§05 Gaming Tech

Matchmaking Latency Fairness: The Hard Constraints of Distance and Time

In multiplayer games, matchmaking is an optimization problem. Studios can trade longer queue times for fairer skill-based games, yet latency is harder to control due to the physics of distance and server placement.

Developers aim to find matches as quickly as possible, but queue wait times often range from 30 seconds to five or ten minutes - even more in ranked or high-skill games. This is because extending queue times allows for broader player skill bands to be matched together.

While widening the acceptable skill levels can improve match quality, the same trade-off is not possible with latency. Geographical distance imposes real and unchanging constraints on player ping.

For example, the latency policy on Amazon Gamelift Servers will stop a game session if any player exceeds a specified maximum latency. The matchmaking process will search incrementally wider for acceptable locations, but the base costs of geography and server placement cannot be eliminated.

The AWS system gives an example of matchmaking with a 320-second search policy:

120 seconds to find players with all latencies under 50 milliseconds
If that fails, 120 seconds under 100 milliseconds
Finally, 80 seconds under 200 milliseconds

This staged policy allows matchmaking to try to find optimal ping before dropping to less ideal matches. Amazon GameLift also supports more advanced latency-based use cases with its FlexMatch tool.

Latency as the Harder Constraint

While matchmaking seeks to optimize skills, distances are treated very differently. Official guidance from a published matchmaking design states that latency is a near-hard constraint, and that players should be bucketed by region or data center first, before any other factor. Skill is allowed to flex as part of the optimization process, but geography changes "barely at all."

The same source says that only adjacent data centers with acceptable ping should be used as a last resort. This makes sense, as

This means that matchmaking algorithms use geographic data as a hard limit before skill and queue time, and only cross-regional matches when they must. AccelByte's official guidance suggests a policy of widening the acceptable latency window incrementally, for example by 50 milliseconds every 10 seconds, moving into neighboring regions only when strictly necessary.

Stages of Widening

The exact policies vary, but several sources agree that matchmaking uses a staged approach to widening, and that both latency and skill windows grow as queue time extends.

AccelByte provides the example of a 50 millisecond latency increase every 10 seconds, while others take a three-step approach of two minutes each at 50, 100, and 200 millisecond caps. The idea is to allow for fair matches with optimal latencies, but to accept a longer queue for good reason.

Existing implementations do not allow skills and latency thresholds to widen indefinitely, but frame these thresholds as strategic dials that developers can tune. Backfill, where players are added to an ongoing game, is another variable that affects a player's net losses.

Fairness: Region Matching and Latency Difference

For 1v1 games, fairness can be defined by the absolute ping difference between two players. In multiplayer, it's treated as the average per-team latency difference. These definitions are published by Edgegap, a white-label provider of player hosting and connectivity analytics.

Geographical placement is always the first step in modern matchmaking. Skill flex is secondary, and backfill is the final step, with its own mix of skills and latency.

Matchmaking designers have documented the necessary tradeoffs of latency, skills, and queue time, but have not spelled out the link between their specifics across different game genres/pace/backfill.

Beware Broad Statements on Latency Effects

Asserting that always changes results should be avoided. No evidence was found to back the claim that such is a universal threshold for skilled or lower games. While claim that 40 ms fairness changes the outcome, asking for game-specific telemetry, a netcode paper, or a publisher postmortem is essential.

Worked numerical examples to explain the trade-offs of lowering latency caps and extending queue times is another serious area. Further to this, since this operation can define the game’s UX and economies, formats for review are needed. Finally, the system also needs to be able to discern this versus seller loads.