Skip to content

Case study (archive): exasol-personal-ai — Personal + MCP + JSON Tables

Archived — superseded by exasol-quickstart

exasol-personal-ai was the original Personal-based bundle: the host launcher provided the database and the two tools ran in containers reaching up to it via host.docker.internal. It has since been folded into the unified exasol-quickstart front door (the macOS / Personal path, currently experimental). Kept for history.

Repo: github.com/krishna-exasol/exasol-personal-ai

The goal

Same three pieces as exasol-ai — database + MCP Server + JSON Tables in one command — but with Exasol Personal as the database.

Why Personal changes everything

Exasol Personal is not a Docker image. It's installed via a host launcher CLI — exasol install local (macOS Apple-Silicon only) — and the database runs on the host, not in a container.

So this bundle is hybrid:

  • The database runs on the host via the launcher.
  • The two companion tools run as containers, connecting up to the host DB via host.docker.internal:<dbPort>.
  • Connection details are discovered at install time from exasol info --json plus the launcher's secrets.json (host, port, credentials), then injected into the containers.

That split — host DB, containerized tools, dynamic connection discovery — is exactly the orchestration a bare compose file can't do on its own, which is why a script-pipe installer drove it.

Constraints that shaped it

  • The same pyexasol conflict as exasol-ai (MCP >=1,<2 vs JSON Tables >=2.2,<3) → separate environments.
  • JSON Tables needs a Rust toolchain at runtime (no wheel).
  • macOS Apple-Silicon only — inherited from the Personal launcher's platform limit.

The defining caveat — ingest reverse-connection

JSON Tables' bulk ingest uses Exasol's HTTP transport, where the database connects back to the client. With the DB on the host and JSON Tables in a container, that reverse connection is fragile across the host/container boundary — the DB may not reach a port inside the container. A host-mode fallback (running JSON Tables ingest directly on the host) is documented for when it fails. wrap / describe / query (client→DB only) are unaffected.

This is the key reason the unified tool's macOS path co-locates the tools on the host next to the DB — see Personal + JSON Tables + MCP.

Where it went

exasol-personal-ai is now the macOS / Personal path of exasol-quickstart (exasol-quickstart --base personal) — experimental, not yet validated end-to-end. The fully-tested path today is the Nano + Docker bundle.