123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877287828792880288128822883288428852886288728882889289028912892289328942895289628972898289929002901290229032904290529062907290829092910291129122913291429152916291729182919292029212922292329242925292629272928292929302931293229332934293529362937293829392940294129422943294429452946294729482949295029512952295329542955295629572958295929602961296229632964296529662967296829692970297129722973297429752976297729782979298029812982298329842985298629872988298929902991299229932994299529962997299829993000300130023003300430053006300730083009301030113012301330143015301630173018301930203021302230233024302530263027302830293030303130323033303430353036303730383039304030413042304330443045304630473048304930503051305230533054305530563057305830593060306130623063306430653066306730683069307030713072307330743075307630773078307930803081308230833084308530863087308830893090309130923093309430953096309730983099310031013102310331043105310631073108310931103111311231133114311531163117311831193120312131223123312431253126312731283129313031313132313331343135313631373138313931403141314231433144314531463147314831493150315131523153315431553156315731583159316031613162316331643165316631673168316931703171317231733174317531763177317831793180318131823183318431853186318731883189319031913192319331943195319631973198319932003201320232033204320532063207320832093210321132123213321432153216321732183219322032213222322332243225322632273228322932303231323232333234323532363237323832393240324132423243324432453246324732483249325032513252325332543255325632573258325932603261326232633264326532663267326832693270327132723273327432753276327732783279328032813282328332843285328632873288328932903291329232933294329532963297329832993300330133023303330433053306330733083309331033113312331333143315331633173318331933203321332233233324332533263327332833293330333133323333333433353336333733383339334033413342334333443345334633473348334933503351335233533354335533563357335833593360336133623363336433653366336733683369337033713372337333743375337633773378337933803381338233833384338533863387338833893390339133923393339433953396339733983399340034013402340334043405340634073408340934103411341234133414341534163417341834193420342134223423342434253426342734283429343034313432343334343435343634373438343934403441344234433444344534463447344834493450345134523453345434553456345734583459346034613462346334643465346634673468346934703471347234733474347534763477347834793480348134823483348434853486348734883489349034913492349334943495349634973498349935003501350235033504350535063507350835093510351135123513351435153516351735183519352035213522352335243525352635273528352935303531353235333534353535363537353835393540354135423543354435453546354735483549355035513552355335543555355635573558355935603561356235633564356535663567356835693570357135723573357435753576357735783579358035813582358335843585358635873588358935903591359235933594359535963597359835993600360136023603360436053606360736083609361036113612361336143615361636173618361936203621362236233624362536263627362836293630363136323633363436353636363736383639364036413642364336443645364636473648364936503651365236533654365536563657365836593660366136623663366436653666366736683669367036713672367336743675 |
- # Global release test configuration file.
- # All your release test configuration should go here. Adding release tests here
- # will automatically enable them in the Buildkite release testing schedules
- # (except they have frequency: disabled).
- # Here is an example configuration for reference:
- #- name: example_test
- # # Tests with the same group will be grouped in the Buildkite UI
- # group: Example group
- # # Provide the working directory which will be uploaded to the cluster
- # working_dir: example_dir
- #
- # # For release test infra migration, we provide these fields that are populated
- # # in the database
- # legacy:
- # test_name: example_test
- # test_suite: examples
- #
- # # How often to run the tests.
- # # One of [disabled, any, multi, nightly, weekly].
- # frequency: weekly
- # # Owning team. This field will be persisted to the database
- # team: ml
- #
- # # Python version. This optional field determines which Python version to run tests
- # # on. This must be a string!
- # python: "3.7"
- #
- # # Optional location of a bash setup script to run on the driver
- # # when setting up the local environment. Relative to working_dir
- # driver_setup: setup_driver.sh
- #
- # # Cluster information
- # cluster:
- # # Location of cluster env, relative to working_dir
- # cluster_env: cluster_env.yaml
- # # Location of cluster compute, relative to working_dir
- # cluster_compute: cluster_compute.yaml
- # # Autosuspend parameter passed to the cluster.
- # # The cluster will automatically terminate if inactive for this
- # # many minutes. Defaults to 10 if not set.
- # autosuspend_mins: 10
- # # Optional cloud_id to use instead of the default cloud
- # cloud_id: cld_12345678
- # # Alternatively, you can specify a cloud name
- # cloud_name: anyscale_default_cloud
- #
- # # Run configuration for the test
- # run:
- # # Type of test. Can be sdk_command or client (job to be implemented soon).
- # # Uses either Anyscale SDK commands or the Ray client to run the actual
- # # release test.
- # type: sdk_command
- #
- # # File manager to use to transfer files to and from the cluster.
- # # Can be any of [sdk, client, job].
- # file_manager: sdk
- #
- # # If you want to wait for nodes to be ready, you can specify this here:
- # wait_for_nodes:
- # # Number of nodes
- # num_nodes: 16
- # # Timeout for waiting for nodes. If nodes are not up by then, the
- # # test will fail.
- # timeout: 600
- #
- # # Optional prepare script to be run on the cluster before the test script
- # prepare: python prepare.py
- # # The prepare command can have a separate timeout
- # prepare_timeout: 300
- #
- # # Main script to run as the test script
- # script: python workloads/train_small.py
- # # Timeout in seconds. After this time the test is considered as failed.
- # timeout: 600
- #
- # # You can specify smoke test definitions here. If a smoke test is triggered,
- # # it will deep update the main test configuration with the values provided
- # # here. Smoke tests will automatically run with IS_SMOKE_TEST=1 as en
- # # environment variable and receive the --smoke-test flag as a parameter in the
- # # run script.
- # smoke_test:
- # # Smoke tests can have different frequencies. A smoke test is only triggered
- # # when the regular test is not matched.
- # frequency: nightly
- # # Here we adjust the run timeout down and run on less nodes. The test script
- # # remains the same.
- # run:
- # timeout: 300
- # wait_for_nodes:
- # num_nodes: 4
- # timeout: 600
- #
- # # After the test finished, this handler (in alerts/) will process the results.
- # # It can then let the test fail, e.g. if a metric regression is observed.
