Talk 1: Ines Montani > Spacy
-
check
prodigyfrom the same as spacy
S&P case studies
-
real-time commodities trading insights by extracting structured attributes
-
high-security environments
-
used LLM during annotation
Development workflow
custom models (pytorch) > pretrained models (chatgpt) > custom models(spacy prodigy: distill models into smaller and private components)
| human-facing system | machine-facing models |
| –––––––––– | –––––––––––––– |
| chatGPT Google | Chatgpt4 Gemini |
| UI/UX, customization | speed, cost accuracy latency |
| ?? | ?? |
2. Talk: Daniel Hepper
Lessons from processing one billion celery tasks
CTO at Userlike: Customer messaging solution
Why should I use Celery?
-
Execute long running task outside of your request handler
-
gracefully handle workload spikes
-
schedule tasks
Use case: Userlike
each websocket-connection becomes a task
Why task crashes?
-
Exception handling
-
Retries
Why worker crashes?
- Late acknowledge/idempotent tasks
Optimization-Performance
-
only for small and quick task the selection of a broker is important
-
avoid mixing long and short running tasks
-
fairness in multi-tenant applications
-
graceful degradation
Alternative
-
rq
-
django-background-tasks
-
hatchet
Talk 3: UUID
Why?
-
creating decentralized systems
-
concurrent systems
-
global table lock in SQL
OSF DCE/RFC 4122 UUID
- v1: is not monotonic without parsing it, not thread-safe
- v2: DCE security variant
- v3: md5 checksum, non-monotonic
- v4: random, non-monotonic, slow to generate, based on probability
- v5: like v3, but with sha
- v6: based in v1, is monotonic, reversed timestamp
- v7: based on time and randomness, monotonic, thread-safe, faster than v4, universally unique
- v8: user-defined
Why in his use case
-
from multi-schema(one per tenant) Why: incremental id create migrations problems
-
DB indexes like it
-
today: recommendation v7
Takeaways
-
Don’t reinvent the wheel!
-
If you really need to, use UUIDv8
-
Use v7 for everything if you’re not sure
-
Don’t use v4, it’ll slow you down (unless you want to hide the dates)
Talk 4: MTTR Mean time to recovery
Some takeaways
-
CI/CD make it easy and quick
-
code: identify critical parts of your code, use precise logs, take advantages from LLM
Talk 5: Django as an AI Orchestrator
Talk 6: Security python web applications
Basic Rules
-
avoid SQL injections > parameterize queries
-
leaking personal information(API Key, configs)> gitignore, fields encryption, data redaction in API, access control
-
pip-audit??
-
pip-tools??
Scale Up security
-
too many requests for the same user > protect critical endpoints with throttling mechanism
-
managing secrets > use Vault or secrets manager
-
error handling > don’t expose stacktrace to users, configure custom error pages, log all possible errors
Security edge cases
- Managing Stateful servers > use immutable objects, track objects changes, validate object before operating.
OWASP ??
-
track of top security risks
-
tools for improving security ….