3)
  1. [§2.3, §8] The treatment of intra-checkpoint write-to-read forwarding is load-bearing and, as stated, not conservative for the quantitative claims. The paper calls the T_w1 → T_r edge 'conditional' and excludes it, but under Sui's MVCC version-assignment a reader that pins to obj@v1 produced by T_w1 in the same checkpoint has a genuine data dependency on T_w1: no valid schedule can execute T_r before T_w1 has produced v1, even if the scheduler happens to run T_w1 first. The omitted edges are therefore dependencies, not merely timing-dependent waits. The W-only graph remains a lower bound on write-write conflicts, but it is not a faithful lower bound on actual execution serialisations unless the W→R population is measured. Since '30–40% tighter' (Finding RQ2) and the '10–50% sequential value' (Finding RQ4) are computed from W-only LCC/LSP statistics, they could materially understate true contention.
  2. [§4.2, Table 1, Finding RQ2] The headline 30–40% tightening compares W-only numbers from the authors' fixed-stride ~100 cp/day sample with R+W numbers from Biton & Friedman's 1.5M full-trace dataset, and §8 acknowledges that some data may be missing. 'Matched load regimes' is asserted but not precisely defined: the reader is asked to accept that the routine 150M and high-load 66M intervals in [6] correspond to the authors' routine and largest checkpoints, but no detailed matching protocol is given. A load match by checkpoint size or transaction count does not control for object-access mix, checkpoint boundary definitions, or sampling period. Part of the 30–40% narrowing could therefore be a dataset or sampling effect rather than a semantic effect of the W-only model. Please recompute R+W and W-only on the same checkpoints (or on a precisely defined intersection of the two datasets) and report paired distributions, n
  3. [§6, Finding RQ4] The economic-exposure figure 10–50% is presented as the fraction of USD value 'gated by sequential execution paths' and 'currently exposed' to ordering effects. This overstates what is actually measured. The measurement is the fraction of value inside the W-only LCC, which is a lower bound on dependency chains once W→R forwarding is included, and LCC membership is a graph-theoretic proxy, not a per-transaction order-sensitivity test. The paper does say 'in principle' and 'potentially' later in the section, but the finding and the abstract do not carry that caveat. Please present the value share as a bound with the W→R sensitivity analysis included, and remove the implication that the 10–50% range is a proven MEV-exposed share.
minor comments (4)
  1. [§3.1] The fixed global stride of every 3455th checkpoint is justified by pipeline capacity, but the pre-Mysticeti lower checkpoint rate means the sample density changes across eras. The limitation is acknowledged in §8, but it would help to state explicitly that all cross-era comparisons (e.g., the June 2023 spike in §4.3) are subject to this density change.
  2. [Table 1] The row 'Median LSP/χ lower bound' reports 0.667 for W-only. Since LSP is a chain length and χ is a colouring number, a ratio below 1 is possible, but the direction of the 'lower/upper bound' terminology is not defined here and will confuse readers. Please define the exact quantity and which graph parameter is bounded by which side.
  3. [§7, References] There are a few typos: 'Frideman' in §7, 'Mysticei' in §7, and the reference [6] lists 'Biton & Frideman' while elsewhere it is 'Biton & Friedman'. Figure 5's caption says 'leftmost bar' for 0x6 while the text describes it as the top object; please align the caption with the plot.
  4. [§2.3] The four pinned source locations are a strong point, but the paper does not state whether these pinned locations correspond to the exact Sui version that produced the sampled mainnet data. Given that Sui's execution path has changed over time, please specify the version window and whether the serialisation semantics were stable across the full sample.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the W-only analysis is an empirical measurement whose key premise is supported by pinned engine source code, not by a self-referential derivation.

full rationale

The paper's central move—adopting a write-only conflict graph—is definitional in the innocuous sense that it changes the edge predicate from write∩(read∪write) to write∩write, but every derived claim (hub-and-spoke collapse, 30–40% tighter LSP/χ headroom, DeepBook's volume-proportional contention, and the 10–50% sequential-value fraction) is an empirical consequence computed from mainnet checkpoint data, not an input to the model. The load-bearing premise that Sui serializes only on mutable shared access is supported by pinned, inspectable source-code locations (execution_scheduler_impl.rs:72-82, congestion_tracker.rs:100-104, shared_object_version_manager.rs:464, shared_object_congestion_tracker.rs:223) rather than by a chain of unverified self-citations; although two authors are Mysten employees, the cited artifacts are open source code and externally published protocol papers, and the analysis pipeline is provided with a reproducibility repository. The model's known blind spot—exclusion of intra-checkpoint write-to-read forwarding—is explicitly acknowledged in §2.3 and §8, and while it may weaken the lower-bound claim, it is a correctness/scope concern, not circularity. No fitted parameter is renamed as a prediction, and no claimed result is equivalent by construction to an input. Therefore no significant circularity is present.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities or fitted parameters. Its empirical claims rest on the semantic model of Sui's scheduler, the fidelity of the trace-based write-set extraction, the representativeness of the sample, and standard graph-metric proxies for parallelism.

