API Priority and Fairness

Controlling the behavior of the Kubernetes API server in an overload situation is a key task for cluster administrators. The kube-apiserver has some controls available (i.e. the --max-requests-inflight and --max-mutating-requests-inflight command-line flags) to limit the amount of outstanding work that will be accepted, preventing a flood of inbound requests from overloading and potentially crashing the API server, but these flags are not enough to ensure that the most important requests get through in a period of high traffic.

The API Priority and Fairness feature (APF) is an alternative that improves upon aforementioned max-inflight limitations. APF classifies and isolates requests in a more fine-grained way. It also introduces a limited amount of queuing, so that no requests are rejected in cases of very brief bursts. Requests are dispatched from queues using a fair queuing technique so that, for example, a poorly-behaved controller need not starve others (even at the same priority level).

This feature is designed to work well with standard controllers, which use informers and react to failures of API requests with exponential back-off, and other clients that also work this way.

Enabling/Disabling API Priority and Fairness

The API Priority and Fairness feature is controlled by a feature gate and is enabled by default. See Feature Gates for a general explanation of feature gates and how to enable and disable them. The name of the feature gate for APF is "APIPriorityAndFairness". This feature also involves an API Group with: (a) a v1alpha1 version, disabled by default, and (b) v1beta1 and v1beta2 versions, enabled by default. You can disable the feature gate and API group beta versions by adding the following command-line flags to your kube-apiserver invocation:

kube-apiserver \
--feature-gates=APIPriorityAndFairness=false \
--runtime-config=flowcontrol.apiserver.k8s.io/v1beta1=false,flowcontrol.apiserver.k8s.io/v1beta2=false \
 # …and other flags as usual

Alternatively, you can enable the v1alpha1 version of the API group with --runtime-config=flowcontrol.apiserver.k8s.io/v1alpha1=true.

The command-line flag --enable-priority-and-fairness=false will disable the API Priority and Fairness feature, even if other flags have enabled it.

Concepts

There are several distinct features involved in the API Priority and Fairness feature. Incoming requests are classified by attributes of the request using FlowSchemas, and assigned to priority levels. Priority levels add a degree of isolation by maintaining separate concurrency limits, so that requests assigned to different priority levels cannot starve each other. Within a priority level, a fair-queuing algorithm prevents requests from different flows from starving each other, and allows for requests to be queued to prevent bursty traffic from causing failed requests when the average load is acceptably low.

Priority Levels

Without APF enabled, overall concurrency in the API server is limited by the kube-apiserver flags --max-requests-inflight and --max-mutating-requests-inflight. With APF enabled, the concurrency limits defined by these flags are summed and then the sum is divided up among a configurable set of priority levels. Each incoming request is assigned to a single priority level, and each priority level will only dispatch as many concurrent requests as its configuration allows.

The default configuration, for example, includes separate priority levels for leader-election requests, requests from built-in controllers, and requests from Pods. This means that an ill-behaved Pod that floods the API server with requests cannot prevent leader election or actions by the built-in controllers from succeeding.

Seats Occupied by a Request

The above description of concurrency management is the baseline story. In it, requests have different durations but are counted equally at any given moment when comparing against a priority level's concurrency limit. In the baseline story, each request occupies one unit of concurrency. The word "seat" is used to mean one unit of concurrency, inspired by the way each passenger on a train or aircraft takes up one of the fixed supply of seats.

But some requests take up more than one seat. Some of these are list requests that the server estimates will return a large number of objects. These have been found to put an exceptionally heavy burden on the server, among requests that take a similar amount of time to run. For this reason, the server estimates the number of objects that will be returned and considers the request to take a number of seats that is proportional to that estimated number.

Execution time tweaks for watch requests

API Priority and Fairness manages watch requests, but this involves a couple more excursions from the baseline behavior. The first concerns how long a watch request is considered to occupy its seat. Depending on request parameters, the response to a watch request may or may not begin with create notifications for all the relevant pre-existing objects. API Priority and Fairness considers a watch request to be done with its seat once that initial burst of notifications, if any, is over.

The normal notifications are sent in a concurrent burst to all relevant watch response streams whenever the server is notified of an object create/update/delete. To account for this work, API Priority and Fairness considers every write request to spend some additional time occupying seats after the actual writing is done. The server estimates the number of notifications to be sent and adjusts the write request's number of seats and seat occupancy time to include this extra work.

