| 1 |
/* |
| 2 |
Copyright 2021 The Kubernetes Authors. |
| 3 |
|
| 4 |
Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 |
you may not use this file except in compliance with the License. |
| 6 |
You may obtain a copy of the License at |
| 7 |
|
| 8 |
http://www.apache.org/licenses/LICENSE-2.0 |
| 9 |
|
| 10 |
Unless required by applicable law or agreed to in writing, software |
| 11 |
distributed under the License is distributed on an "AS IS" BASIS, |
| 12 |
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 |
See the License for the specific language governing permissions and |
| 14 |
limitations under the License. |
| 15 |
*/ |
| 16 |
|
| 17 |
package request |
| 18 |
|
| 19 |
import ( |
| 20 |
"math" |
| 21 |
"net/http" |
| 22 |
"time" |
| 23 |
|
| 24 |
apirequest "k8s.io/apiserver/pkg/endpoints/request" |
| 25 |
"k8s.io/apiserver/pkg/util/flowcontrol/metrics" |
| 26 |
) |
| 27 |
|
| 28 |
func newMutatingWorkEstimator(countFn watchCountGetterFunc, config *WorkEstimatorConfig, maxSeatsFn maxSeatsFunc) WorkEstimatorFunc { |
| 29 |
estimator := &mutatingWorkEstimator{ |
| 30 |
config: config, |
| 31 |
countFn: countFn, |
| 32 |
maxSeatsFn: maxSeatsFn, |
| 33 |
} |
| 34 |
return estimator.estimate |
| 35 |
} |
| 36 |
|
| 37 |
type mutatingWorkEstimator struct { |
| 38 |
config *WorkEstimatorConfig |
| 39 |
countFn watchCountGetterFunc |
| 40 |
maxSeatsFn maxSeatsFunc |
| 41 |
} |
| 42 |
|
| 43 |
func (e *mutatingWorkEstimator) estimate(r *http.Request, flowSchemaName, priorityLevelName string) WorkEstimate { |
| 44 |
minSeats := e.config.MinimumSeats |
| 45 |
maxSeats := e.maxSeatsFn(priorityLevelName) |
| 46 |
if maxSeats == 0 || maxSeats > e.config.MaximumSeatsLimit { |
| 47 |
maxSeats = e.config.MaximumSeatsLimit |
| 48 |
} |
| 49 |
|
| 50 |
// TODO(wojtekt): Remove once we tune the algorithm to not fail |
| 51 |
// scalability tests. |
| 52 |
if !e.config.Enabled { |
| 53 |
return WorkEstimate{ |
| 54 |
InitialSeats: minSeats, |
| 55 |
} |
| 56 |
} |
| 57 |
|
| 58 |
requestInfo, ok := apirequest.RequestInfoFrom(r.Context()) |
| 59 |
if !ok { |
| 60 |
// no RequestInfo should never happen, but to be on the safe side |
| 61 |
// let's return a large value. |
| 62 |
return WorkEstimate{ |
| 63 |
InitialSeats: minSeats, |
| 64 |
FinalSeats: maxSeats, |
| 65 |
AdditionalLatency: e.config.eventAdditionalDuration(), |
| 66 |
} |
| 67 |
} |
| 68 |
|
| 69 |
if isRequestExemptFromWatchEvents(requestInfo) { |
| 70 |
return WorkEstimate{ |
| 71 |
InitialSeats: minSeats, |
| 72 |
FinalSeats: 0, |
| 73 |
AdditionalLatency: time.Duration(0), |
| 74 |
} |
| 75 |
} |
| 76 |
|
| 77 |
watchCount := e.countFn(requestInfo) |
| 78 |
metrics.ObserveWatchCount(r.Context(), priorityLevelName, flowSchemaName, watchCount) |
| 79 |
|
| 80 |
// The cost of the request associated with the watchers of that event |
| 81 |
// consists of three parts: |
| 82 |
// - cost of going through the event change logic |
| 83 |
// - cost of serialization of the event |
| 84 |
// - cost of processing an event object for each watcher (e.g. filtering, |
| 85 |
// sending data over network) |
| 86 |
// We're starting simple to get some operational experience with it and |
| 87 |
// we will work on tuning the algorithm later. Given that the actual work |
| 88 |
// associated with processing watch events is happening in multiple |
| 89 |
// goroutines (proportional to the number of watchers) that are all |
| 90 |
// resumed at once, as a starting point we assume that each such goroutine |
| 91 |
// is taking 1/Nth of a seat for M milliseconds. |
| 92 |
// We allow the accounting of that work in P&F to be reshaped into another |
