Skip to main content
Version: 2.32.2

Flow Lifecycle

The lifecycle of a flow begins when a service initiates it, requesting a decision from the Aperture Agent. As the flow enters the Aperture Agent, it embarks on a journey through multiple stages before a final decision is made. The following diagram illustrates these stages and the sequence they follow.

While some stages primarily serve to augment the flow with additional labels or provide telemetry, others actively participate in the decision to accept or reject the flow. Should a flow be rejected at any such active stage, its journey through the subsequent stages is halted, and an immediate decision is dispatched back to the service.

flowchart LR classDef Green fill:#56AE89,stroke:#000000,stroke-width:2px; classDef Orange fill:#F8773D,stroke:#000000,stroke-width:2px; subgraph "Service" end class Service Green Service <-- "request\n decision" --> Selectors Selectors <-- "request\n decision" --> Components subgraph "Selectors" Text("Match components") class Text Orange end class Selectors Green subgraph Components Classifiers --> FluxMeters("Flux Meters") --> Samplers Samplers --"accept"--> RateLimiters("Rate Limiters") Samplers --"accept"--> ConcurrencyLimiters("Concurrency Limiters") RateLimiters("Rate Limiters") --"accept"--> ConcurrencySchedulers ConcurrencyLimiters("Concurrency Limiters") --"accept"--> ConcurrencySchedulers ConcurrencySchedulers --"accept"--> Caches Caches --> LoadSchedulers("Load Schedulers") Caches --> QuotaSchedulers("Quota Schedulers") class Classifiers Orange class FluxMeters Orange class Samplers Orange class RateLimiters Orange class ConcurrencyLimiters Orange class Caches Orange class ConcurrencySchedulers Orange class LoadSchedulers Orange class QuotaSchedulers Orange end class Components Green

Outlined below is the sequence of stages in detail, along with their roles:

note

Remember, a flow can bypass certain or all stages if there are no matching components for that stage.

Selection, Classification, and Telemetry

  • Selectors are the criteria used to determine the components that will be applied to a flow in the subsequent stages.
  • Classifiers perform the task of assigning labels to the flow based on the HTTP attributes of the request. However, classifiers are only pertinent for HTTP or gRPC Control Points and do not apply to flows associated with feature Control Points.
  • FluxMeters are employed to meter the flows, generating metrics such as latency, request counts, or other arbitrary telemetry based on access logs. They transform request flux that matches certain criteria into Prometheus histograms, enabling enhanced observability and control.

Rate limiting (fast rejection)

  • Samplers manage load by permitting a portion of flows to be accepted, while immediately dropping the remainder with a forbidden status code. They are particularly useful in scenarios such as feature rollouts.
  • Rate-Limiters proactively guard against abuse by regulating excessive requests in accordance with per-label limits.
  • Concurrency-Limiters enforce in-flight request quotas to prevent overloads. They can also be used to enforce limits per entity such as a user to ensure fair access across users.

Request Prioritization and Cache Lookup

Schedulers offer on-demand queuing based on a limit enforced through a token bucket or a concurrency counter, and prioritize requests using weighted fair queuing. Multiple matching schedulers can evaluate concurrently, with each having the power to drop a flow. There are three variants running at various stages of the flow lifecycle:

  • The Concurrency Scheduler uses a global concurrency counter as a ledger, managing the concurrency across all Agents. It proves especially effective in environments with strict global concurrency limits, as it allows for strategic prioritization of requests when reaching concurrency limits.
  • Caches Look of response and global caches occur at this stage. If a response cache hit occurs, the flow is not sent to the Concurrency and Load Scheduling stages, resulting in an early acceptance.
  • The Quota Scheduler uses a global token bucket as a ledger, managing the token distribution across all Agents. It proves especially effective in environments with strict global rate limits, as it allows for strategic prioritization of requests when reaching quota limits.
  • The Load Scheduler oversees the current token rate in relation to the past token rate, adjusting as required based on health signals from a service. This scheduler type facilitates active service protection.

After traversing these stages, the flow's decision is sent back to the initiating service.