Queuing

Even within a priority level there may be a large number of distinct sources of traffic. In an overload situation, it is valuable to prevent one stream of requests from starving others (in particular, in the relatively common case of a single buggy client flooding the kube-apiserver with requests, that buggy client would ideally not have much measurable impact on other clients at all). This is handled by use of a fair-queuing algorithm to process requests that are assigned the same priority level. Each request is assigned to a flow, identified by the name of the matching FlowSchema plus a flow distinguisher — which is either the requesting user, the target resource's namespace, or nothing — and the system attempts to give approximately equal weight to requests in different flows of the same priority level. To enable distinct handling of distinct instances, controllers that have many instances should authenticate with distinct usernames

After classifying a request into a flow, the API Priority and Fairness feature then may assign the request to a queue. This assignment uses a technique known as shuffle sharding, which makes relatively efficient use of queues to insulate low-intensity flows from high-intensity flows.

The details of the queuing algorithm are tunable for each priority level, and allow administrators to trade off memory use, fairness (the property that independent flows will all make progress when total traffic exceeds capacity), tolerance for bursty traffic, and the added latency induced by queuing.

Exempt requests

Some requests are considered sufficiently important that they are not subject to any of the limitations imposed by this feature. These exemptions prevent an improperly-configured flow control configuration from totally disabling an API server.

Resources

The flow control API involves two kinds of resources. PriorityLevelConfigurations define the available priority levels, the share of the available concurrency budget that each can handle, and allow for fine-tuning queuing behavior. FlowSchemas are used to classify individual inbound requests, matching each to a single PriorityLevelConfiguration. There is also a v1alpha1 version of the same API group, and it has the same Kinds with the same syntax and semantics.

PriorityLevelConfiguration

A PriorityLevelConfiguration represents a single priority level. Each PriorityLevelConfiguration has an independent limit on the number of outstanding requests, and limitations on the number of queued requests.

Concurrency limits for PriorityLevelConfigurations are not specified in absolute number of requests, but rather in "concurrency shares." The total concurrency limit for the API Server is distributed among the existing PriorityLevelConfigurations in proportion with these shares. This allows a cluster administrator to scale up or down the total amount of traffic to a server by restarting kube-apiserver with a different value for --max-requests-inflight (or --max-mutating-requests-inflight), and all PriorityLevelConfigurations will see their maximum allowed concurrency go up (or down) by the same fraction.

When the volume of inbound requests assigned to a single PriorityLevelConfiguration is more than its permitted concurrency level, the type field of its specification determines what will happen to extra requests. A type of Reject means that excess traffic will immediately be rejected with an HTTP 429 (Too Many Requests) error. A type of Queue means that requests above the threshold will be queued, with the shuffle sharding and fair queuing techniques used to balance progress between request flows.

The queuing configuration allows tuning the fair queuing algorithm for a priority level. Details of the algorithm can be read in the enhancement proposal, but in short:

• Increasing queues reduces the rate of collisions between different flows, at the cost of increased memory usage. A value of 1 here effectively disables the fair-queuing logic, but still allows requests to be queued.

• Increasing queueLengthLimit allows larger bursts of traffic to be sustained without dropping any requests, at the cost of increased latency and memory usage.

• Changing handSize allows you to adjust the probability of collisions between different flows and the overall concurrency available to a single flow in an overload situation.

  Note: A larger handSize makes it less likely for two individual flows to collide (and therefore for one to be able to starve the other), but more likely that a small number of flows can dominate the apiserver. A larger handSize also potentially increases the amount of latency that a single high-traffic flow can cause. The maximum number of queued requests possible from a single flow is handSize * queueLengthLimit.

Following is a table showing an interesting collection of shuffle sharding configurations, showing for each the probability that a given mouse (low-intensity flow) is squished by the elephants (high-intensity flows) for an illustrative collection of numbers of elephants. See https://play.golang.org/p/Gi0PLgVHiUg , which computes this table.

FlowSchema

A FlowSchema matches some inbound requests and assigns them to a priority level. Every inbound request is tested against every FlowSchema in turn, starting with those with numerically lowest --- which we take to be the logically highest --- matchingPrecedence and working onward. The first match wins.