- # alert: default
- #########################
- # AIR release tests
- #########################
- - name: long_running_horovod_tune_test
- group: AIR tests
- working_dir: air_tests
- frequency: weekly
- team: ml
- cluster:
- cluster_env: horovod/app_config_master.yaml
- cluster_compute: horovod/compute_tpl.yaml
- run:
- timeout: 36000
- script: python horovod/workloads/horovod_tune_test.py
- long_running: true
- wait_for_nodes:
- num_nodes: 3
- type: sdk_command
- file_manager: job
- smoke_test:
- frequency: disabled
- run:
- timeout: 1800
- alert: default
- #######################
- # XGBoost release tests
- #######################
- - name: xgboost_train_small
- group: XGBoost
- working_dir: xgboost_tests
- legacy:
- test_name: train_small
- test_suite: xgboost_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_small.yaml
- run:
- timeout: 600
- script: python workloads/train_small.py
- wait_for_nodes:
- num_nodes: 4
- type: client
- alert: xgboost_tests
- - name: xgboost_train_moderate
- group: XGBoost
- working_dir: xgboost_tests
- legacy:
- test_name: train_moderate
- test_suite: xgboost_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_moderate.yaml
- run:
- timeout: 600
- script: python workloads/train_moderate.py
- wait_for_nodes:
- num_nodes: 32
- type: sdk_command
- file_manager: sdk
- alert: xgboost_tests
- - name: xgboost_train_gpu
- group: XGBoost
- working_dir: xgboost_tests
- legacy:
- test_name: train_gpu
- test_suite: xgboost_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config_gpu.yaml
- cluster_compute: tpl_gpu_small.yaml
- run:
- timeout: 600
- script: python workloads/train_gpu.py
- wait_for_nodes:
- num_nodes: 5
- type: sdk_command
- file_manager: sdk
- alert: xgboost_tests
- - name: xgboost_distributed_api_test
- group: XGBoost
- working_dir: xgboost_tests
- legacy:
- test_name: distributed_api_test
- test_suite: xgboost_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_small.yaml
- run:
- timeout: 600
- script: python workloads/distributed_api_test.py
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- alert: xgboost_tests
- - name: xgboost_ft_small_elastic
- group: XGBoost
- working_dir: xgboost_tests
- legacy:
- test_name: ft_small_elastic
- test_suite: xgboost_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_small.yaml
- run:
- timeout: 900
- script: python workloads/ft_small_elastic.py
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- alert: xgboost_tests
- - name: xgboost_ft_small_non_elastic
- group: XGBoost
- working_dir: xgboost_tests
- legacy:
- test_name: ft_small_non_elastic
- test_suite: xgboost_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_small.yaml
- run:
- timeout: 900
- script: python workloads/ft_small_non_elastic.py
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- alert: xgboost_tests
- - name: xgboost_tune_small
- group: XGBoost
- working_dir: xgboost_tests
- legacy:
- test_name: tune_small
- test_suite: xgboost_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_small.yaml
- run:
- timeout: 600
- script: python workloads/tune_small.py
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- alert: xgboost_tests
- - name: xgboost_tune_32x4
- group: XGBoost
- working_dir: xgboost_tests
- legacy:
- test_name: tune_32x4
- test_suite: xgboost_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_moderate.yaml
- run:
- timeout: 900
- script: python workloads/tune_32x4.py
- wait_for_nodes:
- num_nodes: 32
- type: sdk_command
- file_manager: sdk
- alert: xgboost_tests
- - name: xgboost_tune_4x32
- group: XGBoost
- working_dir: xgboost_tests
- legacy:
- test_name: tune_4x32
- test_suite: xgboost_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_moderate.yaml
- run:
- timeout: 900
- script: python workloads/tune_4x32.py
- wait_for_nodes:
- num_nodes: 32
- type: sdk_command
- file_manager: sdk
- alert: xgboost_tests
- #######################
- # LightGBM tests
- #######################
- - name: lightgbm_train_small
- group: LightGBM tests
- working_dir: lightgbm_tests
- legacy:
- test_name: train_small
- test_suite: lightgbm_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_small.yaml
- run:
- timeout: 600
- script: python workloads/train_small.py
- wait_for_nodes:
- num_nodes: 4
- type: client
- alert: default
- - name: lightgbm_train_moderate
- group: LightGBM tests
- working_dir: lightgbm_tests
- legacy:
- test_name: train_moderate
- test_suite: lightgbm_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_moderate.yaml
- run:
- timeout: 600
- script: python workloads/train_moderate.py
- wait_for_nodes:
- num_nodes: 32
- type: sdk_command
- file_manager: job
- alert: default
- - name: lightgbm_distributed_api_test
- group: LightGBM tests
- working_dir: lightgbm_tests
- legacy:
- test_name: distributed_api_test
- test_suite: lightgbm_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_small.yaml
- run:
- timeout: 600
- script: python workloads/distributed_api_test.py
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: job
- alert: default
- - name: lightgbm_ft_small_non_elastic
- group: LightGBM tests
- working_dir: lightgbm_tests
- legacy:
- test_name: ft_small_non_elastic
- test_suite: lightgbm_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_small.yaml
- run:
- timeout: 900
- script: python workloads/ft_small_non_elastic.py
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: job
- alert: default
- - name: lightgbm_tune_small
- group: LightGBM tests
- working_dir: lightgbm_tests
- legacy:
- test_name: tune_small
- test_suite: lightgbm_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_small.yaml
- run:
- timeout: 600
- script: python workloads/tune_small.py
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: job
- alert: default
- - name: lightgbm_tune_16x4
- group: LightGBM tests
- working_dir: lightgbm_tests
- legacy:
- test_name: tune_16x4
- test_suite: lightgbm_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_moderate.yaml
- run:
- timeout: 900
- script: python workloads/tune_16x4.py
- wait_for_nodes:
- num_nodes: 32
- type: sdk_command
- file_manager: job
- alert: default
- - name: lightgbm_tune_4x16
- group: LightGBM tests
- working_dir: lightgbm_tests
- legacy:
- test_name: tune_4x16
- test_suite: lightgbm_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_moderate.yaml
- run:
- timeout: 900
- script: python workloads/tune_4x16.py
- wait_for_nodes:
- num_nodes: 32
- type: sdk_command
- file_manager: job
- alert: default
- #######################
- # ML user tests
- #######################
- - name: ml_user_horovod_user_test_latest
- group: ML user tests
- working_dir: ml_user_tests
- legacy:
- test_name: horovod_user_test_latest
- test_suite: ml_user_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: horovod/app_config.yaml
- cluster_compute: horovod/compute_tpl.yaml
- driver_setup: horovod/driver_setup_latest.sh
- run:
- timeout: 1200
- script: python horovod/horovod_user_test.py
- type: client
- alert: default
- - name: ml_user_horovod_user_test_master
- group: ML user tests
- working_dir: ml_user_tests
- legacy:
- test_name: horovod_user_test_master
- test_suite: ml_user_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: horovod/app_config_master.yaml
- cluster_compute: horovod/compute_tpl.yaml
- driver_setup: horovod/driver_setup_master.sh
- run:
- timeout: 1200
- script: python horovod/horovod_user_test.py
- type: client
- alert: default
- - name: ml_user_train_tensorflow_mnist_test
- group: ML user tests
- working_dir: ml_user_tests
- legacy:
- test_name: train_tensorflow_mnist_test
- test_suite: ml_user_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: train/app_config.yaml
- cluster_compute: train/compute_tpl.yaml
- driver_setup: train/driver_setup.sh
- run:
- timeout: 36000
- script: python train/train_tensorflow_mnist_test.py
- type: client
- alert: default
- - name: ml_user_train_torch_linear_test
- group: ML user tests
- working_dir: ml_user_tests
- legacy:
- test_name: train_torch_linear_test
- test_suite: ml_user_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: train/app_config.yaml
- cluster_compute: train/compute_tpl.yaml
- driver_setup: train/driver_setup.sh
- run:
- timeout: 36000
- script: python train/train_torch_linear_test.py
- type: client
- alert: default
- - name: ml_user_xgboost_gpu_connect_latest
- group: ML user tests
- working_dir: ml_user_tests
- legacy:
- test_name: xgboost_gpu_connect_latest
- test_suite: ml_user_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: xgboost/app_config_gpu.yaml
- cluster_compute: xgboost/tpl_gpu_small_scaling.yaml
- run:
- timeout: 1200
- script: python xgboost/train_gpu_connect.py
- type: client
- alert: default
- - name: ml_user_xgboost_gpu_connect_master
- group: ML user tests
- working_dir: ml_user_tests
- legacy:
- test_name: xgboost_gpu_connect_master
- test_suite: ml_user_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: xgboost/app_config_gpu_master.yaml
- cluster_compute: xgboost/tpl_gpu_small_scaling.yaml
- run:
- timeout: 1200
