Upgrade to Pro — share decks privately, control downloads, hide ads and more …

카카오클라우드 MLOps 활용 방안 소개

kakao
November 01, 2024

카카오클라우드 MLOps 활용 방안 소개

#빅데이터 #MLops #AI #Cloud

카카오클라우드 위에서 AI를 위한 데이터 수집부터 모델 개발, 학습, 튜닝, 서빙까지의 파이프라인 구성과 활용 방안에 대해 소개합니다.

발표자 : evan.ejin
카카오클라우드에서 Data 및 Machine Learning 플랫폼 PM을 담당하고 있는 Evan입니다.

kakao

November 01, 2024
Tweet

More Decks by kakao

Other Decks in Programming

Transcript

  1. .BDIJOF-FBOJOH 0QFSBUJPOT ݠन ۞׬ ѐߊҗ ਍৔ਸ ই਋ܰח दझమ ߂ ೐۽ࣁझ

    ؘ੉ఠ੄ ন੉ ӝೞәࣻ੸ਵ۽ ૐоೞҊ ੓Ҋ ੉ী ٮܲ ੹୊ܻ ੿ઁ ؘ੉ఠ ߡ੹ ҙܻ ١੄ ੘স ࠂ੟بо ૐоೞҊ ੓਺ ٩۞׬ਸ ನೣೠ ࠂ੟ೠ ݽ؛੉ ݆੉ ഝਊغݶࢲ ള۲ ಣо పझ౟ ӒܻҊ ߓನ җ੿੄ ࠂ੟بо ૐоೞҊ ੓਺ (16 /16 516١੄ Ҋࢿמ ஹೊ౴ ੗ਗ੉ ݆੉ ഝਊغݶࢲ ੉ ੗ਗਸ ୭੸ച೧ ࢎਊೡ ࣻ ੓ח ਃҳࢎ೦੉ ૐоೞҊ ੓਺ ؘ੉ఠ ࠂ੟ࢿ ૐо ݽ؛੄ ࠂ੟ࢿ ૐо ஹೊ౴ ੗ਗ ୭੸ച .-0QT ۆ
  2. %BUB0QT .BDIJOF-FBSOJOH'MPX ؘ੉ఠ ࣻ૘ ؘ੉ఠ ੹୊ܻ ؘ੉ఠ ߡ੷׬ ؘ੉ఠ ੿ઁ

    ؘ੉ఠ ߸ജ %BUB-BLF 1JQFMJOF ݽ؛ ѐߊ ݽ؛ ೟ण ೞ੉ಌ౵ۄ޷ఠ ౚ׬ ݽ؛ ಣо ݽ؛ ߓನ ࢲࡂ ࢿמ ݽפఠ݂ সؘ੉౟ Ѿ੿ ݽ؛ ߡ੹ ҙܻ 1JQFMJOF .-0QT ۆ
  3. %BUB*OHFTUJPO 1SFQBSBUJPO 데이터수집및 적재 Infrastructure as a Service Management 1VC4VC

    #FZPOE4UPSBHF4FSWJDF 데이터 ETL )BEPPQ&DP %BUB$BUBMPH 데이터분석및 학습 .BOBHFE,BGLB #FZPOE$PNQVUJOH4FSWJDF #FZPOE/FUXPSLJOH4FSWJDF "OBMZUJDT .- *". .POJUPSJOH $MPVE5SBJM "MFSU$FOUFS .POJUPSJOH'MPX ,VCFGMPX %BUB2VFSZ .-0QTܳ *OHFTUJPO 1SFQBSBUJPO "OBMZUJDT .- ױ҅۽ ա׀׮ݶ .-0QT PO,BLBPDMPVE
  4. Object Storage Pub/Sub Managed Kafka Hadoop Eco Data Catalog Data

    Query ؘ੉ఠ ੋࢎ੉౟ 2 *OHFTUJPO 1SFQBSBUJPO "OBMZUJDT 3 Data 1 )VF 4VQFSTFU ;FQQFMJO ഝਊ ࢲ࠺झ ߂ ো҅ 'MPX %BUB*OHFTUJPO 1SFQBSBUJPO "OBMZUJDTױ҅ ૑ਗ 'MPX .-0QT PO,BLBPDMPVE
  5. +VQZUFS /PUFCPPL 5SBJOJOH 0QFSBUPS ,BUJC ,4FSWF .FUBEBUB 1JQFMJOF ,VCFGMPX $MVTUFS

