Skip to content

Latest commit

 

History

History
86 lines (72 loc) · 5.09 KB

api_call_latency.md

File metadata and controls

86 lines (72 loc) · 5.09 KB
Raw

API call latency SLIs/SLOs details

Definition

Status SLI SLO
Official Latency of processing1 mutating2 API calls for single objects for every (resource, verb) pair, measured as 99th percentile over last 5 minutes In default Kubernetes installation, for every (resource, verb) pair, excluding virtual and aggregated resources and Custom Resource Definitions, 99th percentile per cluster-day <= 1s
Official Latency of processing1 non-streaming read-only3 API calls for every (resource, scope4) pair, measured as 99th percentile over last 5 minutes In default Kubernetes installation, for every (resource, scope) pair, excluding virtual and aggregated resources and Custom Resource Definitions, 99th percentile per cluster-day: (a) <= 1s if scope=resource (b) <= 30s5 otherwise (if scope=namespace or scope=cluster)

[1] The SLI only measures latency incurred by the processing time of the request. The processing time of a request is the moment when apiserver gets the request to last byte of response sent to the user, excluding latency incurred by webhooks and priority & fairness queue wait times.

[2] By mutating API calls we mean POST, PUT, DELETE and PATCH.

[3] By non-streaming read-only API calls we mean GET requests without watch=true option set. (Note that in Kubernetes internally it translates to both GET and LIST calls).

[4] A scope of a request can be either

  • resource - if the request is about a single object
  • namespace - if it is about objects from a single namespace
  • cluster - if it spawns objects from multiple namespaces

[5] Historically, the threshold for LISTs with scope=namespace was set to 5 seconds. However, the threshold was chosen when Kubernetes didn't support the scale it supports today and when individual namespace didn't contain tens of thousands (if not more) objects of a given type. We adjusted the limits to accommodate the usage patterns change, given that users are fine with listing tens of thousands of objects taking more than 5 seconds.

User stories

  • As a user of vanilla Kubernetes, I want some guarantee how quickly I get the response from an API call.
  • As an administrator of Kubernetes cluster, if I know characteristics of my external dependencies of apiserver (e.g custom admission plugins, priority & fairness configuration, and webhooks). I want to be able to provide guarantees for API calls latency to users of my cluster.

Other notes

  • We obviously can’t give any guarantee in general, because cluster administrators are allowed to register custom admission plugins, webhooks, and priority and fairness configurations, which we don’t have any control about and they obviously impact API call latencies.
  • As a result, we define the SLIs to be very generic (no matter how your cluster is set up), but we provide SLO only for default installations (where we have control over what apiserver is doing). This doesn’t provide a false impression, that we provide guarantee no matter how the cluster is setup and what is installed on top of it.
  • At the same time, API calls are part of pretty much every non-trivial workflow in Kubernetes, so this metric is a building block for less trivial SLIs and SLOs.
  • The SLO for latency for read-only API calls of a given type may have significant buffer in threshold. In fact, the latency of the request should be proportional to the amount of work to do (which is number of objects of a given type in a given scope) plus some constant overhead. For better tracking of performance, we may want to define purely internal SLI of "latency per object". But that isn't in near term plans.
  • To recall, SLOs are guaranteed only if thresholds defined in thresholds file are satisfied. This is particularly important for this SLO, because it limits the number of objects that are returned by LIST calls.

Caveats

  • The SLO has to be satisfied independently from used encoding in user-originated requests. This makes mix of client important while testing. However, we assume that all core components communicate with apiserver using protocol buffers.
  • In case of GET requests, user has an option opt-in for accepting potentially stale data (being served from cache) and the SLO again has to be satisfied independently of that. This makes the careful choice of requests in tests important.
  • The SLI & SLO excludes latency incurred by factors that are outside our control, specifically from webhooks (1.23+) and API priority & fairness queue wait times (1.27+).

TODOs

  • We may consider treating non-namespaced resources as a separate bucket in the future. However, it may not make sense if the number of those may be comparable with namespaced ones.

Test scenario

TODO: Describe test scenario.