- script: python xgboost/train_gpu_connect.py
- type: client
- alert: default
- - name: ml_user_ray_lightning_user_test_latest
- group: ML user tests
- working_dir: ml_user_tests
- legacy:
- test_name: ray_lightning_user_test_latest
- test_suite: ml_user_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: ray-lightning/app_config.yaml
- cluster_compute: ray-lightning/compute_tpl.yaml
- driver_setup: ray-lightning/driver_setup.sh
- run:
- timeout: 1200
- script: python ray-lightning/ray_lightning_user_test.py
- type: client
- alert: default
- - name: ml_user_ray_lightning_user_test_master
- group: ML user tests
- working_dir: ml_user_tests
- legacy:
- test_name: ray_lightning_user_test_master
- test_suite: ml_user_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: ray-lightning/app_config_master.yaml
- cluster_compute: ray-lightning/compute_tpl.yaml
- driver_setup: ray-lightning/driver_setup.sh
- run:
- timeout: 1200
- script: python ray-lightning/ray_lightning_user_test.py
- type: client
- alert: default
- - name: ml_user_tune_rllib_connect_test
- group: ML user tests
- working_dir: ml_user_tests
- legacy:
- test_name: tune_rllib_connect_test
- test_suite: ml_user_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: ../rllib_tests/app_config.yaml
- cluster_compute: tune_rllib/compute_tpl.yaml
- driver_setup: tune_rllib/driver_setup.sh
- run:
- timeout: 2000
- script: python tune_rllib/run_connect_tests.py
- type: client
- alert: default
- #######################
- # Tune cloud tests
- #######################
- - name: tune_cloud_aws_no_sync_down
- group: Tune cloud tests
- working_dir: tune_tests/cloud_tests
- stable: false
- legacy:
- test_name: aws_no_sync_down
- test_suite: tune_cloud_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_aws_4x2.yaml
- run:
- timeout: 600
- script: python workloads/run_cloud_test.py no_sync_down
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- alert: tune_tests
- - name: tune_cloud_aws_ssh_sync
- group: Tune cloud tests
- working_dir: tune_tests/cloud_tests
- stable: false
- legacy:
- test_name: aws_ssh_sync
- test_suite: tune_cloud_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_aws_4x2.yaml
- run:
- timeout: 600
- script: python workloads/run_cloud_test.py ssh_sync
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- alert: tune_tests
- - name: tune_cloud_aws_durable_upload
- group: Tune cloud tests
- working_dir: tune_tests/cloud_tests
- stable: false
- legacy:
- test_name: aws_durable_upload
- test_suite: tune_cloud_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_aws_4x2.yaml
- run:
- timeout: 600
- script: python workloads/run_cloud_test.py durable_upload --bucket s3://tune-cloud-tests/durable_upload
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- alert: tune_tests
- - name: tune_cloud_aws_durable_upload_rllib_str
- group: Tune cloud tests
- working_dir: tune_tests/cloud_tests
- stable: false
- legacy:
- test_name: aws_durable_upload_rllib_str
- test_suite: tune_cloud_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config_ml.yaml
- cluster_compute: tpl_aws_4x2.yaml
- run:
- timeout: 600
- script: python workloads/run_cloud_test.py durable_upload --trainable rllib_str
- --bucket s3://tune-cloud-tests/durable_upload_rllib_str
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- alert: tune_tests
- - name: tune_cloud_aws_durable_upload_rllib_trainer
- group: Tune cloud tests
- working_dir: tune_tests/cloud_tests
- stable: false
- legacy:
- test_name: aws_durable_upload_rllib_trainer
- test_suite: tune_cloud_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config_ml.yaml
- cluster_compute: tpl_aws_4x2.yaml
- run:
- timeout: 600
- script: python workloads/run_cloud_test.py durable_upload --trainable rllib_trainer
- --bucket s3://tune-cloud-tests/durable_upload_rllib_trainer
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- alert: tune_tests
- - name: tune_cloud_gcp_k8s_no_sync_down
- group: Tune cloud tests
- working_dir: tune_tests/cloud_tests
- stable: false
- legacy:
- test_name: gcp_k8s_no_sync_down
- test_suite: tune_cloud_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_gcp_k8s_4x8.yaml
- cloud_id: cld_k8WcxPgjUtSE8RVmfZpTLuKM # anyscale_k8s_gcp_cloud
- run:
- timeout: 600
- script: python workloads/run_cloud_test.py no_sync_down --cpus-per-trial 8
- type: client
- wait_for_nodes:
- num_nodes: 4
- alert: tune_tests
- - name: tune_cloud_gcp_k8s_ssh_sync
- group: Tune cloud tests
- working_dir: tune_tests/cloud_tests
- stable: false
- legacy:
- test_name: gcp_k8s_ssh_sync
- test_suite: tune_cloud_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_gcp_k8s_4x8.yaml
- cloud_id: cld_k8WcxPgjUtSE8RVmfZpTLuKM # anyscale_k8s_gcp_cloud
- run:
- timeout: 600
- script: python workloads/run_cloud_test.py ssh_sync --cpus-per-trial 8
- type: client
- wait_for_nodes:
- num_nodes: 4
- alert: tune_tests
- - name: tune_cloud_gcp_k8s_durable_upload
- group: Tune cloud tests
- working_dir: tune_tests/cloud_tests
- stable: false
- legacy:
- test_name: gcp_k8s_durable_upload
- test_suite: tune_cloud_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_gcp_k8s_4x8.yaml
- cloud_id: cld_k8WcxPgjUtSE8RVmfZpTLuKM # anyscale_k8s_gcp_cloud
- run:
- timeout: 600
- script: python workloads/run_cloud_test.py durable_upload --cpus-per-trial 8 --bucket gs://tune-cloud-tests/durable_upload
- type: client
- wait_for_nodes:
- num_nodes: 4
- alert: tune_tests
- ########################
- # Tune scalability tests
- ########################
- - name: tune_scalability_bookkeeping_overhead
- group: Tune scalability tests
- working_dir: tune_tests/scalability_tests
- legacy:
- test_name: bookkeeping_overhead
- test_suite: tune_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_1x16.yaml
- run:
- timeout: 1200
- script: python workloads/test_bookkeeping_overhead.py
- type: sdk_command
- file_manager: sdk
- alert: tune_tests
- - name: tune_scalability_durable_trainable
- group: Tune scalability tests
- working_dir: tune_tests/scalability_tests
- legacy:
- test_name: durable_trainable
- test_suite: tune_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_16x2.yaml
- run:
- timeout: 900
- script: python workloads/test_durable_trainable.py --bucket tune-cloud-tests
- wait_for_nodes:
- num_nodes: 16
- type: sdk_command
- file_manager: sdk
- alert: tune_tests
- - name: tune_scalability_long_running_large_checkpoints
- group: Tune scalability tests
- working_dir: tune_tests/scalability_tests
- legacy:
- test_name: long_running_large_checkpoints
- test_suite: tune_tests
- frequency: weekly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_1x32_hd.yaml
- run:
- timeout: 86400
- script: python workloads/test_long_running_large_checkpoints.py
- long_running: true
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: nightly
- run:
- timeout: 3600
- alert: tune_tests
- - name: tune_scalability_network_overhead
- group: Tune scalability tests
- working_dir: tune_tests/scalability_tests
- legacy:
- test_name: network_overhead
- test_suite: tune_tests
- frequency: weekly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_100x2.yaml
- run:
- timeout: 900
- prepare_timeout: 1200
- script: python workloads/test_network_overhead.py
- wait_for_nodes:
- num_nodes: 100
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: nightly
- cluster:
- cluster_compute: tpl_20x2.yaml
- run:
- timeout: 400
- prepare_timeout: 600
- wait_for_nodes:
- num_nodes: 20
- alert: tune_tests
- - name: tune_scalability_result_throughput_cluster
- group: Tune scalability tests
- working_dir: tune_tests/scalability_tests
- legacy:
- test_name: result_throughput_cluster
- test_suite: tune_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_16x64.yaml
- run:
- timeout: 600
- script: python workloads/test_result_throughput_cluster.py
- wait_for_nodes:
- num_nodes: 16
- type: sdk_command
- file_manager: sdk
- alert: tune_tests
- - name: tune_scalability_result_throughput_single_node
- group: Tune scalability tests
- working_dir: tune_tests/scalability_tests
- legacy:
- test_name: result_throughput_single_node
- test_suite: tune_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_1x96.yaml
- run:
- timeout: 600
- script: python workloads/test_result_throughput_single_node.py
- type: sdk_command
- file_manager: sdk
- alert: tune_tests
- - name: tune_scalability_xgboost_sweep
- group: Tune scalability tests
- working_dir: tune_tests/scalability_tests
- legacy:
- test_name: xgboost_sweep
- test_suite: tune_tests
- frequency: weekly
- team: ml
- cluster:
- cluster_env: app_config_data.yaml
- cluster_compute: tpl_16x64.yaml
- run:
- timeout: 3600
- script: python workloads/test_xgboost_sweep.py
- wait_for_nodes:
- num_nodes: 16
- type: sdk_command
- file_manager: sdk
- alert: tune_tests
- ########################
- # Golden Notebook tests
- ########################
- - name: golden_notebook_dask_xgboost_test
- group: Golden Notebook tests
- working_dir: golden_notebook_tests
- legacy:
- test_name: dask_xgboost_test
