Attending Google Cloud Next Tokyo ‘24

Introduction

I attended Google Cloud Next Tokyo ’24 at PACIFICO YOKOHAMA, Japan on August 1–2. In this article, I will share the contents and my impressions of the event.

Keynotes

The keynote sessions featured Google employees discussing key products and future activities, as well as Google Cloud users presenting their use cases. Although this year’s keynote continued to focus on Generative AI (Gen AI) from last year, I felt a strong emphasis on the enthusiastic use of multi-modal agent AIs, which can recognize images, audio, and language, for business applications. Google has been providing a comprehensive environment for developing Gen AI, including:

• Infrastructure: Building data centers in Japan and expanding the submarine cable from the US to Japan.
• Gen AI Models: Offering not only Gemini but also models from other companies through Vertex AI.
• Development Environment: Supporting no-code, low-code, and full-code in Vertex AI.
• Business Use Cases: Integrating Gen AI into business operations through Gemini for Google Workspace.
• Custom AI Agents: Allowing the creation of personalized AI agents through Gems.

Notably, Google’s search engine technology is a crucial feature in the development of Gen AI, providing essential grounding. Additionally, the announcement of the Gemini model in Japan, which can store learning and generating data exclusively within Japan, is highly beneficial for data-sensitive companies.

I believe Gemini for Google Workspace is a significant advantage over other companies. The demonstration showcased the following capabilities:

• Typing @Gmail to call Gmail and asking Gemini to prioritize tasks based on email contents.
• Typing @Google Drive to call Google Drive and asking Gemini to generate comparison charts from documents stored in the drive.
• Directly exporting the comparison chart to Google Docs and calling another Gemini from the side panel of Google Docs to collaborate on other slides and documents.

It’s amazing! Gemini for Google Workspace significantly saves working time, especially when creating documents and slides.

Finally, the following Google Cloud updates were announced:

• Data Preparation for Gemini in BigQuery
• Vector Indexing in BigQuery
• BigQuery to Vertex AI
• Gemini in Looker
• Spanner Graph
• Spanner supports full-text search and vector search
• Spanner Editions
• Bigtable supports SQL

Booths

At the booths, Google, partner companies, and various SaaS providers showcased their demonstrations and services.

BigQuery Racing Insights, brought from the US!

This demonstration collects data from a toy car running on a roller coaster track using a Raspberry Pi with IR LED sensors. The data is then sent to Bigtable, a low-latency NoSQL database service, and finally analyzed using BigQuery or Vertex AI Workbench. This roller coaster was also displayed at Google Cloud Next ’24 in Las Vegas.

Beat Google at Load Balancing, a speed challenge versus Google

In this demonstration, participants experience the difficulty of handling network traffic and compete against Google’s Load Balancer. The game involves pressing the correct button from multiple options as they blink, simulating network traffic management.

Additionally, there was a lounge for attendees with Google Cloud Professional Certifications. This lounge was equipped with tables and chairs, allowing attendees to work, charge their laptops, and enjoy free beverages. This year, Google provided a special gift for those who hold all certifications. I achieved acquiring all certifications last month, and I was honored to receive the gift.

In the company booths, I visited several SaaS providers to update my knowledge. I was surprised to see that there were more overseas SaaS companies than Japanese ones. Despite Japan’s declining population and the potential shrinking of our market, these companies still regard Japan as one of the big markets in Asia. I feel really complicated about this trend; on one hand, it makes our market more active, but on the other hand, Japanese SaaS providers should play an important role in the domestic market.
Then, I spoke with representatives from:

• box
• Confluent
• fastly
• Fivetran
• Kyndryl
• LumApps
• MongoDB
• Okta
• SUSE
• Snyk
• ThoughtSpot

Presentation Sessions

In the presentation sessions, various companies shared their case studies. I attended several presentations and will share the content and my impressions here.

Note: The original presentation titles were in Japanese, and I translated them to English. I apologize in advance if there are any mistranslations.

Optimization and Challenges of Infrastructure Before and After the Release of Our Game Titles (Bandai Namco Entertainment Inc.)

Bandai Namco Entertainment Inc. is a Japanese multinational video game publisher. They discussed their famous game titles “MY HERO ULTRA RUMBLE” and “Tekken 8” as examples to introduce their challenges in this presentation.

