Game state

Given five substitutes instead of three, Europe's big clubs mostly used the extra changes to protect leads. Software's incident process, borrowed from wildfire command, says the person in charge is not a resolver either, but the call that mattered on a release that went wrong, P0 now or backlog later, rested on relationships no rulebook hands out.

On 9 March 2022 Real Madrid were two goals down on aggregate to PSG at the Bernabéu when Carlo Ancelotti, at 57 minutes, took off Kroos and Asensio for Camavinga and Rodrygo; Benzema scored at 61, 76 and 78, and neither substitute scored. A month later Chelsea led 3-0 on the night until Rodrygo, on at 78 minutes, scored at 80, and in the semi-final against City he came on at 68 and scored at 90 and 90+1. The other bench that night is the half of the story the genre leaves out: City took De Bruyne off at 72, went ahead through Mahrez a minute later and, with five minutes left, swapped Mahrez for Fernandinho, which is what closing a game out is supposed to look like, and they went out. Ancelotti didn't kick a ball in any of those matches. What he did, as did the managers in the other dugouts, was read the game state, meaning the score and the clock and what they demanded, and decide who should be on the pitch for the next half hour, which is the only lever a manager has once the whistle goes.

Game state is analytics shorthand for the fact that teams play differently depending on the score. StatsBomb's Benjamin Pugsley, working with data from early in the 2013-14 season, found that teams a goal up tend to form a "defensive shell" and drop deeper, and get outshot 46.6% to 53.4%, and Opta now splits Premier League xG by game state. Chasing and closing out are different jobs with different players, and the substitute is how a manager switches between them.

The Premier League ran the cleanest test of what managers do with more of that lever, mostly by accident. It allowed five substitutes for the 2019-20 restart, went back to three for 2020-21 and 2021-22 — the only top European league to drop five — and voted on 31 March 2022 to bring five back from 2022-23, over objections from smaller clubs that it favoured deep squads (the PFA backed it on player workload). Subs per match went from 5.9 in 2019-20 to 7.9 in 2022-23, and substitutes' share of goals went from 12.7% that season to 14.3% partway through 2023-24, a record pace.

To see what the extra changes were spent on, ESPN's Dale Johnson tracked ten big clubs across LaLiga, Serie A, the Bundesliga and Ligue 1 in 2021-22, the season the Premier League sat out, and found they mostly used all five less often than everyone else, and usually made the fifth change when already winning: PSG used five subs eight times in 38 league games, seven of them while ahead. Bayern's points barely moved: 78 in the last three-sub season, then 78 and 77 in the two seasons with five. Atlético were the exception, with 14 of their 24 five-sub games coming while drawing or losing and the biggest net gain of any of the ten. The fear was that the rich would use deeper benches to chase games down; what the clubs with the deepest benches actually did with the extra changes, mostly, was protect what they already had. It's journalism and one season, not a peer-reviewed study, but it fits the shell.

Software's live game is the incident, and its playbook came from somewhere less glamorous than a dugout. Google's SRE book says its incident management system "is based on the Incident Command System", and footnotes FEMA's national incident framework rather than any fireground; ICS itself came out of FIRESCOPE, a Forest Service programme paid for by a special congressional appropriation after a 13-day Southern California fire siege in 1970 killed 16 people and destroyed more than 700 homes, on fires where the agencies sent to help set up rival command posts and couldn't agree on what to call things. The SRE chapter splits the work the same way a match does. The incident commander "holds the high-level state about the incident," the operations team "should be the only group modifying the system," and one of the failure modes it names is freelancing — an engineer making well-meant changes without coordinating, which made "a bad situation far worse." PagerDuty's public response guide is blunter: the commander should "delegate all repair actions, the Incident Commander is NOT a resolver." That's the touchline written down as a role definition.

From my own side of this, on the last large release I was part of that went wrong, the people who led the process and the resolution were the tech lead, the architect and the engineering manager. What mattered was that those three had relationships across the company that let them make the tough calls, and the one I'd single out is deciding whether a given problem gets fixed right now as a P0 or goes onto the backlog. That's a triage call, and it lands on other people's time and other teams' plans; it holds only if the people receiving it accept it. The SRE book has a line for this that's easy to read past: the commander, "if appropriate, can remove roadblocks that prevent Ops from working most effectively." On a large release that roadblock is often another team's priorities, and a role title alone doesn't move those.

That's the closest fit I can find to the substitution. Ancelotti's call at 57 minutes was a decision about who would carry the next phase of the game, made by someone who wasn't going to carry it himself, and the P0-or-backlog call is the same shape: which problem gets the people now, and which waits.

The fit stops at authority. A manager's right to make that change is written into the Laws of the Game, and nobody on the pitch gets a vote on it, whereas nothing equivalent guarantees an incident lead's call once it crosses into another team's plan. On that release, what let those three make the calls was the relationships they already had, more than anyone's ability to fix the code. The game state is different too. A manager always knows the score and the minute, and the whole skill is in reading what they imply; in an incident the score is often the thing in dispute — how bad it is, and whether it's even one problem — and there's no final whistle, which is why the SRE book spends a paragraph on handing command over cleanly at the end of a working day, something no manager has ever had to do at the hour mark. And substitutions are rationed by rule, while the number of P0s a lead can declare is limited only by how many the rest of the company will put up with.

If that last limit is real, it's a budget, and it raises the five-sub question in engineering form. Given more room to make calls, the big clubs mostly spent it protecting a lead. When an incident lead has the standing to push something onto another team's plate, how much of that standing goes on chasing the actual problem, and how much on closing out the release date?

No comments yet