free parameters (1)
  • checkpoint sampling stride = 3455
    Chosen to sample ~100 checkpoints/day in the post-Mysticeti era. The central 30-40% comparison assumes this sample is representative of routine and high-load regimes; a different stride would change which checkpoints are analyzed.
assumptions (4)
  • domain assumption Sui serializes exclusively on mutable shared-object access; read-only shared accesses are no-ops
    Central premise of the W-only model; supported by cited Sui source code (execution_scheduler_impl.rs:72-82, shared_object_version_manager.rs:464), but not proven for all execution paths.
  • domain assumption Checkpoint object_changes traces allow complete static reconstruction of write sets
    The analysis relies on the SQL join over object_changes to form write sets; missing or misattributed writes would bias the graph.
  • domain assumption A fixed-stride sample of ~100 checkpoints/day is representative, and load regimes are comparable across the two papers
    The 30-40% tightening is a cross-paper comparison on different samples; the 'matched load regimes' matching is not precisely defined.
  • domain assumption LSP/χ and LCC/clique gas-weighted ratios are valid proxies for achievable parallelism
    Inherited from prior work on Ethereum and Sui R+W; heuristic estimates for NP-hard quantities are used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Not All Reads Are Conflicts: A Write-Only Analysis of the Sui Blockchain." pith.science (2026). https://pith.science/paper/WXQLEZ4R

@misc{pith2026260726691,
  author       = {Pith},
  title        = {Pith review of: Not All Reads Are Conflicts: A Write-Only Analysis of the Sui Blockchain},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WXQLEZ4R}},
  note         = {Machine review of arXiv:2607.26691}
}
abstract

Sui's object-centric data model enables parallel transaction execution, but realised performance is fundamentally bounded by workload contention. Prior empirical analyses of Sui have relied on "read+write" (R+W) conflict graphs inherited from account-based blockchains. Because Sui's engine serialises only on mutable shared access, R+W graphs contain spurious edges, bounding contention from above. In this paper, we adopt a complementary "write-set-only" (W-only) model in which every edge represents a real write-serialisation event, providing a lower bound on contention. Together, the two models bracket Sui's true execution-dependency structure. Applying the W-only analysis to Sui mainnet data through 2025 yields three primary findings. First, removing read-only dependencies (notably the system clock) causes previously reported "hub-and-spoke" structures to collapse. The remaining contention topology is highly assortative and clique-dominated, with the W-only bound shaving roughly $30$--$40\%$ off the R+W estimate of Sui's optimal-parallelism headroom. Second, via union-find object grouping, we isolate DeepBook (Sui's native central limit order book). While it dominates contention by volume, its underlying logic does not impose disproportionate sequential bottlenecks. Finally, we quantify the economic cost of contention, showing that $10$--$50\%$ of the network's USD-denominated value flows through sequentially constrained execution paths, exposing it to potential ordering effects.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 1 linked inside Pith

  1. [1]

    https :// github . com / M y s t e n L a b s / sui . git

    Listing 1SQL query used to create the WRITE set SELECT t . c h e c k p o i n t _ s e q u e n c eAScheckpoint , H. Tsoi, A. Sonnino, P. Jovanovic 20:21 Table 2checkpoints Field Type sequence_number Integer digest Hash Digest timestamp Timestamp epoch_id Integer user_tx_count Integer system_tx_count Integer Figure 12Concurrent access graph diameter oc . t r...

    2026

  2. [2]

    3 Postgresql: The world’s most advanced open source relational database

    Accessed: 2026-01-24. 3 Postgresql: The world’s most advanced open source relational database. https://www. postgresql.org/,

    2026

  3. [3]

    4 Mohammad Javad Amiri, Divyakant Agrawal, and Amr El Abbadi

    Accessed: 2026-01-24. 4 Mohammad Javad Amiri, Divyakant Agrawal, and Amr El Abbadi. Parblockchain: Leveraging transaction parallelism in permissioned blockchain systems. In2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 1337–1347. IEEE,

    2026

  4. [7]

    Defi and nfts hinder blockchain scalability

    AFT 2026 20:20 Not All Reads Are Conflicts: A Write-Only Analysis of the Sui Blockchain 15 Lioba Heimbach, Quentin Kniep, Yann Vonlanthen, and Roger Wattenhofer. Defi and nfts hinder blockchain scalability. InInternational conference on financial cryptography and data security, pages 291–309. Springer,

    2026

  5. [9]

    18 Mysten Labs

    Pinned to commit 35abd5c201; accessed 2026-05-19. 18 Mysten Labs. Sui source: execution-scheduler match on SharedObjectMutability (execution_scheduler_impl.rs, lines 72–82). https://github.com/MystenLabs/sui/ blob/35abd5c201430ece1826e6fcf96ede6b4a756eeb/crates/sui-core/src/execution_ scheduler/execution_scheduler_impl.rs#L72-L82,