    (16$164FSWFS/PEFT ܻࣗझ ௪ఠ ҙܻ ࢎਊ੗Ӓܛ ӂ ೠ ҙܻ (16.JH ӝמ ঘ࣌ ۽Ӓ ҙܻ ੗ਗ ݽפఠ݂ ӝמ ,BLBP$MPVE ,VCFGMPX • ਕ௼೒۽਋ ױ҅߹ ਬਊೠ ో ߂ ౵੉೐ۄੋ ઁҕ • ஠஠য়௿ۄ਋٘ ௑ࣛ উীࢲ औҊ ࡅܰѱ ҳࢿ оמ • ੗ਗ ௪ఠ ӂೠ ҙܻ ݽפఠ݂ ١ ݒפ૚ ӝמ ઁҕ ,VCFGMPX ౠ૚ ,BLBP$MPVE ,VCFGMPX੄ ׮নೠ ஹನք౟ٜਸ ഝਊೞৈ .BDIJOF-FBSOJOH'MPXܳ ૑ਗೣ .-0QT PO,BLBPDMPVE
  6. -#"DDFTT-PHؘ੉ఠܳ ഝਊೠ ౟ې೗ ৘ஏ ݽ؛ ҳഅਸ ా೧ दझమ ੉࢚ ૚റܳ

    पदрਵ۽ ఐ૑ೞҊ ؀਽ ࣘبܳ ೱ࢚ दఃҊ੗ ೣ ୶૓ ߓ҃
  7. ౟ې೗ ৘ஏ ݽ؛ ҳഅ 'MPX -#ؘ੉ఠ ਃ୒ ߂ ਽׹ ۽Ӓ

    ౟ې೗ ੿ࠁ ী۞ ߂ ࢚క ௏٘ ֎౟ਕ௼ ࢿמ ݫ౟ܼ ؘ੉ఠ ੹୊ܻ ݽ؛ ѐߊ ݽ؛ ߓನ %BUB$BUBMPH )BEPPQ&DP /PUFCPPL 5FOTPSCPBSE *OHFTUJPO 1SFQBSBUJPO .BDIJOF-FBSOJOH ೞ੉ಌ ౵ۄ޷ఠ ౚ׬ ,4FSWF ,BUJC ഝਊ ৘ઁ
  8. *OHFTUJPO 1SFQBSBUJPO .BDIJOF-FBSOJOH Load Balancer Object Storage Hadoop Eco Load

    & ETL Managed Kafka Pub/Sub Hadoop Eco Kubeflow ML CI/CD Test&Analyze Train Deploy Pre-Process Dashboard Data Catalog Hadoop Eco Serving ഝਊ ৘ઁ ݽ؛ ҳഅ 'MPX߹ ഝਊ ࢲ࠺झ
  9. ݽ؛ ҳഅ 'MPX߹ ഝਊ ࢲ࠺झ -PBE#BMBODFS 0CKFDU4UPSBHF )BEPPQ&DP -PBE&5- .BOBHFE,BGLB

    1VC4VC )BEPPQ&DP ,VCFGMPX .-$*$% 5FTU"OBMZ[F 5SBJO %FQMPZ 1SF1SPDFTT Dashboard %BUB$BUBMPH )BEPPQ&DP 4FSWJOH 2 1 *OHFTUJPO 1SFQBSBUJPO .BDIJOF-FBSOJOH ഝਊ ৘ઁ
  10. parser = argparse.ArgumentParser() parser.add_argument("--end_time", type=str, required=True, help="End time in 'yyyy/MM/dd

    HH:mm:ss' format") args = parser.parse_args() time_format = "%Y-%m-%d %H:%M:%S" end_time = args.end_time start_time = (datetime.strptime(end_time, time_format) - timedelta(minutes=30)).strftime(time_format) spark.sql("USE kclogs") df = spark.sql(""” SELECT FROM_UNIXTIME(UNIX_TIMESTAMP(`time`, 'yyyy/MM/dd HH:mm:ss:SS')) AS `time`, client_port AS source, REGEXP_EXTRACT(client_port, '(.*?)(?=:)') AS source_ip, REGEXP_EXTRACT(client_port, ':(.*)') AS source_port, target_port AS destination, REGEXP_EXTRACT(target_port, '(.*?)(?=:)') AS destination_ip, REGEXP_EXTRACT(target_port, ':(.*)') AS destination_port FROM tb_alb UNION ALL SELECT FROM_UNIXTIME(UNIX_TIMESTAMP(`time`, 'yyyy/MM/dd HH:mm:ss')) AS `time`, client_port AS source, REGEXP_EXTRACT(client_port, '(.*?)(?=:)') AS source_ip, REGEXP_EXTRACT(client_port, ':(.*)') AS source_port, destination_port AS destination, REGEXP_EXTRACT(destination_port, '(.*?)(?=:)') AS destination_ip, REGEXP_EXTRACT(destination_port, ':(.*)') AS destination_port FROM tb_nlb """) filtered_df = df.filter(F.col("time").between(start_time, end_time)) 1SFQBSBUJPOª %BTICPBSE &YUSBDU 5SBOTGPSN -PBE 4QBSL ؘ੉ఠ ੸੤ %SVJE
  11. 2"