Mlsys Workflow

MLSys Workflow

Use this as the project-management skill for a Conference on Machine Learning and Systems submission. MLSys runs one cycle per year, so a missed gate costs twelve months or a reroute to another venue. All dates below are the verified 2026-cycle timetable (access 2026-07-08) — treat them as shape, and replace every one with the current mlsys.org dates page before committing a plan.

MLSys is a conference with rotating per-edition leadership (2026: General Chair Luis Ceze; Program Chairs Zhihao Jia and Aakanksha Chowdhery) and no APC — proceedings publish open on proceedings.mlsys.org, and the money side is registration, where accepted-paper authors get a two-week reserved-ticket window.

The 2026-cycle shape

Backward plan from the paper deadline

The distinctive constraint versus ML-venue planning: the evaluation competes for shared hardware. Reserve cluster time for weeks 6-4 when the matrix runs, and again for the January response window, where two rebuttal experiments may need machines on two days' notice.

Standing team rituals

Weekly (weeks 12->2):
  - experiment-matrix review: rows launched / landed / failed, GPU budget burned
  - claims board: abstract claim -> supporting figure -> status (red/yellow/green)
  - baseline watch: new releases of compared systems since last week?
Once (week 3): mock review — one systems-culture and one ML-culture reader,
  scored on workload realism, baseline fairness, attribution, and tails.
Response-week protocol (pre-agreed in December):
  - who reads reviews hour 0, who owns machines, who drafts, who signs off

Failure modes by stage

• Week 6 matrix slip: the sweep launches late, so ablations get cut — and ablation absence is precisely what MLSys reviews punish. Cut breadth (hardware variants) before cutting attribution (ablations).
• Baseline rot: a compared system ships a major release in September; deciding whether to re-run everything at week 4 needs a pre-agreed rule ("re-run headline comparisons if the release notes claim >10% on our metric").
• Response-window surprise: the four-day window found nobody available in January. It is on the calendar from the day of submission.
• Post-acceptance pile-up: camera-ready, AE package, and conference logistics all land in February-March; assign three different owners at notification time.
• Reroute paralysis after rejection: the systems-venue calendar (OSDI/SOSP/NSDI/ ASPLOS/ATC deadlines) should be listed in the postmortem doc before the decision arrives.

The reroute decision, prepared in advance

Because the cycle is annual, "revise and resubmit to next MLSys" costs a year. Write the decision rule before the notification arrives:

Maintain the alternative-deadline list (OSDI, SOSP, NSDI, ASPLOS, ATC, EuroSys, SC) with dates in the project doc from week 12, so a January rejection flows into a concrete February plan instead of a stall.

Ownership map (minimum viable)

• Evaluation owner: matrix, logs, figure regeneration.
• Baseline owner: versions, tuning parity, release watch.
• Writing owner: 10-page budget, style-kit compliance, claims board.
• Compliance owner: anonymity, appendix upload, OpenReview fields, deadlines.
• (Post-acceptance) AE owner: package, badges, evaluator on-call.

One person may hold two hats; no hat may be unassigned. Industrial-track submissions add a sixth hat that academic teams forget: the internal-approval owner, who clears company names, trace releases, and benchmark numbers with legal/PR — a process whose latency routinely exceeds the writing itself, so it starts at week 10, not week 1.

Cycle-volatility warning

Every date above, the track structure, the response-window length, and the AE mechanics are 2026 facts. The 2027 CFP was not yet visible at access time (待核实, all of it) — rebuild this table from mlsys.org before planning a real cycle.

Output format

[Current stage] fit / building / evaluating / writing / submitted / response / decided / AE
[Next official gate] <date + source URL>
[Critical path] <three tasks, with GPU-time dependencies flagged>
[Claims board] <claim -> figure -> status>
[Ownership] <evaluation/baseline/writing/compliance/AE -> person>
[Risk register] <matrix slip / baseline rot / response availability / logistics pile-up>