Reference · Chapter 33
TechFabric Toolkit for Databricks, the Chrome extension
Drafted with AI assistance and reviewed by TechFabric engineers. How we write this guide →
The TechFabric Toolkit for Databricks is a free, MIT-licensed Chrome extension that adds a drawer to Databricks workspace pages on AWS, Azure and Google Cloud. On a table, job, run, cluster, SQL warehouse, pipeline, model or serving endpoint it says what the object is, whether it's healthy and what to do next, and it writes the SQL, Python, bundle YAML and CLI for it. Install it from the Chrome Web Store.
It's an independent tool by TechFabric, not affiliated with or endorsed by Databricks. This page lists every feature, where it shows up, what it needs, and the chapter of this guide each one relates to. Last verified against version 0.10.0 on 26 October 2026.
The short version
Three kinds of feature live in the extension. Generators and the error decoder run in your browser with nothing sent anywhere. Live details (lineage, grants, run history, the last 30 days of cost) read your own workspace, read-only, and only after you press Connect for that workspace. Ask, code review and monthly estimates send a redacted request to this guide's MCP server, and only after you turn the Field Guide on. Most of the guide's advice on cost, Unity Catalog and jobs shows up next to the object it is about.
How do I install the extension? #
- Add it from the Chrome Web Store. It needs Chrome 116 or later.
- Open any Databricks workspace page:
*.cloud.databricks.com(AWS),*.azuredatabricks.net(Azure) or*.gcp.databricks.com(Google Cloud). A small tab appears on the right edge.Alt+Shift+Tshows or hides the drawer, and the toolbar icon opens the same tools in Chrome's side panel. - For live details, press Connect on the Now tab. The extension uses your browser session, or a personal access token if the workspace won't accept the session for its REST API. For Ask, code review and estimates, turn the Field Guide on in Options > Live workspace.
Developers can load the source instead. Download the 0.10.0 archive from techfabric.com/databricks-toolkit, unzip it, open chrome://extensions, turn on Developer mode, press Load unpacked and pick the folder. The extension is plain JavaScript with no build step, so the folder you load is the code that runs.
The drawer has four tabs. Now shows the live card and suggestions for the page you're on, Build holds the generators, Learn is the Field Guide hub, and Saved keeps pins, bookmarks and recent pages.
| Shortcut | Action |
|---|---|
Alt+Shift+T |
Show or hide the drawer |
Alt+Shift+L |
Copy a clean link to the current page |
Ctrl/⌘+Shift+K |
The command palette, which searches every generator, snippet and action (Alt+Shift+K inside a code editor, where the other key deletes a line) |
1 to 6 |
Copy a suggestion on the Now tab |
| The question mark key | Every shortcut |
What runs where #
Every feature falls into one of four places, and the difference decides what you have to switch on.
- In the browser. The generators, the error decoder, the pre-flight check and the formatters work offline. Nothing is requested or sent.
- Your workspace, read-only. Live details call that workspace's REST API from the page you're on. The client sends
GETrequests, plus a short list of requests it builds itself: singleDESCRIBEorSHOWstatements, fixed read-only queries on system tables for cost, cluster CPU and memory, and table activity, and MLflow's run search with a fixed body. Any other request is refused before it is made. A button that needs a SQL warehouse asks before it starts a stopped one, because starting a warehouse costs money. - This guide's MCP server. Ask, Review with the Field Guide and Monthly estimate call
https://databricks.techfabric.com/mcpwith emails, tokens, secrets, storage paths, hosts and long ids removed. Code is previewed before it goes. - A model in your workspace. Explain with AI and Draft with AI call a chat model served in your own workspace, at the endpoint you picked for it, so your admins' AI Gateway permissions, rate limits and guardrails apply. Its output is shown as text and never run.
