Admission

Admission control regulator that prevents scheduler overload by measuring queuing delays.

The Admission regulator protects the system from overload by monitoring scheduler queuing delays and selectively rejecting tasks when delays increase. It maintains an admission percentage that adjusts dynamically based on measured delays.

==Queuing Delay Measurement==

The regulator probes for queuing delays by periodically submitting special timing tasks into the scheduler and measuring how long they wait before execution. A probe task simply measures the time between its creation and execution. High variance or increasing delays in these measurements indicate scheduler congestion, triggering reductions in the admission rate to alleviate pressure.

==Rejection Mechanism==

Tasks are rejected using a deterministic hashing mechanism that provides stable and consistent admission decisions within time windows. Each task key (string or integer) is hashed using a large prime number multiplication and the current time window to generate a value between 0-99. Tasks with hash values above the current admission percentage are rejected.

The rotation mechanism is particularly important for fairness when task keys represent user identifiers:

• Without rotation, users whose IDs hash to high values would face persistent rejection during system pressure
• This could lead to poor user experience where some users are consistently locked out while others maintain access
• Rotation ensures rejection patterns shift periodically, giving previously rejected users opportunities for admission
• The rotation window duration can be tuned to balance stability and fairness

For example, with a 60-minute rotation window:

• User A might be rejected in the first window due to their ID's hash value
• In the next window, the time-based component changes the hash calculation
• User A now has a different effective priority and may be admitted while others are rejected
• This prevents any user from experiencing extended denial of service

This approach ensures:

• Consistent decisions within each rotation window
• Even distribution of rejections across the key space
• Periodic rotation to prevent unfair persistent rejection
• Load balancing through prime number distribution
• Fair access patterns over time for all users

==Load Shedding Pattern==

The system responds to increasing pressure through a gradual and predictable load shedding pattern:

• As pressure increases and admission percentage drops, tasks with highest hash values are rejected
• Additional tasks are rejected if pressure continues building
• System stabilizes at lower load with a stable subset of traffic
• During recovery, admission percentage gradually increases
• Previously rejected tasks may be admitted in new time windows

==Backpressure Characteristics==

This design creates an effective backpressure mechanism with several key characteristics. Load reduces predictably as the admission percentage drops, avoiding the oscillation patterns common with random rejection strategies. The system maintains a stable subset of flowing traffic within each time window, while providing natural queue-like behavior for rejected requests.

==Distributed Systems Context==

The admission control mechanism is particularly effective in microservices architectures. Consistent rejection patterns help downstream services manage their own load effectively, while enabling client libraries to implement intelligent backoff strategies. The predictable degradation and recovery patterns make it easier to maintain system stability across service boundaries.