- test_suite: golden_notebook_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: dask_xgboost_app_config.yaml
- cluster_compute: compute_tpl.yaml
- run:
- timeout: 1800
- script: python workloads/dask_xgboost_test.py --num-actors 4 --cpus-per-actor 4 --num-actors-inference 16 --cpus-per-actor-inference 1
- type: client
- wait_for_nodes:
- num_nodes: 4
- alert: default
- - name: golden_notebook_modin_xgboost_test
- group: Golden Notebook tests
- working_dir: golden_notebook_tests
- legacy:
- test_name: modin_xgboost_test
- test_suite: golden_notebook_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: modin_xgboost_app_config.yaml
- cluster_compute: compute_tpl.yaml
- run:
- timeout: 1800
- script: python workloads/modin_xgboost_test.py --num-actors 4 --cpus-per-actor 4 --num-actors-inference 16 --cpus-per-actor-inference 1
- type: client
- wait_for_nodes:
- num_nodes: 4
- alert: default
- - name: golden_notebook_torch_tune_serve_test
- group: Golden Notebook tests
- working_dir: golden_notebook_tests
- legacy:
- test_name: torch_tune_serve_test
- test_suite: golden_notebook_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: torch_tune_serve_app_config.yaml
- cluster_compute: gpu_tpl.yaml
- run:
- timeout: 1800
- script: python workloads/torch_tune_serve_test.py
- type: client
- alert: default
- #######################
- # Long running tests
- #######################
- - name: long_running_actor_deaths
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: actor_deaths
- test_suite: long_running_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_1.yaml
- run:
- timeout: 86400
- prepare: ray stop
- script: python workloads/actor_deaths.py
- long_running: true
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_apex
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: apex
- test_suite: long_running_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: ../rllib_tests/app_config.yaml
- cluster_compute: tpl_cpu_3.yaml
- run:
- timeout: 86400
- script: python workloads/apex.py
- long_running: true
- wait_for_nodes:
- num_nodes: 3
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_impala
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: impala
- test_suite: long_running_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config_np.yaml
- cluster_compute: tpl_cpu_1_large.yaml
- run:
- timeout: 86400
- script: python workloads/impala.py
- long_running: true
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_many_actor_tasks
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: many_actor_tasks
- test_suite: long_running_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_1.yaml
- run:
- timeout: 86400
- prepare: ray stop
- script: python workloads/many_actor_tasks.py
- long_running: true
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_many_drivers
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: many_drivers
- test_suite: long_running_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_1.yaml
- run:
- timeout: 86400
- prepare: ray stop
- script: python workloads/many_drivers.py --iteration-num=4000
- long_running: true
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_many_ppo
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: many_ppo
- test_suite: long_running_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: ../rllib_tests/app_config.yaml
- cluster_compute: many_ppo.yaml
- run:
- timeout: 86400
- script: python workloads/many_ppo.py
- long_running: true
- wait_for_nodes:
- num_nodes: 1
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_many_tasks
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: many_tasks
- test_suite: long_running_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_1.yaml
- run:
- timeout: 86400
- prepare: ray stop
- script: python workloads/many_tasks.py
- long_running: true
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_many_tasks_serialized_ids
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: many_tasks_serialized_ids
- test_suite: long_running_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_1.yaml
- run:
- timeout: 86400
- prepare: ray stop
- script: python workloads/many_tasks_serialized_ids.py
- long_running: true
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_node_failures
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: node_failures
- test_suite: long_running_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_1.yaml
- run:
- timeout: 86400
- prepare: ray stop
- script: python workloads/node_failures.py
- long_running: true
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_pbt
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: pbt
- test_suite: long_running_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: ../rllib_tests/app_config.yaml
- cluster_compute: tpl_cpu_1.yaml
- run:
- timeout: 86400
- prepare: ray stop
- script: python workloads/pbt.py
- long_running: true
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_serve
- group: Long running tests
- working_dir: long_running_tests
- legacy:
- test_name: serve
- test_suite: long_running_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_1.yaml
- run:
- timeout: 86400
- prepare: ray stop
- script: python workloads/serve.py
- long_running: true
- type: sdk_command
- file_manager: job
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- - name: long_running_serve_failure
- group: Long running tests
- working_dir: long_running_tests
- stable: false
- legacy:
- test_name: serve_failure
- test_suite: long_running_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_cpu_1.yaml
- run:
- timeout: 86400
- prepare: ray stop
- script: python workloads/serve_failure.py
- long_running: true
- type: sdk_command
- file_manager: job
- smoke_test:
- frequency: disabled
- run:
- timeout: 600
- alert: long_running_tests
- - name: long_running_distributed_pytorch_pbt_failure
- group: Long running tests
- working_dir: long_running_distributed_tests
- legacy:
- test_name: pytorch_pbt_failure
- test_suite: long_running_distributed
- frequency: weekly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl.yaml
- run:
- timeout: 86400
- script: python workloads/pytorch_pbt_failure.py
- long_running: true
- type: sdk_command
- file_manager: job
- smoke_test:
- frequency: disabled
- run:
- timeout: 3600
- alert: long_running_tests
- ########################
- # Jobs tests
- ########################
- - name: jobs_basic_local_working_dir
- group: Jobs tests
- working_dir: jobs_tests
- legacy:
- test_name: jobs_basic_local_working_dir
- test_suite: jobs_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_4_xlarge.yaml
- run:
- timeout: 600
- script: python workloads/jobs_basic.py --working-dir "workloads"
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: job
- alert: default
- - name: jobs_basic_remote_working_dir
- group: Jobs tests
- working_dir: jobs_tests
- legacy:
- test_name: jobs_basic_remote_working_dir
- test_suite: jobs_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_4_xlarge.yaml
- run:
- timeout: 600
- script: python workloads/jobs_basic.py --working-dir "https://github.com/anyscale/job-services-cuj-examples/archive/refs/heads/main.zip"
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: job
- alert: default
- ########################
- # Runtime env tests
- ########################
- - name: runtime_env_rte_many_tasks_actors
- group: Runtime env tests
- working_dir: runtime_env_tests
- legacy:
- test_name: rte_many_tasks_actors
- test_suite: runtime_env_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: rte_small.yaml
- run:
- timeout: 600
- script: python workloads/rte_many_tasks_actors.py
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: job
- alert: default
- - name: runtime_env_wheel_urls
- group: Runtime env tests
- working_dir: runtime_env_tests
- legacy:
- test_name: wheel_urls
- test_suite: runtime_env_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: rte_minimal.yaml
- run:
- timeout: 9000
- script: python workloads/wheel_urls.py
- wait_for_nodes:
- num_nodes: 1
- type: sdk_command
- file_manager: job
- alert: default
- - name: runtime_env_rte_ray_client
- group: Runtime env tests
- working_dir: runtime_env_tests
- legacy:
- test_name: rte_ray_client
- test_suite: runtime_env_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: rte_minimal.yaml
- run:
- timeout: 600
- script: python workloads/rte_ray_client.py
- wait_for_nodes:
- num_nodes: 1
- type: client
- alert: default
- ########################
- # Serve tests
- ########################
- - name: serve_single_deployment_1k_noop_replica
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: single_deployment_1k_noop_replica
- test_suite: serve_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_32_cpu.yaml
- run:
- timeout: 7200
- long_running: false
- script: python workloads/single_deployment_1k_noop_replica.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: serve_multi_deployment_1k_noop_replica
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: multi_deployment_1k_noop_replica
