6 Errors I Hit Connecting Databricks Apps to Genie AI

The setup: everything looked perfect

We had just wrapped a successful hackathon. A complete data platform on Databricks:

• Bronze → Silver → Gold medallion architecture

• Data quality enforcement

• Dimensional modelling

• A Gen AI insights layer

Now came the final challenge: build a Databricks App that connects our Gold layer to Genie AI for natural-language querying.

"How hard could it be?" — famous last words.

What followed was six hours of debugging across six different errors, each one teaching me something the documentation didn't make obvious. Here's everything I wish someone had told me before I started.

Error 1: "App crashed" on startup

What I saw in the logs:

AttributeError: st.session_state has no attribute "messages"

What I thought: Simple initialisation bug.

What I did: Fixed the session state initialisation, moved st.set_page_config() to the top of app.py, redeployed.

What happened: Still crashed.

The fix

st.set_page_config() must be the first Streamlit command in the script — before any import of session-state-touching modules or any other st.* call. And every key you read from st.session_state must be initialised before first use:

python

import streamlit as st

# This MUST come first
st.set_page_config(page_title="Genie Chat", layout="wide")

# Initialise every session_state key you'll read later
if "messages" not in st.session_state:
    st.session_state.messages = []

Lesson learned: a Streamlit crash on startup is almost always one of these two things. Check both before going deeper.

Error 2: "App not available" after a successful deploy

What I saw:

• Deployment status: ✅ Successful

• App URL: ❌ Currently unavailable

What I learned: Deployment success ≠ app accessibility.

The app deploys as a separate identity — a service principal with a name like app-xxxxx. That service principal is not you, and it has none of your permissions by default. Until you explicitly grant the service principal access to your Genie space, the app can boot but it can't connect to anything useful.

The fix

In the Genie space → Share → add the app's service principal with CAN RUN.

This is the first of many places where the app's identity needs to be granted access. It will not be the last.

Error 3: MessageStatus.FAILED — the one that haunted me

What I saw:

Error: failed to reach COMPLETED, got MessageStatus.FAILED

The cruel twist: Genie worked perfectly in the UI. Same question, same data, same warehouse. But via the App API? Failed every time.

What I tried:

• Redeployed the app — 5 times

• Cleared browser cache — 3 times

• Rewrote app.py end-to-end — 4 versions

• Re-checked the Genie space configuration — multiple times

Nothing worked.

This is where I learned the most important debugging rule of the whole project: the Genie UI and the Genie API use different permission paths. Working in the UI tells you nothing about whether the API will work. They run as different identities, with different access checks.

The fix lived inside Error 4.

Error 4: the permission rabbit hole

After hours of guessing, I finally opened Genie Monitoring — and there it was, the real error message that never showed up in my app logs:

PERMISSION_DENIED: Failed to fetch tables for the space.
 — No access to table 'insurancesdp.mansigold.claimsanalysis'
 — No access to table 'insurancesdp.mansigold.claimsbyregion'
 — No access to table 'insurancesdp.mansigold.factclaims'

Wait — I had already granted SELECT permissions! Here's what I had done:

sql

GRANT SELECT ON TABLE insurancesdp.mansigold.claimsanalysis
TO `5a59870a-e666-4c16-985f-eaa011cca1b4`;

Here's what I was missing:

sql

GRANT USE CATALOG ON CATALOG insurancesdp
TO `5a59870a-e666-4c16-985f-eaa011cca1b4`;

GRANT USE SCHEMA ON SCHEMA insurancesdp.mansigold
TO `5a59870a-e666-4c16-985f-eaa011cca1b4`;

Unity Catalog is hierarchical. SELECT on a table is not enough — the service principal also needs USE CATALOG on the parent catalog and USE SCHEMA on the parent schema. Without those, the table grant is invisible.

The complete permission checklist (save this)

Everything the app service principal needs, in one place:

Miss any one of these and you'll see a different error somewhere. Grant all of them, and you've eliminated 90% of the failure surface.

Error 5: "SQL warehouse stopped"

After fixing the permissions, I tried again. Still failing.

What I saw: The SQL warehouse status was Stopped.

The lesson: You can have every permission in the universe. If the warehouse isn't running, nothing works. Genie needs a warehouse to execute the SQL it generates, and the app needs that warehouse to be CAN USE-able by the service principal and actually running.

The fix

• Start the warehouse manually before testing

• Or enable serverless / auto-start so it spins up on demand

• And confirm the app's service principal has CAN USE on that specific warehouse

Error 6: the silent killer — permission propagation delay

What I learned the hard way: Unity Catalog permissions don't apply instantly. They take 5 to 15 minutes to propagate across services.

What I had been doing in a loop:

1. Grant permissions

2. Test immediately

3. Fail

4. Assume the grant was wrong

5. Grant again

6. Test

7. Fail

8. Lose another hour

What I should have been doing:

1. Grant permissions

2. Wait 10 minutes

3. Test

4. Works

This single rule, on its own, would have saved me hours.

