How engineering works at Uber (Category Engineering)
After memories When I was at Charles-Axel Dein's Uber in the past, I became interested in how the engineering part of Uber works today, and I collected some of that information.
The engineering wing of the company has 4k+ engineers (into 2023 theirs was 4k) It's distributed all over the world. Engineering teams fall into two categories: “program” and “platform.” Program teams fully own part of the product (both the client part in the application and the necessary backend) And the metrics of his success. Platform teams specialize in the development of universal technology components and infrastructure used by all product groups.
In 2023 During the year, Uber had a big change in IT - separate branches were created for these programs and platforms. Pravin Naga has become a CTO in Mobility and Delivery, which oversees the development of all Uber user products and applications.
- Albert Greenberg became a VP in platform engineering (He previously worked as a VP in Azure.). It is responsible for the overall technical platform, infrastructure, data and tools.
When it comes to architecture and infrastructure, 2016 year saw a monolith for many microservices, each responsible for its function (Car search, price calculation, user profiles, payments, cards, notifications, etc.). This distributed architecture allowed the teams to develop different parts of the system in parallel and scale them independently. However, it also required cutting-edge solutions for orchestration, data storage and monitoring – they created their own cloud stack inside the company. For example, in the same 2016 year told How they developed "Schemaless," their datastore using MySQL. Most services migrated to the cloud, such as the Fulfillment platform. 2021 year moved on the storage engine as Google Spanner in GCP.
Before you start. 2023 year 95% of loads lived inhouse, and then went to a large-scale move to the public cloud, which was announced in the Q1 2023 for investors. Uber has signed strategic agreements with Google Cloud and Oracle Cloud. 7 In order to migrate your server systems to the cloud of these providers. This multi-cloud approach is chosen to increase reliability and flexibility. By the way, in the Oracle cloud, they went to Ampere ARM-processors to reduce bones and already to the end of the day. 2024 Thousands of Uber services a year migrate Arm-architecture in Oracle Cloud, which gave a gain in the price / performance ratio and reduced the dependence on x86.
Backend architecture is built around containerization and orchestration, allowing developers to deliver updates hundreds of thousands of times a week without downtime. Moving to the Uber Cloud substitute self-painting (µDeploy) A new multicloud platform called Up. Thanks to this, 2022 I moved to the cloud for a year. 2+ mln 2 cores, as well as most service teams, abstracted from low-level parts (That is, for developers, the microservice depot is now the same regardless of the environment: inhouse or cloud.).
The development of AI can be divided into eras (more article 2024 year):
- Predictive ML (2016–2019) - emphasis on tabular tasks (ETA, risk, pricing) with XGBoost and nascent DL cases (3D mapping, autopilot)Launch of the centralized Michelangelo platform and Palette fichestor
- Scaling Deep Learning (2019–2023) Transition to DL as first-class citizen, unification of instruments at Michelangelo 2.0Increased use of DL models (1+ mln rps)
- Gen AI stage (on 2023) LLM/GenAI for customer experience and internal productivity, development of LLMOps and GenAI Gateway with access to external and proprietary models while maintaining a single platform and MLOps processes.
In 2023 The company launched the Uber AI Solutions initiative, which offers its corporate customers a service to collect and mark up data for model training, as well as testing AI systems. Uber does not develop autonomous cars itself, but partners with external companies.
#Architecture #Management #Software #Leadership