Cost features #
These sit on the objects that spend money. The What It Actually Costs chapter explains the DBU arithmetic behind them, and Cost and Performance what to change once you see the number.
| Feature | Where it shows up | What it does | Needs |
|---|---|---|---|
| What did this cost? | Job, SQL warehouse and serving endpoint pages | The last 30 days of DBUs and their list-price estimate for that one object, from system.billing.usage joined to system.billing.list_prices. It runs only when you press the button. |
Workspace and a SQL warehouse |
| Monthly estimate | SQL warehouse, cluster and serving endpoint pages | A monthly range for 2 to 24 hours a day (for a model endpoint, 1M to 1B input tokens a month), with the hourly price, the formula and the guide section each figure comes from. Only the shape is sent, never a name or an id. | Field Guide |
| Seven days of CPU and memory | Cluster pages | A button reads p50 and p95 CPU and memory from system.compute.node_timeline. Under 30% CPU and 50% memory at p95 reads as over-provisioned, and over 85% as saturated. |
Workspace and a SQL warehouse |
| Compute that never stops | Cluster, warehouse and endpoint pages, and the workspace pulse | Flags clusters and warehouses with no auto-stop, legacy access modes, and endpoints that never scale to zero. | Workspace |
| Query history | The Query history page | Flags queries that spilled to disk, waited over five seconds for compute or a queue, scanned 1 GB or more to return 1,000 rows or fewer, or failed. Analyze slowest query hands the worst one to the slow-query helper. | Workspace |
| Cost explainer | Build > Cost | Writes the system.billing SQL for a job, run, warehouse, pipeline, notebook, endpoint, app, cluster or the whole workspace, at list price times your contract multiplier, plus an alert query. Paste the result back and it writes a paragraph for Slack. |
Nothing |
Governance features #
On Catalog Explorer pages the extension reads what Unity Catalog in practice tells you to check: who owns an object, who can read it and what depends on it.
| Feature | Where it shows up | What it does | Needs |
|---|---|---|---|
| Table details | Table pages | Owner, type, columns and comments, last update and predictive optimization. Flags a missing description, uncommented columns and owners that are people rather than groups, with copyable fixes. | Workspace |
| Who can read this table | Table pages | Effective grants, including those inherited from the schema and the catalog, with a wide group holding SELECT flagged. When the workspace returns no SELECT grant it says so rather than claiming nobody can read the table, because the owner still can. |
Workspace |
| Lineage | Table pages | What the table is built from, what reads it, and the jobs and notebooks in between, by name, with a warning before you change a table others depend on. | Workspace |
| Audit & 90-day activity | Table pages | The last seven days of grant and table changes from system.access.audit, and 90 days of reads from system.access.table_lineage. A table nothing read in 90 days is flagged as a candidate to retire. |
Workspace and a SQL warehouse |
| Metastore, catalog and schema summaries | The Catalog Explorer root, catalog and schema pages | Catalogs by type, and the catalogs, schemas and tables owned by a person or missing a description. | Workspace |
| Delta health check | Table pages, or Build > Data with pasted results | Runs DESCRIBE DETAIL and DESCRIBE HISTORY and explains small files, partitioning against liquid clustering, and the OPTIMIZE and VACUUM history, with fix SQL. Tables and storage has the reasoning. |
Workspace and a SQL warehouse, or nothing when pasted |
| Draft with AI | Tables with uncommented columns | A model in your workspace drafts ALTER ... COMMENT statements for you to review. The extension never runs them. |
A model in your workspace |
The audit view leans on the system tables Security and audit recommends keeping on.
Jobs and troubleshooting #
A failed run is where the extension saves the most clicks. The Orchestration chapter covers the job settings it checks.
