Glossary | Octelion

Data platforms

Lakehouse: Data stored as open files in cloud storage with warehouse-style reliability and SQL on top. Cheap storage, open formats, transactions. Databricks is the platform we run it on.
Delta Lake: The open table format underneath a lakehouse: Parquet files plus a transaction log that gives you reliable writes, schema enforcement and time travel.
Medallion architecture: Three layers with three promises. Bronze lands everything as it arrived. Silver fixes, drops or quarantines by rule. Gold holds one table per business question.
Auto Loader: The Databricks feature that streams new files into a table and remembers what it has already seen, so a republished file is picked up without a manual step.
Expectation: A data quality rule attached to a pipeline step, so each run reports how many rows failed it. Rows that fail can be kept, dropped or quarantined.
Quarantine table: Where rows that fail a rule go, with the reason attached, instead of being deleted silently. The place engineers look when a total is wrong.
Unity Catalog: Databricks' governance layer: who can read which table, lineage from source to report, and an audit log of access. Grants go to groups; writes go to service principals.
Service principal: A non-human identity that pipelines run as, so no person's credentials are in a job and the audit log shows which job did what.
Job cluster: Compute that starts for one scheduled run and terminates when it finishes, as opposed to an all-purpose cluster left running for people to use. The main lever on cost.
Liquid clustering: A way of physically organising a Delta table by the columns people filter on first (usually a date), so queries read less of it.
DuckDB: A fast analytical database that runs inside a process on a laptop or a server and reads files, including Parquet over HTTP. What we used to compute the figures on this site.
Microsoft Fabric: Microsoft's platform that bundles a lakehouse, pipelines and Power BI under one licence. A reasonable choice for teams standardised on Microsoft.

Reports and models

Semantic model: The layer between tables and visuals in Power BI: relationships, measures with written definitions, security roles. One model can serve many reports, which is how everyone shows the same number.
Star schema: One fact table surrounded by clean dimension tables, one hop from any filter to the fact. The shape that makes Power BI fast.
Measure: A calculation that runs for the cells on screen, written in DAX. The alternative, a calculated column, is stored for every row at refresh and is usually the wrong choice.
DAX: The formula language of Power BI measures and calculated columns.
Incremental refresh: A policy that keeps older data as stored partitions and reloads only a recent window each refresh, so a nightly refresh takes minutes instead of hours.
Aggregation table: A summary table the engine answers from when a visual only needs totals by date, store or category, falling through to the detail when it needs more.
Direct Lake: A Fabric mode where Power BI reads Delta tables directly without importing them, sitting between import mode and DirectQuery.
Row-level security: Rules that filter which rows a signed-in person can see, enforced in the model or the database rather than hidden in the interface.

AI assistants and agents

Assistant: An AI system that answers questions, often over your own documents, for a person who decides what to do next.
Agent: An AI system with a fixed set of tools that carries out a defined job, with a person approving anything that changes a system.
Frontier model: One of the most capable general-purpose language models available from a major provider. We build mostly on Claude.
Safety classifier: A small, fast model that checks every input before the main model sees it and flags what must be refused or escalated, with keyword rules as a fallback.
Evals: A written set of test cases an AI system must pass before launch and on every change, run automatically like unit tests.
Retrieval: Finding the relevant pieces of your documents for each request and placing them in the model's context, so answers are grounded in what you have approved and can cite it.
Knowledge base: The documents, policies and approved answers the assistant retrieves from. Corrections from the review queue go back into it.
Review queue: Where a named person reads a sample of what the system said, flags what was wrong, and feeds the correction back. How the system improves.
Approval step: The point where an agent proposes an action that would change something and a person confirms it before it happens.
MCP (Model Context Protocol): A standard way to expose tools and data to a model so any compatible client can use them. Useful when tools are shared across assistants; unnecessary for one agent with three tools.
Guardrail: A check on the model's output that replaces an unsafe or out-of-scope answer with a safe fallback before a person sees it.
Trace: The record of one request end to end: who asked, the full prompt, the answer, the tools that ran, the timing of each stage, what the classifier decided.
Rate limit: A cap on requests per user and per hour, so a loop or an abusive user hits a limit rather than your budget.
Cost ceiling: A daily spending limit that stops the system, with an alert that reaches a person before it is hit.
Retention rule: How long conversations and traces are kept before a scheduled job deletes them, enforced by the database rather than by a promise.

Workflows and monitoring

Data quality gate: Checks that run after a pipeline and before anything is sent: counts in range, no future dates, totals reconciled. If one fails, nothing goes out and a person is told why.
Runbook: One page per alert: what it means, how to confirm it, the likely causes with the query for each, what to do, whom to tell.
Freshness: When a table or report last updated against when it should have. The first thing a monitoring page shows.
Drift: A slow change in data or in an AI system's behaviour that no single run flags. Caught by comparing each run with a rolling baseline.
Power Automate: Microsoft's workflow tool for the last mile inside Microsoft 365: sending, posting, approving, filing. Paired with Python for the logic.
Audit log: Every action by every role, every pipeline run, every grant change and every purge, in one place an administrator can filter. The answer to "what happened" as a query.

A term that is missing?

Tell us which page sent you here and what it should have explained.

hello@octelion.dev