Managing your ML lifecycle with Azure Databricks and Azure ML – BRK3010

RIGHT. I THINK, LET’S GET STARTED WELCOME TO BUILD. THANK YOU SO MUCH FOR BEING HERE. I KNOW THIS IS A CHALLENGING TIME SLOT, JUST AFTER LUNCH, SO DOUBLE THANKS TO YOU GUYS I HOPE YOU GUYS ARE HAVING A GOOD BUILD SO FAR. JUST TO MAKE SURE WE ARE IN THE RIGHT PLACE, THIS IS MANAGING YOUR ML LIFECYCLE WITH AZURE DATABRICK WITH AZURE MACHINE LEARNING. >> GOOD AFTERNOON, MY NAME IS PARASHAR SHAH. >> JUST TO KIND OF — WE’RE NOT BROTHERS, WE SHARE THE SAME LAST NAME, AND WE ARE FROM THE SAME TOWN, BUT WE ARE VERY GOOD FRIENDS. SO JUST TO CLEAR THAT DOUBT, IF YOU HAVE IN YOUR MIND. SO HOW MANY OF YOU HAVE USED AZURE DATABRICKS OR ARE FAMILIAR WITH AZURE DATABRICKS? A QUICK SHOW OF HANDS. >> AND HOW ABOUT AZURE MACHINE LEARNING. >> OKAY. AWESOME, I THINK HOPEFULLY, THIS TALK WILL SORT OF BE VALUABLE TO HOPEFULLY MOST OF YOU. SO IF YOU LOOK AT THE OVERALL MACHINE LEARNING ON AZURE LANDED SCAPE, THE WAY WE THINK ABOUT MACHINE LEARNING IS IN TERMS OF THESE FIVE LAYERS, SO THE TOP-MOST LAYER IS THE SOPHISTICATED SET OF PRETRAINED MODELS THAT WE OFFER OUR CUSTOMERS THROUGH THE SET OF COGNITIVE SERVICES, WHICH YOU CAN USE TO BUILD END TO END, SO THIS DOES NOT ZOOM ANY KIND OF MODEL BUILDING CAPABILITIES, YOU CAN USE THEM OUT OF THE BOX. THE NEXT FOUR LAYERS ARE THE NOTION OF CUSTOM AI, YOU WANT TO BUILD YOUR OWN MACHINE LEARNING MODELS. SO YOU MIGHT WANT TO USE ANY TOOL THAT YOU’RE FAMILIAR WITH, ANY ID THAT YOU’RE COMFORTABLE USING A VARIETY OF DEEP LEARNING AND MACHINE LEARNING FRAMEWORKS SO PSYCHIC LEARN. AND THEN THE SET OF SERVICES THAT HELP YOU REALLY BUILD THESE END TO END SOPHISTICATED MACHINE LEARNING KUSS P APPLICATIONS IS AZURE DATABRICK, MACHINE LEARNING, VMS AND WHAT PARASHAR AND I WILL TALKING ABOUT IS AZURE DATABRICKS AND HOW TO HELP YOU MANAGE YOUR AZURE END TO END LEARNING CYCLE THE FOUNDATIONAL LAYER OF, YOU KNOW, HARDWARE INFRASTRUCTURES, WE’RE SOPHISTICATED IN AZURE, GPUS, F PGAS AND THEY ARE USED FOR ACCELERATED TRAINING AND INFERENCING CAPABILITIES THAT YOU CAN LEVERAGE, YOU KNOW, WHAT WE HAVE OUT OF THE BOX IN AZURE THIS IS PROBABLY THE BEST MACHINE LEARNING PAPER THAT I’VE EVER HAD IT’S MY FAVORITE. AND WHEN I STARTED, YOU KNOW, LEARNING ABOUT MACHINE LEARNING, I WAS UNDER THE IMPRESSION THAT, YOU KNOW, YOU BUILD THE MACHINE LEARNING MODEL AND THEN YOU THROW ANY DATA OUT OF IT AND OUT COMES A BEAUTIFUL MODEL AND BEAUTIFUL SET OF PREDICTIONS. HOW WRONG I WAS. AND BASED ON MY IGNORANCE, AND WHAT WE HAVE REALIZED IS THAT MACHINE LEARNING AS YOU CAN SEE IS JUST A SMALL SORT OF SMALL BOX IN THE ENTIRE SCHEME OF THINGS TLSHGS A BUNCH OF THINGS THAT’S GOING ON, THIS IS WHERE YOU’RE SPENDING 80 OF YOUR TIME, COMBLOUS HAVE HEARD THE STATISTICINGS, WHETHER IT’S FEATURE, DATA PREPARATION, ONCE YOU BUILD THE MODEL, HOW DO YOU DEPLOY TO — HOW DO SYSTEMS HELP YOU ACHIEVE THE WHOLE END TO END PIPELINE IS WHAT PARASHAR AND I WANT TO TALK ABOUT, USING AZURE DATABRICK AND AZURE MACHINE LEARNING WHAT IS AZURE DATABRICKS? SOME OF YOU HAVE USED IT? IT’S A FAST, EASY, COLLABORATIVE, A BUNCH OF SPARK PLATFORM IN AZURE, THIS IS A UNIQUE AZURE SERVICE BECAUSE WE HAVE A PARTNERSHIP WITH THE DATABRICKS COMPANY WHICH ACTUALLY ARE THE CREATORS OF SPARK, SO WHAT WE HAVE DONE IS TAKEN THE BEST OFF DATABRICKS AND THE BEST OF AZURE AND SORT OF COMBINED IT TO CREATE THIS AZURE DATABRICKS FIRST PARTY OFFERING. SO THERE ARE THREE KEY THEMES OF PILLARS TO THE PRODUCT. FUNDAMENTALLY HELP YOU INCREASE PRODUCTIVITY, WHAT THAT MEANS IS THROUGH COLLABORATION, NOTEBOOKS, THROUGH THE ABILITY FOR YOU TO DO DEVOPS, THE ABILITY TO CREATE CLUSTERS, WITHOUT UNDERSTANDING THE NITTY-GRITTY DETAILS OF THE INFRASTRUCTURE, HOW CAN WE HELP DATA SCIENTISTS AND DEVELOPERS, YOU KNOW, BE MUCH MORE PRODUCTIVE ON THIS PLATFORM? THE SECOND IS, YOU KNOW, WE WANT TO BRING TO OUR CUSTOMERS THE SECURITY AND THAT AZURE HAS SECURITY ENTRUST THAT THE CUSTOMERS HAVE COME TO EXPECT OF AZURE. HOW DO WE TAKE DATABRICKS AND HOW DO WE INTEGRATE IT WITH

AZURE SUCH THAT WE CONTINUE TO OFFER AN END TO END SOLUTION FOR OUR CUSTOMERS? AND FINALLY, BEING A BIG DATA SYSTEM, YOU WANT TO BE ABLE TO SCALE IT, YOU KNOW, WITHOUT PRETTY MUCH ANY LIMITS. SO HOW DO WE ENSURE THAT AS YOU THROW MORE AND MORE DATA IT, IT’S ABLE TO SCALE EFFICIENTLY AND YOU CAN MANAGE THE BIG DATA WORKLOADS. AND OF COURSE, WE MAKE ALL OF THIS EASIER THROUGH, YOU KNOW, NATIVE AUTHENTICATION, THROUGH AUTHORIZATION AND A BUNCH OF OTHER SORT OF NEEDS THAT CUSTOMERS HAVE COME TO US WITH. SO IF YOU LOOK AT — IF YOU CAN LIKE JUST OPEN THE BOX A BIT, SO AZURE DATABRICKS PRIMARILY IS THREE LAY YARS AND THE CORE SYSTEM IS THE DATABRICKS RUN TIME ENGINE, WHICH IS, YOU KNOW, WHICH RUNS ABOUT 7-10 TIMES FASTER THAN OPEN SOURCE SPARK. SO OPEN SOURCE SPARK HAS A LOT OF ADDITIONS IN TERMS OF, YOU KNOW, ADVANCED CACHING AND IN TERMS OF DATABRICK, IN TERMS OF SERVERLESS COMPUTING YOUR SPINNING UP AND SPINNING OUT OF CLUSTERS IS AUTO SCALE. COMING BACK TO THE PRODUCTIVITY ASPECT, YOU DON’T HAVE TO WORRY ABOUT MANAGING THESE CLUSTER, THEY ARE MANAGED FOR YOU, THEY ARE SCALED FOR YOU DEPENDING UPON YOUR WORKLOADS, THEY SPIN UP AND DOWN, WHICH HAS COST BENEFITS. THE WHOLE PLATFORM IS ACCESSIBLE THROUGH A SET OF REST API, ALL OF YOUR END TO END SORT OF PROCESSES CAN BE AUTOMATED USING THESE APIS. ONCE WE HAVE THE BASE IN PLACE, YOU KNOW, YOU WANT TO BE ABLE TO DEPLOY YOUR PRODUCTION JOBS. SO YOU WANT TO BE ABLE TO RUN PIPELINES AND THAT’S WHERE THE JOB INFRASTRUCTURE COMES INTO THE PICTURE, SO HOW DO YOU SCHEDULE JOBS? HOW DO YOU RUN JOBS? HOW DO YOU DEPLOY THESE INTO PRODUCTION? SO THAT’S THE LAYER WHICH TALKS ABOUT THE DEPLOYMENT AND PRODUCTION WORKFLOWS. AND FINALLY, IF YOU LOOK AT THE PERSONAS, WE HAVE DATA SCIENTISTS, DATA ENGINEERS AND BUSINESS ANALYSTS AND WHAT WE ARE TRYING TO DO IS BLUR THE DIFFERENCES BETWEEN THE TWO. SO PARTICULARLY DATA ENGINEERS DO A CERTAIN KIND OF FUNCTION, DATA SCIENTISTS DO A CERTAIN KIND OF FUNCTION AND THE PLATFORM WE ARE TRYING TO BLUR THE DIFFERENCES THAT, YOU KNOW, THROUGH ALL OF THESE ADDED CAPABILITIES OF THE DATA SCIENTISTS AND DATA ENGINEERS WORK TOGETHER AND BUSINESS ANALYSTS, YOU KNOW, CAN ACCESS THE SYSTEM TO DO REPORTING SO ON THE LEFT HAND SIDE, YOU’LL SEE WE WANT TO INTEGRATE WITH THE REST OF THE AZURE INFRASTRUCTURES, IT HAS THE ABILITY TO TAKE INPUTS FROM ALL OF THESE DIFFERENT SYSTEMS AND STORAGE SYSTEMS, WAREHOUSES, AND THEN FINALLY, ONCE YOU ARE DONE WITH YOUR — THE BIG DATA ANALYSIS, HOW DO YOU EXPOSE IT TO DOWNSTREAM, YOU KNOW, THE TOOLS, WAREHOUSES, HOW DO YOU BUILD MODELS IN AZURE DATABRICKS AND DEPLOY THEM SUCH THAT YOUR LINE OF BUSINESS APPLICATIONS CAN START USING THEM? SO WHAT IS AZURE MACHINE LEARNING SERVICE? AGAIN, THREE KEY THEMES HERE, ONE, WE WANT TO MAKE DATA SCIENTISTS