A FlowSchema matches a given request if at least one of its rules matches. A rule matches if at least one of its subjects and at least one of its resourceRules or nonResourceRules (depending on whether the incoming request is for a resource or non-resource URL) matches the request.

For the name field in subjects, and the verbs, apiGroups, resources, namespaces, and nonResourceURLs fields of resource and non-resource rules, the wildcard * may be specified to match all values for the given field, effectively removing it from consideration.

A FlowSchema's distinguisherMethod.type determines how requests matching that schema will be separated into flows. It may be either ByUser, in which case one requesting user will not be able to starve other users of capacity, or ByNamespace, in which case requests for resources in one namespace will not be able to starve requests for resources in other namespaces of capacity, or it may be blank (or distinguisherMethod may be omitted entirely), in which case all requests matched by this FlowSchema will be considered part of a single flow. The correct choice for a given FlowSchema depends on the resource and your particular environment.

Defaults

Each kube-apiserver maintains two sorts of APF configuration objects: mandatory and suggested.

Mandatory Configuration Objects

The four mandatory configuration objects reflect fixed built-in guardrail behavior. This is behavior that the servers have before those objects exist, and when those objects exist their specs reflect this behavior. The four mandatory objects are as follows.

• The mandatory exempt priority level is used for requests that are not subject to flow control at all: they will always be dispatched immediately. The mandatory exempt FlowSchema classifies all requests from the system:masters group into this priority level. You may define other FlowSchemas that direct other requests to this priority level, if appropriate.

• The mandatory catch-all priority level is used in combination with the mandatory catch-all FlowSchema to make sure that every request gets some kind of classification. Typically you should not rely on this catch-all configuration, and should create your own catch-all FlowSchema and PriorityLevelConfiguration (or use the suggested global-default priority level that is installed by default) as appropriate. Because it is not expected to be used normally, the mandatory catch-all priority level has a very small concurrency share and does not queue requests.

Suggested Configuration Objects

The suggested FlowSchemas and PriorityLevelConfigurations constitute a reasonable default configuration. You can modify these and/or create additional configuration objects if you want. If your cluster is likely to experience heavy load then you should consider what configuration will work best.

The suggested configuration groups requests into six priority levels:

• The node-high priority level is for health updates from nodes.

• The system priority level is for non-health requests from the system:nodes group, i.e. Kubelets, which must be able to contact the API server in order for workloads to be able to schedule on them.

• The leader-election priority level is for leader election requests from built-in controllers (in particular, requests for endpoints, configmaps, or leases coming from the system:kube-controller-manager or system:kube-scheduler users and service accounts in the kube-system namespace). These are important to isolate from other traffic because failures in leader election cause their controllers to fail and restart, which in turn causes more expensive traffic as the new controllers sync their informers.

• The workload-high priority level is for other requests from built-in controllers.

• The workload-low priority level is for requests from any other service account, which will typically include all requests from controllers running in Pods.

• The global-default priority level handles all other traffic, e.g. interactive kubectl commands run by nonprivileged users.

The suggested FlowSchemas serve to steer requests into the above priority levels, and are not enumerated here.

Maintenance of the Mandatory and Suggested Configuration Objects

Each kube-apiserver independently maintains the mandatory and suggested configuration objects, using initial and periodic behavior. Thus, in a situation with a mixture of servers of different versions there may be thrashing as long as different servers have different opinions of the proper content of these objects.

Each kube-apiserver makes an initial maintenance pass over the mandatory and suggested configuration objects, and after that does periodic maintenance (once per minute) of those objects.

For the mandatory configuration objects, maintenance consists of ensuring that the object exists and, if it does, has the proper spec. The server refuses to allow a creation or update with a spec that is inconsistent with the server's guardrail behavior.

Maintenance of suggested configuration objects is designed to allow their specs to be overridden. Deletion, on the other hand, is not respected: maintenance will restore the object. If you do not want a suggested configuration object then you need to keep it around but set its spec to have minimal consequences. Maintenance of suggested objects is also designed to support automatic migration when a new version of the kube-apiserver is rolled out, albeit potentially with thrashing while there is a mixed population of servers.