    2026

  6. [10]

    19 Mysten Labs

    Pinned to commit 35abd5c201; accessed 2026-05-19. 19 Mysten Labs. Sui source: per-object congestion cost gate (shared_object_congestion_tracker.rs, line 223). https://github.com/MystenLabs/sui/ blob/35abd5c201430ece1826e6fcf96ede6b4a756eeb/crates/sui-core/src/authority/ shared_object_congestion_tracker.rs#L223,

    2026

  7. [11]

    20 Mysten Labs

    Pinned to commit 35abd5c201; accessed 2026-05-19. 20 Mysten Labs. Sui source: shared-object version manager exclusivity filter (shared_object_version_manager.rs, line 464). https://github.com/MystenLabs/sui/ blob/35abd5c201430ece1826e6fcf96ede6b4a756eeb/crates/sui-core/src/authority/ shared_object_version_manager.rs#L464,

    2026

  8. [12]

    21 Mysten Labs

    Pinned to commit35abd5c201; accessed 2026-05-19. 21 Mysten Labs. Sui Source: System-Transaction Kinds and Their Mut- able Shared-Singleton Inputs. https://github.com/MystenLabs/sui/blob/ 35abd5c201430ece1826e6fcf96ede6b4a756eeb/crates/sui-types/src/transaction.rs# L1955-L1985,

    2026

Show all 17 references
  1. [13]

    22 Mysten Labs

    Pinned to commit35abd5c201; crates/sui-types/src/transaction.rs, lines 1955–1985; accessed 2026-05-19. 22 Mysten Labs. Sui source: SharedObjectMutability enum (sui-types/src/transaction.rs, line 4629). https://github.com/MystenLabs/sui/ blob/35abd5c201430ece1826e6fcf96ede6b4a7...

    1955

  2. [14]

    23 Satoshi Nakamoto

    Pinned to commit35abd5c201; accessed 2026-05-19. 23 Satoshi Nakamoto. Bitcoin: A peer-to-peer electronic cash system.https://bitcoin.org/ bitcoin.pdf,

    2026

  3. [16]

    An empirical study of speculative concurrency in ethereum smart contracts.arXiv preprint arXiv:1901.01376,

    25 Vikram Saraph and Maurice Herlihy. An empirical study of speculative concurrency in ethereum smart contracts.arXiv preprint arXiv:1901.01376,

    1901 arXiv

  4. [2008]

    24 Patrick Ocheja, Mikel Cortes-Goicoechea, Tarun Mohandas-Daryanani, Brendan Flanagan, Hiroaki Ogata, Jose Luis Munoz, and Leonardo Bautista-Gomez

    Accessed: 2026-04-17. 24 Patrick Ocheja, Mikel Cortes-Goicoechea, Tarun Mohandas-Daryanani, Brendan Flanagan, Hiroaki Ogata, Jose Luis Munoz, and Leonardo Bautista-Gomez. An analytical study of large blocks on ethereum. InProceedings of the 2024 6th Blockchain and Internet of ...

    2026

  5. [2019]

    Mysticeti: Reaching the limits of latency with uncertified dags.arXiv preprint arXiv:2310.14821,

    5 Kushal Babel, Andrey Chursin, George Danezis, Anastasios Kichidis, Lefteris Kokoris-Kogias, Arun Koshy, Alberto Sonnino, and Mingwei Tian. Mysticeti: Reaching the limits of latency with uncertified dags.arXiv preprint arXiv:2310.14821,

  6. [2023]

    11 Thomas Dickerson, Paul Gazzillo, Maurice Herlihy, and Eric Koskinen

    Accessed: 2026-01-24. 11 Thomas Dickerson, Paul Gazzillo, Maurice Herlihy, and Eric Koskinen. Adding concurrency to smart contracts. InProceedings of the ACM Symposium on Principles of Distributed Computing, pages 303–312,

    2026

  7. [2024]

    2 Networkx: Network analysis in python.https://networkx.org/,

    Accessed: 2026-01-24. 2 Networkx: Network analysis in python.https://networkx.org/,

    2026

  8. [2025]

    Sui lutris: A blockchain combining broadcast and consensus

    8 Sam Blackshear, Andrey Chursin, George Danezis, Anastasios Kichidis, Lefteris Kokoris- Kogias, Xun Li, Mark Logan, Ashok Menon, Todd Nowacki, Alberto Sonnino, et al. Sui lutris: A blockchain combining broadcast and consensus. InProceedings of the 2024 on ACM SIGSAC Conferenc...

    2024

  9. [2026]

    17 Mysten Labs

    Accessed: 2026-05-22. 17 Mysten Labs. Sui source: checkpoint-level congestion filter dropping InputConsensusObject::ReadOnly (congestion_tracker.rs, lines 100–104). https: //github.com/MystenLabs/sui/blob/35abd5c201430ece1826e6fcf96ede6b4a756eeb/ crates/sui-core/src/congestion...

    2026

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.