Svoboda | Graniru | BBC Russia | Golosameriki | Facebook
SlideShare a Scribd company logo

Discover the Unseen: Tailored Recommendation of Unwatched Content

•Download as PPTX, PDF•
•

The session shares how JioCinema approaches ""watch discounting."" This capability ensures that if a user watched a certain amount of a show/movie, the platform no longer recommends that particular content to the user. Flawless operation of this feature promotes the discover of new content, improving the overall user experience. JioCinema is an Indian over-the-top media streaming service owned by Viacom18.

1 of 21

1

Discover the Unseen: Tailored Recommendation of Unwatched Content Harshit Jain & Charan Kamal

2

Harshit Jain ■ Software Engineer at JioCinema ■ Works with the Personalization team ■ 5 years of experience building large scale distributed systems ■ Passionate about technology, a dedicated Golang enthusiast, and an avid traveller Your photo goes here, smile :)

3

■ About JioCinema ■ Recommendations - how do we ensure freshness? ■ Scale and challenges ■ ScyllaDB on steroids with Bloom Filters Agenda

4

About JioCinema JioCinema is an OTT streaming platform that offers free and subscription-based video on demand and live streaming content. ■ More than 10M daily active unique users. ■ Streams prominent cricket tournaments, notably IPL, widely acknowledged as the most-watched cricket league worldwide. ■ Home to one of the biggest Football leagues in Europe (LaLiga). ■ Offers Video-on-Demand (VOD) content in more than 10 Indian languages.

5

Unlocking Engagement The Power and Significance of “Personalization”

6

Let’s look at some examples

7

Optimizing Personalization Challenges with Managing Redundancy in Personalization

8

The Challenge! ■ Customer has already watched “House of the Dragon” ■ Recommending "House of the Dragon" in the personalized tray constitutes an inefficient allocation of valuable real estate and resources

9

The Solution:“Watch Discounting” Watch Discounting refers to the practice of removing content that customers have already watched ■ Importance ■ Efficient Real Estate Utilization ■ Improved Content Discovery ■ Enhanced Customer Experience

10

Watch Discounting in action ! After Watch Discounting “House of Dragons”

11

Navigating Technical Challenges of Watch Discounting

12

Hurdles in Fueling Watch Discounting ■ Scale: Managing already-watched content for more than 10M daily active customers poses a considerable challenge. ■ Concurrency: Handling user interactions becomes challenging with an average of 20 million of concurrent users during high concurrency events. ■ Latency: Maintaining a smooth user interaction on JioCinema necessitates keeping latency within SLA, regardless of scale and concurrency challenges.

13

Charan Kamal ■ Software Engineer at JioCinema ■ Works with the Personalisation Team ■ 6 years of experience in developing expansive distributed system at scale ■ Passionate about highly scalable system, loves Go and Java, in free time you can find me either playing fifa or drawing

14

Bloom filters to the rescue! ■ Bloom filters are space-efficient probabilistic data structures designed for rapid membership lookup in a set. ■ What makes them suitable for our use case ? ■ Trade-off with False Positives Acceptable: Given that it's a recommendation system, tolerating a few false positives is acceptable, as it allows for more efficient memory usage without significant compromises. ■ Reduced Storage Requirements: In our context, the critical aspect is minimizing storage due to the presence of a very large dataset.

15

Why in-memory or redis Bloom Filters won’t work ■ In-memory Bloom filters do exhibit relatively lower latency, they come with drawbacks that are not conducive to this particular use case which are: ■ Data Volatility ■ Replicating the data across all the application pods is costly ■ Bloom filters in Redis and ScyllaDB share similar purposes, distinct differences make Redis unsuitable for this specific use case: ■ Cost consideration - Redis charges us on each operation whereas scylla has a fixed cost.

16

Serving Fresh Content!

17

■ ScyllaDB tuning ■ Using Mini pods ■ Using High cardinality of partition key ■ Using TTLs to make sure older and irrelevant data gets removed ■ Using LOCAL_QUORUM to read from local data centres only ScyllaDB + Bloom filters for the win!

18

Time for some Statistics

19

■ Statistics from a recently concluded high scale event ■ At the onset of the match, a hockey stick pattern of requests was noted. Here Comes Scale!

20

■ Latency within SLA ■ Healthy CPU utilisation Scale Handled!