Maintenance of a suggested configuration object consists of creating it --- with the server's suggested spec --- if the object does not exist. OTOH, if the object already exists, maintenance behavior depends on whether the kube-apiservers or the users control the object. In the former case, the server ensures that the object's spec is what the server suggests; in the latter case, the spec is left alone.

The question of who controls the object is answered by first looking for an annotation with key apf.kubernetes.io/autoupdate-spec. If there is such an annotation and its value is true then the kube-apiservers control the object. If there is such an annotation and its value is false then the users control the object. If neither of those condtions holds then the metadata.generation of the object is consulted. If that is 1 then the kube-apiservers control the object. Otherwise the users control the object. These rules were introduced in release 1.22 and their consideration of metadata.generation is for the sake of migration from the simpler earlier behavior. Users who wish to control a suggested configuration object should set its apf.kubernetes.io/autoupdate-spec annotation to false.

Maintenance of a mandatory or suggested configuration object also includes ensuring that it has an apf.kubernetes.io/autoupdate-spec annotation that accurately reflects whether the kube-apiservers control the object.

Maintenance also includes deleting objects that are neither mandatory nor suggested but are annotated apf.kubernetes.io/autoupdate-spec=true.

Health check concurrency exemption

The suggested configuration gives no special treatment to the health check requests on kube-apiservers from their local kubelets --- which tend to use the secured port but supply no credentials. With the suggested config, these requests get assigned to the global-default FlowSchema and the corresponding global-default priority level, where other traffic can crowd them out.

If you add the following additional FlowSchema, this exempts those requests from rate limiting.

priority-and-fairness/health-for-strangers.yaml

apiVersion: flowcontrol.apiserver.k8s.io/v1beta2
kind: FlowSchema
metadata:
  name: health-for-strangers
spec:
  matchingPrecedence: 1000
  priorityLevelConfiguration:
    name: exempt
  rules:
  - nonResourceRules:
    - nonResourceURLs:
      - "/healthz"
      - "/livez"
      - "/readyz"
      verbs:
      - "*"
    subjects:
    - kind: Group
      group:
        name: system:unauthenticated

Diagnostics

Every HTTP response from an API server with the priority and fairness feature enabled has two extra headers: X-Kubernetes-PF-FlowSchema-UID and X-Kubernetes-PF-PriorityLevel-UID, noting the flow schema that matched the request and the priority level to which it was assigned, respectively. The API objects' names are not included in these headers in case the requesting user does not have permission to view them, so when debugging you can use a command like

kubectl get flowschemas -o custom-columns="uid:{metadata.uid},name:{metadata.name}"
kubectl get prioritylevelconfigurations -o custom-columns="uid:{metadata.uid},name:{metadata.name}"

to get a mapping of UIDs to names for both FlowSchemas and PriorityLevelConfigurations.

Observability

Metrics

When you enable the API Priority and Fairness feature, the kube-apiserver exports additional metrics. Monitoring these can help you determine whether your configuration is inappropriately throttling important traffic, or find poorly-behaved workloads that may be harming system health.

• apiserver_flowcontrol_rejected_requests_total is a counter vector (cumulative since server start) of requests that were rejected, broken down by the labels flow_schema (indicating the one that matched the request), priority_level (indicating the one to which the request was assigned), and reason. The reason label will be have one of the following values:

  • queue-full, indicating that too many requests were already queued,
  • concurrency-limit, indicating that the PriorityLevelConfiguration is configured to reject rather than queue excess requests, or
  • time-out, indicating that the request was still in the queue when its queuing time limit expired.

• apiserver_flowcontrol_dispatched_requests_total is a counter vector (cumulative since server start) of requests that began executing, broken down by the labels flow_schema (indicating the one that matched the request) and priority_level (indicating the one to which the request was assigned).

• apiserver_current_inqueue_requests is a gauge vector of recent high water marks of the number of queued requests, grouped by a label named request_kind whose value is mutating or readOnly. These high water marks describe the largest number seen in the one second window most recently completed. These complement the older apiserver_current_inflight_requests gauge vector that holds the last window's high water mark of number of requests actively being served.

• apiserver_flowcontrol_read_vs_write_request_count_samples is a histogram vector of observations of the then-current number of requests, broken down by the labels phase (which takes on the values waiting and executing) and request_kind (which takes on the values mutating and readOnly). The observations are made periodically at a high rate. Each observed value is a ratio, between 0 and 1, of a number of requests divided by the corresponding limit on the number of requests (queue length limit for waiting and concurrency limit for executing).