- test_suite: serve_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_32_cpu.yaml
- run:
- timeout: 7200
- long_running: false
- script: python workloads/multi_deployment_1k_noop_replica.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: serve_autoscaling_single_deployment
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: autoscaling_single_deployment
- test_suite: serve_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_8_cpu_autoscaling.yaml
- run:
- timeout: 7200
- long_running: false
- script: python workloads/autoscaling_single_deployment.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: serve_autoscaling_multi_deployment
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: autoscaling_multi_deployment
- test_suite: serve_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_8_cpu_autoscaling.yaml
- run:
- timeout: 7200
- long_running: false
- script: python workloads/autoscaling_multi_deployment.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: serve_serve_micro_benchmark
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: serve_micro_benchmark
- test_suite: serve_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_single_node.yaml
- run:
- timeout: 7200
- long_running: false
- script: python workloads/serve_micro_benchmark.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: serve_serve_micro_benchmark_k8s
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: serve_micro_benchmark_k8s
- test_suite: serve_tests
- # TODO(architkulkarni) Reenable after K8s migration. Currently failing
- frequency: disabled
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_single_node_k8s.yaml
- run:
- timeout: 7200
- long_running: false
- script: python workloads/serve_micro_benchmark.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: serve_serve_cluster_fault_tolerance
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: serve_cluster_fault_tolerance
- test_suite: serve_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_single_node.yaml
- run:
- timeout: 7200
- long_running: false
- script: python workloads/serve_cluster_fault_tolerance.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: deployment_graph_long_chain
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: deployment_graph_long_chain
- test_suite: serve_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_single_node_32_cpu.yaml
- run:
- timeout: 3600
- long_running: false
- script: python workloads/deployment_graph_long_chain.py --chain-length=10 --num-clients=4 --local-test=False
- type: sdk_command
- file_manager: job
- alert: default
- stable: False
- - name: deployment_graph_wide_ensemble
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: deployment_graph_wide_ensemble
- test_suite: serve_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_single_node_32_cpu.yaml
- run:
- timeout: 3600
- long_running: false
- script: python workloads/deployment_graph_wide_ensemble.py --fanout-degree=10 --num-clients=4 --local-test=False
- type: sdk_command
- file_manager: job
- alert: default
- stable: False
- - name: serve_handle_long_chain
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: serve_handle_long_chain
- test_suite: serve_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_single_node_32_cpu.yaml
- run:
- timeout: 3600
- long_running: false
- script: python workloads/serve_handle_long_chain.py --chain-length=10 --num-clients=4 --local-test=False
- type: sdk_command
- file_manager: job
- alert: default
- stable: False
- - name: serve_handle_wide_ensemble
- group: Serve tests
- working_dir: serve_tests
- legacy:
- test_name: serve_handle_wide_ensemble
- test_suite: serve_tests
- frequency: nightly
- team: serve
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl_single_node_32_cpu.yaml
- run:
- timeout: 3600
- long_running: false
- script: python workloads/serve_handle_wide_ensemble.py --fanout-degree=10 --num-clients=4 --local-test=False
- type: sdk_command
- file_manager: job
- alert: default
- stable: False
- ########################
- # Train tests
- ########################
- - name: train_horovod_multi_node_test
- group: Train tests
- working_dir: train_tests/horovod
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: compute_tpl.yaml
- run:
- timeout: 3000
- script: python train_horovod_multi_node_test.py
- wait_for_nodes:
- num_nodes: 2
- type: sdk_command
- file_manager: job
- alert: default
- ########################
- # RLlib tests
- ########################
- - name: rllib_learning_tests
- group: RLlib tests
- working_dir: rllib_tests
- legacy:
- test_name: learning_tests
- test_suite: rllib_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: 8gpus_64cpus.yaml
- run:
- timeout: 14400
- script: python learning_tests/run.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: rllib_multi_gpu_learning_tests
- group: RLlib tests
- working_dir: rllib_tests
- legacy:
- test_name: multi_gpu_learning_tests
- test_suite: rllib_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: 8gpus_96cpus.yaml
- run:
- timeout: 7200
- script: python multi_gpu_learning_tests/run.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: rllib_multi_gpu_with_lstm_learning_tests
- group: RLlib tests
- working_dir: rllib_tests
- legacy:
- test_name: multi_gpu_with_lstm_learning_tests
- test_suite: rllib_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: 8gpus_96cpus.yaml
- run:
- timeout: 7200
- script: python multi_gpu_with_lstm_learning_tests/run.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: rllib_multi_gpu_with_attention_learning_tests
- group: RLlib tests
- working_dir: rllib_tests
- legacy:
- test_name: multi_gpu_with_attention_learning_tests
- test_suite: rllib_tests
- frequency: nightly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: 8gpus_96cpus.yaml
- run:
- timeout: 7200
- script: python multi_gpu_with_attention_learning_tests/run.py
- type: sdk_command
- file_manager: job
- alert: default
- - name: rllib_stress_tests
- group: RLlib tests
- working_dir: rllib_tests
- legacy:
- test_name: stress_tests
- test_suite: rllib_tests
- frequency: weekly
- team: ml
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: 4gpus_544_cpus.yaml
- run:
- timeout: 5400
- script: python stress_tests/run_stress_tests.py
- wait_for_nodes:
- num_nodes: 6
- type: sdk_command
- file_manager: job
- smoke_test:
- frequency: nightly
- run:
- timeout: 2000
- alert: default
- ########################
- # Core Nightly Tests
- ########################
- - name: shuffle_10gb
- group: core-multi-test
- team: core
- frequency: multi
- working_dir: nightly_tests
- legacy:
- test_name: shuffle_10gb
- test_suite: nightly_tests
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_single.yaml
- run:
- timeout: 3000
- script: python shuffle/shuffle_test.py --num-partitions=50 --partition-size=200e6
- type: sdk_command
- file_manager: sdk
- - name: shuffle_50gb
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: shuffle_50gb
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_single.yaml
- run:
- timeout: 3000
- script: python shuffle/shuffle_test.py --num-partitions=50 --partition-size=1e9
- type: sdk_command
- file_manager: sdk
- - name: shuffle_50gb_large_partition
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: shuffle_50gb_large_partition
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_single.yaml
- run:
- timeout: 3000
- script: python shuffle/shuffle_test.py --num-partitions=500 --partition-size=100e6
- type: sdk_command
- file_manager: sdk
- - name: shuffle_100gb
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: shuffle_100gb
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_multi.yaml
- run:
- timeout: 3000
- script: python shuffle/shuffle_test.py --num-partitions=200 --partition-size=500e6
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- - name: non_streaming_shuffle_100gb
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: non_streaming_shuffle_100gb
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_multi.yaml
- run:
- timeout: 3000
- script: python shuffle/shuffle_test.py --num-partitions=200 --partition-size=500e6
- --no-streaming
- wait_for_nodes:
- num_nodes: 4
- type: sdk_command
- file_manager: sdk
- - name: non_streaming_shuffle_50gb_large_partition
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: non_streaming_shuffle_50gb_large_partition
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_single.yaml
- run:
- timeout: 3000
- script: python shuffle/shuffle_test.py --num-partitions=500 --partition-size=100e6
- --no-streaming
- type: sdk_command
- file_manager: sdk
- - name: non_streaming_shuffle_50gb
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: non_streaming_shuffle_50gb
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_single.yaml
- run:
- timeout: 3000
- script: python shuffle/shuffle_test.py --num-partitions=50 --partition-size=1e9
- --no-streaming
- type: sdk_command
- file_manager: sdk
- - name: stress_test_placement_group
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: stress_test_placement_group
