Downloading release v0.5.2...
Pulling images for both stacks (the notebook image may take a few minutes)...
Starting the user portal...

Iceberg Data Platform 0.5.2 is running.

1. Sign in to the administration portal
   http://localhost:3000/#guide
   Username  platform-admin
   Password  ••••••••••••  (temporary: you choose a new one)

2. Create a team and a user, then sign in as that user
   http://localhost:3002/#guide
   Or skip the setup: the demo-company skill below does it

Optional
   Skills for coding agents such as Claude Code, Codex and GitHub Copilot:
     quick-share    Share the portals for a class or demo
     demo-company   Create a demo company, one account per participant

• Linux, macOS and Windows. Only Docker with Compose is needed: no source checkout, Python or Node.js.
• The installer verifies the bundle and writes random credentials to .env. It ends with who signs in first, and where.
• It pulls pinned images and starts both stacks.
• It puts iceberg_connect.py for your own tools and the agent skills in the same folder.
• To update, run it again. .env and the data stay.

• It is not a query engine and does not execute SQL, so you pick DuckDB, Trino, Spark, Snowflake, or all of them.
• It is not a storage system: it sits on the S3 compatible object storage you already have.
• It is not a security layer by itself. Permissions and credential vending live in the catalog and in your identity provider.
• It is not maintenance free. Compaction, snapshot expiry and orphan cleanup are yours to schedule.

• Operations are your responsibility: upgrades, backups of Postgres and the buckets, monitoring, HTTPS and secrets.
• Table maintenance is yours as well. Compaction, snapshot expiration and orphan file cleanup need a schedule.
• Security boundaries are explicit. This build is a local development platform with in-memory sessions on a trusted Docker host, so read the security model before running it anywhere else.
• Fewer vendors means tighter control over the architecture, but we have to be realistic: it also pushes more integration decisions onto a small team.
• Performance across data products is harder. Combining many Iceberg data products and querying them fast takes more work than inside one integrated platform.

CREATE TABLE lakehouse.iceberg_v3.sensor_events (
    event_id     BIGINT,
    device_id    VARCHAR,
    site         VARCHAR,
    location     GEOMETRY,
    measured_at  TIMESTAMP_NS,
    payload      VARIANT,
    synthetic    BOOLEAN
) WITH ('format-version' = 3);

ALTER TABLE … ADD COLUMN firmware VARCHAR DEFAULT 'v1.0';
UPDATE … SET firmware = 'v2.0' WHERE device_id = 'SENSOR-001';
DELETE FROM … WHERE payload.kind::VARCHAR = 'fault';

• VARIANT holds a different structure per event, without a fixed schema.
• TIMESTAMP_NS and GEOMETRY are native column types.
• Column defaults mean that adding firmware rewrites none of the 480 rows.
• Deletion vectors let updates and deletes mark row positions in a small Puffin file.
• Row lineage keeps a stable ID per row, together with the commit that last changed it.

from pyiceberg.catalog import load_catalog

catalog = load_catalog(
    "shared", type="rest",
    uri="https://catalog.example/api/catalog",
    warehouse="db-…",
    credential="<client-id>:<client-secret>",
    scope="PRINCIPAL_ROLE:ALL",
    **{"header.X-Iceberg-Access-Delegation":
       "vended-credentials"},
)

table = catalog.load_table(
    "synthetic.neighborhood_electricity")
table.scan().to_arrow()         # 26,880 rows

catalog.list_namespaces()       # Forbidden
catalog.list_tables("synthetic")  # Forbidden
catalog.load_table("synthetic.other")  # Forbidden

• A share holds no list privilege, so the recipient loads only the shared objects, by full name.
• Storage follows the grant: vended S3 credentials are read-only and stop at the prefix of the shared table.
• A view is a definition rather than a filter, so the tables it reads are shared with it and readable in full.
• Revoking is immediate for catalog tokens; storage credentials already handed out expire on their own.
• Outside one machine this needs your own TLS proxy in front of Polaris and RustFS.

claude mcp add --transport http \
  --client-id iceberg-mcp --callback-port 3010 \
  iceberg-user http://localhost:3002/mcp

• The agent signs in through Keycloak as you: a public client with PKCE, no secret to store.
• Every call runs with your Polaris grants, so the agent sees exactly what you see and nothing more.
• A shared database stays read-only for the agent too.
• Destructive tools are marked as such, and the agent is told to ask before using them.

claude mcp add --transport http \
  --client-id iceberg-mcp --callback-port 3010 \
  iceberg-admin http://localhost:3000/mcp

• The platform-admin role is checked in the token on every call, not once per session.
• Irreversible tools are marked as destructive; delete_database refuses without the database's exact name.
• Connection details never include credentials.
• A new user's one-time password lands in the agent transcript, so treat that transcript as a secret.