The two modes
A live connection issues a query to the source every time the view changes. Data is always current, and the source database does the work.
An extract is a compressed columnar snapshot stored in .hyper format. Tableau queries the extract instead of the source. It's usually dramatically faster, works offline, and takes load off production systems — at the cost of freshness.
Why extracts are fast
Hyper is a columnar store built for analytic queries. It reads only the columns a query touches, compresses aggressively, and is designed around the memory hierarchy — hot data in RAM, the rest on disk.
Operational databases are usually row stores tuned for transactions, so they're doing the opposite of what a dashboard needs.
Trimming an extract
You don't have to extract everything. In the extract dialog you can add filters, hide unused fields, aggregate to a coarser level, or take a sample of the top N rows.
Hiding unused fields before extracting is the cheapest performance win in Tableau. Fewer columns means a smaller, faster extract.
Refresh strategy
A full refresh rebuilds the whole extract. An incremental refresh appends only rows where a chosen column — typically a timestamp or ID — is greater than the highest value already present.
Incremental is much cheaper but blind to updates and deletes of existing rows. A common pattern is incremental nightly, full weekly. On Tableau Server or Cloud you set these as scheduled tasks.