집계 파이프라인과 복잡한 쿼리

Simple find queries return documents as-is. But analytics endpoints often need grouped, computed, and reshaped data: total revenue per month, average rating per product, top 10 active users.

MongoDB's aggregation pipeline runs this work inside the database, so you ship only the final result over the wire instead of pulling thousands of documents into Python and looping.

• Pipeline = an ordered list of stages
• Each stage takes a stream of documents in and emits documents out
• Beanie exposes it through Document.aggregate(pipeline)

Throughout this lesson we use an Order document. Each order has a customer, a status, a total amount, and a created timestamp. We'll build analytics endpoints on top of it.

Beanie documents subclass beanie.Document, which is itself a Pydantic model bound to a MongoDB collection.

$match filters documents, exactly like a find query. Place it as early as possible so later stages process fewer documents and any indexes can be used.

This pipeline keeps only paid orders. Beanie's aggregate takes a plain list of dicts and runs it against the collection.

$group is the heart of analytics. It buckets documents by an _id expression and computes accumulators over each bucket.

• $sum — total of a field (or count with $sum: 1)
• $avg, $min, $max
• $push / $addToSet — collect values into an array

Here we compute total revenue and order count per customer.

Inside aggregation expressions, a string starting with $ is a field path (read the value of that field). A plain string is a literal.

• "$total" → the value of the total field
• "total" → the literal string "total"
• {"$sum": 1} → add the literal 1 for every document = a count

Mixing these up is the #1 beginner mistake. $sum: "$total" sums amounts; $sum: 1 counts rows.

Add $sort and $limit after grouping to build a leaderboard. Sort uses 1 for ascending and -1 for descending.

This returns the top 5 customers by revenue — a classic analytics endpoint payload.

$project chooses which fields to keep and lets you rename or compute new ones. After a $group the bucket key lives in _id, which is rarely the name your API consumers expect.

Here we rename _id to customer_id and drop the default _id from the output.

Aggregation returns raw dicts, not Order documents (the shape changed). Pass a projection_model so Beanie validates each row into a typed Pydantic model — perfect for a FastAPI response_model.

For time-series analytics, group by a formatted date. $dateToString turns a timestamp into a string bucket like "2026-06" for monthly revenue.

The result is ideal for charting endpoints: one row per month, sorted chronologically.

Put the pipeline behind an async route. Because the aggregation runs in MongoDB, the handler stays tiny and fast even over millions of orders.

Using projection_model as the response_model gives you automatic validation and OpenAPI docs.

The pipeline pattern — match, group, sum — is just data transformation. Here is the same logic in plain Python so you can see what MongoDB does internally: filter, bucket by key, accumulate a sum.

In production MongoDB does this far faster and with indexes, but understanding the shape helps you write correct stages.

You want total revenue per customer, counting only paid orders. Which pipeline is correct?

You can now build analytics-style endpoints with Beanie aggregation pipelines:

• $match early to filter and use indexes
• $group with accumulators ($sum, $avg, $sum: 1 for counts)
• Remember field paths need a leading $; plain strings are literals
• $sort + $limit for leaderboards, $project to reshape and rename _id
• $dateToString for time-series buckets
• Pass projection_model to map raw rows into typed Pydantic models for clean FastAPI response_model output

The pipeline runs inside MongoDB, keeping your handlers small and your endpoints fast.