Vldb writing style

VLDB Writing Style

Use this during revision passes. The PVLDB reader is a systems builder deciding, inside the first page, whether your design decision is worth their evening. Style at this venue is not ornament — it is claim hygiene.

The page-one contract

By the end of page one the reader must know:

1. The workload that breaks current systems — concrete, not "growing data volumes."
2. The design decision — the bet your system makes that others do not.
3. What was built — real system, extension of a named engine, or simulation; readers calibrate trust to this immediately.
4. The headline measurement, scoped — metric, scale, baseline, hardware class, all in one sentence.

A running example introduced early and reused through design and evaluation sections is the genre's most effective device; pick one query, one tenant, one failure trace and let it carry the mechanism.

Claim scoping table

Unscoped (review bait)	Scoped (survives)
"significantly faster than X"	"2.1x median speedup over X (v3.4, tuned per §6.1) on workload W at 1TB"
"scales to large clusters"	"near-linear to 64 nodes; efficiency drops to 71% at 128 (Fig. 9)"
"negligible overhead"	"adds 3-5% CPU on write-heavy YCSB-A; worst case 11% at 99% writes"
"state-of-the-art performance"	delete; name the systems and the regime

Every superlative either gains a number, a workload, and a baseline — or dies.

Trade-off candor

Systems reviewers assume every design pays somewhere. A paper that never says what its mechanism costs reads as unmeasured, not as flawless.

• State the price of the design in the introduction (memory, write path latency, generality) and point to the section that measures it.
• Put the loss cases in the main evaluation, not a footnote: the workload where the baseline wins is your credibility purchase.
• Prefer "we chose A over B because C, accepting cost D" over passive descriptions that hide the decision.

Budget and terminology discipline

• One name per concept for all 12 pages: a "segment" on page 3 cannot become a "chunk" on page 9. Builder-readers parse terminology as architecture.
• Architecture figure by page 3; reviewers reconstruct systems visually.
• Since references are uncapped but appendices are not, spend prose on mechanism and measurement, and let citations carry the history compactly.
• Single-blind means writing "we extend our earlier system [12]" is fine — clarity about lineage beats coy third-person contortions.

Self-edit sweep

grep-worthy passes before submission:
  "significant"      -> replace with the number or delete
  "novel"            -> the contribution list should prove it instead
  "to the best of our knowledge" -> keep at most one, verified
  "very large"       -> state the size
  passive "was designed" -> name the decision and its reason
  every Figure N     -> referenced in text, at least one sentence of reading

Output format

[Page-one contract] met / missing elements listed
[Unscoped claims] <count, worst three quoted>
[Trade-off candor] price stated in intro? loss cases in eval?
[Terminology drift] <concept -> competing names found>
[Running example] present / absent (suggested candidate)
[Edit order] <highest-leverage rewrites first>