The full solution that finally worked

After everything, here's the exact sequence that fixed it:

1. Grant every permission level

sql

-- Catalog level
GRANT USE CATALOG ON CATALOG insurancesdp
TO `5a59870a-e666-4c16-985f-eaa011cca1b4`;

-- Schema level
GRANT USE SCHEMA ON SCHEMA insurancesdp.mansigold
TO `5a59870a-e666-4c16-985f-eaa011cca1b4`;

-- Table level (for each table)
GRANT SELECT ON TABLE insurancesdp.mansigold.claimsanalysis
TO `5a59870a-e666-4c16-985f-eaa011cca1b4`;

GRANT SELECT ON TABLE insurancesdp.mansigold.claimsbyregion
TO `5a59870a-e666-4c16-985f-eaa011cca1b4`;

GRANT SELECT ON TABLE insurancesdp.mansigold.factclaims
TO `5a59870a-e666-4c16-985f-eaa011cca1b4`;

Then via the UI:

• SQL Warehouses → test-sql → Permissions → Add app → Can Use

• Genie → Space → Share → Add app → Can Run

2. Wait 10 to 15 minutes

Seriously. Wait. Don't test immediately.

3. Use this app.yaml

yaml

command: ["streamlit", "run", "app.py"]
env:
  - name: GENIE_SPACE_ID
    valueFrom: genie-space
runtime: python-3.10

The command matters. ["streamlit", "run", "app.py"] is correct. ["python", "app.py"] will boot a Python process that immediately exits because Streamlit was never started.

4. Clear everything and test

• Clear the chat in the app

• Hard refresh the browser (Cmd + Shift + R / Ctrl + Shift + R)

• Try a simple query first ("How many rows are in the claims table?") before complex ones

7 things I wish someone had told me on day one

1. Genie UI working ≠ Genie API working. They use different permission paths. Test both.

2. USE CATALOG is mandatory. You cannot read a table, even with SELECT, without it.

3. Permissions take time. Don't test for 10 minutes after granting.

4. Check Genie Monitoring. This is where the real error messages live — not in your app logs.

5. The app service principal is not you. It's a separate identity that needs its own permissions everywhere.

6. The warehouse must be running. Sounds obvious. Easy to miss.

7. app.yaml command matters. Use ["streamlit", "run", "app.py"], not ["python", "app.py"].

Quick-reference: error to fix

The real talk

Building on Databricks is powerful. The platform is genuinely production-grade once you get there. But the permission model is layered, and the error messages don't always tell you the real problem.

MessageStatus.FAILED alone could mean any of:

• Missing table permissions

• Missing catalog or schema permissions

• A stopped warehouse

• A misconfigured Genie space

• Permission propagation delay

• The app service principal not added to the Genie space

Good luck figuring out which one — unless you check Genie Monitoring first.

Once it works, though? It's magic. Natural-language queries running on a Gold-layer dimensional model, with Gen AI insights on top, all behind a clean Streamlit app. That's the payoff.

The pain is real. The payoff is worth it.

FAQs

Why does my Databricks App show "Currently unavailable" after a successful deploy? Deployment success means the app started; it does not mean the app's service principal can reach the resources it needs. The most common cause is missing access to the Genie space. Grant the app's service principal CAN RUN on the Genie space via Genie → Share.

Why does Genie work in the UI but fail with MessageStatus.FAILED from my app? The Genie UI runs as you. The Genie API, when called from a Databricks App, runs as the app's service principal. They have different permissions. Open Genie Monitoring to see the real error — it's almost always a Unity Catalog permission the service principal is missing.

What permissions does a Databricks App service principal need to query Gold layer tables through Genie? Six things, in order: USE CATALOG on the catalog, USE SCHEMA on the schema, SELECT on each table, CAN USE on the SQL warehouse, and CAN RUN on the Genie space. Miss any one and you'll see a different error.

Why isn't my GRANT SELECT working on the table? Unity Catalog is hierarchical. SELECT on a table requires USE CATALOG on the parent catalog and USE SCHEMA on the parent schema. Without those, the table grant is effectively invisible.

How long do Unity Catalog permission grants take to apply? Typically 5 to 15 minutes. Test only after waiting. Granting the same permission twice does not speed it up.

Should my app.yaml use python or streamlit? For a Streamlit app, use command: ["streamlit", "run", "app.py"]. Using ["python", "app.py"] will start a Python process that exits immediately because the Streamlit server is never launched.

Resources that would have saved me hours

• Databricks Apps Documentation

• Unity Catalog Permissions Guide

• Genie Spaces API Reference

• Streamlit on Databricks Apps

Have you hit similar permission nightmares on Databricks? Drop a note — I'm building a running list of edge cases that aren't obvious from the docs.

Built with Databricks, Streamlit, Genie AI, and a lot of coffee.