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: stress_tests/stress_tests_app_config.yaml
- cluster_compute: stress_tests/placement_group_tests_compute.yaml
- run:
- timeout: 7200
- script: python stress_tests/test_placement_group.py
- type: sdk_command
- file_manager: sdk
- - name: shuffle_1tb_1000_partition
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: shuffle_1tb_1000_partition
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_large_scale.yaml
- run:
- timeout: 3000
- script: python shuffle/shuffle_test.py --num-partitions=1000 --partition-size=1e9
- wait_for_nodes:
- num_nodes: 20
- type: sdk_command
- file_manager: sdk
- - name: non_streaming_shuffle_1tb_1000_partition
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: non_streaming_shuffle_1tb_1000_partition
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_large_scale.yaml
- run:
- timeout: 3000
- script: python shuffle/shuffle_test.py --num-partitions=1000 --partition-size=1e9
- --no-streaming
- wait_for_nodes:
- num_nodes: 20
- type: sdk_command
- file_manager: sdk
- - name: shuffle_1tb_5000_partitions
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: shuffle_1tb_5000_partitions
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_large_scale.yaml
- run:
- timeout: 9000
- script: python shuffle/shuffle_test.py --num-partitions=5000 --partition-size=200e6
- wait_for_nodes:
- num_nodes: 20
- type: sdk_command
- file_manager: sdk
- - name: decision_tree_autoscaling
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: decision_tree_autoscaling
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: decision_tree/decision_tree_app_config.yaml
- cluster_compute: decision_tree/autoscaling_compute.yaml
- run:
- timeout: 3000
- script: python decision_tree/cart_with_tree.py
- type: sdk_command
- file_manager: sdk
- - name: decision_tree_autoscaling_20_runs
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: decision_tree_autoscaling_20_runs
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: decision_tree/decision_tree_app_config.yaml
- cluster_compute: decision_tree/autoscaling_compute.yaml
- run:
- timeout: 9600
- script: python decision_tree/cart_with_tree.py --concurrency=20
- type: sdk_command
- file_manager: sdk
- - name: autoscaling_shuffle_1tb_1000_partitions
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: autoscaling_shuffle_1tb_1000_partitions
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_autoscaling.yaml
- run:
- timeout: 4000
- script: python shuffle/shuffle_test.py --num-partitions=1000 --partition-size=1e9
- --no-streaming
- type: sdk_command
- file_manager: sdk
- - name: pg_long_running_performance_test
- group: core-multi-test
- working_dir: nightly_tests
- legacy:
- test_name: pg_long_running_performance_test
- test_suite: nightly_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: placement_group_tests/app_config.yaml
- cluster_compute: placement_group_tests/long_running_test_compute.yaml
- run:
- timeout: 3600
- script: python placement_group_tests/long_running_performance_test.py --num-stages
- 2000
- wait_for_nodes:
- num_nodes: 2
- type: sdk_command
- file_manager: sdk
- - name: microbenchmark
- group: core-daily-test
- team: core
- frequency: nightly
- working_dir: microbenchmark
- python: "3.7"
- legacy:
- test_name: microbenchmark
- test_suite: microbenchmark
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_64.yaml
- run:
- timeout: 1800
- script: OMP_NUM_THREADS=64 RAY_ADDRESS= python run_microbenchmark.py
- - name: microbenchmark_38
- group: core-daily-test
- team: core
- frequency: nightly
- working_dir: microbenchmark
- python: "3.8"
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: tpl_64.yaml
- run:
- timeout: 1800
- script: OMP_NUM_THREADS=64 RAY_ADDRESS= python run_microbenchmark.py
- - name: dask_on_ray_10gb_sort
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: dask_on_ray_10gb_sort
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: dask_on_ray/dask_on_ray_app_config.yaml
- cluster_compute: dask_on_ray/dask_on_ray_sort_compute_template.yaml
- run:
- timeout: 7200
- script: python dask_on_ray/dask_on_ray_sort.py --nbytes 10_000_000_000 --npartitions
- 50 --num-nodes 1 --ray --data-dir /tmp/ray --file-path /tmp/ray
- type: sdk_command
- file_manager: sdk
- - name: dask_on_ray_100gb_sort
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: dask_on_ray_100gb_sort
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: dask_on_ray/dask_on_ray_app_config.yaml
- cluster_compute: dask_on_ray/dask_on_ray_sort_compute_template.yaml
- run:
- timeout: 7200
- script: python dask_on_ray/dask_on_ray_sort.py --nbytes 100_000_000_000 --npartitions
- 200 --num-nodes 1 --ray --data-dir /tmp/ray --file-path /tmp/ray
- type: sdk_command
- file_manager: sdk
- - name: dask_on_ray_large_scale_test_no_spilling
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: dask_on_ray_large_scale_test_no_spilling
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: dask_on_ray/large_scale_dask_on_ray_app_config.yaml
- cluster_compute: dask_on_ray/dask_on_ray_stress_compute.yaml
- run:
- timeout: 7200
- script: python dask_on_ray/large_scale_test.py --num_workers 20 --worker_obj_store_size_in_gb
- 20 --error_rate 0 --data_save_path /tmp/ray
- wait_for_nodes:
- num_nodes: 21
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: multi
- cluster:
- app_config: dask_on_ray/large_scale_dask_on_ray_app_config.yaml
- cluster_compute: dask_on_ray/large_scale_dask_on_ray_compute_template.yaml
- run:
- timeout: 7200
- script: python dask_on_ray/large_scale_test.py --num_workers 4 --worker_obj_store_size_in_gb
- 20 --error_rate 0 --data_save_path /tmp/ray
- wait_for_nodes:
- num_nodes: 5
- - name: dask_on_ray_large_scale_test_spilling
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: dask_on_ray_large_scale_test_spilling
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: dask_on_ray/large_scale_dask_on_ray_app_config.yaml
- cluster_compute: dask_on_ray/dask_on_ray_stress_compute.yaml
- run:
- timeout: 7200
- script: python dask_on_ray/large_scale_test.py --num_workers 150 --worker_obj_store_size_in_gb
- 70 --error_rate 0 --data_save_path /tmp/ray
- wait_for_nodes:
- num_nodes: 21
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: multi
- cluster:
- app_config: dask_on_ray/large_scale_dask_on_ray_app_config.yaml
- cluster_compute: dask_on_ray/large_scale_dask_on_ray_compute_template.yaml
- run:
- timeout: 7200
- script: python dask_on_ray/large_scale_test.py --num_workers 32 --worker_obj_store_size_in_gb
- 70 --error_rate 0 --data_save_path /tmp/ray
- wait_for_nodes:
- num_nodes: 5
- - name: stress_test_many_tasks
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: stress_test_many_tasks
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: stress_tests/stress_tests_app_config.yaml
- cluster_compute: stress_tests/stress_tests_compute.yaml
- run:
- timeout: 7200
- script: python stress_tests/test_many_tasks.py
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: multi
- cluster:
- app_config: stress_tests/stress_tests_app_config.yaml
- cluster_compute: stress_tests/smoke_test_compute.yaml
- run:
- timeout: 3600
- script: python stress_tests/test_many_tasks.py --num-nodes=4 --smoke-test
- - name: stress_test_dead_actors
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: stress_test_dead_actors
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: stress_tests/stress_tests_app_config.yaml
- cluster_compute: stress_tests/stress_tests_compute.yaml
- run:
- timeout: 7200
- script: python stress_tests/test_dead_actors.py
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: multi
- cluster:
- app_config: stress_tests/stress_tests_app_config.yaml
- cluster_compute: stress_tests/smoke_test_compute.yaml
- run:
- timeout: 3600
- script: python stress_tests/test_dead_actors.py --num-nodes=4 --num-parents=3
- --num-children=3
- # The full test is not stable, so run the smoke test only.
- # See https://github.com/ray-project/ray/issues/23244.
- - name: threaded_actors_stress_test
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: threaded_actors_stress_test
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: stress_tests/stress_tests_app_config.yaml
- cluster_compute: stress_tests/smoke_test_compute.yaml
- run:
- timeout: 3600
- script: python stress_tests/test_threaded_actors.py --test-runtime 1800 --kill-interval_s
- 30
- wait_for_nodes:
- num_nodes: 5
- # - name: threaded_actors_stress_test
- # group: core-daily-test
- # working_dir: nightly_tests
- # legacy:
- # test_name: threaded_actors_stress_test
- # test_suite: nightly_tests
- #
- # frequency: nightly
- # team: core
- # cluster:
- # cluster_env: stress_tests/stress_tests_app_config.yaml
- # cluster_compute: stress_tests/stress_test_threaded_actor_compute.yaml
- #
- # run:
- # timeout: 7200
- # script: python stress_tests/test_threaded_actors.py --test-runtime 3600 --kill-interval_s
- # 60
- #
- # wait_for_nodes:
- # num_nodes: 201
- # timeout: 600
- #
- # type: sdk_command
