Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Feat/event based instance replacement - Merged all "new" commit from …
…master (#433) * Allow flavor to be customised (#359) * fix typo (#360) * Fix typo (#361) * UserData wrappers for CloudFormation helper scripts when using Beanstalk (#366) * Support custom role for cfn-init in Beanstalk UserData * Wrappers & refactoring * Docs * Docs fixes * More docs fixes * Docs fixes * Yet more docs fixes * Lambda & Kubernetes config * AutoSpottingElasticBeanstalk managed policy + Rename beanstalk_cfn_init_role to beanstalk_cfn_wrappers * Update param description * Kubernetes config fix * Rename Beanstalk variable + Move test data * Add missing permission for AutoSpottingBeanstalk role * Bring Go to 1.13 (#367) * merge fixes * Begin - port regional CloudFormation stack changes to the main template so it can also run from a StackSet * Merge fix * Fix to template * Multiple fixes to handle lambda concurrency As instaces can be launched in concurrency/overlapping we have the problems related to multiple lambdas acting on the same ASG * Removed commented code * Begin working on: #354 (comment) * Progress #354 Created Queue in CF template Begin sending message * Progress #354 * Progress #354 * Progress #354 * Progress #354 * Progress #354 * * Use inline python lambda to increase ASG Use it only if AttachInstances method fails for wrong ASG max size. * * Use inline python lambda to increase ASG progress * * Use inline python lambda to increase ASG progress * * Improvements on event based instance replacement - Ported to StackSet deploy mode. - Fix to "ScalingActivityInProgress" if launched ondemand instance is terminated before going "inservice": Created method "waitForInstanceStatus", that wait, until a max retry (5), that an instance belonging to an ASG is in the desired status (InService). (Sleep time is 5*retry) - Fix to multiple problems caused by lambda concurrency changing ASG MaxSize: Created another Lambda (LambdaManageASG) in the same region of the main one; code python3.7, inline in template. Lambda concurrency is set to one. Is function is to change ASG MaxSize by the ammount specified. Used a "try-catch approach": if AttachInstances return error code "ValidationError" and string "update the AutoScalingGroup sizes" is present in error message, means that we need to increase ASG MaxSize. So we execute method changeAutoScalingMaxSize that invoke LambdaManageASG. If invoke return error "ErrCodeTooManyRequestsException", means that multiple Main lambdas are executing LambdaManageASG, we sleep for a random interval part of seconds and retry. Method attachSpotInstance now return an int that represent the ammount of change to the ASG MaxSize, so that in swapWithGroupMember we can defer the call to changeAutoScalingMaxSize to decrease the ASG MaxSize of the previously increased ammount. We use "waitForInstanceStatus" in attachSpotInstance before returning to be sure that spot instance has been attached and are InService before beginning to terminate ondemand one. * * Improvements on event based instance replacement - Add logPrefix to better identify lambda actions in case of concurent executions. TE = Spot terminate event ST = Instance start event SC = Schedule event TE and ST are followed by instanceID, SC is followed by activation time. * * Improvements on event based instance replacement - fix/improvement on Suspending/Resuming Termination process: suspend/resume is now handled by LambdaManageASG too. When i successfully suspend termination, i add a tag "autospotting_suspend_process_by" with value equals to the instanceId event that triggered the main lambda. When i try to resume termination first check if the value of the tag above equals the current one. If not, means that another main lambda, different from the one who suspended it, is trying to resume the process; in that case i do not resume it. Considered that LambdaManageASG has a concurrent execution limit set to 1 we can have the following cases: 1) *) A main lambda suspend the process. *) No other lambda suspend it. *) Main lambda resume the process. *) Another main lambda suspend the process *) ....and so on 2) *) A main lambda suspend the process. *) Before the first one resume the process another one suspend it (so replacing the tag value) *) First lambda do not resume the process (tag value differ) *) Second lambda resume the process In the wrost scenario of a Lambda dying before resuming the process, it will be resumed after another one will suspend it. * * Improvements on event based instance replacement code cosmetic fix * * Improvements on event based instance replacement - for rand seed instead of using time.Now().UnixNano() we build a seed based on the instanceId that triggered the event. The seed is build this way: for every char of instanceId (starting from third char) we get his rune "representation" and sum it to previous one. We use it as a temporary seed and get a random number between 0 and 9. The final seed is the concatenation of the generated random numbers. This way we have a seed number (int64) that depend from the instanceId and of the same lenght. * * Improvements