Skip to main content
Challenges
of Real-Time
AI at Scale
Felipe Mendes, Vibe Coder
Tim Koopmans, Token Churner
Felipe Mendes, Technical Director at ScyllaDB
+ Published Author on Linux and Databases
+ Uses LLMs to help with college homework (no cheating!)
+ Vibe code all the things, just because… Why not?
Tim Koopmans, Senior Director at ScyllaDB
+ Fought latency and (sometimes) won
+ Trained ML models a gazillion times and still can't get it right
+ Built my very own Bloomberg terminal
Introductions
2
Powering India's top
social media platform
Video recommendation
management
Real-time fraud
detection
Seamless experiences
across content + devices
Network security
threat detection
Content personalization &
recommendation platform
Mobile Growth &
Monetization Platform
Inventory hub for
retail operations
Property listings
and updates
Cryptocurrency
exchange app
Real-time auctions
advertising platform
Predictable performance
for on sale surges
Online gaming ad
targeting
Media streaming
for 45M+ subscribers
Bridging AI to IT Service
Management
Real-time ML-driven
recommendations
Real-time endpoint threat
detection and security
Real-time personalized
recommendations
World leading beauty
platform behind Avon
Real-time AI decisioning
for digital advertisers
AI-centric customer
research platform
Powering Unreal Engine
real-time asset distribution
Real-time interactions
at massive scale
Always-on e-commerce
platform for millions of fans
3
ScyllaDB Users
Agenda + Problems with Scaling AI
+ ScyllaDB for Real-time AI
+ Takeaways
4
How relevant is AI for
your organization today?
6
“Most AI pipelines work at
a prototype level but collapse
under production scale”
Latency Spikes
7
“You can’t dig yourself
out of tail latency”
8
9
Feature Freshness
“Stale features kill accuracy”
10
11
Vector Index Degradation
“Vectors indexes need
maintenance”
12
13
Resource Contention
“You gotta keep’em
separated”
14
15
Model Retraining Impact
“Retraining is inevitable”
16
Coupled Failure Loops
17
Which of the following
challenges do you anticipate
within the next year?
Why ScyllaDB for Real-time AI?
+ Performance-first distributed architecture
+ Highly concurrent, lock-free, multi-writer design
+ Just in time elastic scaling
App
App
App
App
App
App
Active/Active, replicated, auto-sharded
AI Trilemma
20
Performance
Cost Quality
Low recall
No high availability
High memory usage
Multiple replicas
>100 ms tail latency
Vector Similarity Feature Store Highly Scalable Database
within an AI Stack
Dissimilar
Dissimilar
Similar Results
(Matches)
Vector Space
Query
Vector
Fast, Accurate
Matching of Complex Data
Raw Data
Sources
Real-time
Event Streams
ML Training
ML Serving
(Interference)
Central
Feature Store
(ScyllaDB)
Ingestion
Processing
& Training
Serving &
Inference
High Throughput & Low Latency Access
21
AI Umbrella
Vector Similarity Functions
22
SELECT commenter, comment,
similarity_cosine(comment_vector, [0.12, 0.34, 0.56, 0.78, 0.91])
FROM myapp.comments
ORDER BY comment_vector ANN OF [0.12, 0.34, 0.56, 0.78, 0.91]
LIMIT 5;
+ similarity_cosine
+ General semantic search (incl. caching)
+ RAG applications
+ similarity_dot_product
+ Blending scores (eg: e-commerce product search)
+ similarity_euclidean
+ Anomaly detection
+ Image processing
23
ScyllaDB Vector Search
Vector Database
Saved in Searched in
Embeddings Embeddings
Document
Image
Audio
Transformed
into
User’s
Search Query
Transformed
into
1 2
3
Perform Similarity Search
Accurate Search Result
To Provide
24
ScyllaDB Vector Search Architecture
Scaling to 1 Billion Vectors
25
Low Latency, Moderate Recall
Slightly Higher Latency, High Recall
The following params impact the recall %
Tuning recall % @ ScyllaDB
26
Controls the quality of the index during
construction, Higher values → better recall but
slower and more memory-intensive indexing.
efconstruct
(Index Quality)
efsearch
(Search Accuracy vs. Latency)
M
(Links per Node)
Controls how many bi-directional links each node
has in the index graph.
Higher values → more connectivity → better
recall but more memory and slower build time,
also impact search time.
Controls accuracy vs. latency during search,
Higher values → higher recall but slower
queries. Have the highest impact on recall% but
also on query performance / latency (the higher
efsearch the lower QPS/vCPU).
Recall % Recommended M
90% 32
95% 48
99% 64
Recall % Recommended efconstruct
90% 128
95% 256
99% 350
Recall % Recommended efsearch
90% 128
95% 256
99% 500
What is a feature?
https://sharechat.com/blogs/artificial-intelligence/how-sharechat-built-a-scalable-cost-efficient-ml-feature-system 27
Feature Store Pipeline
28
29
https://www.scylladb.com/tech-talk/inside-tripadvisors-real-time-personalization-scylladb-aws/
Recommendation system
+ ML models drive 'For you', Daily
Digests features
+ Hot path
+ Low-latency, high throughput
needs
30
Scaling AI with ScyllaDB
31
Trilemma: Freshness, Latency, Recall
+ Network fan-out gets expensive
+ Top-k candidates → Separate DB
+ Thundering Herd problem
+ Streaming vs Batched updates
+ Fresher signals vs Delayed personalization
+ Popularity bias
+ High-signal content often favored
+ Leading to Cold Start
Image Ref: https://docs.databricks.com/aws/en/machine-learning/train-recommender-models
What actually matters at scale?
32
+ Data freshness
+ Tail latency consistency
+ Resource isolation
+ Index health
+ Upgrade safety
+ Cost predictability
Model accuracy alone is not sufficient.
What do you need to do?
33
+ Monitor freshness, backlog, tail latency, index health
+ Load test beyond steady state
+ Isolate workloads aggressively
+ Design for zero-downtime upgrades
+ Choose storage engines that tolerate write pressure
“Real-time AI is a distributed
systems problem under a
costume.”
34
“ScyllaDB solved that
problem many years ago”
https://deathbyclawd.com/?url=scylladb.com
35
What’s your primary
AI use case today?
37
Must-Read Books for High-
Performance System Design
Access 4 PDFs on building and optimizing data-intensive applications
lp.scylladb.com/book-library-offer
Thank you
for joining us today.
@scylladb scylladb/
slack.scylladb.com
@scylladb company/scylladb/
scylladb/