Canva: How Canva collects 25 billion events per day

Canva writes about its event collection infrastructure capabilities, handling 25 billion events per day (800 billion events per month) with 99.999% uptime.

At our team’s inception, a key decision we made, one we still believe to be a big part of our success, was that every collected event must have a machine-readable, well-documented schema.

The principle is the key motivation for me to write Schemata. If you want to adopt similar principles that Canva follows out of the box, you can use Schemata.

https://www.canva.dev/blog/engineering/product-analytics-event-collection/

Discord: How Discord Uses Open-Source Tools for Scalable Data Orchestration & Transformation

Discord writes about its migration journey from a homegrown orchestration engine to Dagster. The blog highlights the reasoning behind selecting dbt and Dagster and some of the key improvements while adopting them, such as handling race conditions in dbt incremental update and bulk backfilling with Dagster.

https://discord.com/blog/how-discord-uses-open-source-tools-for-scalable-data-orchestration-transformation

Philip Rathle: The GraphRAG Manifesto - Adding Knowledge to GenAI

It is one of the most fascinating reads about using Graphs as a structure to add knowledge to GenAI. The author highlights that there are two ways to represent knowledge.

1. Vectors

2. Graphs

The vector representation is an array of numbers. In a RAG context, it is useful when you want to identify how similar one handful of words is to another. The author is making a case where if you want to make sense of what’s inside of a vector, understand what’s around it, get a handle on the things represented in your text, or understand how any of these fit into a larger context, then vectors as a representation just aren’t able to do that. The graph is an appropriate model to represent knowledge.

https://neo4j.com/blog/graphrag-manifesto/

Sponsored: 7/25 Amazon Bedrock Data Integration Tech Talk

Streamline & scale data integration to and from Amazon Bedrock for generative AI applications. Feat. Sathya B. (Senior Solutions Architect at AWS)

Learn about:

• Efficient methods to feed unstructured data into Amazon Bedrock without intermediary services like S3.

• Techniques for turning text data and documents into vector embeddings and structured data.

• Practical insights into scaling data integration for generative AI with Nexla and Amazon Bedrock.

• Real-world applications of Nexla’s RAG data flow capabilities in enhancing AI deployment.

https://nexla.com/resource/how-to-scale-data-integration-with-nexla-and-amazon-bedrock-for-generative-ai/

Wasteman.codes: Engineering Principles for Building Financial Systems

Any system that involves moving money or counting money is always complex to get it correct. I used to joke the entire Walstreet runs on unknown SQL codes that no one understands. The author highlights the engineering best practices such as,

1. Immutability and Durability of data

2. Data should be represented at the smallest grain

3. The code should be Idempotent

I’m a bit confused by the recommendation to use preferred integers to represent financial amounts since precision is critical in all financial computing. It would be helpful if the author added more context to the recommendation.

https://substack.wasteman.codes/p/engineering-principles-and-best-practices

Mark Raasveldt: Memory Management in DuckDB

Memory management is one of the hardest parts of building in-memory database engines while supporting large-scale data processing. The author writes an in-depth overview of key parts of DuckDB memory optimization techniques.

1. Streaming execution to process a small chunk of data at a time.

2. Intermediate spilling to disk while computing aggregations.

3. The buffer manager caches as many pages as possible.

https://duckdb.org/2024/07/09/memory-management

Ibis: Querying 1TB on a laptop with Python dataframes

Ibis is an open-source dataframe framework to interact with multiple database engines. Ibis published a 1TB benchmark processing MacBook Pro with 96 GiB of RAM comparing Ibis, Pandas & variations of Polaris.

https://ibis-project.org/posts/1tbc/

Max Zheng: Contact & Organization Enrichment

Finding your champion product advocate within your user base is critical in enterprise selling. Metabase writes about why they build a continuous enrichment pipeline for customer contacts pipeline and its system design. The blog provided a nice comparison summary of various 3rd-party data providers in this space and their capabilities.

https://metabase.notion.site/Contact-Organization-Enrichment-dc974a4092674d2dab4da1fc01e57458

Meta: Meta’s approach to machine learning prediction robustness

Meta writes about its approach to machine learning prediction robustness and the challenges of ensuring reliable ML predictions. The authors identify several factors that make this difficult, including the stochastic nature of ML models and the constant updates to models and features. Meta’s approach to addressing these challenges involves a systematic framework incorporating preventative measures, fundamental understanding, and technical fortifications.

https://engineering.fb.com/2024/07/10/data-infrastructure/machine-learning-ml-prediction-robustness-meta/

Dropbox: Bringing AI-powered answers and summaries to file previews on the web

Dropbox writes about building AI-powered features such as Q&A and summaries on unstructured data. The blog highlights optimization techniques to build embedding to enable such features.

The lessons learned section highlights the key factors to consider while building, especially around clustering & segmentation, chunk priority calculation, and cached embeddings.

https://dropbox.tech/machine-learning/bringing-ai-powered-answers-and-summaries-to-file-previews-on-the-web

Pinterest: Building Pinterest Canvas, a text-to-image foundation model

Pinterest writes about Canvas, its homegrown text-to-image foundation model designed to generate images from textual descriptions.

The blog details the process of generating high-quality training datasets to build the base model and fine-tune the model further to create personalized experiences.

https://medium.com/pinterest-engineering/building-pinterest-canvas-a-text-to-image-foundation-model-aa34965e84d9

All rights reserved ProtoGrowth Inc, India. I have provided links for informational purposes and do not suggest endorsement. All views expressed in this newsletter are my own and do not represent current, former, or future employer” opinions.