>>> [!note] Migrated issue <!-- Drupal.org comment --> <!-- Migrated from issue #3615674. --> Reported by: [mably](https://www.drupal.org/user/3375160) >>> <p><code>Together::spread()</code> hands a whole scope (one row with gaps, then an area, then a section) to <code>SeatingShareOut::shareOut()</code>. Where the requested tariffs price different grades that goes through <code>shareOutByMatching()</code>, which walks the places in house order and takes the earliest eligible one for each tariff. It answers the question it was written for, which is feasibility, and it has no notion of how far apart the places it picked end up.</p> <p>So a party across several tariffs is seated in whichever part of the scope each grade happens to appear first, and the rung reports its closeness from the scope rather than from the result.</p> <h3>Reproduced</h3> <p>A party of 4 + 4 + 4 across three tariffs pricing three grades, in a section whose row A carries grade 5 at positions 1 to 6 and grade 6 at positions 7 to 13, was seated on positions 1, 2, 3, 4 and 7, 8, 9, 10, leaving positions 5 and 6 empty between the two groups. Taking positions 3, 4, 5, 6 for the first tariff gives one unbroken run of eight for the same two tariffs. Both answers report the same closeness.</p> <h3>Proposed fix</h3> <p>Candidates already arrive ordered by section, row and position, so a contiguous window of that list is a compact set of seats and the search space is windows rather than subsets.</p> <ul> <li>Sweep two pointers for the minimum window covering the per-tariff grade demand, keeping a running count per grade so a window is accepted or rejected in a few integer comparisons.</li> <li>Rank the surviving windows by extent: rows spanned first, then places spanned, then house order to break ties.</li> <li>Run the existing matcher inside the winner, on a party's worth of places instead of a hall's worth.</li> <li>Keep the whole-scope call as the fallback. A tight window can be grade-infeasible where a scattered pick works, and feasibility must never regress: refusing an arrangement that exists is the defect the matcher was written to avoid.</li> </ul> <p><code>adjacent()</code> is left alone: a true contiguous run is already the compact answer and is tried first. <code>nearTheParty()</code> is left alone too, because it already passes a distance-ordered list and the matcher honours input order. This is specifically the last-resort rung, which is the one that was not compactness-aware.</p> <h3>Without degrading performance</h3> <p>Today's path is O(party), not O(hall): both share-out paths stop as soon as every slot is filled, so when the first eligible places work they touch a party's worth of candidates and return. A window sweep is O(hall), so replacing the current call outright would be a regression on the common case even though the constant is small.</p> <ul> <li>Compute the cheap answer first, exactly as today.</li> <li>Measure its extent over the places it returned, which is O(party).</li> <li>Only sweep when that answer is not already tight. An answer sitting in one row cannot be beaten at this rung, since <code>adjacent()</code> has already ruled out a contiguous run, so it is returned without reading the scope again.</li> <li>Short-circuit the sweep on the first window whose extent is a single row.</li> </ul> <p>That leaves the common path with one extent measurement added, and spends a linear pass only on the case that is currently wrong. No new query either way: this all runs on the candidate array the rung already holds.</p> <p>To be settled by the propose benchmark (A/B in one process, no database) over a full hall across a single tariff at quantity 1 and 4, a mixed 4 + 4 + 4, and a nearly full hall where this rung is actually reached, with the statement count asserted as a maximum rather than an equality since it has to hold on both SQLite and MySQL. If the swept path measures worse than today, the design changes rather than shipping with a caveat.</p> <h3>Note on the promise</h3> <p>Where a venue's grades are separate zones, a party across three of them cannot sit together and the promise can only be "as close as the price bands allow". This is worth doing for a hall that interleaves grades within a row, which the bundled example does.</p>