BE PRODUCTIVE, SO THE IDEA IS WHICHEVER ID YOU’RE COMFORTABLE WITH, YOU CAN USE THAT TO LEVERAGE AZURE MACHINE LEARNING SERVICES, YOU KNOW, THROUGH SDK. WHEN YOU’RE BUILDING OUT A MACHINE LEARNING MODEL, IT’S NOT JUST ABOUT TRAINING THE MODEL, IT’S ABOUT HOW DO YOU EXPERIMENT? HOW DO YOU COME UP WITH THE BEST MODEL, HOW DO YOU MAKE YOUR TRAINING PROCESSES REPRODUCEABLE IS UCH THAT WHEN YOU ARRIVE AT A PARTICULAR MODEL, YOU KNOW HOW YOU ARRIVED AT THAT PARTICULAR MODEL. THE RATE OF EXPERIMENTATION IS IN DIFFERENT COMPUTE INFRASTRUCTURES POSSIBLE THROUGH AZURE MACHINE LEARNING AND PARASHAR WILL WALK YOU THROUGH THAT LATER. AND FINALLY, THE MOST POWERFUL CONCEPT HERE IS THAT YOU CAN DEPLOY, MANAGE AND MONITOR YOUR MODELS, USING AZURE MACHINE LEARNING SERVICES. AND TYPICALLY IN A MACHINE LEARNING LIFECYCLE, THE END MILE OF DEPLOYMENT IS THE MOST SORT OF DIFFICULT AND PAINFUL PART. YOU KNOW, IN TERMS OF LIKE ONCE I HAVE A MODEL, HOW DO I MAKE IT USABLE THROUGH THE ENTERPRISES? THAT’S WHERE AZURE MACHINE LEARNING IS USEFUL AND PARASHAR WILL WALK YOU THROUGH DEMOS. SO IF YOU LOOK AT THE MACHINE LEARNING LIFECYCLE, USING AZURE DATABRICKS AND AZURE ML, THERE ARE THREE — THIS IS SOMETHING THAT YOU MUST ALREADY BE FAMILIAR WITH. THERE ARE THREE MAIN SORT OF COMPONENTS TO BUILDING AN END TO END LIFECYCLE, THE FIRST ONE IS, YOU KNOW, HAVE YOU TO PREPARE DATA, COMING BACK TO THAT SLIDE WHICH I SHOWED YOU EARLIER, 80 OF YOUR TIME IS USED IN SORT OF THIS DATA PREPARATION PHASE. AND WE’RE GOING TO SHOW YOU HOW AZURE DATABRICKS HELPS YOU TO PREPARE THAT DATA ONCE WE HAVE THE DATA THAT WE WANT TO BUILD AND TRAIN YOUR MODELS, SO YOU WANT TO USE A VARIETY OF TRAINING FRAMEWORKS. YOU WANT TO PICK THE BEST MODEL. AS YOU’RE BUILDING YOUR MODEL, YOU WANT TO DO EXPERIMENTATION, YOU WANT TO MAKE SURE THAT THE MODEL IS REPRODUCEABILITY. ALL OF THE FACILITIES AND HOW AZURE MACHINE LEARNING HELPS YOU TO ACHIEVE THAT,

WE WILL GO INTO DETAIL ON THAT FINALLY, AS I MENTIONED, HOW DEPLOY, AND MANAGE MODELS AND HOW DO YOU DEPLOY MODELS IN DIFFERENT SURFACE AREAS. THE CLOUD AND THE EDGE? TO MAKE IT USEFUL FOR DOWNSTREAM APPLICATIONS SO IN DATA PREPARATION, WHERE AZURE DATABRICKS REALLY SHINES, IS THAT AS YOU GUYS MIGHT KNOW IN SPARK HAS MORE THAN 50 CONNECT R TOS TO DIFFERENT DATA SOURCES AND WHEN YOU’RE BUILDING MACHINE LEARNING MODELS, YOU WANT TO MAKE SURE THAT YOU’RE LEVERAGING AS MANY DATA SOURCES AS YOU CAN, AS MANY DATA FORMATS AS YOU CAN, BECAUSE YOU MIGHT HAVE SOME LEGACY DATA HANGING AROUND IN DIFFERENT FORMATS AND DIFFERENT SYSTEMS SO AZURE DATABRICKS REALLY HELPS YOU TO COLLECT ALL OF THIS DATA TOGETHER AND BE ABLE TO DO LOT SCALE PROCESSING ON THE DATA FOR YOU TO BE ABLE TO BUILD MACHINE LEARNING MODELS OUT OF THAT DATA AND INTEGRATION WITH AZURE DATA FACTORY IS A VERY COOL INTEGRATION AND WHAT IT ALLOWS YOU TO DO IS ABLE TO TAKE YOUR DATA FROM LEGACY SYSTEMS OR ORACLE OR — WHICHEVER SYSTEMS YOU MIGHT HAVE, AND SORT OF MOVE THOSE — THAT DATA TO THE CLOUD AND THEN YOU CAN USE AZURE DATABRICKS TO THEN PROCESS THIS DATA IN A MEANINGFUL WAY FOR YOU TO THEN TO BE ABLE TO DO MACHINE LEARNING MODELLING LATER ON. IN THE BUILD AND TRAIN PHASE, YOU HAVE A VARIETY OF FRAMEWORKS TO CHOOSE FROM. TODAY, YOU KNOW, YOU CAN USE, YOU KNOW, SO MANY DIFFERENT FRAME WORKS TO BUILD MACHINE LEARNING MODELS AND WE WANT TO MAKE SURE THAT WE SUPPORT EVERY FRAMEWORK SO YOU USE AZURE DATABRICKS NOTEBOOKS TO THEN USE THESE FRAMEWORKS TO BUILD MODELS. AND THEN FINALLY, YOU KNOW, YOU USE THE CAPABILITIES OF AUTOMATED MACHINE LEARNING, WHICH IS A COOL DEMO THAT PARASHAR WILL SHOW YOU, HOW TO USE AUTOMATED MACHINE LEARNING TO REALLY HELP PICK THE BEST MODEL FOR YOU. IN AZURE DATABRICKS SO THAT YOU CAN THEN DEPLOY THAT MODEL AND YOU CAN, YOU KNOW, USE AND WE TALK ABOUT ML FLOW AND THE AZURE MACHINE LEARNING INTEGRATION, WHICH IS HADDING SMG WE RELEASED AT — WHICH IS SOMETHING WE RELEASED AT SPARK SUMMIT LAST WEEK. THIS IS THE MOST POWERFUL THING WHERE ONCE YOU HAVE YOUR MODEL, HOW CAN YOU THEN DEPLOY IT? SO IN AZURE MACHINE LEARNING, IS YOU FIRST OF ALL, ONCE YOU HAVE A MODEL AND IT COULD BE ANY MODEL. USING ANY KIND OF FRAMEWORK. AND THEN YOU REGISTER THE MODEL IN THE REGISTRY AND THEN YOU HAVE A SCORE FILE, WHICH IS AN APPLICATION THAT THEN USES THAT MODEL TO SCORE ANY — TO SCORE ANY NEW INCOMING DATA. AND THEN YOU CAN DEPLOY THOSE MODELS ON THE CLOUD OR IN AZURE COMMUNITY SERVICE OR ON THE HEAVY EDGE OR ON THE LIGHT EDGE. YOU CAN — HEAVY EDGE IS LIKE A CAR, AND THE LIGHT EDGE IS LIKE A SENSOR, THE IDEA IS ONCE YOU HAVE YOUR MODEL, HOW YOU CAN LEVERAGE IT AND DEPLOY IT TO DIFFERENT SURFACE AREAS, FOR YOU TO BE ABLE TO THEN START INFERENCING ON NEW DATA THAT’S COMING IN. AND THIS IS VERY, VERY USEFUL FOR IOT TYPE SCENARIOS, ALSO AND THE LAST ONE IS REALLY, ONCE YOU DEPLOY YOUR MODEL, HOW DO YOU KNOW YOUR MODEL IS PERFORMING WELL? MONITORING OF MODELS IS REALLY IMPORTANT WHERE YOU MIGHT WANT TO, YOU KNOW, CHECK TO SEE IF YOUR MODEL IS BEHAVING THE WAY IT’S SUPPOSED TO, AS THE NEW DAT IS COMING IN, IS YOUR MODEL ABLE TO PREDICT AS ACCURATELY AS YOU HAD EXPECTED AND IF NOT, HOW DO YOU GO BACK AND RETRAIN YOUR MODELS AND DEPLOY WHERE YOU’RE GETTING THE BUSINESS VALUE OF WHATEVER YOU HAVE BUILT? I’LL HAND IT OVER TO PARASHAR, WHO IS GOING TO WALK YOU THROUGH DEMOS, SO. >> YEAH, THANK YOU, PREMAL. AND SO, WHAT PREMAL EXPLAINED IS HOW THE REGULAR ML LIFECYCLE WORKS, SO YOU START WITH DATA PREPARATION, AND YOU BEGIN TO CHOOSE A MODEL AND THEN ONCE YOU ARE HAPPY WITH THE MODEL, YOU START DEPLOYING IT. AND SO THE IDEA WHICH I’M GOING TO TALK ABOUT NOW IS LIKE THERE ARE DIFFERENT WAYS IN WHICH YOU CAN DO THIS. SO DEPENDING ON THE SCENARIO OR THE COME FORT LEVEL OF THE TEAM, YOU CAN BEGIN TO CHOOSE WHICH ONE WORKS FOR YOU SO IN AZURE, JUST LIKE A LOT OF SERVICES YOU HAVE MULTIPLE OPTIONS, AND IT ALWAYS TAKES A LITTLE BIT OF EFFORT IN ORDER TO FIGURE OUT WHEN TO USE WHAT. WE ARE GOING TO TRY AND SIMPLIFY IT FOR YOU. AND IF YOU STILL HAVE QUESTIONS, THE MIC IS THERE, O WE CAN TALK ABOUT IT AT THE END. LET’S WALK THROUGH THIS EXAMPLE OF MANUAL MACHINE LEARNING, RIGHT, IN WHICH CASE, YOU ARE TRYING TO PREDICT THE PRICE OF A CAR. SO I’M GOING TO MAKE IT A LITTLE BIT MORE INTERACTIVE. SO HOW MANY PEOPLE

