brycewang-stanford SIGMOD Related Work
Database research has unusually long memory and unusually fast turnover at the same time: reviewers expect a submission to know both the 1980s ancestor of its technique and the PVLDB paper from three months ago. Positioning at SIGMOD is therefore a two-clock problem, and the related-work section is where a PC of system builders first tests whether the authors know the field they are claiming to advance.
The five literature lines to cover
| Line | What it contributes | Cost of missing it |
|---|---|---|
| SIGMOD/PACMMOD + PVLDB | The current research frontier, updated quarterly/monthly | "Unaware of concurrent work" reviews |
| ICDE and sibling conferences | Adjacent engineering results | Novelty delta overstated |
| TODS and journal literature | Consolidated, definitive treatments | Comparing against superseded versions |
| PODS / theory venues | Lower bounds, complexity, formal semantics | Claiming possibility where impossibility is known |
| Industrial systems + docs | What deployed engines already do | "Postgres has done this since 9.5" comments |
The last line is venue-distinctive: at SIGMOD, a feature shipped in an open-source or commercial engine counts as prior art in reviewers' minds even when it was never formally published. Cite documentation, source code, or system papers for deployed behavior, and position against it explicitly.
Deep-history diligence
Many "new" ideas in data management are reinventions — adaptive indexing, fractional cascading, semantic caching, and differential update propagation have each been rediscovered more than once. Before claiming firstness:
- Walk the technique's lineage backwards at least to its earliest SIGMOD/ VLDB/TODS appearance; the classics are findable via the ACM DL and dblp.
- Search under the field's older vocabulary (e.g., "materialized view maintenance," not only "incremental computation").
- If the ancestor exists, the honest and stronger claim is the delta: changed hardware assumptions, changed workload regime, new guarantee.
Concurrency in a rolling world
Your round-based submission competes with PVLDB's monthly acceptances and other rounds' PACMMOD issues appearing during your own review. Practical protocol: sweep new PACMMOD/PVLDB issues at submission, again before the feedback phase, and again when preparing a revision; cite genuinely concurrent work as concurrent, with a factual one-line distinction. A revision letter that quietly ignores a widely-seen concurrent paper invites the reviewer to raise it for you, on their terms.
Exclusivity interacts with positioning
PACMMOD's rule bars submitting work with substantial overlap in scientific content elsewhere while under review. This binds related-work strategy for your own pipeline: a companion paper from the same project must be different enough in results — not just in prose — and when it is, cite the anonymized companion and state the division of contributions plainly. Reviewers who smell salami-slicing check overlap aggressively.
Contrast-form writing
The section should read as a set of distinctions, not an annotated bibliography:
Weak: "Smith et al. [12] proposed an LSM compaction scheduler.
Jones et al. [17] studied write amplification."
Strong: "Prior schedulers [12, 17] choose compaction targets from
size ratios alone; neither observes query access patterns,
so both degrade under the skewed reads we target (Sec 2).
Our scheduler is the first to condition on both, and Sec 6.3
compares against tuned versions of each."
Every cited cluster gets: what they established, the precise dimension where this paper departs, and — for the nearest neighbors — a pointer to the experiment where the departure is measured. A nearest neighbor cited but not benchmarked is the most predictable feedback-phase question at this venue.
A repeatable search protocol
- Enumerate the paper's techniques and, for each, its older names — ask a senior colleague or trace citation chains backwards for vocabulary.
- Sweep the ACM DL for SIGMOD/PACMMOD and TODS hits, dblp for PVLDB and ICDE, and the PODS series for impossibility results touching the claim.
- Check the documentation and changelogs of the three engines nearest your contribution for shipped equivalents.
- Log every sweep with its date; the concurrency protocol above reuses this log at feedback and revision time.
- File each hit as: ancestor / nearest neighbor / concurrent / deployed prior art — the four categories that demand different handling in text.
Placement tactics
SIGMOD tolerates related work early (Section 2) or late (before the conclusion). Choose early when the contribution is only legible against prior designs; choose late when the running example teaches faster than contrast does. Either way, the introduction still owes one paragraph of sharp positioning — reviewers form the novelty judgment on page one.
Surveys and tutorials as leverage
Data management maintains strong survey infrastructure — ACM Computing Surveys pieces, SIGMOD/VLDB tutorials, and the specialized book series many subfields keep current. Citing the definitive survey does double duty: it compresses background (protecting the 12-page budget) and signals that the firstness claim was made knowing the mapped terrain. When a tutorial by a likely PC member exists in your area, assume your reviewer wrote or read it, and position accordingly.
Output format
[Line coverage] five-line audit with gaps named
[Lineage check] earliest ancestor found; firstness claim adjusted
[Concurrent sweep] last swept date; papers requiring mention
[Nearest neighbors] work -> distinction -> comparing experiment
[Self-overlap] companion submissions and division-of-contribution note
[Rewrite targets] bibliography-style passages to convert to contrast form