| Feature | Where it shows up | What it does | Needs |
|---|---|---|---|
| Job summary | Job pages | The schedule in words, a run history strip, success rate and median duration. Flags flaky jobs, missing failure notifications and missing timeouts. | Workspace |
| Why did the last run fail? | Job and run pages | Reads the failed task's error and stack trace and decodes them into a likely cause and fixes, with a summary to paste into Slack. | Workspace |
| What changed since the last good run? | Failed run pages | Compares the run with the newest successful one: parameters, Git commit, compute, environment, libraries, tasks and durations. A changed value that the error message names is flagged first. | Workspace |
| Notify me when it finishes | Running run pages, and job pages whose current run is still going | One desktop notification when the run ends, even with several tabs open. The extension checks the run every 30 seconds. | Workspace |
| Workspace pulse | Jobs, compute and other list pages, and the workspace home | What is running now, what will never stop by itself, and the runs that failed in the last 24 hours. | Workspace |
| Where a job comes from | Bundle-deployed job and pipeline pages | The bundle, target and mode, Git branch and commit, and the YAML file linked at that commit. Warns about bundle jobs still editable in the UI and development-mode deploys. A UI-built job gets the databricks bundle generate and bind commands to move it into code, as CI/CD recommends. |
Workspace |
| Review this job | Job pages | The job's settings, written as bundle YAML, go through the guide's code review after a preview. Copy as bundle YAML gives you the file. | Field Guide |
| Explain with AI | Failed runs | A model in your workspace explains the failure. | A model in your workspace |
| Error decoder and triage | Build > Jobs and Build > Triage | Paste an error, or a databricks jobs get-run output, and get likely causes from a built-in list of common Databricks errors, repair commands, and a message for the run's owner with secrets and emails redacted. |
Nothing |
Pipeline features #
| Feature | Where it shows up | What it does | Needs |
|---|---|---|---|
| Pipeline state | Pipeline and update pages | State, health, mode and recent errors, decoded. | Workspace |
| Expectations | Pipeline pages | Pass and fail counts from the pipeline's own events, and which rule started failing in the latest update. | Workspace |
| Is this table behind its job? | Table pages | Names the job that writes the table and whether it's manual, paused or scheduled. Warns when the last write is behind the schedule, when the job succeeded but the table didn't change, or when the last run failed. | Workspace |
| Starters | Build > Data | Auto Loader in Python and SQL, a bronze and silver Lakeflow pipeline, COPY INTO, Auto CDC for SCD type 1 or 2, and data-quality checks as a query, Delta constraints and Lakeflow expectations. |
Nothing |
Freshness and expectations are the checks Observability asks for, and the starters follow Medallion architecture and CDC and real-time data.
MLflow, serving and Genie #
| Feature | Where it shows up | What it does | Needs |
|---|---|---|---|
| Registered models | Unity Catalog model pages | Aliases, versions with their logged metrics, stuck or failed registrations, and which endpoint serves which version, with a warning when it isn't @champion. Code to load, batch-score and serve it. |
Workspace |
| Experiment leaderboard | MLflow experiment pages | Picks the metric to rank by and its direction, shows the top runs, flags a best run with no logged model, and offers Compare #2 with the best run. | Workspace |
| Why is this run worse? | MLflow run pages | The run's rank and how it differs from the best run: metric deltas, then training data, parameters, Git commit and notebook revision. | Workspace |
| Serving endpoints | Serving endpoint pages | The model and its price per million tokens, the AI Gateway settings (usage tracking, inference table, rate limits, guardrails, fallbacks), and calling code for a notebook, SQL ai_query and curl. |
Workspace |
| Vector Search | Vector index and endpoint pages | Ready state, rows indexed, source table, the embedding model and whether its endpoint is up, last and failed syncs, and query and sync code. | Workspace |
| Genie readiness | Genie space pages | Whether the space has a description, whether its warehouse auto-stops, and which curated tables lack a description or have fewer than half their columns commented. | Workspace |
| AI/ML templates | Build > AI/ML templates | AI Functions SQL, a RAG starter, a Genie setup checklist and an agent governance checklist. | Nothing |
The chapters behind these are MLflow, Foundation Model APIs, RAG and AI Search, Genie and Agents.
Lakebase #
On a Lakebase database instance page (/compute/database-instances/), the live card shows the instance's state, capacity, node count, read/write DNS name, point-in-time recovery window and creator. It warns when the recovery window is under seven days, and when a single node backs what may be a production app. The Lakebase chapter covers branching and recovery in depth.