THINK THIS IS A CLASSIFICATION PROBLEM? HOW MANY OF YOU THINK IT’S A REGRESSION PROGRAM? MOST OF YOU. THAT’S GREAT YEAH, SO IT’S AN REGRESSION NUMBER, YOU ARE TRYING TO PREDICT A NUMBER, INSTEAD OF A YES OR A NO OR TRUE OR FALSE. IF YOU ARE GOING TO DEVELOP THIS MODEL FROM SCRATCH, WHAT ARE SOME OF THE FEATURES THAT YOU THINK ARE IMPORTANT? HOW ABOUT BRAND? DO YOU THINK BRAND IS IMPORTANT WHEN YOU WANT TO PREDICT THE PLIES OF A CAR. MAKE, MODEL, THINGS LIKE THAT. AND THEN, LIKE, AFTER YOU PICK THE FEATURES, YOU START WITH A PARTICULAR ALGORITHM. SO YOU COULD START WITH CLASSICAL MACHINE LEARNING, SOMETHING FROM PSYCHIC LEARN WORD, OR YOU COULD START EVEN WITH DEEP LEARNING, IF YOU WANT TO MAKE IT FANCY. IN EITHER CASE, YOU PICK AND CHOOSE AN ALGORITHM. AND THEN, YOU SET THE CERTAIN PARAMETERS, YOU CAN GO WITH DEFAULT PARAMETERS, LIKE EVERY ALGORITHM WILL HAVE, FOUR PARAMETERS, BUT NORMALLY, YOU ARE REQUIRED TO TUNE IT. SO THE IDEA BEHIND THIS IS WHEN YOU START WITH IT, THE FIRST TIME, YOU MIGHT GET A RESULT AND IN CASE, SAY FOR EXAMPLE, YOU GET A MODEL THAT IS HAVING AN ODD SQUARE OF 0. 3. SO 30 KIND OF REASONABLY — OKAY, LET ME PUT IT — IS IT A GOOD MODEL OR A BAD MODEL? HOW MANY OF YOU THINK IT’S A GOOD MODEL? HOW MANY THINK IT’S A BAD MODEL? AND WHY? IS IT LIKE FLIP OF A COIN THING, RIGHT? I WANT TO MAKE IT MORE INTERACTIVE AND ALL OF YOU DON’T HAVE A MIC SO BUT A RANDOM FLIP OF A COIN WOULD HAVE GIVEN YOU 50 PROEBLTH, YOU DON’T NEED TO DO MACHINE LEARNING IF THE MACHINE IS GIVING YOU 50 ACTIVATE, IN THIS, YOU PROBABLY CHANGE THE FEATURES. SO YOU BRING IN MORE DATA. YOU PROBABLY JOINED SOME DATA SETS, AND THIS IS TO HIS EARLIER POINT, 80 OF THE TIME GOES INTO PRETEARING THE DATA — PREPARING THE DATA. YOU MIGHT HAVE A MACHINE LEARNING PROBLEM, BUT DO YOU HAVE THE RIGHT TYPE OF DATA SO SOLVE THAT PROBLEM? YOU CAN CHANGE THE ALGORITHM, CHANGE THE PARAMETERS AND YOU KEEP ON DOING THIS AND FINALLY, YOU GET A MODEL WHICH ACTUALLY IS GOOD. AND BY THEN, YOU WILL HAVE PROBABLY SPENT SOME WEEKS OR MONTHS AND OF COURSE A LOT OF MONEY AND TIME. RIGHT. SO AUTOMATED ML IS TRYING TO ADDRESS THIS. SO THE IDEA BEHIND AUTOMATED ML IS HOW CAN WE AUGMENT THE PRODUCTIVITY OF THE DATA SCIENTIST? SO AT LEAST IN THE BITEN VERSION, WHICH WE ARE GOING TO SEE TODAY ON DATABRICK, WE ARE NOT TRYING TO REPLACE A DATA SCIENTIST BUT AUGMENT THE WORK. HOW DO WE DO THAT? IT’S ESSENTIALLY, YOU GIVE US WHAT PROBLEM, OR YOU TELL US, AS IN THE SERVICE, YOU TELL THE SERVICE, WHAT PROBLEM YOU’RE TRYING TO SOLVE. AND THEN, YOU’LL DEFINE THE CONSTRAINTS AROUND IT. SO CONSTRAINTS COULD BE LIKE SAY IF YOU ARE THINKING OF THAT THIS MACHINE LEARNING MODEL IS GOING TO HELP YOU SAVE $1, 000 FOR EXAMPLE. SO YOU LEARN WANT TO SPEND MAYBE MORE THAN $100 ON TRAINING IT. YOU ARE LOOKING AT TEN TIMES THAT FOR EXAMPLE. YOU TRANSLATE THAT $100 IN TERMS OF THE TRAINING TIME THAT PROBABLY YOU WOULD WANT TO SPEND. SO THESE ARE SOME OF THE CONSTRAINTS THAT YOU CAN GIVE. OKAY I WANT TO DO A CLASSIFICATION OR I WANT TO DO THE REGRESSION OR FORECASTS, THAT’S WHAT WE SUPPORT WITH AUTOMATED ML AND I WANT TO SPEND ONLY 10 HOURS IN TRAINING WITH THIS MUCH OF COMPUTE SO IN THE DEMO THAT I’M GOING TO SHOW YOU NOW, IT’S GOING TO BE AZURE DATABRICKS AS YOUR COMPUTE, BUT YOU WOULD USE OTHER COMPUTE AS WELL THESE ARE EXAMPLES OF THE CONSTRAINT AND THEN YOU CAN ALSO GIVE LIKE YOU CAN WRITE CERTAIN ALGORITHMS, YOU CAN BLACKLIST CERTAIN ALGORITHMS, BUT YOU DO HAVE TO MANUALLY PICK AND CHOOSE THE ALGORITHM, WE DO THAT FOR YOU. WE JUST DON’T LIMIT OURSELVES TO SELECTING AN ALGORITHM, WE ALSO DO PREPROCESSING, AND WE ALSO DO HYPER PARAMETER TUNING SO IT’S MULTIPLE THINGS THAT WE TAKE CARE OF AS PART OF AUTOMATED ML, OKAY. AND SO IT’S MORE OF A — THINK OF A MACHINE LEARNING MODEL THAT WE HAVE TRAINED WITH OPEN DATA SETS IN ORDER TO HELP YOU DO MACHINE LEARNING. SO THAT’S HOW YOU THINK OF IT. OKAY. AND ONCE YOU — ONCE YOU HAVE THAT, THEN ESSENTIALLY, YOU CAN DEPLOY THE MODEL THE SAME WAY THAT THEY REGISTER IT IN AZURE ML AND DEPLOY IT IN AN — LET’S LOOK AT THE DEMO. AND ONLY A FEW OF YOU RAISED THE HAND FOR AUTOMATED ML, THERE WAS A SESSION IN THE MORNING WHICH MY TEAMMATES 8:30 A. M. , IN CASE YOU WANT TO LOOK AT THE WHOLE SESSION, YOU CAN GO TO THAT

AS WELL. T BUT IF THERE ARE QUESTIONS LATER, LET ME KNOW. WE ARE GOING TO START WITH THE AZURE PORTAL HOW MANY OF HAVE ACCESS TO THE AZURE PORTAL? SO THE FIRST THING THAT — BECAUSE WE ARE GOING TO RUN AUTOMATED ML ON AZURE DATABRICKS, THE FIRST THING THAT YOU NEED IS AN AZURE DATABRICKS WORK SPACE. OKAY. SO — >> A QUICK QUESTION HERE. IS YOU SAID AUTOMATED ML ON AZURE DATABRICKS, WHY IS AZURE DATABRICKS GOOD FOR AUTOMATED ML? WHAT’S THE — IS THERE A SPECIFIC ADVANTAGE YOU GET RUNNING ON AZURE DATABRICKS? >> YEAH, YEAH, SO THAT’S A GOOD QUESTION, THANKS SO THE REASON AZURE DATABRICKS IS PERFORMING VERY WELL WITH AUTOMATED ML IN ADDITION TO THE COMPUTE IS LIKE THERE ARE A LOT OF CUSTOMERS THAT WE WORK WITH THAT ALREADY HAVE THEIR DATA IN DATABRICKS, SO EITHER IT COULD BE IN DBSS OR CON NENTH TORS, WHICH ARE — CONNECTORS WHICH ARE THERE. SO THEY DON’T WANT TO MOVE THEIR DATA OUTSIDE. OR THEY HAVE DATABRICKS AS OTHER THINGS E WANT TO STAY WITH THE DATABRICKS AND WANT TO DO AUTOMATED ML, THEY DON’T WANT TO DO AUTOMATED MACHINE LEARNING. THAT IS ONE. AND WHEN USING AUTOMATED ML ON AZURE DATABRICKS, WE ALSO LEVERAGE SOME OF THE SPARK TECHNOLOGY, YOU CAN DO TRAINING CHSHGS CHSHGS — WHICH IS THERE, AND YOU WILL SEE THAT. >> THANK YOU. >> LET’S FIRST START WITH CREATING AN AZURE DATABRICKS WORK SPACE SO I ALREADY HAVE IT PREDICATED, BUT FOR PEOPLE WHO HAVEN’T SEEN IT BEFORE, SO THE WORK SPACE IS WHERE ESSENTIALLY YOUR DATABRICKS COMPUTE OR THE CLUSTER AND THE NOTEBOOK — HOW MANY OF UNDERSTAND NOTEBOOK? JUPITER OR — OKAY. SO IT’S THE SAME NOTEBOOK, THE ENVIRONMENT IS A LITTLE BIT DIFFERENT THAN JUPITER, BUT YOU START WITH CREATING THIS WORK SPACE BY PROVIDING THE LOCATION, RESOURCE GROUP, AND THE SUBSCRIPTION YOU CAN ALSO DEPLOY THIS CLUSTER IN YOUR OWN V NET IF YOU WANT TO OKAY? I’M NOT GOING TO CREATE THE WORK SPACE, BECAUSE I ALREADY HAVE IT. ONCE YOU CREATE IT, YOU SEE SOMETHING LIKE THIS. SO IT LOADS WITH THIS HOME PAGE. AND THEN, IT’S GOING TO SHOW YOU DIFFERENT OPTIONS ON THE LEFT. THE FIRST THING THAT YOU WILL DO, YOU ARE THE I. T. DEPARTMENT IN YOUR COMPANY OR ANYONE WHO HAS THE PERMISSION TO CREATE CLUSTERS CAN CREATE DATABRICKS CLUSTER IN THEIR OWN TIME. NOW, YOU CAN SELECT HIGH CURRENCY, YOU CAN SELECT THE LATEST RUN TIME THAT YOU HAVE. VERSION IS 3. AND YOU CAN ALSO SELECT AUTOMATION IF YOU WANT TO, BASICALLY, WHEN YOU’RE NOT USING THE CLUSTER, IT WILL SHUT DOWN AUTOMATICALLY. AND IN THIS EXPERIMENT, AT LEAST, I WANT TO RUN TRAINING — SO TO YOUR EARLIER QUESTION, I’M GOING TO FIX THE NUMBER WORKERS, YOU CAN GIVE ANY NUMBER OF WORKERS BASED ON WHAT YOU WANT. THE PREFERENCE FOR USING AUTOMATED ML ON DATABRICKS IS THAT YOU USE OPTIMIZE VMS, SOME OF THESE R THINGS ARE INTENSIVE. YOU CAN USE CPU, BUT THE PREFERENCE IS THAT AND WE HAVE GOOD DOCUMENTATION AS WELL TO TALK ABOUT IT. SO BASICALLY, HOW TO SET IT UP. SO IF YOU GO TO THE AZURE DOCS, YOU CAN GO THROUGH ALL OF THE DETAILS IN TERMS OF THE CONFIG. NOW, AFTER YOU HAVE CREATED THE CLUSTER, WHAT YOU DO IS JUST ONE STEP, WHICH IS INSTALL THE SDK SO THIS IS THE AUTO SDK, AND YOU CAN INSTALL IN THE LIBRARY AND IT’S