on event based instance replacement - no more need to "revert attach/detach order when running on minimum capacity". Defer changeAutoScalingMaxSize and use same logic of swapWithGroupMember. * * Improvements on event based instance replacement - Use suspendResumeProcess for schedule replaceOnDemandInstanceWithSpot too. We use it as a trick to avoid rare cases of concurrency between scheduled and event lambdas. As lambda that handle "suspendResumeProcess" have a concurrency limit of one, scheduled and event lambdas, if concurrent, will be "time shifted" by a random value. This way they will not execute attachSpotInstance at the same time. In case of scheduled lambda we add the "S" char to the instanceId used for the randSeed to avoid that it resume process suspended by event lambda. - Fix in swapWithGroupMember: defer asg.suspendResumeProcess for resume so that it will be executed even if function swapWithGroupMember return error. * * gofmt cosmetic changes * Fix to rand.Intn parameter to include even 9 * fix terminateRandomSpotInstanceIfHavingEnough need another condition to avoid terminating valid spot instance in case ASG have minOnDemand > 0 if all ASG instances are in state running and Min OnDemand instance equals total ondemand running and all instances running equals desired capacity means that i do not need to terminate a spot instance need some testing * fix terminateRandomSpotInstanceIfHavingEnough fixes * fix terminateRandomSpotInstanceIfHavingEnough changed allInstancesRunning to return ondemand instances running too * Specify/override multiplier for the on-demand price on a group level As i.price already have been multiplied by the global value, if specified, i need first to divide it by the same value and then multiply it by the multiplier specific to the ASG. We need to do this for both scheduled and event actions. For event we act in function belongsToEnabledASG. For schedule we act in launchSpotReplacement. Need deep testing... * Specify/override multiplier for the on-demand price on a group level - fix in loadConfOnDemandPriceMultiplier need to use a.config.OnDemandPriceMultiplier in place of a.region.conf.OnDemandPriceMultiplier this way a.region.conf.OnDemandPriceMultiplier will conserve the original global value * Misc fixes to be able to run go tests * Misc fixes to be able to run go tests - continue * Misc fixes to be able to run go tests - end for now * For Test_autoScalingGroup_terminateRandomSpotInstanceIfHavingEnough added contion: "spot capacity is correct, skip termination" * Merge Updating AWS SDK to latest version - [0ee95c8] * Merge Relicense to OSL-3 - [972bc61] * Merge Update readme, mention the relicensing to OSL-3 - [9b438dc] * Merge Fix `make archive` on macOS [c929391] * Merge Move some logs to debug [fa0c27b] * Merge Delete old Gopkg files [5504e34] * Merge Update dependencies [0e14a7b] * Merge Support spot price buffer percentage of 0 [26ca955] * Merge Don't segfault when spot instance doesn't belong to ASG [f313c6d] * Merge Allow specifying GOOS and GOARCH [8d67d6e] * Merge Use /bin/bash for shell [0eed8cd] * Merge Ignore terminating spot instances that don't belong to AutoSpotting (master) and Enable Spot Termination ASG Checking (event) [33a444c] * Merge Revert Use /bin/bash for shell [6e45440] * Merge Pass --abort-on-container-exit to docker-compose [8f90fba] * Merge Rename travisci make targets to ci [2d76a9a] * Merge Create FUNDING.yml + Update FUNDING.yml [dde6d85,ca81828] * Merge Use paginated version of DescribeSpotPriceHistory [9a770b3] * Merge Delete DescribeSecurityGroups mock [3bfd542] * Merge Move config loading out of main and add tests for it [30a4392] * Fixes to Merge Move config loading out of main and add tests for it * Merge Move logs about incompatible instance types to debug [985d675] * Merge Remove incorrect Makefile conditionals [a01ee26] * Merge Actually fail the build if gofmt fails [c391d69] * Merge Add tools to go.mod [8a9a90b] * Merge No longer enforce the name of the ElasticBeanstalk IAM policy [3313298] * Merge Added spot premium [ecf31a5] * Merge Update how bid price is calculated for premium instances [8d13fb0] * Merge Update dependencies [caf373d] * Merge Cron timezone [8e04942] * Merge Spleling [a37aafc] * Merge Update README.md [0109a0b,fb15aa4,4e03db7] * Merge Use the larger of min-OD-instances and (min-OD-percent * current) [25ac31f] * gofmt changes * cronEventAction - fix to logic need to invert test for needReplaceOnDemandInstances and onDemandInstance == nil currently if onDemandInstance == nil methods return and execution stop. This way the check to terminate a spotInstace if their number is more than required is never done. Assume that ASG scale down and terminate the onDemand instances that autospotting is not terminating [autospotting_min_on_demand_number]. On the next runs onDemandInstance will be nil and spot instances in excess will not be terminated. Co-authored-by: Chris <chris.farrenden@domain.com.au> Co-authored-by: 0x11 <14269809+codenoid@users.noreply.github.com> Co-authored-by: Jawad <jawad.stouli@gmail.com> Co-authored-by: Gábor Lipták <gliptak@gmail.com>
- Loading branch information