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

Ikyu’s Marketing: Over One Billion Yen in Month...

Ikyu’s Marketing: Over One Billion Yen in Monthly Transactions

Masaru Kanno (Ikyu / Data Science Department / Data Platform Engineer)

https://tech-verse.me/ja/sessions/218
https://tech-verse.me/en/sessions/218
https://tech-verse.me/ko/sessions/218

Tech-Verse2022

November 17, 2022
Tweet

More Decks by Tech-Verse2022

Other Decks in Technology

Transcript

  1. Agenda - About Ikyu - Understanding Users’ Behavior - Cases

    - Building Data Infrastructure for High-demand Data Load - Communication with Users - Future Direction - Summary
  2. About Ikyu Corporate Culture ・Ideas and discussions are always based

    on user-first - User-first ・Promote data democratization so that all employees can execute SQL for data analysis infrastructure ・CEO, by himself, create weekly report based on data analysis infrastructure ・Other executives extract and utilize data by themselves for business promotion - Data-driven
  3. Understanding Users’ Behavior Implementation background - Non-face-to-face business due to

    the Internet-based corporation - Users’ behavior can be recognized only through data - Access data required - Implemented Google Analytics - Challenges after implementation
  4. Understanding Users’ Behavior Challenges ・Data as of yesterday, due to

    daily data linkage - Real-time nature ・About 10% of reservation behavior data is lost - Data loss ・Huge delay in data linkage - Stability for linkage time Lower degree of freshness Lower degree of precision Lower stability
  5. Understanding Users’ Behavior After implementation ・ Enabled behavior logging analysis

    within one second - Real-time nature ・Enabled acquiring 100% of reservation behavior data - Data loss ・No latency for real-time linkage - Stability for linkage time Higher degree of freshness Higher degree of precision Improved stability
  6. Understanding Users’ Behavior Difficulties ・Time required until fully recognized by

    the product side - Cultural aspect ・Huge impact by data loss and latency, as the data is so important ・Early detection of alerts from Datadog by Slack as the metrics monitoring - Enhanced monitoring ・Very little allowance for latency, due to perpetuation against RDB and maintaining real-time policy at the same time - Technical aspect
  7. Cases Difficulties ・Real-time policy requires securing resources at all time

    ・Resolved with SQL Server Resource Governor - Difficult to maintain stability ・One of the functions of SQL Server Enterprise Edition ・Enables control over CPU, memory, physical I/O against specific log-in information - Resource Governor ・Daily monitoring and tuning as needed ・Automatic tuning is the next challenge - Cost performance of coupons
  8. Cases Top page of personalization Best selling accommodations within two

    hours from Tokyo Best selling accommodations with hot springs in Hakone, Atami, and Izu
  9. Building Data Infrastructure for High-Demand Data Load Background ・Showing the

    same result at the morning and in the afternoon ・Want to change from daily execution to hourly execution - To improve freshness of recommendation and personalization ・Takes four hours, even with parallel processing, from creation to deployment on production ・Needs new infrastructure - Computer resource issues
  10. Building Data Infrastructure for High-Demand Data Load Challenges of S3

    synchronization ・Cannot scale parallel processing due to resource restriction on input side ・Serial processing causes huge delay - Want to daily synchronize with S3 without significant delay
  11. Building Data Infrastructure for High-Demand Data Load Resolution ・Enable multiple

    output for single input ・Within ten minutes delay after implementation - Implement embulk-output-multi plug-in
  12. Building Data Infrastructure for High-Demand Data Load Difficulties - Not

    much information about the world - Need to adjust page_size setting of embulk execution argument
  13. Building Data Infrastructure for High-Demand Data Load Difficulties ・Limitation of

    3,500 PUT/COPY/POST/DELETE requests per second and 5,500 GET/HEAD requests per second - Amazon S3 503 Slow Down error ・Verification results showed potential limitation by the Bucket ・Avoid for a process to reach its upper limit by executing with its dedicated Bucket * Not mentioned in the reference - Verification results ・Recommend to retry and execute some time later - Official position ・ Resolved by Lambda execution with USING EXTERNAL FUNCTION - Want to use a use-defined function on Athena
  14. Building Data Infrastructure for High-Demand Data Load After countermeasure ・Provide

    more fresh data in the morning and the afternoon ・Shift from Daily execution to Hourly execution (Processing time: 10 minutes) - Resulted in higher freshness for recommendation and personalization ・Five to ten percent improvement on above-mentioned business achievement - Achievements
  15. Have you ever experienced this? Useful information Useful information Useful

    information 2022/11/01 2022/11/02 2022/11/03 ・ ・ ・
  16. About Communication With Users Communication method ・Mail ・Pop-up ・Browser notification

    ・LINE delivery - Channel ・Issues after implementation - Implementing external services
  17. About Communication With Users Challenges ・Deliver mails of the same

    contents only ・Deliver only through engineers - Democratization of 1 to 1 CRM ・Requires development for each channel ・Requires development for each extension - Multi-channel enablement ・Requires data linkage with CSVs which include personal information - Security risks Lower quality of information Lower productivity Lower scalability Lower security
  18. About Communication With Users Marketing application Search by campaign name

    Status Channel Services Valid Draft Invalid Expired Mail Pop-up Browser notification LINE delivery Upgrading Not set Common Accommodation Yahoo! Travel Restaurants Spa Vic Delivery channel Services Status Campaign Timing Priority Execute on Valid from Valid to Number of delivery Updated on Updated by Common Common Common Valid Draft Draft Search by campaign name Status Channel Services Valid Draft Invalid Expired Mail Pop-up Browser notification LINE delivery Upgrading Not set Common Accommodation Yahoo! Travel Restaurants Spa Vic Delivery channel Services Status Campaign Timing Priority Execute on Valid from Valid to Number of delivery Updated on Updated by Common Common Common Valid Draft Draft
  19. About Communication With Users After countermeasure ・Add useful information for

    users ・Anyone can deliver - Democratization of 1 to 1 CRM ・Easy to scale - Multi-channel enablement ・Internal development leads no external leak - Security risks Higher quality of information Higher productivity Higher scalability Higher security
  20. About Communication With Users After countermeasure - More than fifty

    percent of open rate for mails from Ikyu (Generally between fifteen to twenty-five percent) - Reason for high open rate is not-sending unnecessary mails - Notify useful information for receiving users
  21. Future Direction - Improvement of accommodation search experiences - Reservation

    without searches - Provide administration pages for promotion by service providers
  22. Summary - User-first commitment - Results come along - Flexible

    and agile response enabled by not relying on external services - Resolving issues with data-driven approach and quickly delivering values to users are sources of the growth of Ikyu