STRAIGHTFORWARD THAT YOU SELECT THE LIBRARY TYPE AND THEN YOU JUST GIVE AZURE SDK OR AZURE DATABRICKS THIS LIBRARY IS AVAILABLE ON THIS CLUSTER ALL THE TIME. AND EVERY TIME THAT THE CLUSTER RESTARTS, YOU CAN GET THE LATEST VERSION NOW, YOU CAN FREEZE IT BY DOING EQUAL TO EQUAL TO A PARTICULAR VERSION, BUT I OTHERWISE, EVERY TWO WEEKS, WE RELEASE A NEW SDK VESHGS, — VERSION, IT’S THAT SIMPLE HOW TO GET STARTED WITH IT. NOW, ONCE YOU HAVE THAT, YOU GO AHEAD AND SELECT YOUR NOTEBOOK. SO I HAVE THIS NOTEBOOK NOW, FOR PEOPLE WHO KNOW JUPITER, THIS ENVIRONMENT IS SIMILAR TO JUPITER, EXCEPT IT’S SERVERLESS, SO WITH JUPITER, WHENEVER YOU WANT TO READ YOUR NOTEBOOK, YOU HAVE TO ACTUALLY SPIN UP THE SERVER, IN THIS CASE, YOU CAN READ THE NOTEBOOK AND THE CACHE DATA, EVEN IF THE CLUSTER IS SHUT DOWN. IN MY CASE, I ALREADY HAVE A CLUSTER RUNNING AND I HAVE ALREADY ATTACHED IT. I’M GOING TO RUN AN AUTOMATED ML EXPERIMENT IN THIS EXPERIMENT, WHAT WE ARE TRYING TO DO IS TO PREDICT THE REMAINING USEFUL LIFE OFF OF A TURBO FAN ENGINE, SO THIS DATA SET IS A REAL-LIFE

DATA SET, IT’S FROM NASA. AND THIS IS BASICALLY CENSORED DATA, WHICH IS THERE. AND WE ARE GOING TO FIGURE OUT BASED ON THE SENSOR DATA INFORMATION, HOW MANY DAYS BEFORE I NEED TO DO MAINTENANCE OF AN ENGINE. NOW, IF YOU DO IT TOO EARLY, THEN IT’S DOWN TIME, RIGHT? SO IT’S LOSS OF REVENUE IF YOU DO IT TOO LATE, IT’S RISKY WE ARE GOING TO USE AUTOMATED ML ON DATABRICKS TO DO THAT. SO I FIRST — I HAVE THE CLUSTER AND I HAVE THIS SDK START, SO I’M GOING TO DOWNLOAD THE DATA SET OVER HERE, I’VE ALREADY DOWNLOADED IT, AROUND 100MB. IT’S A DATA SET. THE FIRST THING I DO IS TO DERIVE THE COLUMNS, SO LABEL COLUMN, WHICH IS THE COLUMN — THAT’S THE ONE THAT WE ARE TRYING TO PREDICT. OTHER THAN THAT, I DON’T HAVE TO PRETTY MUCH DO ANYTHING, JUST PROVIDE THE DATA FRAME THAT I KIND OF DATA, AND I’M GOING TO USE THE AZURE MACHINE LEARNING WORK SPACE. AND SUBMIT THE TRAINING EXPERIMENT TO IT. SO THE INPUT TO AUTOMATED ML IS OUR DATA FRAME. YOU BASICALLY HAVE A PAINLESS DATA FRAME AND IT TAKES THAT DATA FRAME IN THE FORM OF A DATA FLOW AND DOES THE TRAINING, PREPROCESSING OF THE DATA AND DOES HYPER PARAMETER TUNING, ALL THINGS NOW, HOW MANY OF YOU KNOW AZURE MACHINE LEARNING WORK SPACE, BECAUSE MANY OF YOU RAISED YOUR HAND FOR AZURE MACHINE LEARNING? SOME OF YOU. OKAY. SO LET ME EXPLAIN WHAT IT IS. BUT AZURE MACHINE LEARNING WORK SPACE IS A SINGLE PLACE OR A CENTRAL PLACE FOR YOUR TEAM TO COLLABORATE. NOW, IF YOU RECALL WHAT PARASHAR MENNING POLICE DEPARTMENT EARLIER, THERE ARE THREE DIFFERENT PHASES IN TRADITIONAL MACHINE LEARNING, THE PREPARATION OF DATA, TRAINING AND THEN THE DEPLOYMENT. NOW, OFTEN TIMES, THE DATA SCIENTIST TEAM WHO IS ACTUALLY — WHO ARE TRAINING THIS MODELS, IS — ARE NOT THE SAME WO ARE PROBABLY DEPLOYING IT TO PRODUCTION, SO THAT’S A SEPARATE DEVOPS DEVELOPMENT TEAM THAT IS DOING THE DEPLOYMENT OF THE MODEL WHEN I SAY DEPLOYMENT, I MEAN CREATING A WEB SERVICE AND END POINT, OR SOMETHING LIKE THAT SO IT CAN BE CONSUMED IN YOUR SMART APPLICATION RIGHT? SO THEN, IF THESE DIFFERENT TEAMS, IF THEY WANT TO COLLABORATE WITH EACH OTHER, IT BECOMES TRICKY AND THAT’S WHERE THE AZURE MACHINE LEARNING WORK SPACE COMES IN. SO THIS WORK SPACE, IT’S DIFFERENT FROM THE DATABRICKS WORK SPACE WE TALKED ABOUT EARLIER. SO THE DATABRICKS WORK SPACE IS HAVING YOUR CLUSTER AND THE NOTEBOOK WHEREAS THE AZURE MACHINE LEARNING WORK SPACE IS HAVING ALL OF YOUR EXPERIMENTS, ALL OF THE TRAINED MODEL, DIFFERENT VERSIONS, AND ALL OF YOUR DATA SOURCES AND EVERYTHING NOW, THE RELATIONSHIP CAN BE MANY TO MANY, BECAUSE YOU CAN USE AZURE DATABRICKS FOR MACHINE LEARNING, BUT YOU COULD ALSO USE IT FOR ANALYTICS AND WITH AZURE MACHINE LEARNING WORK SPACE, YOU CAN USE DATABRICKS FOR DOING MACHINE LEARNING OR YOU COULD USE ML COMPUTE, OR EVEN YOUR LOCAL LAPTOP IN ORDER TO DO THAT SO IT’S A MANY TO MANY RELATIONSHIP BUT WE ARE LAUNCHING SOMETHING NEW TODAY, AND THAT IS THE CAPABILITY THAT WE HAVE INTEGRATED THESE TWO WORK SPACES. SO EAGER YOU CAN CREATE THE WORK SPACE FROM CODE, LIKE YOU CAN BITEN CODE, OR WHEN YOU HAVE YOUR DATABRICKS WORK SPACE, SO THIS IS MY DATABRICKS WORK SPACE WHICH WE CREATED, FROM THERE, ITSELF, NOW, I CAN CREATE AN AZURE MACHINE LEARNING WORK SPACE, AND WHEN I DO THAT, IT LINKS ITSELF TO THIS SO WE ARE COUPLING IT. AND THE REASON WE ARE DOING THIS IS TO DECREASE THE TIME IT TAKES FOR YOU TO GET STARTED WITH MACHINE LEARNING. RIGHT? SO IT’S SIMPLIFICATION OF THIS WHOLE PROCESS SO THAT’S HOW IT IS. AND NOW, ONCE YOU HAVE THE WORK SPACE, EITHER CREATED FROM BITEN OR FROM THE PORTAL, YOU JUST CALL THAT WORK SPACE AND SUBMIT THE EXPERIMENT. SO HERE, I’M SUBMITTING AN EXPERIMENT, AS YOU CAN SEE. AND THE CONFIGURATION IS PRETTY MUCH STRAIGHTFORWARD IT JUST TAKES WHAT IS THE EXPERIMENT TYPE, SO BASICALLY, WHETHER IT’S CLASSIFICATION, OR REGRESSION OR FORECASTING, AND THEN, THE METRIC, SO WE SUPPORT CERTAIN METRICS FOR EVERY EXPERIMENT TYPE. SO IT COULD BE FOR EXAMPLE, ACCURACY IF YOU ARE DOING CLASSIFICATION, OR IT COULD BE YOUR EDITOR, SO BASICALLY, MINUTE NICING THE EDITOR, IF YOU ARE TRYING TO DO THE REGRESSION >> MAYBE THE AUDIENCE HAS A QUESTION, BUT YOU MENTIONED ABOUT BINDERS, DATA FRAMERS THAT — [ INAUDIBLE ] SO WHAT IF YOUR DATA IS IN — BECAUSE TYPICALLY IN SPARK, IN SPARK DATA FRAMES, YOU KNOW, CAN SPAN ACROSS MULTIPLE NODE, WINDOWS IS ONE NODE. >> UH-HUH. >> IN YOUR

EXPERIENCE, IS SPAN DATA FRAME ENOUGH TO BE ABLE TO SOLVE THESE PROBLEMS OR DO YOU HAVE ANY OTHER GUIDANCE AROUND THAT? >> YEAH, THAT’S A GOOD QUESTION, THANKS, PREMAL. SO THE QUESTION IS, WHEN THE SPAN DATA TO FRAME — WORK. >> ABSOLUTELY RIGHT. THE THING YOU HAVE TO NOTE IS WHEN YOU ARE PROVIDING THIS DATA FRAME, YOU ARE PROVIDING FEATURES, YOU ARE NOT PROVIDING RAW DATA SO AS THE DOMAIN EXPERALTA, YOU KNOW CERTAIN — EXPERT, YOU KNOW THAT CERTAIN ARE IMPORTANT. A STRING COLUMN, FOR EXAMPLE, OR A CATEGORY COLUMN. FROM THE DOMAIN PERSPECTIVE RIGHT. SO WHAT WE HAVE SEEN IS WHEN YOU LOOK AT THE FEATURE DATA, THE DATA SET SIZE DECREASES, EVEN THEY YOU MIGHT HAVE TERABYTES OF RAW DATA, WHEN YOU GET FEATURE DATA SET, IT’S SIGNIFICANTLY SMALLER AND THEN WE ALSO SUPPORT SUB SAMPLINGS, SO IT’S LIKE AS LONG AS YOUR DATA SET IS A PRESENTED SAMPLE OF WHATEVER PROBLEM YOU’RE TRYING TO SOLVE, YOU CAN GET GOOD RESULTS. SO THESE ARE SOME OF THE THINGS. AND WE DO HAVE PLANS TO SUPPORT EVEN SPARK DATA FRAME, IT’S NOT THERE RIGHT NOW. SO RIGHT NOW, IT’S BINDERS NOW, COMING BACK TO THE EXPERIMENT THIS IS WHAT YOU WOULD DO. SO BASICALLY, WHEN YOU’RE RUNNING IT ON DATABRICKS, YOU PROVIDE HOW MANY YOU CAN RUN, AND THIS MEANS HOW MANY WORKER NODES YOU HAVE. SO BECAUSE THIS STREAMING IS HAPPENING IN — SO AUTOMATED ML IS LOOKING AT YOUR DATA SET AND TRYING TO DO PREPROCESSING MODEL SELECTION AND HYPER PARAMETER TUNING, THAT