OLTP workload is when you create/read one order at a time, or a short list of them.
OLAP workload is when you want to aggregate sales numbers across all orders for a given broad criteria, without particular interest in any one order. For example sales by customer by month, or sales by region by date.
The two cases benefit from different data structures and algorithms to achieve their goals.
A system for OLTP may have a table of orders T0 (id, date-time, customer, $total, address, ship method, contact person, ...) where any one record can be fetched efficiently. However reading all of them will require a lot of disk IO.
A system for OLAP can reduce this table to T1 (id, date-time, customer, $total) or even just T2 (date-time, customer, $total) which is good enough to answer most statistical queries and yet consuming less than half the disk bandwidth.
Further yet, you likely do not care about individual order made on a given date by a customer, so you can add up all orders for a given day before storing the data: T3 (day, customer, $total-for-the-day-for-this-customer), or for the week: T4 (week, customer, $total-for-the-week-for-this-customer) or for the months: T5 (month, customer, $total-for-the-month-for-this-customer). Alternatively if you're a big boss you may not care about individual customers, but may care about regions, so the table you will want to see is T6 (week, region, $total-for-the-week-for-all-customers-in-region).
These smaller tables will answer your questions several orders of magnitudes faster compared to T0. And no, they do not have to fit into the memory - to the contrary the entire purpose of the exercise is to make it work efficiently with disk.
One could also chose to not compute e.g. T6 and derive it on the fly from T4, it's slower than having T6 pre-computed, but allows more flexibility in e.g. computing sales by ZIP code or by proximity to major bodies of water, etc.