- # file_manager: sdk
- #
- # smoke_test:
- # frequency: nightly
- # cluster:
- # app_config: stress_tests/stress_tests_app_config.yaml
- # cluster_compute: stress_tests/smoke_test_compute.yaml
- #
- # run:
- # timeout: 3600
- # script: python stress_tests/test_threaded_actors.py --test-runtime 1800 --kill-interval_s
- # 30
- #
- # wait_for_nodes:
- # num_nodes: 5
- # timeout: 600
- - name: dask_on_ray_1tb_sort
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: dask_on_ray_1tb_sort
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: dask_on_ray/dask_on_ray_app_config.yaml
- cluster_compute: dask_on_ray/1tb_sort_compute.yaml
- run:
- timeout: 7200
- script: python dask_on_ray/dask_on_ray_sort.py --nbytes 1_000_000_000_000 --npartitions
- 1000 --num-nodes 31 --ray --data-dir /tmp/ray --s3-bucket core-nightly-test
- wait_for_nodes:
- num_nodes: 32
- type: sdk_command
- file_manager: sdk
- - name: many_nodes_actor_test
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: many_nodes_actor_test
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: many_nodes_tests/app_config.yaml
- cluster_compute: many_nodes_tests/compute_config.yaml
- run:
- timeout: 7200
- script: python many_nodes_tests/actor_test.py
- wait_for_nodes:
- num_nodes: 251
- type: sdk_command
- file_manager: sdk
- - name: pg_autoscaling_regression_test
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: pg_autoscaling_regression_test
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: placement_group_tests/app_config.yaml
- cluster_compute: placement_group_tests/compute.yaml
- run:
- timeout: 1200
- script: python placement_group_tests/pg_run.py
- type: sdk_command
- file_manager: sdk
- - name: placement_group_performance_test
- group: core-daily-test
- working_dir: nightly_tests
- legacy:
- test_name: placement_group_performance_test
- test_suite: nightly_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: placement_group_tests/app_config.yaml
- cluster_compute: placement_group_tests/pg_perf_test_compute.yaml
- run:
- timeout: 1200
- script: python placement_group_tests/placement_group_performance_test.py
- wait_for_nodes:
- num_nodes: 5
- type: sdk_command
- file_manager: sdk
- #########################
- # Core Scalability Tests
- #########################
- - name: single_node
- group: core-scalability-test
- working_dir: benchmarks
- legacy:
- test_name: single_node
- test_suite: benchmark_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: single_node.yaml
- run:
- timeout: 12000
- prepare: sleep 0
- script: python single_node/test_single_node.py
- type: sdk_command
- file_manager: sdk
- - name: object_store
- group: core-scalability-test
- working_dir: benchmarks
- legacy:
- test_name: object_store
- test_suite: benchmark_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: object_store.yaml
- run:
- timeout: 3600
- script: python object_store/test_object_store.py
- wait_for_nodes:
- num_nodes: 50
- type: sdk_command
- file_manager: sdk
- - name: many_actors
- group: core-scalability-test
- working_dir: benchmarks
- legacy:
- test_name: many_actors
- test_suite: benchmark_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: distributed.yaml
- run:
- timeout: 3600
- script: python distributed/test_many_actors.py
- wait_for_nodes:
- num_nodes: 65
- type: sdk_command
- file_manager: sdk
- - name: many_actors_smoke_test
- group: core-scalability-test
- working_dir: benchmarks
- legacy:
- test_name: many_actors_smoke_test
- test_suite: benchmark_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: distributed_smoke_test.yaml
- run:
- timeout: 3600
- script: SMOKE_TEST=1 python distributed/test_many_actors.py
- wait_for_nodes:
- num_nodes: 2
- type: sdk_command
- file_manager: sdk
- - name: many_tasks
- group: core-scalability-test
- working_dir: benchmarks
- legacy:
- test_name: many_tasks
- test_suite: benchmark_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: distributed.yaml
- run:
- timeout: 3600
- script: python distributed/test_many_tasks.py --num-tasks=10000
- wait_for_nodes:
- num_nodes: 65
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: multi
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: distributed_smoke_test.yaml
- run:
- timeout: 3600
- script: python distributed/test_many_tasks.py --num-tasks=100
- wait_for_nodes:
- num_nodes: 2
- - name: many_pgs
- group: core-scalability-test
- working_dir: benchmarks
- legacy:
- test_name: many_pgs
- test_suite: benchmark_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: distributed.yaml
- run:
- timeout: 3600
- script: python distributed/test_many_pgs.py
- wait_for_nodes:
- num_nodes: 65
- type: sdk_command
- file_manager: sdk
- - name: many_pgs_smoke_test
- group: core-scalability-test
- working_dir: benchmarks
- legacy:
- test_name: many_pgs_smoke_test
- test_suite: benchmark_tests
- frequency: multi
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: distributed_smoke_test.yaml
- run:
- timeout: 3600
- script: SMOKE_TEST=1 python distributed/test_many_pgs.py
- wait_for_nodes:
- num_nodes: 2
- type: sdk_command
- file_manager: sdk
- - name: many_nodes
- group: core-scalability-test
- working_dir: benchmarks
- legacy:
- test_name: many_nodes
- test_suite: benchmark_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: many_nodes.yaml
- run:
- timeout: 3600
- script: python distributed/test_many_tasks.py --num-tasks=1000
- wait_for_nodes:
- num_nodes: 250
- type: sdk_command
- file_manager: sdk
- - name: scheduling_test_many_0s_tasks_single_node
- group: core-scalability-test
- working_dir: benchmarks
- legacy:
- test_name: scheduling_test_many_0s_tasks_single_node
- test_suite: benchmark_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: scheduling.yaml
- run:
- timeout: 3600
- script: python distributed/test_scheduling.py --total-num-task=1984000 --num-cpu-per-task=1
- --task-duration-s=0 --total-num-actors=1 --num-actors-per-nodes=1
- wait_for_nodes:
- num_nodes: 32
- type: sdk_command
- file_manager: sdk
- - name: scheduling_test_many_0s_tasks_many_nodes
- group: core-scalability-test
- working_dir: benchmarks
- legacy:
- test_name: scheduling_test_many_0s_tasks_many_nodes
- test_suite: benchmark_tests
- frequency: nightly
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: scheduling.yaml
- run:
- timeout: 3600
- script: python distributed/test_scheduling.py --total-num-task=1984000 --num-cpu-per-task=1
- --task-duration-s=0 --total-num-actors=32 --num-actors-per-nodes=1
- wait_for_nodes:
- num_nodes: 32
- type: sdk_command
- file_manager: sdk
- # - name: scheduling_test_many_5s_tasks_single_node
- # group: core-scalability-test
- # working_dir: benchmarks
- # legacy:
- # test_name: scheduling_test_many_5s_tasks_single_node
- # test_suite: benchmark_tests
- # frequency: nightly
- # team: core
- # cluster:
- # cluster_env: app_config.yaml
- # cluster_compute: scheduling.yaml
- # run:
- # timeout: 3600
- # script: python distributed/test_scheduling.py --total-num-task=1984000 --num-cpu-per-task=1
- # --task-duration-s=5 --total-num-actors=1 --num-actors-per-nodes=1
- # wait_for_nodes:
- # num_nodes: 32
- # timeout: 600
- # type: sdk_command
- # file_manager: sdk
- # stable: false
- # - name: scheduling_test_many_5s_tasks_many_nodes
- # group: core-scalability-test
- # working_dir: benchmarks
- # legacy:
- # test_name: scheduling_test_many_5s_tasks_many_nodes
- # test_suite: benchmark_tests
- # frequency: nightly
- # team: core
- # cluster:
- # cluster_env: app_config.yaml
- # cluster_compute: scheduling.yaml
- # run:
- # timeout: 3600
- # script: python distributed/test_scheduling.py --total-num-task=1984000 --num-cpu-per-task=1
- # --task-duration-s=5 --total-num-actors=32 --num-actors-per-nodes=1
- # wait_for_nodes:
- # num_nodes: 32
- # timeout: 600
- # type: sdk_command
- # file_manager: sdk
- # stable: false
- ###############
- # Dataset tests
- ###############
- - name: inference
- group: core-dataset-tests
- working_dir: nightly_tests/dataset
- legacy:
- test_name: inference
- test_suite: dataset_test
- frequency: multi
- team: core
- cluster:
- cluster_env: app_config.yaml
- cluster_compute: inference.yaml
- run:
- timeout: 600
- script: python inference.py
- wait_for_nodes:
- num_nodes: 2
- type: sdk_command
- file_manager: sdk
- - name: shuffle_data_loader
- group: core-dataset-tests
- working_dir: nightly_tests/dataset
- legacy:
- test_name: shuffle_data_loader
- test_suite: dataset_test
- frequency: multi
- team: core
- cluster:
- cluster_env: shuffle_app_config.yaml
- cluster_compute: shuffle_compute.yaml
- run:
- timeout: 1800
- script: python dataset_shuffle_data_loader.py
- type: sdk_command
- file_manager: sdk
- - name: parquet_metadata_resolution
- group: core-dataset-tests
- working_dir: nightly_tests/dataset
- legacy:
- test_name: parquet_metadata_resolution
- test_suite: dataset_test
- frequency: multi
- team: core
- cluster:
- cluster_env: pipelined_training_app.yaml
- cluster_compute: pipelined_training_compute.yaml
- run:
- timeout: 1200