IS WHAT IS HAPPENING. IN THIS — IF YOU HAVE EIGHT WORKER NODES, IT CAN DO EIGHT PARALLEL TRAININGS AND THAT’S HOW YOU THINK OF IT AND THERE ARE THE PARAMETERS LIKE EXIT SCORE AND THINGS LIKE THAT AND SO THAT’S — IT’S THAT SIMPLE IN TERMS OF THE CONFIG. NOW, LET ME SUBMIT THE TRAINING. GO TO SUBMIT IT. AND MURPHY, HE IS OKAY, SHOULD BE FINE. SO I’M GOING TO SUBMIT THE EXPERIMENT. AND WHAT IS HAPPENING RIGHT NOW, IS IT’S GOING TO TAKE THE DATA SET, WHICH I SHOWED YOU EARLIER, AND THEN, GOING TO TRAIN IT ON THIS DATABRICKS CLUSTER, SO WE ARE NOT USING THE DRIVE NODE, BUT WE ARE USING THE WORKER NODES THAT’S HOW IT HP HAPPENS. I CAN SHOW YOU THE RESULT FROM MY PREVIOUS TRAINING. THE FIRST STEP IT’S GOING TO DO IS CREATE THE SETUP RUN, SETUP IS ALL OF THIS PREPROCESSING THAT I MENTIONED. AND THEN IT’S GOING TO START THE TRAINING. SO CERTAIN MACHINE LEARNING ALGORITHMS EXPECT TO INPUT DATA IN A CERTAIN FORMAT AND WE FIGURED OUT ALL OF THOSE THINGS FOR YOU AS WELL. IN TERMS OF WHAT OTHER PREPROCESSING WE DO, WE HAVE A NICE DOCUMENTATION, SO — WHERE DID IT GO — WE DO SOME OF THESE THINGS LIKE DROPPING AND MISSING VALUE — AND GENERATE FEATURES, DAT COLUMN, CAN WE KEEP ON ADDING MORE CAPABILITIES IN HERE. >> AND ONE QUESTION I HAVE IS THAT WITHIN WHAT KIND OF FRAMEWORKS SUPPORTED IN AUTOMATED ML? IF MAYBE YOU CAN GIVE US AN IDEA OF THE LIST OF ALGORITHMS THAT SUPPORT THAT MIGHT HELP US UNDERSTAND? >> YEAH, YEAH, THAT’S A GREAT QUESTION. SO ON THE ALGORITHMS, WHEN YOUING AUTOMATED ML ON DATABRICKS OR ANY OTHER MODE, WE SUPPORT THESE ALGORITHMS WHICH ARE THERE. IF YOU HAVE YOUR FAVORITE ALLEGE RHYTHM, YOU CAN RIGHT LIST IT OR IF YOU KNOW A CERTAIN THING DOESN’T WORK, YOU CAN BLACKLIST IT. THAT’S HOW IT IS. AND WE KEEP ON, AGAIN, ADDING THINGS TO IT. BUT THIS IS THE CURRENT LIST. LET’S GO BACK TO THE EXPERIMENT WHICH WE CREATED. SO WHAT IS HAPPENING HERE, I’M SUBMITTING THIS FROM AZURE DATABRICKS TO MY AZURE MACHINE LEARNING WORK SPACE. AND THEN IT IS DOING ALL OF THE TRAINING. OKAY. AND THE TRAINING IS HAPPENING IN PARALLEL SO I’LL GO BACK. THE TRAINING IS HAPPENING OVER HERE. AND YOU CAN SEE RUNNING. SO WHAT IT DOES IS NOW, IT IS USING MY WORKER NODES AND ACTUALLY STARTING THIS TRAINING YOU CAN SEE WHICH ALGORITHM I HAVE PICKED. LET ME ZOOM IT OUT. YOU CAN SEE THE ALGORITHM, YOU CAN SEE THE PREPROCESSER THAT I’M USING AND THE ASSOCIATED METRIC. I HAVE SUBMITTED 13 RUNS, IT TAKES A FEW MINUTES, 10 MINUTES, WE ARE NOT GO TO SPEND SO MUCH TIME. I’M GOING TO SHOW YOU A PRERUN RESULT FROM

THE SAME WORK SPACE. I RAN IT A FEW MINUTES BACK. THIS IS WHAT IT IS. SO DATABRICKS WAS THE COMPUTE AND AZURE MACHINE LEARNING IS THE SDK RUNNING ON TOP OF IT USING ML CAPABLE. GIVE ME THE RESULT AND THIS RUN OF 29, USING STAGNANT ENSEMBLE MOD SL THE BEST MODEL, IT DID THE PREPROCESSING. I CAN GO AHEAD AND TAKE THIS MODEL AND ACTUALLY START DEPLOYING IT. SO THIS IS THE — I CAN ALSO LOOK AT HYPER PARAMETERS, IF YOU WANT TO, RIGHT. AND THEN, I’M GOING TO START DEPLOYING IT SO THE DEPLOYMENT FLOW, IS AGAIN, YOU ARE TRAINING USING AZURE DATABRICKS, BUT YOU MAY NOT WANT TO OR DO NOT WANT TO DEPLOY IT IN AZURE DATABRICKS, BECAUSE THIS COMPUTE IS RUNNING ALL THE TIME, BUT YOU PROBABLY WANT ANOTHER SCALEABLE COMPUTE, WHICH IS TAKING THE WEB SERVICE. SO YOU TAKE THAT AUTOML MODEL BUT CHANGE TO A KUBERNETES NODE OR KUBERNETES CLUSTER, WE HELP YOU DO THAT. THERE IS A RUN TIME, WHICH WE HELP YOU CREATE. AND WE MANAGE THE SERVICE FOR YOU, SO IF YOU ARE DEPLOYING IT TO AZURE, THEN YOU GET TO SEE ALL OF THE METRICS AND EVERYTHING AROUND IT. OKAY. SO THAT IS HOW IT WORKS. AND THE DOCKER CONTAINER IS PRETTY MUCH STAY PROBABLY LIKE YOU JUST TELL US HOW MANY — HOW MUCH MEMORY AND CPU YOU WANT AND THEN WITH JUST ONE CALL, LIKE SAYING DEPLOY FROM MODEL, WE WILL DO THAT FOR YOU. SO THAT’S HOW THE DEPLOYMENT HAPPENS. I ALREADY HAVE PREDEPLOYED, LET’S SEE WHAT IT’S PREDICTING THIS GOING TO BE MY WEB SERVICE AND THEN I’M GOING SEE IF IT PREDICTS SO I’M SENDING IT SOME DATA. OKAY AND MY ACTUAL DATA WAS SAYING THAT MY DOMAIN AND USE OF LIFE OF THE ENGINE WAS 191 DAYS. AND THE RESULT THAT CAME OUT FROM THE AUTO ML PREDICTED MODEL IS 191 DAYS, WHICH IS REASONABLY GOOD. WE SPENT, 10-15 MINUTES ON SUBMITTING THE TRAINING. NOW, THIS IS A SAMPLE DATA SET. BUT THIS NOTEBOOK IS AVAILABLE ON GITHUB AND ON GITHUB AND YOU CAN ACTUALLY MODIFY IT TO A DAPT TO YOUR SCENARIO. SO BECAUSE PREDICTIVE DOMAINS CAN BE A COMMON THING THAT YOU WOULD HAVE. SO THIS IS HOW WE SAW HOW TO RUN ONE PART OR AUTOMATED ML ON AZURE DATABRICKS, NOW LET US GO TO THE OTHER ONE, WHICH IS ML FLOW. SO HOW MANY OF YOU KNOW ML FLOW? AT LEAST SOME OF YOU. SO ML FLOW HAS CAPABILITIES TO DO TRACKING OFF OF YOUR EXPERIMENT SO WHATEVER YOU ARE RUNNING. RIGHT AND IT’S AN OPEN SOURCE PROJECT AND THEN IT ALLOWS YOU TO CREATE PROJECTS AND THEN IT ALSO ALLOWS YOU TO TRAIN THE MODEL. SO THIS IS JUST AN ALTERNATIVE WAY THAT WE ARE SWITCHING. LIKE WE STARTED THE TALK, WE WANTED TO KEEP IT OPEN SO THAT WE WANTED TO HAVE HELP YOU DECIDE RATHER THAN DECIDING IT FOR YOU. >> THERE’S ONE THING I’M NOT CLEAR ON. SO YOU’RE SAYING THAT ML FLOW IS THE API R. >> UH-HUH >> AND AZURE MACHINE LEARNING IS THE ENGINE UNDERNEATH THE ML FLOW SO WHY WOULD I WANT TO DO THAT OR WHAT’S THE BENEFIT FOR ME TO DO THAT? >> OKAY. YEAH, I THAT’S A GOOD QUESTION, WHY WOULD A WANT TO USE ML FLOW? SO PROBABLY ONE THING IS OF COURSE I NEED TO — IN ADDITION TO THAT, WITH WE HEARD FROM CUSTOMERS IS THEY LIKE TO USE OPEN SOURCE. RIGHT. SO RATHER THAN COMPETING WITH OPEN SOURCE, WE PART OF EMBRACING IT. IT’S JUST ANOTHER SET OF APIS, BUT THE UNDERLYING IS STILL THE SAME AZURE MACHINE LEARNING AS SDK AND THE AZURE MACHINE LEARNING WORK SPACE THAT I SHOWED YOU EARLIER, SO THE CAPABILITIES LIKE THE EXPERIMENT TRACKING, SO YOU SAW IN THE PORTAL I SHOWED YOU THE TRACKING OF EXPERIMENTS WHENEVER YOU RUN IT. AND THE PROJECTS WHICH ARE THERE. AND THE DEPLOYED MODEL SO WHAT YOU SAW EARLIER IN THE AUTOMATED EXAMPLE THAT THE MODEL GETS DEPLOYED, SO WHEN YOU’RE USING ML FLOW, NOW IT CAN USE THOSE THINGS AS WELL SO THE DEPLOYMENT PIECE WAS THERE FOR AROUND SIX MONTHS, BUT WHAT WE LAUNCHED LAST WEEK, YOU TOUCHED ON IT EARLIER WAS THE SPARK SUMMIT, WE LAUNCHED THAT MICROSOFT IS OFFICIALLY JOINING THE ML FLOW OPEN SOURCE PROJECTMENT SO OUR TEAM IS OFFICIALLY PART OF IT. AND WHEN YOU USE ML FLOW IN AZURE, YOU CAN LEVERAGE THE SAME MACHINE LEARNING WORK SPACE AND THE IDEA BEHIND THIS IS YOU CAN USE ML FLOW FOR THINGS THAT YOU LIKE. AND YOU CAN USE OTHER AZURE MACHINE LEARNING CAPABILITIES, BUT INDEPENDENT OF WHICH API YOU USE, ALL OF YOUR TRAINING AND DEPLOYMENT

