On Monday, 17th April, Delivery Hero was again delighted to join 1700 Pythonistas at the sold-out PyCon DE and PyData Berlin 2023 event as sponsors, speakers and attendees!
During the welcome talk, it was encouraging to learn that a majority of attendees were first-time visitors to this largest volunteer-run event for the Python programming language in Germany. After a big round of applause thanking the organizers and volunteers, it was time to start the talks.
Cooking up an ML Platform: Growing pains and lessons learned
For the very first session of the conference, our colleague Cole Bailey (ML Platform Lead) described Delivery Hero’s journey building an ML Platform for our Logistics department. Machine Learning is an extremely important ingredient in the Delivery Hero customer journey being used for everything from forecasting food preparation times, delivery times, customer demand, rider workforce requirements and detecting fraud.
The workloads handled by this internal platform include over 30 models covering both batch ‘offline’ predictions and real-time predictions over API that handle up to 130 million predictions a day for 70+ countries, with P95 latency <10ms.
Taking inspiration from Team Topology’s “Thinnest Viable Platform” concept, we got a glimpse into the lightweight stack Cole’s team has implemented: a standardized set of tools that helps reduce cognitive load for the ~30 data scientists and ML engineers working on machine learning models for Logistics use cases.
Figure showing the high-level architecture of Delivery Hero’s ML Platform for Logistics: including Kafka and Apache Flink for our internal feature store called “Franz”, Metaflow for training jobs and Kubernetes to serve the trained models in production.
A few other talks we liked
Exploring the Power of Cyclic Boosting
Cyclic Boosting is a recently published supervised machine learning algorithm designed for retail demand forecasting. Using coordinate descent instead of gradient descent, it is also fully “explainable”, which means business stakeholders can understand the factors contributing to the final forecast.
Driving Down the Mamray Lane
Most Python engineers will relate to a feeling of dread after experiencing “MemoryError” or OOM pod kills if you’re running applications on Kubernetes. This talk was about profiling and debugging a Python program’s memory usage using Bloomberg’s recently open-sourced memory profiler, Memray.
Building Chat Assistants with Haystack and GPT
Large Language Models will revolutionize many aspects of Delivery Hero customers’ journey, but preventing hallucination and facilitating access to real-time data remain unsolved problems. This talk discussed one method of tackling those concerns using the popular Haystack NLP framework in combination with ChatGPT to provide real-time ‘factual’ information from a custom corpus.
Towards Learned Database Systems
Any engineer using a relational database will have experienced slow queries that can be fixed with a CREATE INDEX statement. But surprisingly, even this basic software engineering task may soon be disrupted by AI! Prof. Dr. Carsten Binning presented the latest research from his group demonstrating how deep learning techniques can outperform traditional cardinality estimation heuristics for join order optimization.
Ask-A-Question: an FAQ-answering service for when there’s little to no data
We are heroes because we care so it’s always inspirational to hear stories about technology being used for good. In this talk, we learned about an open-source library called “faqt” that uses word embeddings and neural networks to power a maternal and child health support service in South Africa.
PyLadies Panel Session and PyLadies Workshop
Diversity and Inclusion is a top priority for the Delivery Hero tech team. During the PyLadies panel of minority background tech workers, we heard a clear definition of diversity: the freedom and safety to show up at work as our authentic selves. The privileged majority were encouraged to vocally address inequity when they see it. In the subsequent follow-up workshop, attendees were able to learn their employee rights in the case of discrimination or contract termination.
Monorepos with Python
This has been a recurring topic within the Delivery Hero engineering team and I’m sure we’ll discuss it again! We currently have a mix of polyrepos and monorepos. It was useful to learn about some new techniques for structuring monorepos and the difference between ‘package-based’ (Nx) and integrated approaches (pantsbuild).
Conclusion and thanks
Phew, after a packed three-day schedule, talking to 100s of engineers we left the conference with a lot of new ideas and contacts! Thanks again to the organizers and speakers, we’re already looking forward to next year’s conference. See you there!
Would you like to become a Hero? We have good news for you, join our Talent Community to stay up to date with what’s going on at Delivery Hero and receive customised job alerts!