- script: python parquet_metadata_resolution.py --num-files 915
- wait_for_nodes:
- num_nodes: 15
- type: sdk_command
- file_manager: sdk
- - name: dataset_random_access
- group: core-dataset-tests
- working_dir: nightly_tests/dataset
- stable: false
- frequency: multi
- team: core
- cluster:
- cluster_env: pipelined_training_app.yaml
- cluster_compute: pipelined_training_compute.yaml
- run:
- timeout: 1200
- script: python dataset_random_access.py
- wait_for_nodes:
- num_nodes: 15
- type: sdk_command
- file_manager: sdk
- - name: pipelined_training_50_gb
- group: core-dataset-tests
- working_dir: nightly_tests/dataset
- legacy:
- test_name: pipelined_training_50_gb
- test_suite: dataset_test
- frequency: multi
- team: core
- cluster:
- cluster_env: pipelined_training_app.yaml
- cluster_compute: pipelined_training_compute.yaml
- run:
- timeout: 4800
- script: python pipelined_training.py --epochs 1
- wait_for_nodes:
- num_nodes: 15
- type: sdk_command
- file_manager: sdk
- - name: pipelined_ingestion_1500_gb
- group: core-dataset-tests
- working_dir: nightly_tests/dataset
- legacy:
- test_name: pipelined_ingestion_1500_gb
- test_suite: dataset_test
- frequency: nightly
- team: core
- cluster:
- cluster_env: pipelined_ingestion_app.yaml
- cluster_compute: pipelined_ingestion_compute.yaml
- run:
- timeout: 9600
- script: python pipelined_training.py --epochs 2 --num-windows 2 --num-files 915
- --debug
- wait_for_nodes:
- num_nodes: 21
- type: sdk_command
- file_manager: sdk
- - name: datasets_ingest_train_infer
- group: core-dataset-tests
- working_dir: nightly_tests/dataset
- legacy:
- test_name: datasets_ingest_train_infer
- test_suite: dataset_test
- frequency: multi
- team: core
- cluster:
- cluster_env: ray_sgd_training_app.yaml
- cluster_compute: ray_sgd_training_compute.yaml
- run:
- timeout: 14400
- script: python ray_sgd_training.py --address auto --use-s3 --num-workers 16 --use-gpu
- --large-dataset
- wait_for_nodes:
- num_nodes: 66
- type: sdk_command
- file_manager: sdk
- smoke_test:
- frequency: multi
- cluster:
- app_config: ray_sgd_training_app.yaml
- cluster_compute: ray_sgd_training_smoke_compute.yaml
- run:
- timeout: 3600
- script: python ray_sgd_training.py --address auto --use-s3 --num-workers 8 --use-gpu
- wait_for_nodes:
- num_nodes: 8
- - name: datasets_preprocess_ingest
- group: core-dataset-tests
- working_dir: nightly_tests/dataset
- legacy:
- test_name: datasets_preprocess_ingest
- test_suite: dataset_test
- frequency: nightly
- team: core
- cluster:
- cluster_env: ray_sgd_training_app.yaml
- cluster_compute: ray_sgd_training_compute_no_gpu.yaml
- run:
- timeout: 7200
- script: python ray_sgd_training.py --address auto --use-s3 --num-workers 16 --use-gpu
- --large-dataset --debug
- wait_for_nodes:
- num_nodes: 21
- type: sdk_command
- file_manager: sdk
- - name: datasets_ingest_400G
- group: core-dataset-tests
- working_dir: nightly_tests/dataset
- legacy:
- test_name: datasets_ingest_400G
- test_suite: dataset_test
- frequency: multi
- team: core
- cluster:
- cluster_env: ray_sgd_training_app.yaml
- cluster_compute: dataset_ingest_400G_compute.yaml
- run:
- timeout: 7200
- script: python ray_sgd_runner.py --address auto --use-gpu --num-epochs 1
- type: sdk_command
- file_manager: sdk
- - name: dataset_shuffle_random_shuffle_1tb
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: dataset_shuffle_random_shuffle_1tb
- test_suite: dataset_test
- stable: true
- frequency: nightly
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_large_scale.yaml
- run:
- timeout: 7200
- script: python dataset/sort.py --num-partitions=1000 --partition-size=1e9 --shuffle
- wait_for_nodes:
- num_nodes: 20
- type: sdk_command
- file_manager: sdk
- - name: dataset_shuffle_sort_1tb
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: dataset_shuffle_sort_1tb
- test_suite: dataset_test
- stable: true
- frequency: nightly
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/shuffle_compute_large_scale.yaml
- run:
- timeout: 7200
- script: python dataset/sort.py --num-partitions=1000 --partition-size=1e9
- wait_for_nodes:
- num_nodes: 20
- type: sdk_command
- file_manager: sdk
- - name: dataset_shuffle_random_shuffle_1tb_small_instances
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: dataset_shuffle_random_shuffle_1tb_small_instances
- test_suite: dataset_test
- frequency: nightly
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml
- run:
- timeout: 7200
- script: python dataset/sort.py --num-partitions=1000 --partition-size=1e9 --shuffle
- wait_for_nodes:
- num_nodes: 20
- type: sdk_command
- file_manager: sdk
- - name: dataset_shuffle_sort_1tb_small_instances
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: dataset_shuffle_sort_1tb_small_instances
- test_suite: dataset_test
- frequency: nightly
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml
- run:
- timeout: 7200
- script: python dataset/sort.py --num-partitions=1000 --partition-size=1e9
- wait_for_nodes:
- num_nodes: 20
- type: sdk_command
- file_manager: sdk
- - name: dataset_shuffle_push_based_random_shuffle_1tb
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: dataset_shuffle_push_based_random_shuffle_1tb
- test_suite: dataset_test
- stable: false
- frequency: nightly
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml
- run:
- timeout: 7200
- script: RAY_DATASET_PUSH_BASED_SHUFFLE=1 python dataset/sort.py --num-partitions=1000 --partition-size=1e9 --shuffle
- wait_for_nodes:
- num_nodes: 20
- type: sdk_command
- file_manager: sdk
- - name: dataset_shuffle_push_based_sort_1tb
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: dataset_shuffle_push_based_sort_1tb
- test_suite: dataset_test
- stable: false
- frequency: nightly
- team: core
- cluster:
- cluster_env: shuffle/shuffle_app_config.yaml
- cluster_compute: shuffle/datasets_large_scale_compute_small_instances.yaml
- run:
- timeout: 7200
- script: RAY_DATASET_PUSH_BASED_SHUFFLE=1 python dataset/sort.py --num-partitions=1000 --partition-size=1e9
- wait_for_nodes:
- num_nodes: 20
- type: sdk_command
- file_manager: sdk
- ##################
- # Core Chaos tests
- ##################
- - name: chaos_many_tasks_no_object_store
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: chaos_many_tasks_no_object_store
- test_suite: chaos_test
- frequency: multi
- team: core
- cluster:
- cluster_env: chaos_test/app_config.yaml
- cluster_compute: chaos_test/compute_template.yaml
- run:
- timeout: 3600
- wait_for_nodes:
- num_nodes: 10
- prepare: python setup_chaos.py --no-start
- script: python chaos_test/test_chaos_basic.py --workload=tasks
- type: sdk_command
- file_manager: sdk
- - name: chaos_many_actors
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: chaos_many_actors
- test_suite: chaos_test
- frequency: multi
- team: core
- cluster:
- cluster_env: chaos_test/app_config.yaml
- cluster_compute: chaos_test/compute_template.yaml
- run:
- timeout: 3600
- wait_for_nodes:
- num_nodes: 10
- prepare: python setup_chaos.py --no-start
- script: python chaos_test/test_chaos_basic.py --workload=actors
- type: sdk_command
- file_manager: sdk
- - name: chaos_dask_on_ray_large_scale_test_no_spilling
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: chaos_dask_on_ray_large_scale_test_no_spilling
- test_suite: chaos_test
- frequency: nightly
- team: core
- cluster:
- cluster_env: chaos_test/dask_on_ray_app_config_reconstruction.yaml
- cluster_compute: dask_on_ray/chaos_dask_on_ray_stress_compute.yaml
- run:
- timeout: 7200
- wait_for_nodes:
- num_nodes: 21
- prepare: python setup_chaos.py --node-kill-interval 100
- script: python dask_on_ray/large_scale_test.py --num_workers 20 --worker_obj_store_size_in_gb
- 20 --error_rate 0 --data_save_path /tmp/ray
- type: sdk_command
- file_manager: sdk
- - name: chaos_dask_on_ray_large_scale_test_spilling
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: chaos_dask_on_ray_large_scale_test_spilling
- test_suite: chaos_test
- frequency: nightly
- team: core
- cluster:
- cluster_env: chaos_test/dask_on_ray_app_config_reconstruction.yaml
- cluster_compute: dask_on_ray/dask_on_ray_stress_compute.yaml
- run:
- timeout: 7200
- wait_for_nodes:
- num_nodes: 21
- prepare: python setup_chaos.py --node-kill-interval 100
- script: python dask_on_ray/large_scale_test.py --num_workers 150 --worker_obj_store_size_in_gb
- 70 --error_rate 0 --data_save_path /tmp/ray
- type: sdk_command
- file_manager: sdk
- - name: chaos_pipelined_ingestion_1500_gb_15_windows
- group: core-dataset-tests
- working_dir: nightly_tests
- legacy:
- test_name: chaos_pipelined_ingestion_1500_gb_15_windows
- test_suite: chaos_test
- frequency: nightly
- team: core
- cluster:
- cluster_env: dataset/pipelined_ingestion_app.yaml
- cluster_compute: dataset/pipelined_ingestion_compute.yaml
- run:
- timeout: 7200
- wait_for_nodes:
- num_nodes: 21
- prepare: ' python setup_chaos.py --node-kill-interval 300'
- script: python dataset/pipelined_training.py --epochs 1 --num-windows 15 --num-files
- 915 --debug
- type: sdk_command
- file_manager: sdk
|