CAN STILL HAPPEN FROM THE SAME WORK SPACE. >> THE BEST OF BOTH WORLDS >> THE BEST OF BOTH WORLDS AND IT’S A CENTRAL PLACE SO YOU DON’T HAVE TO WORRY ABOUT IT. OKAY. NOW, LET’S GO AHEAD AND LOOK AT THE DEMO FOR THIS. ON HOW TO USE THAT. SO IN THIS CASE, I’M GOING TO SHOW YOU TWO DEMO PARTLY SUNNY, SO ONE IS USING ML FLOW IN AZURE DATABRICKS, AND THE OTHER IS USING ML FLOW IN JUPITER NOTEBOOK ENVIRONMENT. SO IT’S REALLY STRAIGHTFORWARD TO START WITH THIS. SO BASICALLY, YOU HAVE YOUR DATABRICKS CLUSTER, AND IN THAT CLUSTER, BASICALLY, WHAT YOU DO IS YOU INSTALL THE LIBRARY AS A INSTALLABLE LIBRARY, AND THEN, ON THE AZURE MACHINE LEARNING WORK SPACE THAT YOU CREATE, RIGHT? YOU JUST HAVE TO USE TWO LINES OF CODE AND I’LL SHOW YOU THAT. BUT IT’S AS SIMPLE AS THAT. YOU SAID THE URI, YOU DON’T HAVE TO DO ANYTHING ELSE. SO OUR GOAL IS TO MEET YOU ARE, OTHER THAN YOU HAVING TO LEARN SOMETHING ELSE. THESE TWO LINES OF CODE, WHENEVER YOU ARE RUNNING THE ML FLOW EXPERIMENT ON AZURE DATABRICKS, IT’S ACTUALLY LOGGING ALL OF THE RESULTS TO THE SAME AZURE MACHINE LEARNING WORK SPACE. OKAY SO NOW, IN THIS DEMO, WHAT WE ARE GOING TO SHOW IS THE EXAMPLE, HOW MANY OF YOU KNOW AM NEST. IT’S A TRADITIONAL DAT SET WHERE YOU HAVE — THE ORIGINAL PROBLEM STARTED WITH TRYING TO IDENTIFY NUMBERS LIKE THE ZIP CODES FOR POSTAL SERVICES RIGHT, AND THEN IT BECAME DATA SET WHERE BASED ON NUMBERS, YOU ARE TRYING TO PREDICT WHICH NUMBER IT IS. IN THIS CASE, THIS IS A DEEP LEARNING EXAMPLE. OKAY. SO EARLIER, WE SAW AUTOMATED ML, WHICH WAS USING A CERTAIN FRAMEWORKS, NOW, WE ARE GOING TO GO INTO THE DEEP LEARNING PART OF IT. AGAIN, WE ARE STILL IN THE TRAINING DOMAIN. AND THEN DEPLOY. BUT WE ARE GOING TO USE PIE CHARTS ON TOP OF DATABRICKS, AND WE ARE GOING TO USE ML FLOW FOR TRACKING YOUR EXPERIMENTS. AND ML FLOW FOR DEPLOYMENT. BUT PLEASE REMEMBER WHEN WE DOING THAT, IT IS STILL LEVERAGING THE SAME AZURE MACHINE LEARNING SERVICE AND AZURE MACHINE LEARNING WORK SPACE. BECAUSE OF THE INTEGRATION TO WE HAVE THAT OKAY. SO IN THIS CASE, I HAVE AN EXPERIMENT THAT IS CREATED. AND THIS IS ALL PIE CHARTS CODE, WHERE I’M BRINGING IN THE DATA AND HERE, I’M LOGGING THE METRIC, SO I’M TRYING TO LOG THE LOSS, WHICH IS ASSOCIATED WITH THE EXPERIMENT. AND I’M GOING TO RUN THIS. AND THIS IS GOING TO LEVERAGE THIS SAME DATABRICKS CLUSTER SO THIS IS JUST LIKE AUTOMATED ML, WHICH WAS RUNNING ON YOUR DATABRICKS CLUSTER, IN THIS CASE, ML FLOW IS DOING TRAINING ON DATABRICKS CLUSTER, LOGGING THROUGH THE API, BUT IT’S ACTUALLY GETTING REGISTERED THROUGH THE AZURE MACHINE LEARNING WORK SPACE. SO LET’S SEE HOW IT DOES THAT. AND I HAVE THIS EXPERIMENT OPEN. SO THIS ONE. SO THIS IS MY WORK SPACE. AND WHEN IT FINISHES THE TRAINING, I THINK IT’S JUST STARTING, YEAH. SO, YEAH, SO WHEN IT FINISHES THE TRAINING, YOU CAN SEE THE ACTUAL LOGGING HAPPENING IN THE SAME WORK SPACE. SO THIS IS THE SAME WORK SPACE WHERE IN WHICH I SHOWED YOU THE AUTOMATED ML EXPERIMENT AS WELL. RIGHT. SO THERE WAS THIS AUTOMATED ML EXPERIMENT WHICH DID NOT USE ML FLOW. AND THEN THERE IS THIS PIE CHART EXPERIMENT, WHERE I’M USING ML FLOW. SO NOW THINK OF WRITING YOUR OWN SCENARIO IN YOUR ORGANIZATION WHERE THERE ARE DIFFERENT TEAMS OF DATA SCIENTISTS, THEY PROBABLY WANT TO USE DIFFERENT FRAME WORKS, FOR EXAMPLE, ONE TEAM MIGHT WANT TO USE AUTOMATED ML, ANOTHER TEAM MIGHT WANT TO USE DEEP LEARNING WITH ML FLOW, OR SDK, BUT YOU WANT A CENTRAL PLACE WHERE ALL OF THE MODELS ARE TRAINED. SO THIS IS ALLOWING YOU DID DO THAT NOW AND YEAH. SO THEY HAVE STARTED TRAINING I THINK WE SHOULD BE ABLE TO SEE THE RESULTS IN REAL TIME IF I’M NOT — IF MY DEMO GODS ARE WITH ME. IT’S ALWAYS THE NERVOUS MOMENT IT’S RUNNING HERE. AND I JUST MINIMIZE THIS. AND YOU CAN SEE THEY ARE GETTING UPDATED IN REAL TIME. IT’S ACTUALLY DOING THE TRAINING AND IT TAKES AROUND 2-3 MINUTES SO I JUST LEAVE IT THERE. BUT THIS IS REAL TIME LOGGING OF THAT FROM ML FLOW TO AZURE MACHINE LEARNING WEB SERVICE THE SECOND DEMO THAT I’M GOING TO SHOW YOU IS A VARIATION OF THAT SO IN CASE, WHAT I’M GOING TO SHOW YOU IS YOU STILL WAB WANT TO USE ML FLOW, LIKE PROBABLY, LIKE I SAID, YOU LIKE OPEN SOURCE OR SOMETHING

ELSE. SOME OTHER REASON FOR WHICH YOU WANT TO HAVE A ML FLOW COE SYSTEM, BUT YOU WANT TO HAVE AZURE — BUT YOU DON’T WANT TO USE AZURE DATABRICKS NOTEBOOK. I HAVE A JUPITER LAB THAT IS THERE, OR YOU CAN HAVE THE SAME APPROACH IN JUPITER. I HAVE STARTED ML FLOW ON TOP OF IT. AND I’M USING THE NEWLY LAUNCHED NOTEBOOKS MANY, VM, HOW MANY OF YOU KNOW THE NEW PRODUCT WE LAUNCHED THIS WEEK? ONLY A COUPLE OF YOU. SO THIS IS BASICALLY A HOSTED EXPERIENCE FOR JUPITER LAB. SO THE WAY IT WORKS IS IN THE SAME WORK SPACE THAT I SHOWED YOU EARLIER, I’M GOING TO GO BACK, YOU SEE ON THE LEFT-HAND SIDE, THERE IS AN OPTION CALLED NOTEBOOK VMS AND YOU CAN HAVE THIS VM AND YOU CAN PICK THE VM TYPE, SO YOU CAN HAVE A CPU OR GPU BASED VM, YOU CAN PICK ANY OF THE VMS, AND THIS IS GOING TO YOUR SERVER, BUT WE MANAGE THE SERVER FOR YOU. SO THE OUTCOME IS BASICALLY, JUPITER LAB BASE SERVER, WHICH IS RUNNING, AND THEN NOW YOU CAN ENABLE IT FOR PEOPLE TO USE IT. SO THIS IS WHAT WE LAUNCHED THIS WEEK IN PRIVATE PREVIEW. AND I ALREADY HAVE ML FLOW INSTALLED AS PART OF IT. IT COMES PREINSTALLED WITH AZURE ML FLOW SO YOU DON’T HAVE TO DO ANYTHING. SO THE DIFFERENCE HERE IS BECAUSE THIS IS NOT A CLUSTER ENVIRONMENT, RIGHT, THIS IS A SINGLE VIRTUAL MACHINE, AND SINGLE VM NODE, WHICH IS NOW RUNNING THE JUPITER LAB, BUT WE MANAGE IT FOR YOU. SO INSTEAD OF LIKE RUNNING THE TRAINING ON THAT NODE, WE ARE ACTUALLY SUBMITTING IT TO A REMOTE COMPUTE. SO NOW IF YOU CONTRAST THE TWO SCENARIOS WE TALKED ABOUT. IN THE PREVIOUS SCENARIO, WE LEVERAGED THE SAME DATABRICKS CLUSTER, AZURE — WE ARE USING JUPITER LAB, BUT NOW WE ARE SUBMITTING THE TRAINING TO A REMOTE COMPUTE CLUSTER, WHICH IS AN ALTERNATIVE. IF YOU ARE SPARK — BITEN PROBABLY USE AML COMPUTE. BUT THE END RESULTS ARE THE SAME. BECAUSE EVEN THIS ONE IS GOING TO LOG YOUR RESULTS IN THE SAME MACHINE LEARNING WORK SPACE. AND YOU DEPLOY IT FROM THE SAME. OKAY. WE ARE GOING TO SEE THAT. SO TO CODE REMAINS EXACTLY THE SAME IN TERMS OF THE ACTUAL TRAINING BUT WHAT CHANGES IS THIS, IN. I’M GOING TOHO Y– TO SHOW YOU, IT’S A LITTLE BIT OF SCROLLING, I’LL ZOOM IT. THIS PART. SO I HAVE A COMPUTE THAT IS PROVISIONED, AND THIS IS WHAT IS GOING TO ACTUALLY TAKE THE TRAINING. SO WE ARE GOING TO RUN IT. AND AGAIN, I’M AT THE MERCY OF THE DEMO GODS. I’M GOING TO RUN AND THIS I’M GOING TO DO RUN. NOW, WHAT IT IS DOING IS IT IS TAKING THAT DATA SET, BECAUSE THE COMPUTE WAS CREATED, IT IS SUBMITTING IT, IF THE COMPUTE DOESN’T EXIST, THEN IT WILL CREATE THE COMPUTE FOR YOU AND THEN SUBMIT THE TRAINING AND THEN SHUT DOWN THE COMPUTE SO YOU CAN SET THAT MINIMUM NODE 0, MAXIMUM NODE X, BASED ON THE UTILIZATION, IT WILL GO UP AND DOWN IN THIS CASE, LET’S SEE WHAT HAPPENED WHO DID THE TRAINING. TAKES — YEAH SO I HAVE THE TRAINING NOW. AND WHEN I OPEN THIS, IT NO LONGER GOES TO THE PORTAL. BUT WE HAVE — WE BROUGHT THE