| 93 |
// rectangle of equal area for practical reasons. |
| 94 |
var finalSeats uint64 |
| 95 |
var additionalLatency time.Duration |
| 96 |
|
| 97 |
// TODO: Make this unconditional after we tune the algorithm better. |
| 98 |
// Technically, there is an overhead connected to processing an event after |
| 99 |
// the request finishes even if there is a small number of watches. |
| 100 |
// However, until we tune the estimation we want to stay on the safe side |
| 101 |
// an avoid introducing additional latency for almost every single request. |
| 102 |
if watchCount >= int(e.config.WatchesPerSeat) { |
| 103 |
// TODO: As described in the KEP, we should take into account that not all |
| 104 |
// events are equal and try to estimate the cost of a single event based on |
| 105 |
// some historical data about size of events. |
| 106 |
finalSeats = uint64(math.Ceil(float64(watchCount) / e.config.WatchesPerSeat)) |
| 107 |
finalWork := SeatsTimesDuration(float64(finalSeats), e.config.eventAdditionalDuration()) |
| 108 |
|
| 109 |
// While processing individual events is highly parallel, |
| 110 |
// the design/implementation of P&F has a couple limitations that |
| 111 |
// make using this assumption in the P&F implementation very |
| 112 |
// inefficient because: |
| 113 |
// - we reserve max(initialSeats, finalSeats) for time of executing |
| 114 |
// both phases of the request |
| 115 |
// - even more importantly, when a given `wide` request is the one to |
| 116 |
// be dispatched, we are not dispatching any other request until |
| 117 |
// we accumulate enough seats to dispatch the nominated one, even |
| 118 |
// if currently unoccupied seats would allow for dispatching some |
| 119 |
// other requests in the meantime |
| 120 |
// As a consequence of these, the wider the request, the more capacity |
| 121 |
// will effectively be blocked and unused during dispatching and |
| 122 |
// executing this request. |
| 123 |
// |
| 124 |
// To mitigate the impact of it, we're capping the maximum number of |
| 125 |
// seats that can be assigned to a given request. Thanks to it: |
| 126 |
// 1) we reduce the amount of seat-seconds that are "wasted" during |
| 127 |
// dispatching and executing initial phase of the request |
| 128 |
// 2) we are not changing the finalWork estimate - just potentially |
| 129 |
// reshaping it to be narrower and longer. As long as the maximum |
| 130 |
// seats setting will prevent dispatching too many requests at once |
| 131 |
// to prevent overloading kube-apiserver (and/or etcd or the VM or |
| 132 |
// a physical machine it is running on), we believe the relaxed |
| 133 |
// version should be good enough to achieve the P&F goals. |
| 134 |
// |
| 135 |
// TODO: Confirm that the current cap of maximumSeats allow us to |
| 136 |
// achieve the above. |
| 137 |
if finalSeats > maxSeats { |
| 138 |
finalSeats = maxSeats |
| 139 |
} |
| 140 |
additionalLatency = finalWork.DurationPerSeat(float64(finalSeats)) |
| 141 |
} |
| 142 |
|
| 143 |
return WorkEstimate{ |
| 144 |
InitialSeats: 1, |
| 145 |
FinalSeats: finalSeats, |
| 146 |
AdditionalLatency: additionalLatency, |
| 147 |
} |
| 148 |
} |
| 149 |
|
| 150 |
func isRequestExemptFromWatchEvents(requestInfo *apirequest.RequestInfo) bool { |
| 151 |
// Creating token for service account does not produce any event, |
| 152 |
// but still serviceaccounts can have multiple watchers. |
| 153 |
if requestInfo.Resource == "serviceaccounts" && requestInfo.Subresource == "token" { |
| 154 |
return true |
| 155 |
} |
| 156 |
return false |
| 157 |
} |