Notebook and query pages #
| Feature | Where it shows up | What it does | Needs |
|---|---|---|---|
| Editor companion | Notebook and SQL editor pages | The interactive clusters and SQL warehouses running now, flagging any that never stop, and recent queries. On a notebook it lists the widgets and %run includes, and flags widgets with empty defaults and table names with an environment written in (prod_sales.orders). |
Workspace |
| Review with the Field Guide | Notebook and saved query pages, and a paste box in Learn | The code goes to the guide's review after a preview of how many lines go and what was replaced. Each finding names the cell and line, why it matters, the fix and the guide section. | Field Guide |
| Pre-flight check | Build > Safety, and offered when you copy risky SQL on a workspace labelled PROD | Finds DROP, TRUNCATE, DELETE or UPDATE without WHERE, VACUUM under seven days, INSERT OVERWRITE, CREATE OR REPLACE, MERGE ... DELETE and more, says whether each can be undone, and writes read-only queries that show the impact. Nothing is run. |
Nothing |
The pre-flight check pairs with Disaster recovery, which covers what to do when a destructive statement already ran.
The offline generators #
The Build tab holds the generators, grouped by job and ordered for the page you're on, with names filled in from it. They run in the browser, and every one is also available through this guide's MCP server as toolkit_generate, with the same code.
| Group | Generators |
|---|---|
| SQL & DDL | Table helper: SELECT, CREATE TABLE, INSERT, MERGE, GRANT, REVOKE and column comments from a column list |
| Data | Delta health check from pasted results, schema diff to ALTER statements, Auto Loader and Lakeflow starters, Auto CDC, data-quality checks, cost and lineage queries on system tables, a time travel and restore wizard |
| Jobs | Error decoder, run history to a Slack or Jira digest, failed run to a Slack or Jira message |
| AI/ML templates | MLflow and Unity Catalog model helper, Model Serving calls, Vector Search index and query, AI Functions, RAG starter, Genie checklist, agent governance checklist |
| Bundles & IaC | Declarative Automation Bundle databricks.yml for a job, pipeline, app or serving endpoint with its CLI commands, a Unity Catalog permissions matrix with a rollback script, a table starter kit (tags, constraints, liquid clustering, row filter, column mask), synthetic test data, Terraform, secret scopes and external locations |
| Tools | Unity Catalog name helper, SQL formatter and minifier, deep-link builder, diff-friendly formatter for SQL and notebook source |
| Cost, Triage, Safety | Cost explainer, failed-run triage, slow-query analysis, pre-flight check |
| Accelerators | Links and commands for TechFabric's accelerators |
The bundle generator follows CI/CD with bundles and Terraform, and the permissions matrix follows the group-based grant model in Unity Catalog in practice.
Learn and Ask #
Learn is a Field Guide hub inside the drawer. The chapters of this guide for the kind of page you're on, a cheat sheet, and SQL and PySpark side by side are built in and work offline. Once the Field Guide is on, it adds an Ask box, a paste box for code review and What's new. Its docs links go to this guide first, with one official Databricks docs link at the end.
Ask (the sparkle button in the header) answers a Databricks question from this guide, with the question redacted and no workspace data sent, and links the sections it used. If you also turn on a model in your workspace, that model writes the answer grounded on the guide's excerpts and cites them.
Feedback goes from the thumbs under answers, decoded errors and run comparisons. The sheet shows exactly what will be sent, and nothing goes until you press Send. How to learn Databricks is the chapter Learn points a newcomer to.
Everyday features #
- Links and names. A clean link to the page, and the object's name in Markdown, Slack and Jira formats and with or without backticks. A workspace switcher moves you to another saved workspace, keeping the Catalog Explorer path.
- Suggested for this page. The three to six actions most useful on the page, each copied with a number key.
- Saved. Pinned snippets, bookmarks, recent pages and a notebook basket that becomes a notebook source file to import.