PORTAL INSIDE OF YOUR JUPITER LAB EXTENSION. INSIDE OF YOUR JUPITER LAB ENVIRONMENT ITSELF SO NOW WE CAN SEE IN REAL TIME THE SAME TRAINING THAT YOU SAW EARLIER, NOW IT’S HAPPENING HERE, AND YOU’RE ABLE TO SEE THE RESULTS HERE. BUT YOU GO TO THE AZURE PORTAL AND YOU CAN STILL SEE THE SAME. SO IT’S INTEGRATING EVERYTHING. SO YOUR ML LIFECYCLE OF DATA PREPARATION, WHICH PROBABLY STARTED WITH AZURE DATABRICKS, DOING THE DATA CLEANUP AND THE PROCESSING OF YOUR DATA JOINING AND THEN USING TRAINING WITH EITHER AML COMPUTE, AZURE DATABRICKS NOW IT’S GENERATING MODELS THAT ARE ALL LOGGED IN THE SAME WORK SPACE. AND ONCE YOU HAVE THAT, YOU CAN ACTUALLY NOW DEPLOY IT FROM HERE SO THIS IS THE SAME APPROACH, AGAIN, RIGHT, BUT INSTEAD OF NOW THE AUTOMATED ML OR SOME MANUAL TRAINED MODEL WE ARE DEPLOYING THE SAME MODEL AND THE REST OF THE FLOW IS THE SAME. YOU REGISTER THE MODEL, CREATE THE DOCKER AND THEN DEPLOY IT TO — AND THEN THE RESULT, LET’S SEE WHAT IT PREDICTS. SO I’M GOING TO — SO IN THE DATA SET, I HAVE 1, 000 IMAGE, I’M GOING TO PICK IMAGE LIKE SAY TODAY’S DATE, RIGHT, SO 508, HOW ABOUT THAT? SO I’M GOING TO JUST TAKE THE IMAGE NUMBER 508,

LET’S SEE WHAT THAT IS, 6. THIS IS THE DATA SET. I HAVE THIS MODEL DEPLOYED USING AZURE MACHINE LEARNING AND ISLE GOING TO SEE WHAT IT PREDICTED IT PREDICTED RIGHT. SO IF MY MODEL HAD DONE SOME GOOF UP HERE, IT WOULD HAVE BEEN A TERRIBLE MODEL, BECAUSE I MEAN, THE RESEARCH THAT HAS BEEN DONE ON THE DATA SET, PEOPLE ARE PREDICTING MORE THAN 99 ACCURACY, SO, YEAH, BUT IT’S DOING WELL. NOW, YOU CAN SEE THE SAME RESULT IN THE WORK SPACE AS WELL. SO IF YOU GO TO THE WORK SPACE, WHICH I HAD EARLIER, THE SAME EXPERIMENT, I SHOULD BE RUNNING HERE, SO THIS IS THE PIE TORCH EXAMPLE, AND IT COMPUTED JUST NOW, BUT THIS WAS THE SCRIPT RUN YOU CAN SEE THE SAME RESULTS IN EVERYTHING. ONE SINGLE PLACE TO DO EVERYTHING. SO LET ME SUMMARIZE WHAT WE SAW IN THE DEMO. IN THE THREE DIFFERENT DEMONTHS THAT WE SHOWED YOU, WE SHOWED HOW TO HAVE YOUR MACHINE LEARNAL LIFECYCLE, RIGHT FROM USING DATABRICKS FOR DATA REPARATION, USING AUTOMATED MACHINE LEARNING OR MANUAL MACHINE LEARNING TRAINING, TO TRAIN THE MODEL. AND AGAIN, WE SAW OPTIONS LIKE EITHER USING AZURE DATABRICKS AS THE COMPUTE OR EVEN USING AML AS THE COMPUTE, YOU’RE STILL TRACKING, LOGGING, REGISTERING, WHATEVER YOU WANT TO CALL IT, IN THE SAME MACHINE LEARNING WORK SPACE AND THEN YOU ARE READY TO DEPLOY IT TO PRODUCTION FROM THERE. RIGHT. SO IF YOU DON’T KNOW DOCKER, LIKE I DON’T KNOW DOCKER, I COULD CITY DEPLOY THE MODEL, SO YEAH, WE SIMPLIFY IT FOR YOU, AND ONCE IT’S DEPLOYED IN AZURE, YOU CAN ACTUALLY TRACK THE METRICS, SO THE LATENCY, HOW MANY USERS ARE ACTUALLY HITTING THE WEB SERVICE AND THE WEB SERVICE SCALES AUTOMATICALLY SO IF LIKE SAY I WORK WITH RETAIL CUSTOMERS MANY TIMES, SO BLACK FRIDAY AND ALL OF THOSE THINGS, AND SO THEY DON’T HAVE TO WORRY ABOUT IT, THEY CAN HAVE THIS CONFIGURATION SET UP, SO THE SERVICE WILL AUTOMATICALLY SCALE IN A FEW MINUTES, SO SCALE UP OR SCALE DOWN, BASED ON THE TRAFFIC THAT IS HITTING IT. AND YOU CAN ACTUALLY — YOU WANT TO DRIFT ANALYSIS, SO THE ONCE THE MODEL HAS BEEN DEPLOYED OVER THERE, YOU CAN ACTUALLY HAVE TRIGGERS FOR THAT YOU KNOW WHEN TO RETRAIN IT AND YOU CAN ALSO AUTOMATE THE RETRAINING PART BY INTEGRATING IT WITH PIPELINES. SO WE ARE NOT COVERING THAT PART, THAT WAS ANOTHER SESSION THAT WAS I THINK YESTERDAY OR TODAY. BUT HOW TO ORCHESTRATE THIS ENTIRE WORKFLOW IN TERMS OF LIFECYCLE. SO THIS IS — THIS IS WHAT WE HAVE IN TERMS OF THE DIFFERENT APPROACHES WHEN USING AZURE DATABRICKS CAN WE WERE TRYING TO ANSWER THE QUESTIONS, OKAY, IS DABS COMPETING WITH AZURE MACHINE LEARNING SDK, I GET THIS QUESTION A LOT WHEN I WORK WITH CUSTOMERS, SO I HOPE YOU ALSO PROBABLY AGREE WITH ME THAT THE SANS — ANSWER IS NO. WE ARE INTEGRATING IT TOGETHER. IT’S A BETTER TOGETHER STORY, RATHER THAN ONE VERSUS THE OTHER. I HAND IT OVER TO PREMAL TO SUMMARIZE AND >> THANKS, PARASHAR FOR A GREAT SET OF DEMOS, IN SUMMARY, THE THREE THINGS WE WANT YOU TO TAKE BACK WITH YOU, THE FIRST THING IS, THE BETTER TOGETHER STORY BETWEEN AZURE DATABRICKS AND AZURE MACHINE LEARNING AS PARASHAR WAS SAYING, IT’S A COMBINATION THAT SORT OF MAKES YOU BUILD THESE COMPELLING APPLICATIONS. THE OTHER THING IS THAT THE IMPORTANCE OF THIS AZURE ML WORK SPACE AS THE SINGLE POINT WHERE YOU SORT OF COLLECTING ALL OF YOUR EXPERIMENTS AND ALL OF YOUR MODELS, SO FOR ALL OF YOUR ML AND AZURE, THE AZURE ML WORK SPACE IS THE SIMPLE PLACE WHERE YOU’RE COLLECTING MODELS AND REGISTERING MODELS THROUGHOUT THE ORGANIZATIONS, SO THAT’S KIND OF THE CENTRAL PIECE AND THAT’S A GREAT VALUE PART OF AZURE MACHINE LEARNING HERE, ALSO, THAT’S WHERE THE ML FLOW AZURE MACHINE LEARNING SORT OF INLT GRAGS MAKES A LOT OF SENSE THERE, TOO. SO KEY UPDATES IS THE GOOD NEWS IS, A BUNCH OF OTHER NEW FEATURES, YOU USE IT START GIVING US FEEDBACK. AND ML FLOW WITH AZURE ML IS JUST RELEASED LAST WEEK AT — I THINK LAST WEEK AT THE SPARK SUMMIT. AGAIN, YOU WILL SEE A LOT OF DEVELOPMENTS IN THIS INTEGRATION, SO WE ARE CONTINUING TO STRENGTHEN THE BETTER TOGETHER STORY ACROSS ALL OF THESE PRODUCTS, SO THAT, AGAIN, CUSTOMERS GET THE BENEFIT OF ALL OF THESE WORKING TOGETHER SEAMLESSLY. AND THE LAST THING IS THE AZURE MACHINE LEARNING WORK SPACE INTEGRATION WITH AZURE DATABRICKS WORK SPACE, A LOT OF THE CUSTOMERS ARE USING AZURE DATABRICKS AND THEY WANT TO USE AZURE MACHINE LEARNING, AND WHAT WE ARE TRYING TO DO IS MAKE THE EXPERIENCE REALLY SIMPLE. RIGHT NOW, IF YOU USE IT

WITH THIS INLT — INLT — INTEGRATION THE WHOLE STORY IS SIMPLER, YOU HAVE TO WRITE FEWER LINES OF CODE, YOU CAN GET UP AND RUNNING WITH AZURE ML, MUCH, MUCH MORE QUICKLY, DUE TO THIS INTEGRATION, AND WE’D LOVE YOUR PARTICIPATION IN THE PRIVATE PREVIEW FOR THE NEW FEATURES, PLEASE REGISTER AND THEN YOL WE CAN DEFINITELY REACH OUT TO YOU AND HELP YOU ON BOARD ON TO THIS. >> I JUST HAD ONE POINT TO ADD. SO AUTOMATED ML, IT HAS CAPABILITIES LIKE MODEL EXPLAINABILITY AND FORECASTING AND ALL OF THOSE THINGS, WHICH WE DID NOT SHOW IN OUR SESSION, BUT THAT’S WHY THE DRUM ROLL IS THERE. BASICALLY, WHEN WE SAY GA, NOW, IT’S PRODUCTION READY FOR YOU AND YOU CAN LEVERAGE ALL OF THE FEATURES ON AZURE DATABRICKS, WHICH WAS NOT THE CASE. WE WERE IN PREVIEW FROM THE PAST FIVE MONTHS, DECEMBER — DECEMBER IS WHEN WE LAUNCHED THE FIRST TIME, THE PUBLIC PREVIEW FOR OUR AUTOMATED ML ON AZURE DATABRICKS AND IT MADE THIS WEEK, WE ARE DOING GA, IT IS PRODUCTION READY FOR YOU WITH THE INTEGRATION THE REMAINING FEATURES ARE THE ONES FOR WHICH — YOU NEED TO PROVIDE YOUR AZURE SUBSCRIPTION AND PROBABLY YOU NEED AN NDA IF YOU DON’T. YEAH, THANK YOU. >> THERE’S SOME ADDITIONAL RESOURCES THAT YOU CAN LEVERAGE TO KNOW MORE OF WHAT WE HAVE JUST COVERED. AND FINALLY, YOU KNOW, ANY QUESTIONS FROM THE AUDIENCE, ANYTHING