• apiserver_flowcontrol_read_vs_write_request_count_watermarks is a histogram vector of high or low water marks of the number of requests (divided by the corresponding limit to get a ratio in the range 0 to 1) broken down by the labels phase (which takes on the values waiting and executing) and request_kind (which takes on the values mutating and readOnly); the label mark takes on values high and low. The water marks are accumulated over windows bounded by the times when an observation was added to apiserver_flowcontrol_read_vs_write_request_count_samples. These water marks show the range of values that occurred between samples.

• apiserver_flowcontrol_current_inqueue_requests is a gauge vector holding the instantaneous number of queued (not executing) requests, broken down by the labels priority_level and flow_schema.

• apiserver_flowcontrol_current_executing_requests is a gauge vector holding the instantaneous number of executing (not waiting in a queue) requests, broken down by the labels priority_level and flow_schema.

• apiserver_flowcontrol_request_concurrency_in_use is a gauge vector holding the instantaneous number of occupied seats, broken down by the labels priority_level and flow_schema.

• apiserver_flowcontrol_priority_level_request_count_samples is a histogram vector of observations of the then-current number of requests broken down by the labels phase (which takes on the values waiting and executing) and priority_level. Each histogram gets observations taken periodically, up through the last activity of the relevant sort. The observations are made at a high rate. Each observed value is a ratio, between 0 and 1, of a number of requests divided by the corresponding limit on the number of requests (queue length limit for waiting and concurrency limit for executing).

• apiserver_flowcontrol_priority_level_request_count_watermarks is a histogram vector of high or low water marks of the number of requests (divided by the corresponding limit to get a ratio in the range 0 to 1) broken down by the labels phase (which takes on the values waiting and executing) and priority_level; the label mark takes on values high and low. The water marks are accumulated over windows bounded by the times when an observation was added to apiserver_flowcontrol_priority_level_request_count_samples. These water marks show the range of values that occurred between samples.

• apiserver_flowcontrol_priority_level_seat_count_samples is a histogram vector of observations of the utilization of a priority level's concurrency limit, broken down by priority_level. This utilization is the fraction (number of seats occupied) / (concurrency limit). This metric considers all stages of execution (both normal and the extra delay at the end of a write to cover for the corresponding notification work) of all requests except WATCHes; for those it considers only the initial stage that delivers notifications of pre-existing objects. Each histogram in the vector is also labeled with phase: executing (there is no seat limit for the waiting phase). Each histogram gets observations taken periodically, up through the last activity of the relevant sort. The observations are made at a high rate.

• apiserver_flowcontrol_priority_level_seat_count_watermarks is a histogram vector of high or low water marks of the utilization of a priority level's concurrency limit, broken down by priority_level and mark (which takes on values high and low). Each histogram in the vector is also labeled with phase: executing (there is no seat limit for the waiting phase). The water marks are accumulated over windows bounded by the times when an observation was added to apiserver_flowcontrol_priority_level_seat_count_samples. These water marks show the range of values that occurred between samples.

• apiserver_flowcontrol_request_queue_length_after_enqueue is a histogram vector of queue lengths for the queues, broken down by the labels priority_level and flow_schema, as sampled by the enqueued requests. Each request that gets queued contributes one sample to its histogram, reporting the length of the queue immediately after the request was added. Note that this produces different statistics than an unbiased survey would.

  Note: An outlier value in a histogram here means it is likely that a single flow (i.e., requests by one user or for one namespace, depending on configuration) is flooding the API server, and being throttled. By contrast, if one priority level's histogram shows that all queues for that priority level are longer than those for other priority levels, it may be appropriate to increase that PriorityLevelConfiguration's concurrency shares.

• apiserver_flowcontrol_request_concurrency_limit is a gauge vector holding the computed concurrency limit (based on the API server's total concurrency limit and PriorityLevelConfigurations' concurrency shares), broken down by the label priority_level.

• apiserver_flowcontrol_request_wait_duration_seconds is a histogram vector of how long requests spent queued, broken down by the labels flow_schema (indicating which one matched the request), priority_level (indicating the one to which the request was assigned), and execute (indicating whether the request started executing).

