The document discusses building a high-performance autosuggestion feature using a trie data structure to store keywords. It evaluates storing keywords in MySQL, Elasticsearch, and an in-memory trie structure. Performance tests found the in-memory trie structure provided the fastest response time at around 10ms on average and highest throughput of 1100 requests per second using minimal memory of 25MB. The proposed solution is to build the trie structure from the keywords database on application start up to allow for sub-10ms response times without compromising user experience.

1. Search query assistance:
Autosuggestion
Vitalii Melnychuk, Software Engineer

2. Intro
10 members
50 servers
6 TB of data
Mysql
Elasticsearch
NODEJS
Docker
AWS
Terraform
KIbana
Jenkins
Grafana

3. Highload Project - highload solution
1
Highload begins when one physical server
becomes unable to handle data processing.
2
If a single instance serves 10,000
connections simultaneously - it’s highload.
3
Your project is highload if it processes 100+
dynamic requests per second.
4
Usage of Lambda Architecture and Kafka
makes the system highload.

4. Understanding
the problem

5. Expectations?

6. How Fast Should A Website Load?
QUOTE: “I wouldn’t worry about it
too much. Make it as fast as you
reasonably can.” Gary Illyes,
Google 2016

7. What we have to build?

8. Solutions analysis
- Frontend side filter
- Mysql `LIKE` search
- Elasticsearch Completion Suggester
- Own Data structures

10. name order
management
analyst
37
manager 5
management
assistant
35
... ...
manag...

12. Trie Structure
1. With Trie, we can insert and find strings in O(L) time where L
represent the length of a single word.
2. We can easily print all words in alphabetical order which is not easily
possible with hashing.
3. We can efficiently do prefix search with Trie.

13. Performance platform
- 100 threads
- 50 req/sec
- 50 000 samples
- c3.large

14. Mysql Elasticsearch In-memory
Average response 100ms 130ms 10ms
Throughput ~650 req/sec 500 req/sec ~1100 req/sec
Errors % 2% 5% 0%
Performance tests

15. Vision
1th
2th
Memory-optimized
structure
3th
4th
5th
Trie sorted by alphabet
Prioritize keywordsBuild data before
application started
Store only addresses

16. Process
0
1
Download all keywords 0
2
Build a trie
0
3
Respond by using
in-memory trie
structure

17. Disadvantages:
● Self-managed solution
● Cannot analyze full text
● Complicated in implementation

18. Advantages:
● Scalable and high-load search
engine
● response time
: Advantages

19. Not alphabetical search
Google
Analytics
DB
Trie

20. Make some modifications
Array of keywords
d
a
he
l е
l

21. https://github.com/melnychukvitaliy/trie

22. Proposed solution
You can save time in the place that
doesn’t affect user experience and
respond quickly.

23. Results
50K
Keywords in our mysql
databases sorted by priority
25mb
Memory used in order to build
trie structure
<10ms
Response

24. Thank you

25. Questions?