visiting Codemotion 2018 (tech talk for colleagues)

In terms of this tech talk for my colleagues, I spoke of the experience of visiting Codemotion 2018 and tools and practice we could use in development.

Transcript

1. Codemotion Berlin ‘18 experience
3. Sponsors
4. Further referred talks
5. “Classic development revolves around if / else clauses and loops. Machine Learning revolves around hyperparameters and data pipelines”. AI with devops mindset
7. Why should we care about data pipelines?
8. “Continuous Integration and continuous deployment in Machine Learning is necessary and it's not easy. But continuous evaluation is more important than CI / CD”
10. Continuous evaluation of ML models • Daily (?) time series charting of used metric (AUC, Gini and etc.) - to track model performance, seasonality trends • Constantly compare distribution of features to initial dataset • A/B testing • Data aggregation runtime, model runtime, deploy bugs
11. “Sometimes, the model training procedure may use a different metric than the evaluation. This can happen when we are reappropriating a model for a different task than it was designed for.”
14. Bias and drifts are Machine Learning bugs!
15. ML extra problems • Data scientist != Software engineer
16. ML extra problems • Data pipeline won’t scale • Local development & local data & lack of versioning • A lot of rework “Local data science is dead.”
17. Pitfalls in microservice integration “System failures are not customers problems so they require concise message”. synchronous message - for user, asynchronous error message for the system
18. Pitfalls in microservice integration • Idempotency of services - calling the service several times in a row is ok • Using messages like “don't close browser, window etc” is bad practice
19. Deadly sins of web security • Penetration testing is a must! (especially with big data) • Developers should be attentive as well https://snyk.io/ - check libs & packages versions for vulnerabilities
20. Has anyone else seen your code? • Senior review is a choke point
21. Has anyone else seen your code? “Do not rely solely on senior dev code review. Encourage juniors to review your code, you may learn something too.”
22. Code reviews ++ • Code reviews make everybody know the code, improve coding culture, status updates get faster with time • Code reviews provide shared responsibility - when something happens blame is spread and you can go straight to fixing the problem
23. Technical leadership “Never compromise culture and quality in hiring new employees, better to be slow than that.” • Transparency and trust is a must - the team should know what goes wrong • Distributed responsibilities - frequent releases but from different teams, n% of time for refactoring and side projects • All team members are part of a road map for being involved in creating a real product
24. Diversity & technical leadership “Diversity is about being invited to the party. Inclusion is being welcomed here”. Diverse teams bring more broad perspectives to problems solutions. Diversity is not only about HR team it’s about everyone in the company. Quotas for hiring women are just the means for the goal but it doesn’t mean that diversity is completed.
25. Дякую за увагу.
26. Additional resources (DS & ML) Data pipelines for data science projects from scratch Data pipelines in Serverless Era Evaluating ML model 3 ways biased data can ruin your model Dashboards analytics Building DS team
27. Additional resources How to implement business transactions using Microservices. Saga pattern Tech blog about microservice architectures Idempotency in service-oriented architecture How diversity makes us smarter