  Note: Since each FlowSchema always assigns requests to a single PriorityLevelConfiguration, you can add the histograms for all the FlowSchemas for one priority level to get the effective histogram for requests assigned to that priority level.

• apiserver_flowcontrol_request_execution_seconds is a histogram vector of how long requests took to actually execute, broken down by the labels flow_schema (indicating which one matched the request) and priority_level (indicating the one to which the request was assigned).

• apiserver_flowcontrol_watch_count_samples is a histogram vector of the number of active WATCH requests relevant to a given write, broken down by flow_schema and priority_level.

• apiserver_flowcontrol_work_estimated_seats is a histogram vector of the number of estimated seats (maximum of initial and final stage of execution) associated with requests, broken down by flow_schema and priority_level.

• apiserver_flowcontrol_request_dispatch_no_accommodation_total is a counter vec of the number of events that in principle could have led to a request being dispatched but did not, due to lack of available concurrency, broken down by flow_schema and priority_level. The relevant sorts of events are arrival of a request and completion of a request.

Debug endpoints

When you enable the API Priority and Fairness feature, the kube-apiserver serves the following additional paths at its HTTP[S] ports.

• /debug/api_priority_and_fairness/dump_priority_levels - a listing of all the priority levels and the current state of each. You can fetch like this:

  kubectl get --raw /debug/api_priority_and_fairness/dump_priority_levels
  

  The output is similar to this:

  PriorityLevelName, ActiveQueues, IsIdle, IsQuiescing, WaitingRequests, ExecutingRequests,
  workload-low,      0,            true,   false,       0,               0,
  global-default,    0,            true,   false,       0,               0,
  exempt,            <none>,       <none>, <none>,      <none>,          <none>,
  catch-all,         0,            true,   false,       0,               0,
  system,            0,            true,   false,       0,               0,
  leader-election,   0,            true,   false,       0,               0,
  workload-high,     0,            true,   false,       0,               0,
  

• /debug/api_priority_and_fairness/dump_queues - a listing of all the queues and their current state. You can fetch like this:

  kubectl get --raw /debug/api_priority_and_fairness/dump_queues
  

  The output is similar to this:

  PriorityLevelName, Index,  PendingRequests, ExecutingRequests, VirtualStart,
  workload-high,     0,      0,               0,                 0.0000,
  workload-high,     1,      0,               0,                 0.0000,
  workload-high,     2,      0,               0,                 0.0000,
  ...
  leader-election,   14,     0,               0,                 0.0000,
  leader-election,   15,     0,               0,                 0.0000,
  

• /debug/api_priority_and_fairness/dump_requests - a listing of all the requests that are currently waiting in a queue. You can fetch like this:

  kubectl get --raw /debug/api_priority_and_fairness/dump_requests
  

  The output is similar to this:

  PriorityLevelName, FlowSchemaName, QueueIndex, RequestIndexInQueue, FlowDistingsher,       ArriveTime,
  exempt,            <none>,         <none>,     <none>,              <none>,                <none>,
  system,            system-nodes,   12,         0,                   system:node:127.0.0.1, 2020-07-23T15:26:57.179170694Z,
  

  In addition to the queued requests, the output includes one phantom line for each priority level that is exempt from limitation.

  You can get a more detailed listing with a command like this:

  kubectl get --raw '/debug/api_priority_and_fairness/dump_requests?includeRequestDetails=1'
  

  The output is similar to this:

  PriorityLevelName, FlowSchemaName, QueueIndex, RequestIndexInQueue, FlowDistingsher,       ArriveTime,                     UserName,              Verb,   APIPath,                                                     Namespace, Name,   APIVersion, Resource, SubResource,
  system,            system-nodes,   12,         0,                   system:node:127.0.0.1, 2020-07-23T15:31:03.583823404Z, system:node:127.0.0.1, create, /api/v1/namespaces/scaletest/configmaps,
  system,            system-nodes,   12,         1,                   system:node:127.0.0.1, 2020-07-23T15:31:03.594555947Z, system:node:127.0.0.1, create, /api/v1/namespaces/scaletest/configmaps,
  

What's next

For background information on design details for API priority and fairness, see the enhancement proposal. You can make suggestions and feature requests via SIG API Machinery or the feature's slack channel.