• Reducing Server Cost: In MY HERO ULTRA RUMBLE, they used Google Kubernetes Engine (GKE) and Agones for the matching server, initially using n1-standard-4 as a node. This node could handle one match per 1.8 cores, so they switched to T2D-standard-4, which can handle one match per 0.8 cores. This change allowed them to conduct four matches per server, significantly reducing server costs. They were initially concerned about changing the chip of the vCPU, but it turned out to be manageable.
• Worldwide Matching: In Tekken 8, they also used GKE and Agones (though not confirmed) with servers located in North America, Europe, and Asia (Japan) to handle players from multiple countries. Physical distances significantly affect latency in gaming. For instance, while the Asia server is located in Japan, Singapore is still a bit far from Japan, so placing another server in an Asian country might be better.
• Development Tips: It’s essential to fix the number of players in a pod during the design phase because this affects the number of vCPUs and infrastructure costs. In some cases, not using Agones can be a cheaper and better option. Additionally, egress fees can be surprisingly high and are included in the Compute Engine fee, so careful monitoring is recommended.
• Using Spanner: Profiling plays an important role in troubleshooting issues. To profile your game, you should use tools such as APM, Cloud Trace, and Datadog. Additionally, warming up the cluster is needed before releasing a new version of the game because re-sharding is not applied when you increase the nodes of Spanner. Load testing should be conducted to re-shard Spanner before releasing updates.

As a video game enthusiast, it was an amazing experience to listen to the background of these games while imagining their gameplay. In my field, retail, we implement multi-regional strategies to enhance availability, whereas worldwide game companies need multi-regional strategies to reduce latency due to physical distances. This insight into the differences between business fields was particularly valuable to me.

Achieving Rapid Establishment of In-House Software Development Teams Using Google Cloud (EARTHBRAIN Ltd.)

EARTHBRAIN Ltd. is a Japanese company that develops software for the construction industry. In this presentation, they shared their mindset and strategies for achieving in-house software development.

• Software Controllability: In-house development plays a crucial role in acquiring software controllability. This approach enables the enhancement of key functions in the software, reduces development time, and ensures the team possesses the right expertise.
• Key Points of In-House Software Development: In-house development should focus on core business or technology, such as data platform and 3D processing capabilities in their company, to maintain a competitive advantage. Utilizing fully managed services and serverless technology allows the team to concentrate on core business activities without developing non-core infrastructure. It is crucial to decide what not to develop. Efficiently allocating resources to core business areas is important, especially in the challenging Japanese IT job market where hiring multi-skilled IT engineers is difficult.
• Conducting Kaizen after In-House Development: Even after achieving in-house development, ongoing improvements (Kaizen) are necessary. DORA’s four key metrics play a crucial role in managing these improvements. Achieving the Elite level for each index has varying degrees of difficulty, so it’s important to prioritize which areas to improve.

This presentation was particularly impressive for me because I tend to prefer developing all aspects of software myself, which isn’t always practical. The concept of “decide what you don’t do” is something I should keep in mind during software design and development. We should focus on competing in our core business areas.

AI/LLM Utilization Cases by Andpad Accelerating DX in the Architecture and Construction Industry (Andpad Inc.)

Andpad Inc. is a Japanese company that develops centralized management software for the architecture and construction industry. In this presentation, they discussed their use case of AI and LLM (Gemini) for their software.

• Changing Traditional Methods: Traditionally, workers keep records of construction sites by writing information on a blackboard and taking a picture with it. Andpad aims to change this with their software, Mame-zu AI and Kokuban AI. Mame-zu refers to construction detail diagrams, and Kokuban means blackboard.
• Handling Unstructured Data: The industry deals with a lot of unstructured data, such as diagrams, documents, construction site pictures, and chat history. Additionally, many special construction symbols are used in diagrams. While OCR technology has been used to extract data from unstructured sources, Gemini has become a game changer in this field.
• Comparing Gemini and OCR: Andpad compared the performance of fine-tuned Gemini and fine-tuned open-source OCR in recognizing construction symbols. Gemini achieved significantly better results. They explained that LLMs, including Gemini, can maintain the context of the diagram and use this information to recognize symbols accurately. However, the area of the diagram given to the LLM affects the number of requests and computation costs. It’s crucial to find a balance between accuracy and cost.

Comparing Gemini and OCR was particularly interesting for me. When I was an AI software developer, I also developed OCR to recognize characters. At that time, increasing the accuracy of OCR often led to recognizing unnecessary characters in the document. I’m amazed that LLMs can maintain the document’s context, allowing them to recognize special characters with high accuracy. Of course, OCR is more cost-effective than LLMs, so it’s important to balance these technologies.

Summary

In this article, I shared the contents and my impressions of Google Cloud Next Tokyo ’24. Through this event, I felt that Google is rapidly integrating Google Cloud and Google Workspace with Gemini. I believe that Gemini will become a valuable partner in our working environment.

Regarding Google Cloud, managed services are becoming the main infrastructure, and Gemini allows us to create non-managed infrastructure services without complicated settings. As a DevOps and Cloud engineer, I keenly feel the need to consider how to build my career in the future. On the other hand, LLMs enable us to develop full-stack applications with minimal knowledge. You only live once, so pursue what you like with the help of LLMs.

Attending the event allowed us to communicate with engineers from other fields and update our knowledge. I highly recommend attending these events in person!