- Environment label. A coloured bar on workspaces your own rules call PROD, STAGE or DEV. On a PROD workspace, copying risky SQL offers the pre-flight check.
- Heads-up cards. Optional hints for failed runs, tables with few column comments, the PROD guard and slow queries, each off until you turn it on in Options.
Permissions it asks for #
| Permission | Why |
|---|---|
storage |
Settings, pins and bookmarks in Chrome sync storage; recent pages, object names and live connections in local storage on this computer only; personal access tokens in the service worker's own storage on this computer |
clipboardWrite |
The Copy buttons |
sidePanel |
The side panel the toolbar icon opens |
scripting |
Putting the new version into open Databricks tabs after an update |
notifications |
Notify me when it finishes |
alarms |
The timer that checks watched runs |
Host access to *.cloud.databricks.com, *.azuredatabricks.net, *.gcp.databricks.com |
It runs on Databricks workspace pages only |
Optional: app.airlift.fabric.pro, *.databricksapps.com |
Only if you turn on the optional Airlift tab |
Privacy and data #
There are no analytics, no remote code and no accounts. Nothing goes to TechFabric unless you ask: a question to the Field Guide once you've turned it on, code you send for review after its preview, a warehouse or cluster's shape for an estimate, and feedback when you press Send. Each is redacted first. The Field Guide logs a tool's name, never the code.
Live details stay off until you press Connect for a workspace. Personal access tokens stay on that computer, never synced, in the extension's service worker storage where page scripts can't read them, and are sent only to the workspace they were saved for. Use a token with a short lifetime. Settings, pins and bookmarks sit in Chrome sync storage, so they follow your Chrome profile to your other browsers when Chrome sync is on. Options clears any of it, and removing the extension clears everything.
Using it from an AI assistant #
This guide's MCP server at https://databricks.techfabric.com/mcp answers questions about the extension and runs its generators, for an assistant connected to it:
toolkit_infosays what the extension does feature by feature, where each one shows up, what it needs, and how to get the same thing from the server.toolkit_infowith a question such as "does it show table lineage" returns the features that answer it.toolkit_generateruns any Build tab generator with the extension's own code, so a grant script or a bundle file matches what the extension writes. Called with no generator, it lists them all with their inputs.- What needs your workspace (lineage, grants, run history, billing) can't run on an anonymous server. For those,
toolkit_infopoints to the closest tool:estimate_costfor cost,review_databricks_codefor a job's YAML or a notebook's source, andexplore_conceptfor how grant inheritance or Lakebase branching works.
Questions people ask #
Is the TechFabric Toolkit for Databricks free? #
Yes. It's free and open source under the MIT license, with no account to create.
Can the extension change anything in my workspace? #
It doesn't change data, grants or settings. Its client sends read requests and a short list of read-only statements it builds itself, and refuses anything else before the request is made. The one thing it can start is a stopped SQL warehouse, for the health check or a cost query, and it asks first. Generated SQL, including the fix statements and the comments a model drafts, is text for you to review and run yourself.
Does it work on Azure Databricks and Google Cloud? #
Yes. It runs on workspace pages at *.cloud.databricks.com, *.azuredatabricks.net and *.gcp.databricks.com. Monthly estimates use the guide's AWS list prices, and the extension says so on Azure and Google Cloud.
Which features work without connecting a workspace? #
Every generator in the Build tab, the error decoder, the pre-flight check, the formatters, the snippets, the links and names, Saved, and Learn's chapter links, cheat sheet and SQL and PySpark pairs. Live details and the Delta health check through a warehouse need Connect.
Does it send my data to TechFabric? #
Only what you send. Questions, code for review, a compute shape for an estimate and feedback go to this guide's server, redacted, after you turn the Field Guide on. There's no analytics or tracking.
Can I use its generators without the extension? #
Yes. Connect an AI assistant to this guide's MCP server and call toolkit_generate. It runs the same generator code, without the page context the extension fills in, so you pass the names yourself.
What's popular