Research Insights Made Simple #6: Interview with Nikolay Golov about data platforms
An issue on how to design and develop a data platform in 2025: from federalization and data products to OLTP/MPP limitations and the evolution of cloud analytics.
Source
Telegram: book_cube
Original post with release announcement and brief abstract.
About the release
This interview continues the series Research Insights Made Simple and focuses on the practical evolution of the data platform. The conversation examines real engineering compromises: how to combine centralization and team autonomy, how not to drown in data product chains, and why the architectural limitations of old approaches are increasingly noticeable in 2025.
Guest and regalia
Nikolay Golov
- Head of Data Engineering at ManyChat.
- Ex Head of Data Platform in Avito.
- Practitioner in building OLTP/OLAP systems and database teacher.
Issue format: interviews as part of the section #Data.
Key topics of conversation
Guest's career path
How Nikolay Golov developed in different companies and taught databases at the same time.
Platform organization models
Centralized, hybrid and decentralized approaches to data platform development.
Data Mesh in theory and practice
Why the principles of federalization are useful, but in reality they often turn into a data mash.
Company-wide approaches
Differences in platform design for startups, medium-sized and large companies in 2025.
Limitations of classic OLTP DBMS
Where Postgres and similar systems come down to workload, cost, and architectural trade-offs.
Limitations of MPP solutions
When Vertica, Greenplum and ClickHouse do not solve the problem entirely and require additional layers.
Data product chains
Why is it important to isolate the base data domain to shorten dependent product chains?
Why cloud DW is so fast
Columnar storage and storage/compute separation as the basis of modern analytical speed.
Medallion architecture
A practical Bronze/Silver/Gold scheme as a way to standardize the data path within the platform.
Related topic
Specifics of designing data systems
Chapter with Medallion, data flows and system trade-offs.
Practical conclusions for 2025
- There is no universal organizational model for a data platform: the structure depends on the stage of the company and domain boundaries.
- Data mesh is useful as a direction, but without a clear platform layer and governance it quickly degrades.
- OLTP and MPP systems are needed, but alone they do not cover end-to-end analytics and data products scenarios.
- Key engineering bet in 2025: storage/compute separation, open formats and a strong self-service platform.
- The main goal of the platform: to reduce time-to-data for product teams without losing control of quality and cost.
Links to materials
YouTube
Full video version of the interview.
Telegram post
Brief summary and announcement of the release.
Ya Music
Audio version of the podcast.
Podster.fm
Alternative audio platform.
To dive deeper into related topics: Introduction to Data Storage, DDIA, Streaming Data, Big Data, Learning DDD (including Data Mesh).