21

Stay in Touch Harshit Jain harshit.jain@viacom18.com https://github.com/modestlearner https://www.linkedin.com/in/harshit-jain-911003148/ Charan Kamal charan.kamal@viacom18.com https://github.com/ckstudy2021 https://www.linkedin.com/in/charan-kamal- 1303ba16a

More Related Content

Discover the Unseen: Tailored Recommendation of Unwatched Content

  • 1. Discover the Unseen: Tailored Recommendation of Unwatched Content Harshit Jain & Charan Kamal
  • 2. Harshit Jain ■ Software Engineer at JioCinema ■ Works with the Personalization team ■ 5 years of experience building large scale distributed systems ■ Passionate about technology, a dedicated Golang enthusiast, and an avid traveller Your photo goes here, smile :)
  • 3. ■ About JioCinema ■ Recommendations - how do we ensure freshness? ■ Scale and challenges ■ ScyllaDB on steroids with Bloom Filters Agenda
  • 4. About JioCinema JioCinema is an OTT streaming platform that offers free and subscription-based video on demand and live streaming content. ■ More than 10M daily active unique users. ■ Streams prominent cricket tournaments, notably IPL, widely acknowledged as the most-watched cricket league worldwide. ■ Home to one of the biggest Football leagues in Europe (LaLiga). ■ Offers Video-on-Demand (VOD) content in more than 10 Indian languages.
  • 5. Unlocking Engagement The Power and Significance of “Personalization”
  • 6. Let’s look at some examples
  • 7. Optimizing Personalization Challenges with Managing Redundancy in Personalization
  • 8. The Challenge! ■ Customer has already watched “House of the Dragon” ■ Recommending "House of the Dragon" in the personalized tray constitutes an inefficient allocation of valuable real estate and resources
  • 9. The Solution:“Watch Discounting” Watch Discounting refers to the practice of removing content that customers have already watched ■ Importance ■ Efficient Real Estate Utilization ■ Improved Content Discovery ■ Enhanced Customer Experience
  • 10. Watch Discounting in action ! After Watch Discounting “House of Dragons”
  • 12. Hurdles in Fueling Watch Discounting ■ Scale: Managing already-watched content for more than 10M daily active customers poses a considerable challenge. ■ Concurrency: Handling user interactions becomes challenging with an average of 20 million of concurrent users during high concurrency events. ■ Latency: Maintaining a smooth user interaction on JioCinema necessitates keeping latency within SLA, regardless of scale and concurrency challenges.
  • 13. Charan Kamal ■ Software Engineer at JioCinema ■ Works with the Personalisation Team ■ 6 years of experience in developing expansive distributed system at scale ■ Passionate about highly scalable system, loves Go and Java, in free time you can find me either playing fifa or drawing
  • 14. Bloom filters to the rescue! ■ Bloom filters are space-efficient probabilistic data structures designed for rapid membership lookup in a set. ■ What makes them suitable for our use case ? ■ Trade-off with False Positives Acceptable: Given that it's a recommendation system, tolerating a few false positives is acceptable, as it allows for more efficient memory usage without significant compromises. ■ Reduced Storage Requirements: In our context, the critical aspect is minimizing storage due to the presence of a very large dataset.
  • 15. Why in-memory or redis Bloom Filters won’t work ■ In-memory Bloom filters do exhibit relatively lower latency, they come with drawbacks that are not conducive to this particular use case which are: ■ Data Volatility ■ Replicating the data across all the application pods is costly ■ Bloom filters in Redis and ScyllaDB share similar purposes, distinct differences make Redis unsuitable for this specific use case: ■ Cost consideration - Redis charges us on each operation whereas scylla has a fixed cost.
  • 17. ■ ScyllaDB tuning ■ Using Mini pods ■ Using High cardinality of partition key ■ Using TTLs to make sure older and irrelevant data gets removed ■ Using LOCAL_QUORUM to read from local data centres only ScyllaDB + Bloom filters for the win!
  • 19. ■ Statistics from a recently concluded high scale event ■ At the onset of the match, a hockey stick pattern of requests was noted. Here Comes Scale!
  • 20. ■ Latency within SLA ■ Healthy CPU utilisation Scale Handled!
  • 21. Stay in Touch Harshit Jain [email protected] https://github.com/modestlearner https://www.linkedin.com/in/harshit-jain-911003148/ Charan Kamal [email protected] https://github.com/ckstudy2021 https://www.linkedin.com/in/charan-kamal- 1303ba16a