THAT WE CAN HELP ANSWER, I THINK YOU WILL NEED TO COME TO THE MIC FOR — SINCE IT’S BEEN RECORDED >> YEAH. SO WE STILL HAVE FIVE MINUTES, SO YEAH, ANY QUESTIONS, WELCOME >> WE WILL BE OUTSIDE AFTER THE TALK IF YOU GUYS HAVE MORE QUESTIONS, WE’LL BE HAPPY TO. >> SURE. >> THE MIC FOR A QUESTION. >> IS IT ON? >> YES. >> YEAH. >> [ INAUDIBLE ] . >> THE QUESTION IS: THE ML FLOW WITH AZURE MACHINE LEARNING AND INTEGRATED WORK SPACE, WHICH ARE IN PRIVATE PREVIEW, GA, IS THAT CORRECT? NORMALLY, THE WAY WE KEEP IT IS WE TRY TO GATHER YOUR FEEDBACK, AND THEN FEW MONTHS IS WHAT WE LOOK AT. SO TO GIVE YOU AN EXAMPLE, AUTOMATED ML ON DATABRICKS, IT TOOK US FIVE MINUTES TO GA, I’M HOPING THIS WILL BE SHORTER, BUT HOW GOOD THE CUSTOMERS ARE THINKING, SO NORMALLY, IT’S LESS THAN SIX MONTHS, BUT IT VARIES, YEAH. >> [ INAUDIBLE ]. >> LET ME REPEAT THE QUESTION FOR THE AWED UDIENCE. THERE’S CONFUSION BETWEEN ML FLOW AND ML WORK SPACE? SO THE IDEA IS THAT ML FLOW IS THE API ML FLOW IS THE EXPERIMENTATION BUT UNDER THE COVERS, SETTING THOSE TWO PARAMETERS IS AML HELPS YOU TO LOG ALL OF THE ML FLOW, THE EXPERIMENTS, IN THE CENTRAL AML WORK SPACE. IF YOU LIKE ML FLOW, YOU CAN USE IT FROM ANYWHERE YOU WANT. SO YOU CAN USE IT ON THE DESKTOP, YOU CAN USE IT IN OTHER SYSTEMS BUT THEN USING A. M. , YOU’RE ABLE TO WRITE TO THE ML FLOW — IN THE CENTRAL WORK SPACE FOR YOU TO BE ABLE TO LEVERAGE >> YEAH. NO, THIS WAS GREAT. AND I JUST WANT TO — WE SHOWCASE THE ML FLOW, BECAUSE WE WERE TALKING ABOUT AZURE DATABRICKS, AND A LOT OF AZURE DATABRICKS CUSTOMER ASK US WHAT IS THE ML FLOW STORY? OTHERWISE, EVERYTHING THAT WE SHOWED CAN RUN PURELY WITH AZURE MACHINE LEARNING SDK, YOU DON’T NEED ML FLOW IF YOU DON’T WANT TO. WE SHOWED THAT, OKAY, IF YOU ARE USING ML FLOW, YOU CAN STILL LEVERAGE AZURE MACHINE LEARNING >> THANK YOU. >> I SHOW THE DEPLOYMENT MODEL — THE DEPLOYED MODEL — [ INAUDIBLE ] THE BACK OF THE PROCESSER — >> SO THE QUESTION IS: WHEN YOU HAD THE AUTO ML TRAINED MODEL — THE ANSWER IS YES, SO IF WE DID SOME PREPROCESSING AS PART OF THE TRAINING, YOU DON’T HAVE TO WORRY ABOUT IT IN — IN THE DEPLOYMENT BUT IF YOU DID SOMETHING OUTSIDE OF WHAT AUTO ML DOES. BUT WE TAKE CARE OF IT AUTOMATICALLY. >> YES, IT GETS AUTOMATICALLY. >> ANY OTHER

QUESTIONS? YEAH. >> YEAH, SO THIS FEEDBACK LINK IS NOT JUS FOR OUR SESSION, IT’S FOR ALL AZURE. SO YOU CAN GIVE US ANY FEEDBACK, WE MONITOR IT. IT COULD BE FEATURES, IT COULD BE BUGS, WHATEVER YOU WANT TO, SO WE HAVE THAT. >> ONE LINK THAT WE DON’T HAVE THERE IS THE AZURE DATABRICKS DOCUMENTATION, SO IT’S DOCS. AZURE DATABRICKS NET, IT’S REALLY SIMPLE. THAT’S WHERE ALL OF THE AZURE DATABRICKS DOCUMENTATION IS, HAVE A LOOK AT THAT ON SORT OF — >> WE HAVE GOOD DOCUMENTATION, SO IF YOU SEARCH FOR IT, YOU’YOU’LL GET IT. >> WHAT ARE THE — [ INAUDIBLE ] WHERE DO YOU SUPPORT IT LANDING INTO? >> SORRY, REPEAT THE QUESTION? >> WHEN YOU HAVE THE DOCKER CONTAINER BUILT OUT, WHAT KIND OF TOOLING DO YOU SUPPORT? IS THERE ANYTHING BEYOND AKS. >> THE QUESTION IS, WHATEVER DOCKER CONTAINER THE MACHINE LEARNING SERVICE CREATES, WHICH ARE PLACES YOU CAN DEPLOY? YES, SO YOU CAN USE THE SAME AS — AZURE KUBERNETESES SERVICE, WHICH IS THE PRODUCTION LEVEL SERVICE. OR AZURE CONTAINER INSTANCE, WHICH IS YOUR DEV TEST SINGLE SCENARIO, BUT YOU COULD DEPLOY TO ANY DOCKER. SO YOU CAN EXPORT THAT BECAUSE EVERY TIME YOU CREATE A MACHINE LEARNING WORK SPACE, YOU ALSO GET AN AZURE CONTAINER REGISTRY AND FROM THERE, YOU CAN PULL THAT DEPLOY TO ANY DOCKER RUN TIME YOU WANT TO. NOW, THE ONLY DIFFERENCE IS WHEN YOU DEPLOY IT IN A NONMANAGED RUN TIME, THEN YOU ARE RESPONSIBLE FOR TRACKING AND DOING ALL OF THOSE THINGS. IF YOU DEPLOY IT IN AKS IN PARTICULAR, WE ACTUALLY MANAGE IT FOR YOU. >> ANY PLANS TO SUPPORT APP SERVICES AS WELL AS THE RUN TIME? >> I DO NOT KNOW THAT. YOU COULD GIVE US A FEEDBACK. I WROTE IT DOWN AS WELL. BUT, YEAH, RIGHT NOW, I DON’T KNOW. >> THAT’S A GOOD QUESTION. >> GOOD QUESTION. THANKS >> HOW IS IT GOING? SO I’M NEWER TO THIS ML STUFF AND I’M TRYING TO CONCEPTUALIZE DATABRICKS, AND THE ONE THING I’LL I’M GETTING CAUGHT ON IN MY HEAD IS TREATING IT AS A COMPUTING RESOURCE, IS THAT FAIR TO DO THAT? OR IS IT MORE THAN JUST THAT? I’M TRYING TO SOLIDIFY IT IN MY HEAD? MY QUESTION IS, IS SHOULD I BE APPROACHING IT AS A COMPUTING RESOURCE, AND IF SO, IS THERE MORE TO IT THAN THAT? >> UH-HUH. OKAY SO. >> I THINK DO YOU HAVE A FEW MINUTES AFTER THIS? >> YEAH, YEAH, DEFINITELY. >> THIS IS A LONGER DISCUSSION. >> ALL RIGHT. >> I’D LOVE TO TALK TO YOU ABOUT THIS >> YEAH, I’LL HANG AROUND. >> MAYBE AFTER. >> YEAH, I’LL HANG AROUND >> OKAY. >> MOST OF THE TIME WHEN WE SAY DEPLOY MACHINE LEARNING MODEL, WE ARE TALKING ABOUT DEPLOYING IT AS A WEB SERVICE. >> YES. >> WHAT IF I WANT TO USE THE MO MO MODEL DIRECTLY IN C SHARP PROGRAM, HOW DO I DO THAT? >> YEAH, SO THE QUESTION IS, INSTEAD OF A WEB SERVICE, IF YOU WANTED TO DEPLOY THE MODEL AS SOMETHING — SO WE DO SUPPORT ONE AS A BATCH SCENARIO, SO YOU CAN DO THAT. AND WE ALSO SUPPORT ML NET SCENARIO. INSTEAD OF USING THE BITEN BASED SDK, WHICH YOU SAW TODAY, ON DATABRICKS, YOU CAN USE ML — YOU HAVE TO USE A DIFFERENT OPTION, BUT YOU CAN DEPLOY IT IN YOUR NET ENVIRONMENT AS WELL. THE SAME AS SDK, AND WE DID A — I THINK THE LAUNCH HAPPENED YESTERDAY, SO THERE MIGHT BE SOME BLOGS OR THERE ARE SOME DOCS WHICH I’M AWARE OF — I DON’T EXACTLY RECALL WHEN, BUT THERE WAS SOME DOCS DURING BUILD, WHICH WE LAUNCHED ML BUILD, WHEN C SHARP, AND BENEFIT FROM WHATEVER WE SHOWED YOU. >> THE MODEL WOULD BE SCORING FILE, OR SOME KIND OF FORM. >> YEAHS, BL — AND HAAS HOW Y– THAT’S HOW YOU CAN GET IT. >> ML. NET, ABLE TO USE THAT SCORING FILE AND — >> YES, IN FACT, EVEN IN THIS ONE, ONCE THE MODEL IS DEPLOYED AS A WEP SERVICE, YOU ARE NOT RESTRICTED TO BITEN. FROM C SHARP OR FROM JAVA OR ANYWHERE. IT’S INDEPENDENT. BUT WHAT I WAS TRYING TO ADDRESS IS IF YOU WANT TO DO THE TRAINING, ITSELF, IN C SHARP INSTEAD OF BITEN, WE STILL SUPPORT IT WITH ML. >> ALL RIGHT. >> ANOTHER QUESTION, ALL OF YOUR DEMO PROGRAM, ARE THEY AVAILABLE ON THE GITHUB? >> YES >> DO YOU HAVE THE LINKS? >> YEAH, YEAH. I CAN SHOW THAT. SO THE ML FLOW DEMO IS NOT ON GIB — GITHUB BUT THE AUTOMATED ML DEMO IS THERE ON MY GITHUB, SO THIS IS MY GITHUB, THE URL, THIS IS ONE, AND THEN WE HAVE A LOT OF NOTEBOOKS IN — WE

PUBLISH IT IN AZURE ITSELF. SO WE PUBLISH A LOT OF NOTEBOOKS OVER HERE. SO YOU CAN SEE HOW TO USE THE AZURE ML. AND IN THAT, NOT ONLY AUTO ML, EVERYTHING IS PUBLISH FD OVER HERE. SO IT’S ALL PUBLIC AND IT’S AVAILABLE FOR YOU TO USE. AND YEAH, AND IF YOU GO TO THAT AZURE DOCS LINK THAT WE SHOWED IN THE REFERENCE, IT WILL ALSO TAKE YOU THERE. >> THERE’S ALSO SORT OF NOTEBOOKS THAT — ARE THERE AZURE MACHINE LEARNING INTEGRATION WHICH YOU COULD NOT COVER IN THIS TALK — >> WE KEEP UPDATING IT ALL THE TIME. >>