Coming from a psychology background with almost zero technical foundation, I used "mental schemas" in my last post to understand what a metadata catalog is. Today is Day 2, and my goal was simple: get Apache Gravitino running locally on my MacBook, and actually understand what real-world problem it solves.

1. From "Error Panic" to Running Gravitino 1.3.0

To avoid wrestling with complex Java setups on my laptop, I used OrbStack to run Docker. After typing the startup command, my terminal flooded with Pull complete messages. What it was actually doing was pulling the official Image—kind of like grabbing a "mooncake mold" that already has Linux, Java 17, and Gravitino packed inside, ready to stamp out a running Container anytime.

I did hit a small bump: when I opened http://localhost:8090 in my browser, a red error popup greeted me: Unable to connect to Gravitino.

After checking the container, I realized nothing was actually broken—the backend was just still "warming up." Once docker ps -a showed the status as Up, I refreshed the page, and the Apache Gravitino 1.3.0 UI finally loaded.

From there, I built the core three-tier hierarchy right in the UI:

  • Metalake (Top-level container): day2_learning
  • Model Catalog (AI model directory): my_ai_models
  • Schema (Sub-folder): nlp_experiments

2. Pausing to Ask "Why": Why Split Data into 4 Categories?

After clicking through those buttons, I suddenly felt a bit lost—if I’m just blindly following a recipe, I still don't know why companies need these "drawers" in the first place.

So I paused and chatted with AI, asking it to explain the 4 tabs in the top-left corner of the Gravitino 1.3.0 UI from a psychology perspective. Everything clicked immediately:

  1. Relational (Structured Data): Just like the survey tables I organize in SPSS or Excel. Whether it’s 1–5 Likert scale scores or text labels like "Experimental / Control Group," as long as it fits neatly into fixed rows and columns, it belongs here.
  2. Fileset (Unstructured Data): Think of interview recordings, experiment videos, EEG scans, or full PDF papers. You can't cram them into spreadsheet cells, yet they are the exact raw fuel needed to train modern AI.
  3. Messaging (Data Streams, e.g., Apache Kafka): This isn't about what the data looks like, but its state of motion. Just like an eye-tracker or EEG cap streaming continuous signals to a screen every second during an experiment, clicks and orders in an app flow in real time on a "conveyor belt."
  4. Model (AI Models): The machine learning and AI models trained by different teams across a company.

Only after setting up this standardized filing cabinet (Metalake → Catalog → Schema) can a company uniformly apply Tags and Policies to keep its data under control.

3. Zooming Out: Polaris vs. Unity Catalog vs. Gravitino

After getting the local environment running, I read two recent comparison articles on data catalogs and finally made sense of how the "Big Three" position themselves:

  • Apache Polaris (initiated by Snowflake): Like an "Iceberg specialist butler." Lightweight and vendor-neutral, focused on managing Apache Iceberg tables really well.
  • Unity Catalog (by Databricks): Like the "Apple ecosystem butler." Super feature-rich, but deeply designed around Databricks' own ecosystem, which can limit flexibility with external legacy systems.
  • Apache Gravitino: Positioned as the "cross-cloud, cross-brand universal butler." It doesn't play favorites with any single vendor—it's built to unify scattered legacy tables, real-time streams, unstructured files, and AI models under one roof.

Final Thoughts

My biggest takeaway today wasn't actually those two Docker commands—it was finally connecting technical actions to real business pain points. If you're also learning data or AI infrastructure from scratch, or have hands-on experience choosing between these catalogs, I'd love to hear your thoughts in the comments!

P.S. It suddenly got chilly today, yet I spotted this little flower on campus still adorably clueless and blooming away.